Reincluding Numbers

.github/workflows/release.yaml

@@ -304,11 +304,6 @@ jobs:
           write-host "Installing plugin"
           & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
           write-host "plugin installed"
-      - name: remove unwanted mingw dll.a files
-        run: |
-          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libpthread.dll.a" -File | Remove-Item -Force
-          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libwinpthread.dll.a" -File | Remove-Item -Force
-          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libstdc++.dll.a" -File | Remove-Item -Force
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
@@ -442,7 +437,6 @@ jobs:
     env:
       OLLAMA_SKIP_IMAGE_BUILD: '1'
       PUSH: '1'
-      GH_TOKEN: ${{ github.token }}
     steps:
       - uses: actions/checkout@v4
       - name: Set Version
@@ -466,20 +460,15 @@ jobs:
           ls -lh dist/
           (cd dist; sha256sum * > sha256sum.txt)
           cat dist/sha256sum.txt
-      - name: Create or update Release
-        run: |
-          echo "Looking for existing release for ${{ env.RELEASE_VERSION }}"
-          OLD_TAG=$(gh release ls --json name,tagName | jq -r ".[] | select(.name == \"${{ env.RELEASE_VERSION }}\") | .tagName")
-          if [ -n "$OLD_TAG" ]; then
-            echo "Updating release ${{ env.RELEASE_VERSION }} to point to new tag ${GITHUB_REF_NAME}"
-            gh release edit ${OLD_TAG} --tag ${GITHUB_REF_NAME}
-          else
-            echo "Creating new release ${{ env.RELEASE_VERSION }} pointing to tag ${GITHUB_REF_NAME}"
-            gh release create ${GITHUB_REF_NAME} \
-              --title ${{ env.RELEASE_VERSION }} \
-              --draft \
-              --generate-notes \
-              --prerelease
-          fi
-          echo "Uploading artifacts for tag ${GITHUB_REF_NAME}"
-          gh release upload ${GITHUB_REF_NAME} dist/* --clobber
+      - uses: ncipollo/release-action@v1
+        with:
+          name: ${{ env.RELEASE_VERSION }}
+          allowUpdates: true
+          artifacts: 'dist/*'
+          draft: true
+          prerelease: true
+          omitBodyDuringUpdate: true
+          generateReleaseNotes: true
+          omitDraftDuringUpdate: true
+          omitPrereleaseDuringUpdate: true
+          replacesArtifacts: true

.github/workflows/test.yaml

@@ -34,13 +34,13 @@ jobs:
             git diff-tree -r --no-commit-id --name-only \
               $(git merge-base ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }}) \
               ${{ github.event.pull_request.head.sha }} \
-              | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
+              | xargs python3 -c "import sys; print(any([x.startswith('$1') for x in sys.argv[1:]]))"
           }
 
           {
-            echo GENERATE=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
-            echo GENERATE_CUDA=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
-            echo GENERATE_ROCM=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo GENERATE=$(changed llm/)
+            echo GENERATE_CUDA=$(changed llm/)
+            echo GENERATE_ROCM=$(changed llm/)
           } >>$GITHUB_OUTPUT
 
   generate:
@@ -58,7 +58,6 @@ jobs:
     runs-on: ${{ matrix.os }}
     env:
       GOARCH: ${{ matrix.arch }}
-      CGO_ENABLED: '1'
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
@@ -80,7 +79,6 @@ jobs:
       - run: go generate -x ./...
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
         name: 'Unix Go Generate'
-      - run: go build .
       - uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
@@ -126,7 +124,7 @@ jobs:
     strategy:
       matrix:
         rocm-version:
-          - '6.1.1'
+          - '6.0.2'
     runs-on: linux
     container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
     steps:
@@ -271,9 +269,9 @@ jobs:
           mkdir -p llm/build/darwin/$ARCH/stub/bin
           touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'macos-') }}
-      - uses: golangci/golangci-lint-action@v6
+      - uses: golangci/golangci-lint-action@v4
         with:
-          args: --timeout 8m0s -v ${{ startsWith(matrix.os, 'windows-') && '' || '--disable gofmt --disable goimports' }}
+          args: --timeout 8m0s -v
   test:
     strategy:
       matrix:
@@ -289,8 +287,6 @@ jobs:
       GOARCH: ${{ matrix.arch }}
       CGO_ENABLED: '1'
       OLLAMA_CPU_TARGET: 'static'
-      OLLAMA_SKIP_CPU_GENERATE: '1'
-      OLLAMA_SKIP_METAL_GENERATE: '1'
     steps:
       - uses: actions/checkout@v4
         with:

.golangci.yaml

@@ -9,26 +9,9 @@ linters:
     - contextcheck
     - exportloopref
     - gocheckcompilerdirectives
-    # conditionally enable this on linux/macos
+    # FIXME: for some reason this errors on windows
     # - gofmt
     # - goimports
-    - intrange
     - misspell
     - nilerr
-    - nolintlint
-    - nosprintfhostport
-    - testifylint
-    - unconvert
     - unused
-    - wastedassign
-    - whitespace
-    - usestdlibvars
-severity:
-  default-severity: error
-  rules:
-    - linters:
-        - gofmt
-        - goimports
-        - intrange
-        - usestdlibvars
-      severity: info

Dockerfile

@@ -2,7 +2,7 @@ ARG GOLANG_VERSION=1.22.1
 ARG CMAKE_VERSION=3.22.1
 # this CUDA_VERSION corresponds with the one specified in docs/gpu.md
 ARG CUDA_VERSION=11.3.1
-ARG ROCM_VERSION=6.1.1
+ARG ROCM_VERSION=6.0.2
 
 # Copy the minimal context we need to run the generate scripts
 FROM scratch AS llm-code
@@ -70,12 +70,12 @@ RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
 RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
 
-FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
+FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS

README.md

@@ -6,7 +6,7 @@
 
 [![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)
 
-Get up and running with large language models.
+Get up and running with large language models locally.
 
 ### macOS
 
@@ -53,8 +53,8 @@ Here are some example models that can be downloaded:
 | Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
 | Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
 | Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
-| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`            |
-| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`        |
+| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
+| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
 | Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
@@ -182,12 +182,6 @@ $ ollama run llama3 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
-### Show model information
-
-```
-ollama show llama3
-```
-
 ### List models on your computer
 
 ```
@@ -291,8 +285,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [macai](https://github.com/Renset/macai) (macOS client for Ollama, ChatGPT, and other compatible API back-ends)
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
-- [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
-- [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
 
 ### Terminal
 
@@ -315,7 +307,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ShellOracle](https://github.com/djcopley/ShellOracle)
 - [tlm](https://github.com/yusufcanb/tlm)
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
-- [gollama](https://github.com/sammcj/gollama)
 
 ### Database
 
@@ -333,13 +324,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
-- [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
-- [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
 - [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
 - [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
@@ -357,7 +346,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
 - [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
 - [LlamaScript](https://github.com/Project-Llama/llamascript)
-
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
@@ -390,9 +378,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
 - [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
-- [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
-
-### Supported backends
 
+### Supported backends 
 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.
 

api/client.go

@@ -23,9 +23,11 @@ import (
 	"net"
 	"net/http"
 	"net/url"
+	"os"
 	"runtime"
+	"strconv"
+	"strings"
 
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/version"
 )
@@ -63,7 +65,10 @@ func checkError(resp *http.Response, body []byte) error {
 // If the variable is not specified, a default ollama host and port will be
 // used.
 func ClientFromEnvironment() (*Client, error) {
-	ollamaHost := envconfig.Host
+	ollamaHost, err := GetOllamaHost()
+	if err != nil {
+		return nil, err
+	}
 
 	return &Client{
 		base: &url.URL{
@@ -74,6 +79,52 @@ func ClientFromEnvironment() (*Client, error) {
 	}, nil
 }
 
+type OllamaHost struct {
+	Scheme string
+	Host   string
+	Port   string
+}
+
+func GetOllamaHost() (OllamaHost, error) {
+	defaultPort := "11434"
+
+	hostVar := os.Getenv("OLLAMA_HOST")
+	hostVar = strings.TrimSpace(strings.Trim(strings.TrimSpace(hostVar), "\"'"))
+
+	scheme, hostport, ok := strings.Cut(hostVar, "://")
+	switch {
+	case !ok:
+		scheme, hostport = "http", hostVar
+	case scheme == "http":
+		defaultPort = "80"
+	case scheme == "https":
+		defaultPort = "443"
+	}
+
+	// trim trailing slashes
+	hostport = strings.TrimRight(hostport, "/")
+
+	host, port, err := net.SplitHostPort(hostport)
+	if err != nil {
+		host, port = "127.0.0.1", defaultPort
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
+			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
+		}
+	}
+
+	if portNum, err := strconv.ParseInt(port, 10, 32); err != nil || portNum > 65535 || portNum < 0 {
+		return OllamaHost{}, ErrInvalidHostPort
+	}
+
+	return OllamaHost{
+		Scheme: scheme,
+		Host:   host,
+		Port:   port,
+	}, nil
+}
+
 func NewClient(base *url.URL, http *http.Client) *Client {
 	return &Client{
 		base: base,
@@ -304,8 +355,8 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 }
 
 // List running models.
-func (c *Client) ListRunning(ctx context.Context) (*ProcessResponse, error) {
-	var lr ProcessResponse
+func (c *Client) ListRunning(ctx context.Context) (*ListResponse, error) {
+	var lr ListResponse
 	if err := c.do(ctx, http.MethodGet, "/api/ps", nil, &lr); err != nil {
 		return nil, err
 	}

api/client_test.go

@@ -1,9 +1,11 @@
 package api
 
 import (
+	"fmt"
+	"net"
 	"testing"
 
-	"github.com/ollama/ollama/envconfig"
+	"github.com/stretchr/testify/assert"
 )
 
 func TestClientFromEnvironment(t *testing.T) {
@@ -33,7 +35,6 @@ func TestClientFromEnvironment(t *testing.T) {
 	for k, v := range testCases {
 		t.Run(k, func(t *testing.T) {
 			t.Setenv("OLLAMA_HOST", v.value)
-			envconfig.LoadConfig()
 
 			client, err := ClientFromEnvironment()
 			if err != v.err {
@@ -45,4 +46,40 @@ func TestClientFromEnvironment(t *testing.T) {
 			}
 		})
 	}
+
+	hostTestCases := map[string]*testCase{
+		"empty":               {value: "", expect: "127.0.0.1:11434"},
+		"only address":        {value: "1.2.3.4", expect: "1.2.3.4:11434"},
+		"only port":           {value: ":1234", expect: ":1234"},
+		"address and port":    {value: "1.2.3.4:1234", expect: "1.2.3.4:1234"},
+		"hostname":            {value: "example.com", expect: "example.com:11434"},
+		"hostname and port":   {value: "example.com:1234", expect: "example.com:1234"},
+		"zero port":           {value: ":0", expect: ":0"},
+		"too large port":      {value: ":66000", err: ErrInvalidHostPort},
+		"too small port":      {value: ":-1", err: ErrInvalidHostPort},
+		"ipv6 localhost":      {value: "[::1]", expect: "[::1]:11434"},
+		"ipv6 world open":     {value: "[::]", expect: "[::]:11434"},
+		"ipv6 no brackets":    {value: "::1", expect: "[::1]:11434"},
+		"ipv6 + port":         {value: "[::1]:1337", expect: "[::1]:1337"},
+		"extra space":         {value: " 1.2.3.4 ", expect: "1.2.3.4:11434"},
+		"extra quotes":        {value: "\"1.2.3.4\"", expect: "1.2.3.4:11434"},
+		"extra space+quotes":  {value: " \" 1.2.3.4 \" ", expect: "1.2.3.4:11434"},
+		"extra single quotes": {value: "'1.2.3.4'", expect: "1.2.3.4:11434"},
+	}
+
+	for k, v := range hostTestCases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", v.value)
+
+			oh, err := GetOllamaHost()
+			if err != v.err {
+				t.Fatalf("expected %s, got %s", v.err, err)
+			}
+
+			if err == nil {
+				host := net.JoinHostPort(oh.Host, oh.Port)
+				assert.Equal(t, v.expect, host, fmt.Sprintf("%s: expected %s, got %s", k, v.expect, host))
+			}
+		})
+	}
 }

api/types.go

@@ -2,6 +2,7 @@ package api
 
 import (
 	"encoding/json"
+	"errors"
 	"fmt"
 	"log/slog"
 	"math"
@@ -159,18 +160,18 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA   bool  `json:"numa,omitempty"`
-	NumCtx    int   `json:"num_ctx,omitempty"`
-	NumBatch  int   `json:"num_batch,omitempty"`
-	NumGPU    int   `json:"num_gpu,omitempty"`
-	MainGPU   int   `json:"main_gpu,omitempty"`
-	LowVRAM   bool  `json:"low_vram,omitempty"`
-	F16KV     bool  `json:"f16_kv,omitempty"`
-	LogitsAll bool  `json:"logits_all,omitempty"`
-	VocabOnly bool  `json:"vocab_only,omitempty"`
-	UseMMap   *bool `json:"use_mmap,omitempty"`
-	UseMLock  bool  `json:"use_mlock,omitempty"`
-	NumThread int   `json:"num_thread,omitempty"`
+	UseNUMA   bool `json:"numa,omitempty"`
+	NumCtx    int  `json:"num_ctx,omitempty"`
+	NumBatch  int  `json:"num_batch,omitempty"`
+	NumGPU    int  `json:"num_gpu,omitempty"`
+	MainGPU   int  `json:"main_gpu,omitempty"`
+	LowVRAM   bool `json:"low_vram,omitempty"`
+	F16KV     bool `json:"f16_kv,omitempty"`
+	LogitsAll bool `json:"logits_all,omitempty"`
+	VocabOnly bool `json:"vocab_only,omitempty"`
+	UseMMap   bool `json:"use_mmap,omitempty"`
+	UseMLock  bool `json:"use_mlock,omitempty"`
+	NumThread int  `json:"num_thread,omitempty"`
 }
 
 // EmbeddingRequest is the request passed to [Client.Embeddings].
@@ -222,7 +223,6 @@ type ShowRequest struct {
 	Model    string `json:"model"`
 	System   string `json:"system"`
 	Template string `json:"template"`
-	Verbose  bool   `json:"verbose"`
 
 	Options map[string]interface{} `json:"options"`
 
@@ -232,16 +232,13 @@ type ShowRequest struct {
 
 // ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
-	License       string         `json:"license,omitempty"`
-	Modelfile     string         `json:"modelfile,omitempty"`
-	Parameters    string         `json:"parameters,omitempty"`
-	Template      string         `json:"template,omitempty"`
-	System        string         `json:"system,omitempty"`
-	Details       ModelDetails   `json:"details,omitempty"`
-	Messages      []Message      `json:"messages,omitempty"`
-	ModelInfo     map[string]any `json:"model_info,omitempty"`
-	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
-	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
+	License    string       `json:"license,omitempty"`
+	Modelfile  string       `json:"modelfile,omitempty"`
+	Parameters string       `json:"parameters,omitempty"`
+	Template   string       `json:"template,omitempty"`
+	System     string       `json:"system,omitempty"`
+	Details    ModelDetails `json:"details,omitempty"`
+	Messages   []Message    `json:"messages,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].
@@ -285,40 +282,19 @@ type PushRequest struct {
 
 // ListResponse is the response from [Client.List].
 type ListResponse struct {
-	Models []ListModelResponse `json:"models"`
+	Models []ModelResponse `json:"models"`
 }
 
-// ProcessResponse is the response from [Client.Process].
-type ProcessResponse struct {
-	Models []ProcessModelResponse `json:"models"`
-}
-
-// ListModelResponse is a single model description in [ListResponse].
-type ListModelResponse struct {
+// ModelResponse is a single model description in [ListResponse].
+type ModelResponse struct {
 	Name       string       `json:"name"`
 	Model      string       `json:"model"`
-	ModifiedAt time.Time    `json:"modified_at"`
+	ModifiedAt time.Time    `json:"modified_at,omitempty"`
 	Size       int64        `json:"size"`
 	Digest     string       `json:"digest"`
 	Details    ModelDetails `json:"details,omitempty"`
-}
-
-// ProcessModelResponse is a single model description in [ProcessResponse].
-type ProcessModelResponse struct {
-	Name      string       `json:"name"`
-	Model     string       `json:"model"`
-	Size      int64        `json:"size"`
-	Digest    string       `json:"digest"`
-	Details   ModelDetails `json:"details,omitempty"`
-	ExpiresAt time.Time    `json:"expires_at"`
-	SizeVRAM  int64        `json:"size_vram"`
-}
-
-type RetrieveModelResponse struct {
-	Id      string `json:"id"`
-	Object  string `json:"object"`
-	Created int64  `json:"created"`
-	OwnedBy string `json:"owned_by"`
+	ExpiresAt  time.Time    `json:"expires_at,omitempty"`
+	SizeVRAM   int64        `json:"size_vram,omitempty"`
 }
 
 type TokenResponse struct {
@@ -330,7 +306,7 @@ type GenerateResponse struct {
 	// Model is the model name that generated the response.
 	Model string `json:"model"`
 
-	// CreatedAt is the timestamp of the response.
+	//CreatedAt is the timestamp of the response.
 	CreatedAt time.Time `json:"created_at"`
 
 	// Response is the textual response itself.
@@ -387,6 +363,8 @@ func (m *Metrics) Summary() {
 	}
 }
 
+var ErrInvalidHostPort = errors.New("invalid port specified in OLLAMA_HOST")
+
 func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
@@ -459,17 +437,6 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					slice[i] = str
 				}
 				field.Set(reflect.ValueOf(slice))
-			case reflect.Pointer:
-				var b bool
-				if field.Type() == reflect.TypeOf(&b) {
-					val, ok := val.(bool)
-					if !ok {
-						return fmt.Errorf("option %q must be of type boolean", key)
-					}
-					field.Set(reflect.ValueOf(&val))
-				} else {
-					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
-				}
 			default:
 				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 			}
@@ -512,7 +479,7 @@ func DefaultOptions() Options {
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   nil,
+			UseMMap:   true,
 			UseNUMA:   false,
 		},
 	}
@@ -609,17 +576,6 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 				case reflect.Slice:
 					// TODO: only string slices are supported right now
 					out[key] = vals
-				case reflect.Pointer:
-					var b bool
-					if field.Type() == reflect.TypeOf(&b) {
-						boolVal, err := strconv.ParseBool(vals[0])
-						if err != nil {
-							return nil, fmt.Errorf("invalid bool value %s", vals)
-						}
-						out[key] = &boolVal
-					} else {
-						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
-					}
 				default:
 					return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
 				}

api/types_test.go

@@ -2,7 +2,6 @@ package api
 
 import (
 	"encoding/json"
-	"fmt"
 	"math"
 	"testing"
 	"time"
@@ -73,13 +72,13 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
 		},
 		{
 			"positive duration",
-			42 * time.Second,
-			42 * time.Second,
+			time.Duration(42 * time.Second),
+			time.Duration(42 * time.Second),
 		},
 		{
 			"another positive duration",
-			42 * time.Minute,
-			42 * time.Minute,
+			time.Duration(42 * time.Minute),
+			time.Duration(42 * time.Minute),
 		},
 		{
 			"zero duration",
@@ -106,105 +105,3 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
 		})
 	}
 }
-
-func TestUseMmapParsingFromJSON(t *testing.T) {
-	tr := true
-	fa := false
-	tests := []struct {
-		name string
-		req  string
-		exp  *bool
-	}{
-		{
-			name: "Undefined",
-			req:  `{ }`,
-			exp:  nil,
-		},
-		{
-			name: "True",
-			req:  `{ "use_mmap": true }`,
-			exp:  &tr,
-		},
-		{
-			name: "False",
-			req:  `{ "use_mmap": false }`,
-			exp:  &fa,
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			var oMap map[string]interface{}
-			err := json.Unmarshal([]byte(test.req), &oMap)
-			require.NoError(t, err)
-			opts := DefaultOptions()
-			err = opts.FromMap(oMap)
-			require.NoError(t, err)
-			assert.Equal(t, test.exp, opts.UseMMap)
-		})
-	}
-}
-
-func TestUseMmapFormatParams(t *testing.T) {
-	tr := true
-	fa := false
-	tests := []struct {
-		name string
-		req  map[string][]string
-		exp  *bool
-		err  error
-	}{
-		{
-			name: "True",
-			req: map[string][]string{
-				"use_mmap": {"true"},
-			},
-			exp: &tr,
-			err: nil,
-		},
-		{
-			name: "False",
-			req: map[string][]string{
-				"use_mmap": {"false"},
-			},
-			exp: &fa,
-			err: nil,
-		},
-		{
-			name: "Numeric True",
-			req: map[string][]string{
-				"use_mmap": {"1"},
-			},
-			exp: &tr,
-			err: nil,
-		},
-		{
-			name: "Numeric False",
-			req: map[string][]string{
-				"use_mmap": {"0"},
-			},
-			exp: &fa,
-			err: nil,
-		},
-		{
-			name: "invalid string",
-			req: map[string][]string{
-				"use_mmap": {"foo"},
-			},
-			exp: nil,
-			err: fmt.Errorf("invalid bool value [foo]"),
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			resp, err := FormatParams(test.req)
-			require.Equal(t, test.err, err)
-			respVal, ok := resp["use_mmap"]
-			if test.exp != nil {
-				assert.True(t, ok, "resp: %v", resp)
-				assert.Equal(t, *test.exp, *respVal.(*bool))
-			}
-		})
-	}
-}

app/lifecycle/logging.go

@@ -5,8 +5,6 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
-	"strconv"
-	"strings"
 
 	"github.com/ollama/ollama/envconfig"
 )
@@ -26,7 +24,6 @@ func InitLogging() {
 		logFile = os.Stderr
 		// TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion
 	} else {
-		rotateLogs(AppLogFile)
 		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
 		if err != nil {
 			slog.Error(fmt.Sprintf("failed to create server log %v", err))
@@ -49,32 +46,3 @@ func InitLogging() {
 
 	slog.Info("ollama app started")
 }
-
-func rotateLogs(logFile string) {
-	if _, err := os.Stat(logFile); os.IsNotExist(err) {
-		return
-	}
-	index := strings.LastIndex(logFile, ".")
-	pre := logFile[:index]
-	post := "." + logFile[index+1:]
-	for i := LogRotationCount; i > 0; i-- {
-		older := pre + "-" + strconv.Itoa(i) + post
-		newer := pre + "-" + strconv.Itoa(i-1) + post
-		if i == 1 {
-			newer = pre + post
-		}
-		if _, err := os.Stat(newer); err == nil {
-			if _, err := os.Stat(older); err == nil {
-				err := os.Remove(older)
-				if err != nil {
-					slog.Warn("Failed to remove older log", "older", older, "error", err)
-					continue
-				}
-			}
-			err := os.Rename(newer, older)
-			if err != nil {
-				slog.Warn("Failed to rotate log", "older", older, "newer", newer, "error", err)
-			}
-		}
-	}
-}

app/lifecycle/logging_test.go

@@ -1,44 +0,0 @@
-package lifecycle
-
-import (
-	"os"
-	"path/filepath"
-	"strconv"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestRotateLogs(t *testing.T) {
-	logDir := t.TempDir()
-	logFile := filepath.Join(logDir, "testlog.log")
-
-	// No log exists
-	rotateLogs(logFile)
-
-	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0644))
-	assert.FileExists(t, logFile)
-	// First rotation
-	rotateLogs(logFile)
-	assert.FileExists(t, filepath.Join(logDir, "testlog-1.log"))
-	assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log"))
-	assert.NoFileExists(t, logFile)
-
-	// Should be a no-op without a new log
-	rotateLogs(logFile)
-	assert.FileExists(t, filepath.Join(logDir, "testlog-1.log"))
-	assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log"))
-	assert.NoFileExists(t, logFile)
-
-	for i := 2; i <= LogRotationCount+1; i++ {
-		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0644))
-		assert.FileExists(t, logFile)
-		rotateLogs(logFile)
-		assert.NoFileExists(t, logFile)
-		for j := 1; j < i; j++ {
-			assert.FileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(j)+".log"))
-		}
-		assert.NoFileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(i+1)+".log"))
-	}
-}

app/lifecycle/paths.go

@@ -16,12 +16,11 @@ var (
 	AppDir     = "/opt/Ollama"
 	AppDataDir = "/opt/Ollama"
 	// TODO - should there be a distinct log dir?
-	UpdateStageDir   = "/tmp"
-	AppLogFile       = "/tmp/ollama_app.log"
-	ServerLogFile    = "/tmp/ollama.log"
-	UpgradeLogFile   = "/tmp/ollama_update.log"
-	Installer        = "OllamaSetup.exe"
-	LogRotationCount = 5
+	UpdateStageDir = "/tmp"
+	AppLogFile     = "/tmp/ollama_app.log"
+	ServerLogFile  = "/tmp/ollama.log"
+	UpgradeLogFile = "/tmp/ollama_update.log"
+	Installer      = "OllamaSetup.exe"
 )
 
 func init() {
@@ -70,6 +69,7 @@ func init() {
 				slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err))
 			}
 		}
+
 	} else if runtime.GOOS == "darwin" {
 		// TODO
 		AppName += ".app"

app/lifecycle/server.go

@@ -15,7 +15,7 @@ import (
 )
 
 func getCLIFullPath(command string) string {
-	var cmdPath string
+	cmdPath := ""
 	appExe, err := os.Executable()
 	if err == nil {
 		cmdPath = filepath.Join(filepath.Dir(appExe), command)
@@ -54,7 +54,7 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 		return nil, fmt.Errorf("failed to spawn server stderr pipe: %w", err)
 	}
 
-	rotateLogs(ServerLogFile)
+	// TODO - rotation
 	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create server log: %w", err)
@@ -65,6 +65,7 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 	if err != nil {
 		if !errors.Is(err, os.ErrNotExist) {
 			return nil, fmt.Errorf("stat ollama server log dir %s: %v", logDir, err)
+
 		}
 
 		if err := os.MkdirAll(logDir, 0o755); err != nil {

app/lifecycle/server_windows.go

@@ -24,8 +24,7 @@ func terminate(cmd *exec.Cmd) error {
 	if err != nil {
 		return err
 	}
-	//nolint:errcheck
-	defer dll.Release()
+	defer dll.Release() // nolint: errcheck
 
 	pid := cmd.Process.Pid
 
@@ -74,8 +73,7 @@ func isProcessExited(pid int) (bool, error) {
 	if err != nil {
 		return false, fmt.Errorf("failed to open process: %v", err)
 	}
-	//nolint:errcheck
-	defer windows.CloseHandle(hProcess)
+	defer windows.CloseHandle(hProcess) // nolint: errcheck
 
 	var exitCode uint32
 	err = windows.GetExitCodeProcess(hProcess, &exitCode)

app/lifecycle/updater.go

@@ -78,7 +78,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	}
 	defer resp.Body.Close()
 
-	if resp.StatusCode == http.StatusNoContent {
+	if resp.StatusCode == 204 {
 		slog.Debug("check update response 204 (current version is up to date)")
 		return false, updateResp
 	}
@@ -87,7 +87,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 		slog.Warn(fmt.Sprintf("failed to read body response: %s", err))
 	}
 
-	if resp.StatusCode != http.StatusOK {
+	if resp.StatusCode != 200 {
 		slog.Info(fmt.Sprintf("check update error %d - %.96s", resp.StatusCode, string(body)))
 		return false, updateResp
 	}
@@ -114,7 +114,7 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
 	if err != nil {
 		return fmt.Errorf("error checking update: %w", err)
 	}
-	if resp.StatusCode != http.StatusOK {
+	if resp.StatusCode != 200 {
 		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
 	}
 	resp.Body.Close()

app/ollama.iss

@@ -88,15 +88,10 @@ DialogFontSize=12
 [Files]
 Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
 Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
+Source: "..\dist\windows-{#ARCH}\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
 Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-#if DirExists("..\dist\windows-amd64\cuda")
-  Source: "..\dist\windows-amd64\cuda\*"; DestDir: "{app}\cuda\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\oneapi")
-  Source: "..\dist\windows-amd64\oneapi\*"; DestDir: "{app}\oneapi\"; Flags: ignoreversion recursesubdirs
-#endif
 #if DirExists("..\dist\windows-amd64\rocm")
   Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
 #endif

app/ollama_welcome.ps1

@@ -4,5 +4,5 @@ write-host "Welcome to Ollama!"
 write-host ""
 write-host "Run your first model:"
 write-host ""
-write-host "`tollama run llama3"
+write-host "`tollama run llama2"
 write-host ""

app/store/store.go

@@ -29,6 +29,7 @@ func GetID() string {
 		initStore()
 	}
 	return store.ID
+
 }
 
 func GetFirstTimeRun() bool {

app/tray/wintray/eventloop.go

@@ -47,6 +47,7 @@ func nativeLoop() {
 		default:
 			pTranslateMessage.Call(uintptr(unsafe.Pointer(m))) //nolint:errcheck
 			pDispatchMessage.Call(uintptr(unsafe.Pointer(m)))  //nolint:errcheck
+
 		}
 	}
 }
@@ -159,8 +160,8 @@ func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam ui
 		lResult, _, _ = pDefWindowProc.Call(
 			uintptr(hWnd),
 			uintptr(message),
-			wParam,
-			lParam,
+			uintptr(wParam),
+			uintptr(lParam),
 		)
 	}
 	return

app/tray/wintray/tray.go

@@ -186,7 +186,7 @@ func (t *winTray) initInstance() error {
 	t.muNID.Lock()
 	defer t.muNID.Unlock()
 	t.nid = &notifyIconData{
-		Wnd:             t.window,
+		Wnd:             windows.Handle(t.window),
 		ID:              100,
 		Flags:           NIF_MESSAGE,
 		CallbackMessage: t.wmSystrayMessage,
@@ -197,6 +197,7 @@ func (t *winTray) initInstance() error {
 }
 
 func (t *winTray) createMenu() error {
+
 	menuHandle, _, err := pCreatePopupMenu.Call()
 	if menuHandle == 0 {
 		return err
@@ -245,7 +246,7 @@ func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title
 	mi := menuItemInfo{
 		Mask:     MIIM_FTYPE | MIIM_STRING | MIIM_ID | MIIM_STATE,
 		Type:     MFT_STRING,
-		ID:       menuItemId,
+		ID:       uint32(menuItemId),
 		TypeData: titlePtr,
 		Cch:      uint32(len(title)),
 	}
@@ -301,10 +302,11 @@ func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title
 }
 
 func (t *winTray) addSeparatorMenuItem(menuItemId, parentId uint32) error {
+
 	mi := menuItemInfo{
 		Mask: MIIM_FTYPE | MIIM_ID | MIIM_STATE,
 		Type: MFT_SEPARATOR,
-		ID:   menuItemId,
+		ID:   uint32(menuItemId),
 	}
 
 	mi.Size = uint32(unsafe.Sizeof(mi))
@@ -424,6 +426,7 @@ func iconBytesToFilePath(iconBytes []byte) (string, error) {
 // Loads an image from file and shows it in tray.
 // Shell_NotifyIcon: https://msdn.microsoft.com/en-us/library/windows/desktop/bb762159(v=vs.85).aspx
 func (t *winTray) setIcon(src string) error {
+
 	h, err := t.loadIconFrom(src)
 	if err != nil {
 		return err
@@ -441,6 +444,7 @@ func (t *winTray) setIcon(src string) error {
 // Loads an image from file to be shown in tray or menu item.
 // LoadImage: https://msdn.microsoft.com/en-us/library/windows/desktop/ms648045(v=vs.85).aspx
 func (t *winTray) loadIconFrom(src string) (windows.Handle, error) {
+
 	// Save and reuse handles of loaded images
 	t.muLoadedImages.RLock()
 	h, ok := t.loadedImages[src]

cmd/cmd.go

@@ -20,7 +20,6 @@ import (
 	"path/filepath"
 	"regexp"
 	"runtime"
-	"slices"
 	"strings"
 	"syscall"
 	"time"
@@ -30,6 +29,7 @@ import (
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
+	"golang.org/x/exp/slices"
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
@@ -162,6 +162,9 @@ func tempZipFiles(path string) (string, error) {
 	}
 	defer tempfile.Close()
 
+	zipfile := zip.NewWriter(tempfile)
+	defer zipfile.Close()
+
 	detectContentType := func(path string) (string, error) {
 		f, err := os.Open(path)
 		if err != nil {
@@ -230,9 +233,6 @@ func tempZipFiles(path string) (string, error) {
 		files = append(files, tks...)
 	}
 
-	zipfile := zip.NewWriter(tempfile)
-	defer zipfile.Close()
-
 	for _, file := range files {
 		f, err := os.Open(file)
 		if err != nil {
@@ -287,12 +287,38 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 }
 
 func RunHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	name := args[0]
+
+	// check if the model exists on the server
+	show, err := client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+	var statusError api.StatusError
+	switch {
+	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
+		if err := PullHandler(cmd, []string{name}); err != nil {
+			return err
+		}
+
+		show, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+		if err != nil {
+			return err
+		}
+	case err != nil:
+		return err
+	}
+
 	interactive := true
 
 	opts := runOptions{
-		Model:    args[0],
-		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]interface{}{},
+		Model:       args[0],
+		WordWrap:    os.Getenv("TERM") == "xterm-256color",
+		Options:     map[string]interface{}{},
+		MultiModal:  slices.Contains(show.Details.Families, "clip"),
+		ParentModel: show.Details.ParentModel,
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -336,38 +362,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.WordWrap = !nowrap
 
-	// Fill out the rest of the options based on information about the
-	// model.
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
+	if !interactive {
+		return generate(cmd, opts)
 	}
 
-	name := args[0]
-	info, err := func() (*api.ShowResponse, error) {
-		showReq := &api.ShowRequest{Name: name}
-		info, err := client.Show(cmd.Context(), showReq)
-		var se api.StatusError
-		if errors.As(err, &se) && se.StatusCode == http.StatusNotFound {
-			if err := PullHandler(cmd, []string{name}); err != nil {
-				return nil, err
-			}
-			return client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-		}
-		return info, err
-	}()
-	if err != nil {
-		return err
-	}
-
-	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
-	opts.ParentModel = info.Details.ParentModel
-	opts.Messages = append(opts.Messages, info.Messages...)
-
-	if interactive {
-		return generateInteractive(cmd, opts)
-	}
-	return generate(cmd, opts)
+	return generateInteractive(cmd, opts)
 }
 
 func errFromUnknownKey(unknownKeyErr error) error {
@@ -580,6 +579,10 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	if len(args) != 1 {
+		return errors.New("missing model name")
+	}
+
 	license, errLicense := cmd.Flags().GetBool("license")
 	modelfile, errModelfile := cmd.Flags().GetBool("modelfile")
 	parameters, errParams := cmd.Flags().GetBool("parameters")
@@ -622,6 +625,8 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 
 	if flagsSet > 1 {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
+	} else if flagsSet == 0 {
+		return errors.New("one of '--license', '--modelfile', '--parameters', '--system', or '--template' must be specified")
 	}
 
 	req := api.ShowRequest{Name: args[0]}
@@ -630,141 +635,22 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	if flagsSet == 1 {
-		switch showType {
-		case "license":
-			fmt.Println(resp.License)
-		case "modelfile":
-			fmt.Println(resp.Modelfile)
-		case "parameters":
-			fmt.Println(resp.Parameters)
-		case "system":
-			fmt.Println(resp.System)
-		case "template":
-			fmt.Println(resp.Template)
-		}
-
-		return nil
+	switch showType {
+	case "license":
+		fmt.Println(resp.License)
+	case "modelfile":
+		fmt.Println(resp.Modelfile)
+	case "parameters":
+		fmt.Println(resp.Parameters)
+	case "system":
+		fmt.Println(resp.System)
+	case "template":
+		fmt.Println(resp.Template)
 	}
 
-	showInfo(resp)
-
 	return nil
 }
 
-func showInfo(resp *api.ShowResponse) {
-	arch := resp.ModelInfo["general.architecture"].(string)
-
-	modelData := [][]string{
-		{"arch", arch},
-		{"parameters", resp.Details.ParameterSize},
-		{"quantization", resp.Details.QuantizationLevel},
-		{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
-		{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
-	}
-
-	mainTableData := [][]string{
-		{"Model"},
-		{renderSubTable(modelData, false)},
-	}
-
-	if resp.ProjectorInfo != nil {
-		projectorData := [][]string{
-			{"arch", "clip"},
-			{"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))},
-		}
-
-		if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok {
-			projectorData = append(projectorData, []string{"projector type", projectorType.(string)})
-		}
-
-		projectorData = append(projectorData,
-			[]string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
-			[]string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
-		)
-
-		mainTableData = append(mainTableData,
-			[]string{"Projector"},
-			[]string{renderSubTable(projectorData, false)},
-		)
-	}
-
-	if resp.Parameters != "" {
-		mainTableData = append(mainTableData, []string{"Parameters"}, []string{formatParams(resp.Parameters)})
-	}
-
-	if resp.System != "" {
-		mainTableData = append(mainTableData, []string{"System"}, []string{renderSubTable(twoLines(resp.System), true)})
-	}
-
-	if resp.License != "" {
-		mainTableData = append(mainTableData, []string{"License"}, []string{renderSubTable(twoLines(resp.License), true)})
-	}
-
-	table := tablewriter.NewWriter(os.Stdout)
-	table.SetAutoWrapText(false)
-	table.SetBorder(false)
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range mainTableData {
-		table.Append(v)
-	}
-
-	table.Render()
-}
-
-func renderSubTable(data [][]string, file bool) string {
-	var buf bytes.Buffer
-	table := tablewriter.NewWriter(&buf)
-	table.SetAutoWrapText(!file)
-	table.SetBorder(false)
-	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range data {
-		table.Append(v)
-	}
-
-	table.Render()
-
-	renderedTable := buf.String()
-	lines := strings.Split(renderedTable, "\n")
-	for i, line := range lines {
-		lines[i] = "\t" + line
-	}
-
-	return strings.Join(lines, "\n")
-}
-
-func twoLines(s string) [][]string {
-	lines := strings.Split(s, "\n")
-	res := [][]string{}
-
-	count := 0
-	for _, line := range lines {
-		line = strings.TrimSpace(line)
-		if line != "" {
-			count++
-			res = append(res, []string{line})
-			if count == 2 {
-				return res
-			}
-		}
-	}
-	return res
-}
-
-func formatParams(s string) string {
-	lines := strings.Split(s, "\n")
-	table := [][]string{}
-
-	for _, line := range lines {
-		table = append(table, strings.Fields(line))
-	}
-	return renderSubTable(table, false)
-}
-
 func CopyHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -860,6 +746,7 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	if wordWrap && termWidth >= 10 {
 		for _, ch := range content {
 			if state.lineLength+1 > termWidth-5 {
+
 				if runewidth.StringWidth(state.wordBuffer) > termWidth-10 {
 					fmt.Printf("%s%c", state.wordBuffer, ch)
 					state.wordBuffer = ""
@@ -868,11 +755,7 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 				}
 
 				// backtrack the length of the last word and clear to the end of the line
-				a := runewidth.StringWidth(state.wordBuffer)
-				if a > 0 {
-					fmt.Printf("\x1b[%dD", a)
-				}
-				fmt.Printf("\x1b[K\n")
+				fmt.Printf("\x1b[%dD\x1b[K\n", runewidth.StringWidth(state.wordBuffer))
 				fmt.Printf("%s%c", state.wordBuffer, ch)
 				chWidth := runewidth.RuneWidth(ch)
 
@@ -1074,11 +957,17 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 }
 
 func RunServer(cmd *cobra.Command, _ []string) error {
+	// retrieve the OLLAMA_HOST environment variable
+	ollamaHost, err := api.GetOllamaHost()
+	if err != nil {
+		return err
+	}
+
 	if err := initializeKeypair(); err != nil {
 		return err
 	}
 
-	ln, err := net.Listen("tcp", net.JoinHostPort(envconfig.Host.Host, envconfig.Host.Port))
+	ln, err := net.Listen("tcp", net.JoinHostPort(ollamaHost.Host, ollamaHost.Port))
 	if err != nil {
 		return err
 	}
@@ -1137,6 +1026,24 @@ func initializeKeypair() error {
 	return nil
 }
 
+//nolint:unused
+func waitForServer(ctx context.Context, client *api.Client) error {
+	// wait for the server to start
+	timeout := time.After(5 * time.Second)
+	tick := time.Tick(500 * time.Millisecond)
+	for {
+		select {
+		case <-timeout:
+			return errors.New("timed out waiting for server to start")
+		case <-tick:
+			if err := client.Heartbeat(ctx); err == nil {
+				return nil // server has started
+			}
+		}
+	}
+
+}
+
 func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -1344,9 +1251,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
 				envVars["OLLAMA_TMPDIR"],
-				envVars["OLLAMA_FLASH_ATTENTION"],
-				envVars["OLLAMA_LLM_LIBRARY"],
-				envVars["OLLAMA_MAX_VRAM"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

cmd/interactive.go

@@ -8,11 +8,11 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
-	"slices"
 	"sort"
 	"strings"
 
 	"github.com/spf13/cobra"
+	"golang.org/x/exp/slices"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
@@ -31,40 +31,65 @@ const (
 )
 
 func loadModel(cmd *cobra.Command, opts *runOptions) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
 	p := progress.NewProgress(os.Stderr)
 	defer p.StopAndClear()
 
 	spinner := progress.NewSpinner("")
 	p.Add("", spinner)
 
-	client, err := api.ClientFromEnvironment()
+	showReq := api.ShowRequest{Name: opts.Model}
+	showResp, err := client.Show(cmd.Context(), &showReq)
 	if err != nil {
 		return err
 	}
+	opts.MultiModal = slices.Contains(showResp.Details.Families, "clip")
+	opts.ParentModel = showResp.Details.ParentModel
 
-	chatReq := &api.ChatRequest{
-		Model:     opts.Model,
-		KeepAlive: opts.KeepAlive,
+	if len(showResp.Messages) > 0 {
+		opts.Messages = append(opts.Messages, showResp.Messages...)
 	}
 
-	return client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
+	chatReq := &api.ChatRequest{
+		Model:    opts.Model,
+		Messages: []api.Message{},
+	}
+
+	if opts.KeepAlive != nil {
+		chatReq.KeepAlive = opts.KeepAlive
+	}
+
+	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
 		p.StopAndClear()
-		for _, msg := range opts.Messages {
-			switch msg.Role {
-			case "user":
-				fmt.Printf(">>> %s\n", msg.Content)
-			case "assistant":
-				state := &displayResponseState{}
-				displayResponse(msg.Content, opts.WordWrap, state)
-				fmt.Println()
-				fmt.Println()
+		if len(opts.Messages) > 0 {
+			for _, msg := range opts.Messages {
+				switch msg.Role {
+				case "user":
+					fmt.Printf(">>> %s\n", msg.Content)
+				case "assistant":
+					state := &displayResponseState{}
+					displayResponse(msg.Content, opts.WordWrap, state)
+					fmt.Println()
+					fmt.Println()
+				}
 			}
 		}
 		return nil
 	})
+	if err != nil {
+		return err
+	}
+
+	return nil
 }
 
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
+	opts.Messages = make([]api.Message, 0)
+
 	err := loadModel(cmd, &opts)
 	if err != nil {
 		return err
@@ -404,7 +429,15 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					showInfo(resp)
+					fmt.Println("Model details:")
+					if len(resp.Details.Families) > 0 {
+						fmt.Printf("Family              %s\n", strings.Join(resp.Details.Families, ", "))
+					} else if resp.Details.Family != "" {
+						fmt.Printf("Family              %s\n", resp.Details.Family)
+					}
+					fmt.Printf("Parameter Size      %s\n", resp.Details.ParameterSize)
+					fmt.Printf("Quantization Level  %s\n", resp.Details.QuantizationLevel)
+					fmt.Println("")
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")

cmd/interactive_test.go

@@ -6,7 +6,6 @@ import (
 	"text/template"
 
 	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
 
 	"github.com/ollama/ollama/api"
 )
@@ -86,11 +85,11 @@ MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 `
 
 	tmpl, err := template.New("").Parse(expectedModelfile)
-	require.NoError(t, err)
+	assert.Nil(t, err)
 
 	var buf bytes.Buffer
 	err = tmpl.Execute(&buf, opts)
-	require.NoError(t, err)
+	assert.Nil(t, err)
 	assert.Equal(t, buf.String(), mf)
 
 	opts.ParentModel = "horseshark"
@@ -108,10 +107,10 @@ MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 `
 
 	tmpl, err = template.New("").Parse(expectedModelfile)
-	require.NoError(t, err)
+	assert.Nil(t, err)
 
 	var parentBuf bytes.Buffer
 	err = tmpl.Execute(&parentBuf, opts)
-	require.NoError(t, err)
+	assert.Nil(t, err)
 	assert.Equal(t, parentBuf.String(), mf)
 }

cmd/start.go

@@ -1,27 +0,0 @@
-//go:build darwin || windows
-
-package cmd
-
-import (
-	"context"
-	"errors"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-func waitForServer(ctx context.Context, client *api.Client) error {
-	// wait for the server to start
-	timeout := time.After(5 * time.Second)
-	tick := time.Tick(500 * time.Millisecond)
-	for {
-		select {
-		case <-timeout:
-			return errors.New("timed out waiting for server to start")
-		case <-tick:
-			if err := client.Heartbeat(ctx); err == nil {
-				return nil // server has started
-			}
-		}
-	}
-}

convert/convert.go

@@ -1,134 +1,200 @@
 package convert
 
 import (
+	"cmp"
+	"encoding/binary"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
 	"os"
 	"path/filepath"
+	"slices"
+	"strings"
 
+	"google.golang.org/protobuf/proto"
+
+	"github.com/ollama/ollama/convert/sentencepiece"
 	"github.com/ollama/ollama/llm"
 )
 
-type Parameters struct {
-	Architectures []string `json:"architectures"`
-	VocabSize     uint32   `json:"vocab_size"`
+const (
+	_ int32 = iota
+	tokenTypeNormal
+	tokenTypeUnknown
+	tokenTypeControl
+	tokenTypeUserDefined
+	tokenTypeUnused
+	tokenTypeByte
+)
+
+type Params struct {
+	Architectures     []string `json:"architectures"`
+	VocabSize         int      `json:"vocab_size"`
+	HiddenSize        int      `json:"hidden_size"`       // n_embd
+	HiddenLayers      int      `json:"num_hidden_layers"` // n_layer
+	ContextSize       int      `json:"max_position_embeddings"`
+	IntermediateSize  int      `json:"intermediate_size"`
+	AttentionHeads    int      `json:"num_attention_heads"` // n_head
+	KeyValHeads       int      `json:"num_key_value_heads"`
+	NormEPS           float64  `json:"rms_norm_eps"`
+	BoSTokenID        int      `json:"bos_token_id"`
+	EoSTokenID        int      `json:"eos_token_id"`
+	HeadDimension     int      `json:"head_dim"`
+	PaddingTokenID    int      `json:"pad_token_id"`
+	RopeFrequencyBase float64  `json:"rope_theta"`
+
+	Experts     int `json:"num_local_experts"`
+	ExpertsUsed int `json:"num_experts_per_tok"`
+
+	PreTokenizer string
+
+	ByteOrder
 }
 
-func (Parameters) KV(t *Tokenizer) llm.KV {
-	kv := llm.KV{
-		"general.file_type":            uint32(1),
-		"general.quantization_version": uint32(2),
-		"tokenizer.ggml.pre":           t.Pre,
-		"tokenizer.ggml.model":         t.Vocabulary.Model,
-		"tokenizer.ggml.tokens":        t.Vocabulary.Tokens,
-		"tokenizer.ggml.scores":        t.Vocabulary.Scores,
-		"tokenizer.ggml.token_type":    t.Vocabulary.Types,
-	}
-
-	if t.Template != "" {
-		kv["tokenizer.chat_template"] = t.Template
-	}
-
-	for _, sv := range t.SpecialVocabulary {
-		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
-		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
-	}
-
-	return kv
+type ByteOrder interface {
+	binary.ByteOrder
+	binary.AppendByteOrder
 }
 
-func (Parameters) specialTypes() []string {
-	return []string{
-		"bos", "eos", "unk", "sep", "pad", "cls", "mask",
-	}
+type ModelArch interface {
+	GetTensors() error
+	LoadVocab() error
+	WriteGGUF(io.WriteSeeker) error
 }
 
-func (Parameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []*llm.Tensor) error {
-	return llm.WriteGGUF(ws, kv, ts)
+type ModelFormat interface {
+	GetLayerName(string) (string, error)
+	GetTensors(string, *Params) ([]llm.Tensor, error)
+	GetParams(string) (*Params, error)
+	GetModelArch(string, string, *Params) (ModelArch, error)
 }
 
-type Converter interface {
-	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) llm.KV
-	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []*llm.Tensor
-
-	// tensorName returns the LLM tensor name for a specific input name
-	tensorName(string) string
-	// specialTypes returns any special token types the model uses
-	specialTypes() []string
-	writeFile(io.WriteSeeker, llm.KV, []*llm.Tensor) error
+type ModelData struct {
+	Path    string
+	Name    string
+	Params  *Params
+	Vocab   *Vocab
+	Tensors []llm.Tensor
+	Format  ModelFormat
 }
 
-func ConvertAdapter(d string, ws io.WriteSeeker) error {
-	c := &adapter{}
-
-	ts, err := parseNPZ(d)
+func GetModelFormat(dirname string) (ModelFormat, error) {
+	files, err := filepath.Glob(filepath.Join(dirname, "*"))
 	if err != nil {
-		return err
+		return nil, err
 	}
 
-	return c.writeFile(ws, c.KV(nil), c.Tensors(ts))
-}
-
-func Convert(d string, ws io.WriteSeeker) error {
-	f, err := os.Open(filepath.Join(d, "config.json"))
-	if err != nil {
-		return err
-	}
-	defer f.Close()
-
-	var p Parameters
-	if err := json.NewDecoder(f).Decode(&p); err != nil {
-		return err
-	}
-
-	if len(p.Architectures) < 1 {
-		return errors.New("unknown architecture")
-	}
-
-	var c Converter
-	switch p.Architectures[0] {
-	case "LlamaForCausalLM", "MistralForCausalLM":
-		c = &llama{}
-	case "MixtralForCausalLM":
-		c = &mixtral{}
-	case "GemmaForCausalLM":
-		c = &gemma{}
-	default:
-		return errors.New("unsupported architecture")
-	}
-
-	bts, err := os.ReadFile(filepath.Join(d, "config.json"))
-	if err != nil {
-		return err
-	}
-
-	if err := json.Unmarshal(bts, c); err != nil {
-		return err
-	}
-
-	t, err := parseTokenizer(d, c.specialTypes())
-	if err != nil {
-		return err
-	}
-
-	if vocabSize := int(p.VocabSize); vocabSize > len(t.Vocabulary.Tokens) {
-		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", p.VocabSize, "actual", len(t.Vocabulary.Tokens))
-		for i := range vocabSize - len(t.Vocabulary.Tokens) {
-			t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i))
-			t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1)
-			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
+	for _, fn := range files {
+		if strings.HasSuffix(fn, ".safetensors") {
+			return &SafetensorFormat{}, nil
+		} else if strings.HasSuffix(fn, ".bin") || strings.HasSuffix(fn, ".pth") {
+			slog.Debug("model is torch")
+			return &TorchFormat{}, nil
 		}
 	}
 
-	ts, err := parseTensors(d)
+	return nil, fmt.Errorf("couldn't determine model format")
+}
+
+// Details on gguf's tokenizer can be found at:
+// https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#tokenizer
+type Vocab struct {
+	Tokens []string
+	Scores []float32
+	Types  []int32
+	Merges []string
+}
+
+func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
+	slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model")))
+	in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model"))
 	if err != nil {
-		return err
+		return nil, err
 	}
 
-	return c.writeFile(ws, c.KV(t), c.Tensors(ts))
+	// To regenerate sentencepiece from the protobufs use:
+	// protoc -I=./ --go_out=./ sentencepiece_model.proto
+	modelProto := &sentencepiece.ModelProto{}
+	if err := proto.Unmarshal(in, modelProto); err != nil {
+		return nil, err
+	}
+
+	v := &Vocab{
+		Tokens: make([]string, 0),
+		Scores: make([]float32, 0),
+		Types:  make([]int32, 0),
+	}
+
+	pieces := modelProto.GetPieces()
+	for _, p := range pieces {
+		v.Tokens = append(v.Tokens, p.GetPiece())
+		v.Scores = append(v.Scores, p.GetScore())
+		t := p.GetType()
+		switch t {
+		case sentencepiece.ModelProto_SentencePiece_UNKNOWN:
+		case sentencepiece.ModelProto_SentencePiece_CONTROL:
+		case sentencepiece.ModelProto_SentencePiece_UNUSED:
+		case sentencepiece.ModelProto_SentencePiece_BYTE:
+		default:
+			t = sentencepiece.ModelProto_SentencePiece_NORMAL
+		}
+		v.Types = append(v.Types, int32(t))
+	}
+
+	slog.Info(fmt.Sprintf("vocab size: %d", len(v.Tokens)))
+
+	// add any additional tokens
+	addIn, err := os.ReadFile(filepath.Join(dirpath, "added_tokens.json"))
+	if os.IsNotExist(err) {
+		return v, nil
+	} else if err != nil {
+		return nil, err
+	}
+
+	slog.Info("reading user defined tokens")
+
+	var extraTokenData map[string]int
+	if err := json.Unmarshal(addIn, &extraTokenData); err != nil {
+		return nil, err
+	}
+
+	type token struct {
+		key string
+		pos int
+	}
+
+	extraTokens := make([]token, 0)
+	for k, id := range extraTokenData {
+		extraTokens = append(extraTokens, token{k, id})
+	}
+
+	slices.SortFunc(extraTokens, func(a, b token) int {
+		return cmp.Compare(a.pos, b.pos)
+	})
+
+	numToks := len(v.Tokens)
+
+	for cnt, t := range extraTokens {
+		// the token id should match the specific index for the total number of tokens
+		if t.pos != cnt+numToks {
+			return nil, fmt.Errorf("token ID '%d' for '%s' doesn't match total token size", t.pos, t.key)
+		}
+		v.Tokens = append(v.Tokens, t.key)
+		v.Scores = append(v.Scores, -1000.0)
+		v.Types = append(v.Types, tokenTypeUserDefined)
+	}
+	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
+
+	if params.VocabSize > len(v.Tokens) {
+		missingTokens := params.VocabSize - len(v.Tokens)
+		slog.Warn(fmt.Sprintf("vocab is missing %d tokens", missingTokens))
+		for cnt := 0; cnt < missingTokens; cnt++ {
+			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
+			v.Scores = append(v.Scores, -1)
+			v.Types = append(v.Types, tokenTypeUserDefined)
+		}
+	}
+
+	return v, nil
 }

convert/convert_adapter.go

@@ -1,56 +0,0 @@
-package convert
-
-import (
-	"io"
-	"strings"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type adapter struct {
-	Parameters
-}
-
-var _ Converter = (*adapter)(nil)
-
-func (p *adapter) writeFile(ws io.WriteSeeker, kv llm.KV, ts []*llm.Tensor) error {
-	return llm.WriteGGLA(ws, kv, ts)
-}
-
-func (p *adapter) KV(t *Tokenizer) llm.KV {
-	// todo - need a way to pass these in
-	kv := llm.KV{
-		"r":     uint32(8),
-		"alpha": uint32(160),
-	}
-	return kv
-}
-
-func (p *adapter) Tensors(ts []Tensor) []*llm.Tensor {
-	var out []*llm.Tensor
-	for _, t := range ts {
-		name := p.tensorName(t.Name())
-
-		out = append(out, &llm.Tensor{
-			Name:     name,
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *adapter) tensorName(n string) string {
-	return strings.NewReplacer(
-		"model.layers", "blk",
-		"self_attn.q_proj", "attn_q.weight",
-		"self_attn.k_proj", "attn_k.weight",
-		"self_attn.v_proj", "attn_v.weight",
-		"self_attn.o_proj", "attn_output.weight",
-		"lora_a", "loraA",
-		"lora_b", "loraB",
-		".npy", "",
-	).Replace(n)
-}

convert/convert_gemma.go

@@ -1,103 +0,0 @@
-package convert
-
-import (
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type gemma struct {
-	Parameters
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	HiddenLayers          uint32  `json:"num_hidden_layers"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	RMSNormEPS            float32 `json:"rms_norm_eps"`
-	HeadDim               uint32  `json:"head_dim"`
-}
-
-var _ Converter = (*gemma)(nil)
-
-func (p *gemma) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
-	kv["general.architecture"] = "gemma"
-	kv["general.name"] = "gemma"
-	kv["gemma.context_length"] = p.MaxPositionEmbeddings
-	kv["gemma.embedding_length"] = p.HiddenSize
-	kv["gemma.block_count"] = p.HiddenLayers
-	kv["gemma.feed_forward_length"] = p.IntermediateSize
-	kv["gemma.attention.head_count"] = p.NumAttentionHeads
-	kv["gemma.attention.head_count_kv"] = p.NumKeyValueHeads
-	kv["gemma.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-	kv["gemma.attention.key_length"] = p.HeadDim
-	kv["gemma.attention.value_length"] = p.HeadDim
-	kv["tokenizer.ggml.eot_token_id"] = uint32(107)
-	kv["tokenizer.ggml.middle_token_id"] = uint32(68)
-	kv["tokenizer.ggml.prefix_token_id"] = uint32(67)
-	kv["tokenizer.ggml.suffix_token_id"] = uint32(69)
-	return kv
-}
-
-func (p *gemma) Tensors(ts []Tensor) []*llm.Tensor {
-	var out []*llm.Tensor
-	for _, t := range ts {
-		name := p.tensorName(t.Name())
-		if strings.HasSuffix(name, "_norm.weight") {
-			t.SetRepacker(p.addOne)
-		}
-
-		out = append(out, &llm.Tensor{
-			Name:     name,
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *gemma) tensorName(n string) string {
-	return strings.NewReplacer(
-		"model.embed_tokens", "token_embd",
-		"model.norm", "output_norm",
-		"model.layers", "blk",
-		"input_layernorm", "attn_norm",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_output",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.down_proj", "ffn_down",
-		"mlp.up_proj", "ffn_up",
-		"post_attention_layernorm", "ffn_norm",
-		"block_sparse_moe.gate", "ffn_inp",
-	).Replace(n)
-}
-
-func (*gemma) addOne(_ string, data []float32, shape []uint64) ([]float32, error) {
-	n := tensor.New(tensor.WithShape(int(shape[0])), tensor.WithBacking(data))
-	ones := tensor.Ones(tensor.Float32, int(shape[0]))
-
-	n, err := n.Add(ones)
-	if err != nil {
-		return nil, err
-	}
-
-	ts, err := native.SelectF32(n, 0)
-	if err != nil {
-		return nil, err
-	}
-
-	var f32s []float32
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-
-	return f32s, nil
-}

convert/convert_llama.go

@@ -1,176 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"fmt"
-	"strings"
-
-	"github.com/ollama/ollama/llm"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-)
-
-type llama struct {
-	Parameters
-	NLayers               uint32  `json:"n_layers"`
-	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
-	NLayer                uint32  `json:"n_layer"`
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	NCtx                  uint32  `json:"n_ctx"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	NEmbd                 uint32  `json:"n_embd"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NInner                uint32  `json:"n_inner"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NHead                 uint32  `json:"n_head"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	RopeTheta             float32 `json:"rope_theta"`
-	RopeScaling           struct {
-		Type   string  `json:"type"`
-		Factor float32 `json:"factor"`
-	} `json:"rope_scaling"`
-	RMSNormEPS       float32 `json:"rms_norm_eps"`
-	LayerNormEPS     float32 `json:"layer_norm_eps"`
-	LayerNormEpsilon float32 `json:"layer_norm_epsilon"`
-	NormEpsilon      float32 `json:"norm_epsilon"`
-}
-
-var _ Converter = (*llama)(nil)
-
-func (p *llama) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
-	kv["general.architecture"] = "llama"
-	kv["general.name"] = "llama"
-	kv["llama.vocab_size"] = p.VocabSize
-
-	kv["llama.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
-
-	if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 {
-		kv["llama.context_length"] = contextLength
-	}
-
-	if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 {
-		kv["llama.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
-	}
-
-	if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 {
-		kv["llama.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
-	}
-
-	if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
-		kv["llama.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
-		kv["llama.rope.dimension_count"] = p.HiddenSize / headCount
-	}
-
-	if p.RopeTheta > 0 {
-		kv["llama.rope.freq_base"] = p.RopeTheta
-	}
-
-	if p.RopeScaling.Type == "linear" {
-		kv["llama.rope.scaling.type"] = p.RopeScaling.Type
-		kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor
-	}
-
-	if p.NumKeyValueHeads > 0 {
-		kv["llama.attention.head_count_kv"] = p.NumKeyValueHeads
-	}
-
-	if p.RMSNormEPS > 0 {
-		kv["llama.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-	}
-
-	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 {
-		kv["llama.attention.layer_norm_epsilon"] = layerNormEpsilon
-	}
-
-	if len(t.Merges) > 0 {
-		kv["tokenizer.ggml.merges"] = t.Merges
-	}
-
-	return kv
-}
-
-func (p *llama) Tensors(ts []Tensor) []*llm.Tensor {
-	var out []*llm.Tensor
-	for _, t := range ts {
-		name := p.tensorName(t.Name())
-		if strings.HasSuffix(name, "attn_q.weight") ||
-			strings.HasSuffix(name, "attn_k.weight") {
-			t.SetRepacker(p.repack)
-		}
-
-		out = append(out, &llm.Tensor{
-			Name:     name,
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *llama) tensorName(n string) string {
-	return strings.NewReplacer(
-		"lm_head", "output",
-		"model.embed_tokens", "token_embd",
-		"model.norm", "output_norm",
-		"model.layers", "blk",
-		"input_layernorm", "attn_norm",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_output",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.down_proj", "ffn_down",
-		"mlp.up_proj", "ffn_up",
-		"post_attention_layernorm", "ffn_norm",
-		// mixtral
-		"block_sparse_moe.gate", "ffn_gate_inp",
-	).Replace(n)
-}
-
-func (p *llama) repack(name string, data []float32, shape []uint64) ([]float32, error) {
-	var dims []int
-	for _, dim := range shape {
-		dims = append(dims, int(dim))
-	}
-
-	var heads uint32
-	if strings.HasSuffix(name, "q_proj.weight") {
-		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, "k_proj.weight") {
-		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
-	} else {
-		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
-	}
-
-	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(dims...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-
-	ts, err := native.SelectF32(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var f32s []float32
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-
-	return f32s, nil
-}

convert/convert_mixtral.go

@@ -1,89 +0,0 @@
-package convert
-
-import (
-	"fmt"
-	"io"
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type mixtral struct {
-	llama
-	NumLocalExperts    uint32 `json:"num_local_experts"`
-	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
-}
-
-var _ Converter = (*mixtral)(nil)
-
-func (p *mixtral) KV(t *Tokenizer) llm.KV {
-	kv := p.llama.KV(t)
-
-	if p.NumLocalExperts > 0 {
-		kv["llama.expert_count"] = p.NumLocalExperts
-	}
-
-	if p.NumExpertsPerToken > 0 {
-		kv["llama.expert_used_count"] = p.NumExpertsPerToken
-	}
-
-	return kv
-}
-
-func (p *mixtral) Tensors(ts []Tensor) []*llm.Tensor {
-	oldnew := []string{
-		"model.layers", "blk",
-		"w1", "ffn_gate_exps",
-		"w2", "ffn_down_exps",
-		"w3", "ffn_up_exps",
-	}
-
-	for i := range p.NumLocalExperts {
-		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
-	}
-
-	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
-	namer := strings.NewReplacer(oldnew...)
-	experts := make(map[string]experts)
-
-	// merge experts into a single tensor while removing them from ts
-	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
-		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
-			return false
-		}
-
-		name := namer.Replace(t.Name())
-		experts[name] = append(experts[name], t)
-		return true
-	})
-
-	var out []*llm.Tensor
-	for n, e := range experts {
-		// TODO(mxyng): sanity check experts
-		out = append(out, &llm.Tensor{
-			Name:     n,
-			Kind:     e[0].Kind(),
-			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
-			WriterTo: e,
-		})
-	}
-
-	return append(out, p.llama.Tensors(ts)...)
-}
-
-type experts []Tensor
-
-func (e experts) WriteTo(w io.Writer) (int64, error) {
-	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
-	for _, t := range e {
-		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
-		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
-		// this accomplishes the same thing by writing each expert tensor in sequence
-		if _, err := t.WriteTo(w); err != nil {
-			return 0, err
-		}
-	}
-
-	return 0, nil
-}

convert/convert_test.go

@@ -1,34 +1,48 @@
+//go:build slow
+
 package convert
 
 import (
-	"bytes"
-	"crypto/sha256"
-	"encoding/json"
-	"errors"
-	"flag"
-	"fmt"
-	"io"
-	"log/slog"
-	"math"
 	"os"
 	"path/filepath"
-	"slices"
 	"testing"
 
 	"github.com/ollama/ollama/llm"
-	"golang.org/x/exp/maps"
 )
 
-func convertFull(t *testing.T, d string) (*os.File, llm.KV, llm.Tensors) {
+func convertFull(t *testing.T, p string) (llm.KV, llm.Tensors) {
 	t.Helper()
 
+	mf, err := GetModelFormat(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	params, err := mf.GetParams(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	arch, err := mf.GetModelArch("", p, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if err := arch.LoadVocab(); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := arch.GetTensors(); err != nil {
+		t.Fatal(err)
+	}
+
 	f, err := os.CreateTemp(t.TempDir(), "f16")
 	if err != nil {
 		t.Fatal(err)
 	}
 	defer f.Close()
 
-	if err := Convert(d, f); err != nil {
+	if err := arch.WriteGGUF(f); err != nil {
 		t.Fatal(err)
 	}
 
@@ -36,200 +50,54 @@ func convertFull(t *testing.T, d string) (*os.File, llm.KV, llm.Tensors) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	t.Cleanup(func() { r.Close() })
+	defer r.Close()
 
-	m, _, err := llm.DecodeGGML(r, math.MaxInt)
+	m, _, err := llm.DecodeGGML(r)
 	if err != nil {
 		t.Fatal(err)
 	}
 
-	if _, err := r.Seek(0, io.SeekStart); err != nil {
-		t.Fatal(err)
-	}
-
-	return r, m.KV(), m.Tensors()
-}
-
-func TestMain(m *testing.M) {
-	var level slog.Level
-	flag.TextVar(&level, "level", slog.LevelInfo, "log level")
-	flag.Parse()
-	slog.SetLogLoggerLevel(level)
-	os.Exit(m.Run())
+	return m.KV(), m.Tensors()
 }
 
 func TestConvertFull(t *testing.T) {
-	cases := []string{
-		"Meta-Llama-3-8B-Instruct",
-		"Mistral-7B-Instruct-v0.2",
-		"Mixtral-8x7B-Instruct-v0.1",
-		"gemma-2b-it",
+	cases := []struct {
+		path    string
+		arch    string
+		tensors int
+		layers  int
+	}{
+		{"Meta-Llama-3-8B-Instruct", "llama", 291, 35},
+		{"Mistral-7B-Instruct-v0.2", "llama", 291, 35},
+		{"Mixtral-8x7B-Instruct-v0.1", "llama", 291, 35},
+		{"gemma-2b-it", "gemma", 164, 20},
 	}
 
-	for i := range cases {
-		tt := cases[i]
-		t.Run(tt, func(t *testing.T) {
-			t.Parallel()
-
-			p := filepath.Join("testdata", tt)
-			if testing.Short() {
-				t.Skip("skipping in short mode")
-			} else if _, err := os.Stat(p); err != nil {
+	for _, tt := range cases {
+		t.Run(tt.path, func(t *testing.T) {
+			p := filepath.Join("testdata", tt.path)
+			if _, err := os.Stat(p); err != nil {
 				t.Skipf("%s not found", p)
 			}
 
-			f, kv, tensors := convertFull(t, p)
-			actual := make(map[string]string)
-			for k, v := range kv {
-				if s, ok := v.(json.Marshaler); !ok {
-					actual[k] = fmt.Sprintf("%v", v)
-				} else {
-					bts, err := json.Marshal(s)
-					if err != nil {
-						t.Fatal(err)
-					}
+			kv, tensors := convertFull(t, p)
 
-					actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
-				}
+			if kv.Architecture() != tt.arch {
+				t.Fatalf("expected llama, got %s", kv.Architecture())
 			}
 
-			for _, tensor := range tensors.Items {
-				sha256sum := sha256.New()
-				sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
-				if _, err := io.Copy(sha256sum, sr); err != nil {
-					t.Fatal(err)
-				}
-
-				actual[tensor.Name] = fmt.Sprintf("%x", sha256sum.Sum(nil))
+			if kv.FileType().String() != "F16" {
+				t.Fatalf("expected F16, got %s", kv.FileType())
 			}
 
-			expectFile, err := os.Open(filepath.Join("testdata", fmt.Sprintf("%s.json", tt)))
-			if err != nil {
-				t.Fatal(err)
+			if len(tensors) != tt.tensors {
+				t.Fatalf("expected %d tensors, got %d", tt.tensors, len(tensors))
 			}
 
-			var expect map[string]string
-			if err := json.NewDecoder(expectFile).Decode(&expect); err != nil {
-				t.Fatal(err)
-			}
-
-			keys := maps.Keys(expect)
-			slices.Sort(keys)
-			for _, k := range keys {
-				if v, ok := actual[k]; !ok {
-					t.Errorf("missing %s", k)
-				} else if v != expect[k] {
-					t.Errorf("unexpected %s: want %s, got %s", k, expect[k], v)
-				}
+			layers := tensors.Layers()
+			if len(layers) != tt.layers {
+				t.Fatalf("expected %d layers, got %d", tt.layers, len(layers))
 			}
 		})
 	}
 }
-
-func TestConvertNPZ(t *testing.T) {
-	cases := []string{
-		"adapters.npz",
-	}
-
-	for _, fn := range cases {
-		ts, err := parseNPZ(filepath.Join("testdata", fn))
-		if err != nil {
-			t.Fatal(err)
-		}
-		if len(ts) != 16*2*2 {
-			t.Errorf("got: %d want: %d total layers", len(ts), 16*2*2)
-		}
-
-		a := adapter{}
-
-		for _, m := range ts {
-			at := m.(adapterTensor)
-			if at.path != filepath.Join("testdata", fn) {
-				t.Errorf("got: %s want: %s", at.path, filepath.Join("testdata", fn))
-			}
-			if at.dtype != "F32" {
-				t.Errorf("got: %s but only F32s are currently supported", at.dtype)
-			}
-			if len(at.tensorBase.shape) != 2 {
-				t.Errorf("got: %d want: %d tensor shape", at.tensorBase.shape, 2)
-			}
-		}
-
-		var ws io.WriteSeeker = &memWriter{}
-		err = llm.WriteGGLA(ws, a.KV(nil), a.Tensors(ts))
-		if err != nil {
-			t.Fatal(err)
-		}
-
-		mw := ws.(*memWriter)
-		slog.Info(fmt.Sprintf("buffer len = %d", len(mw.buf)))
-		if len(mw.buf) == 0 {
-			t.Errorf("ggla layer not written correctly")
-		}
-		rs := bytes.NewReader(mw.buf)
-		ggml, _, err := llm.DecodeGGML(rs, len(mw.buf))
-		if err != nil {
-			t.Fatal(err)
-		}
-		if ggml == nil {
-			t.Fatalf("ggla didn't convert to ggml correctly")
-		}
-
-		kv := ggml.KV()
-		if kv == nil {
-			t.Fatalf("no lora KVs were set")
-		}
-
-		r, ok := kv["r"]
-		if !ok || r != uint32(8) {
-			t.Errorf("lora rank was not set correctly")
-		}
-
-		alpha, ok := kv["alpha"]
-		if !ok || alpha != uint32(160) {
-			t.Errorf("lora alpha was not set correctly")
-		}
-
-		gts := ggml.Tensors()
-		if len(ts) != len(gts.Items) {
-			t.Fatalf("got: %d want: %d tensors in ggla", len(gts.Items), len(ts))
-		}
-	}
-}
-
-type memWriter struct {
-	buf []byte
-	pos int
-}
-
-func (m *memWriter) Write(p []byte) (n int, err error) {
-	minCap := m.pos + len(p)
-	if minCap > cap(m.buf) {
-		buf2 := make([]byte, len(m.buf), minCap+len(p)) // add some extra
-		copy(buf2, m.buf)
-		m.buf = buf2
-	}
-	if minCap > len(m.buf) {
-		m.buf = m.buf[:minCap]
-	}
-	copy(m.buf[m.pos:], p)
-	m.pos += len(p)
-	return len(p), nil
-}
-
-func (m *memWriter) Seek(offset int64, whence int) (int64, error) {
-	newPos, offs := 0, int(offset)
-	switch whence {
-	case io.SeekStart:
-		newPos = offs
-	case io.SeekCurrent:
-		newPos = m.pos + offs
-	case io.SeekEnd:
-		newPos = len(m.buf) + offs
-	}
-	if newPos < 0 {
-		return 0, errors.New("negative result pos")
-	}
-	m.pos = newPos
-	return int64(newPos), nil
-}

convert/gemma.go

@@ -0,0 +1,103 @@
+package convert
+
+import (
+	"fmt"
+	"io"
+	"log/slog"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type GemmaModel struct {
+	ModelData
+}
+
+func addOnes(data []float32, vectorSize int) ([]float32, error) {
+	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
+	ones := tensor.Ones(tensor.Float32, vectorSize)
+
+	n, err := n.Add(ones)
+	if err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+
+	return f32s, nil
+}
+
+func (m *GemmaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
+	for _, l := range t {
+		if strings.HasSuffix(l.Name, "norm.weight") {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.repacker = m.Repack
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *GemmaModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *GemmaModel) Repack(_ string, data []float32, shape []uint64) ([]float32, error) {
+	return addOnes(data, int(shape[0]))
+}
+
+func (m *GemmaModel) WriteGGUF(ws io.WriteSeeker) error {
+	kv := llm.KV{
+		"general.architecture":                   "gemma",
+		"general.name":                           m.Name,
+		"gemma.context_length":                   uint32(m.Params.ContextSize),
+		"gemma.embedding_length":                 uint32(m.Params.HiddenSize),
+		"gemma.block_count":                      uint32(m.Params.HiddenLayers),
+		"gemma.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"gemma.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"gemma.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"gemma.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"gemma.attention.key_length":             uint32(m.Params.HeadDimension),
+		"gemma.attention.value_length":           uint32(m.Params.HeadDimension),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.padding_token_id": uint32(m.Params.PaddingTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(3),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
+}

convert/llama.go

@@ -0,0 +1,158 @@
+package convert
+
+import (
+	"cmp"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type LlamaModel struct {
+	ModelData
+}
+
+func (m *LlamaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			switch m.Format.(type) {
+			case *TorchFormat:
+				wt := l.WriterTo.(torchWriterTo)
+				wt.repacker = m.Repack
+				l.WriterTo = wt
+			case *SafetensorFormat:
+				wt := l.WriterTo.(safetensorWriterTo)
+				wt.repacker = m.Repack
+				l.WriterTo = wt
+			}
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *LlamaModel) LoadVocab() (err error) {
+	pre, ts, merges, err := parseTokens(filepath.Join(m.Path, "tokenizer.json"))
+	if errors.Is(err, os.ErrNotExist) {
+		return nil
+	} else if err != nil {
+		return err
+	}
+
+	m.Vocab = &Vocab{}
+	for _, t := range ts {
+		m.Vocab.Tokens = append(m.Vocab.Tokens, t.Content)
+		m.Vocab.Types = append(m.Vocab.Types, t.Type())
+	}
+
+	m.Vocab.Merges = merges
+	m.Params.PreTokenizer = pre
+	return nil
+}
+
+func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
+	kv := llm.KV{
+		"general.architecture":                   "llama",
+		"general.name":                           m.Name,
+		"llama.vocab_size":                       uint32(len(m.Vocab.Tokens)),
+		"llama.context_length":                   uint32(m.Params.ContextSize),
+		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
+		"llama.block_count":                      uint32(m.Params.HiddenLayers),
+		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
+		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "gpt2",
+
+		"tokenizer.ggml.pre":        m.Params.PreTokenizer,
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+	}
+
+	if len(m.Vocab.Merges) > 0 {
+		kv["tokenizer.ggml.merges"] = m.Vocab.Merges
+	} else {
+		kv["tokenizer.ggml.scores"] = m.Vocab.Scores
+	}
+
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
+}
+
+func (m *LlamaModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}
+
+func llamaRepack(name string, params *Params, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		if dim != 0 {
+			dims = append(dims, int(dim))
+		}
+	}
+
+	var heads int
+	if strings.HasSuffix(name, "attn_q.weight") {
+		heads = params.AttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight") {
+		heads = cmp.Or(params.KeyValHeads, params.AttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor name: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{heads, 2, dims[0] / heads / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/mistral.go

@@ -0,0 +1,79 @@
+package convert
+
+import (
+	"io"
+	"regexp"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type MistralModel struct {
+	ModelData
+}
+
+func (m *MistralModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.repacker = m.Repack
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *MistralModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
+	kv := llm.KV{
+		"general.architecture":                   "llama",
+		"general.name":                           m.Name,
+		"llama.context_length":                   uint32(m.Params.ContextSize),
+		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
+		"llama.block_count":                      uint32(m.Params.HiddenLayers),
+		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+	}
+
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
+}
+
+func (m *MistralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}

convert/mixtral.go

@@ -0,0 +1,87 @@
+package convert
+
+import (
+	"io"
+	"regexp"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type MixtralModel struct {
+	ModelData
+}
+
+func (m *MixtralModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.repacker = m.Repack
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *MixtralModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *MixtralModel) WriteGGUF(ws io.WriteSeeker) error {
+	kv := llm.KV{
+		"general.architecture":          "llama",
+		"general.name":                  m.Name,
+		"llama.block_count":             uint32(m.Params.HiddenLayers),
+		"llama.context_length":          uint32(m.Params.ContextSize),
+		"llama.embedding_length":        uint32(m.Params.HiddenSize),
+		"llama.feed_forward_length":     uint32(m.Params.IntermediateSize),
+		"llama.attention.head_count":    uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
+
+		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+
+		"llama.expert_count":      uint32(m.Params.Experts),
+		"llama.expert_used_count": uint32(m.Params.ExpertsUsed),
+
+		"llama.vocab_size":           uint32(len(m.Vocab.Tokens)),
+		"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+
+		"general.file_type":    uint32(1),
+		"tokenizer.ggml.model": "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
+}
+
+func (m *MixtralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	return llamaRepack(name, m.Params, data, shape)
+}

convert/reader.go

@@ -1,74 +0,0 @@
-package convert
-
-import (
-	"errors"
-	"io"
-	"path/filepath"
-	"strings"
-)
-
-type Tensor interface {
-	Name() string
-	Shape() []uint64
-	Kind() uint32
-	SetRepacker(repacker)
-	WriteTo(io.Writer) (int64, error)
-}
-
-type tensorBase struct {
-	name  string
-	shape []uint64
-	repacker
-}
-
-func (t tensorBase) Name() string {
-	return t.name
-}
-
-func (t tensorBase) Shape() []uint64 {
-	return t.shape
-}
-
-func (t tensorBase) Kind() uint32 {
-	if strings.HasSuffix(t.name, ".block_sparse_moe.gate.weight") {
-		return 0
-	}
-
-	switch len(t.shape) {
-	case 0:
-		panic("invalid tensor shape")
-	case 1:
-		return 0
-	default:
-		return 1
-	}
-}
-
-func (t *tensorBase) SetRepacker(fn repacker) {
-	t.repacker = fn
-}
-
-type repacker func(string, []float32, []uint64) ([]float32, error)
-
-func parseTensors(d string) ([]Tensor, error) {
-	patterns := map[string]func(...string) ([]Tensor, error){
-		"model-*-of-*.safetensors": parseSafetensors,
-		"model.safetensors":        parseSafetensors,
-		"pytorch_model-*-of-*.bin": parseTorch,
-		"pytorch_model.bin":        parseTorch,
-		"consolidated.*.pth":       parseTorch,
-	}
-
-	for pattern, parseFn := range patterns {
-		matches, err := filepath.Glob(filepath.Join(d, pattern))
-		if err != nil {
-			return nil, err
-		}
-
-		if len(matches) > 0 {
-			return parseFn(matches...)
-		}
-	}
-
-	return nil, errors.New("unknown tensor format")
-}

convert/reader_npz.go

@@ -1,140 +0,0 @@
-package convert
-
-import (
-	"encoding/binary"
-	"fmt"
-	"io"
-	"log/slog"
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-	"github.com/sbinet/npyio/npz"
-)
-
-type adapterTensor struct {
-	path  string
-	dtype string
-	*tensorBase
-}
-
-func DetectNPZ(fn string) (bool, error) {
-	f, err := npz.Open(fn)
-	if err != nil {
-		return false, err
-	}
-	defer f.Close()
-
-	if len(f.Keys()) > 0 && strings.HasSuffix(f.Keys()[0], ".npy") {
-		return true, nil
-	}
-
-	return false, nil
-}
-
-func parseNPZ(fn string) ([]Tensor, error) {
-	var ts []Tensor
-
-	f, err := npz.Open(fn)
-	if err != nil {
-		return nil, err
-	}
-	defer f.Close()
-
-	for _, name := range f.Keys() {
-		slog.Info(fmt.Sprintf("reading layer '%s'", name))
-		h := f.Header(name)
-
-		shape := make([]uint64, 2)
-		for cnt, v := range h.Descr.Shape {
-			// llamacpp expects the loraB layer to be reversed
-			if strings.Contains(name, "lora_b") {
-				shape[len(shape)-cnt-1] = uint64(v)
-			} else {
-				shape[cnt] = uint64(v)
-			}
-		}
-
-		dtypeMap := map[string]string{
-			"<f2": "F16",
-			"<f4": "F32",
-		}
-		dtype, ok := dtypeMap[h.Descr.Type]
-		if !ok {
-			return nil, fmt.Errorf("Unknown type '%s' for '%s'", h.Descr.Type, name)
-		}
-
-		ts = append(ts, adapterTensor{
-			path:  fn,
-			dtype: dtype,
-			tensorBase: &tensorBase{
-				name:  name,
-				shape: shape,
-			},
-		})
-	}
-	return ts, nil
-}
-
-func (t adapterTensor) Kind() uint32 {
-	switch t.dtype {
-	case "F32":
-		return 0
-	case "F16":
-		return 1
-	}
-	return 0
-}
-
-func (t adapterTensor) WriteTo(w io.Writer) (int64, error) {
-	f, err := npz.Open(t.path)
-	if err != nil {
-		return 0, err
-	}
-	defer f.Close()
-
-	switch t.dtype {
-	case "F32":
-		var f32s []float32
-		err = f.Read(t.tensorBase.name, &f32s)
-		if err != nil {
-			return 0, err
-		}
-
-		// ggla expects the loraB to be transposed
-		if strings.Contains(t.tensorBase.name, "lora_b") {
-			f32s, err = transpose(f32s, t.tensorBase.shape)
-			if err != nil {
-				return 0, err
-			}
-		}
-
-		return 0, binary.Write(w, binary.LittleEndian, f32s)
-	}
-
-	return 0, fmt.Errorf("unknown data type: %s", t.dtype)
-}
-
-func transpose(f32s []float32, shape []uint64) ([]float32, error) {
-	if len(shape) != 2 {
-		return nil, fmt.Errorf("only 2 dimensions supported for transpose")
-	}
-
-	// the shape is already backward
-	n := tensor.New(tensor.WithShape(int(shape[1]), int(shape[0])), tensor.WithBacking(f32s))
-	if err := n.T(1, 0); err != nil {
-		return nil, err
-	}
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-	ts, err := native.SelectF32(n, 1)
-	if err != nil {
-		return nil, err
-	}
-	f32s = make([]float32, 0)
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-	return f32s, nil
-}

convert/reader_safetensors.go

@@ -1,140 +0,0 @@
-package convert
-
-import (
-	"bytes"
-	"encoding/binary"
-	"encoding/json"
-	"fmt"
-	"io"
-	"os"
-	"slices"
-
-	"github.com/d4l3k/go-bfloat16"
-	"github.com/x448/float16"
-	"golang.org/x/exp/maps"
-)
-
-type safetensorMetadata struct {
-	Type    string   `json:"dtype"`
-	Shape   []uint64 `json:"shape"`
-	Offsets []int64  `json:"data_offsets"`
-}
-
-func parseSafetensors(ps ...string) ([]Tensor, error) {
-	var ts []Tensor
-	for _, p := range ps {
-		f, err := os.Open(p)
-		if err != nil {
-			return nil, err
-		}
-		defer f.Close()
-
-		var n int64
-		if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
-			return nil, err
-		}
-
-		b := bytes.NewBuffer(make([]byte, 0, n))
-		if _, err = io.CopyN(b, f, n); err != nil {
-			return nil, err
-		}
-
-		var headers map[string]safetensorMetadata
-		if err := json.NewDecoder(b).Decode(&headers); err != nil {
-			return nil, err
-		}
-
-		keys := maps.Keys(headers)
-		slices.Sort(keys)
-
-		for _, key := range keys {
-			if value := headers[key]; value.Type != "" {
-				ts = append(ts, safetensor{
-					path:   p,
-					dtype:  value.Type,
-					offset: safetensorsPad(n, value.Offsets[0]),
-					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
-					tensorBase: &tensorBase{
-						name:  key,
-						shape: value.Shape,
-					},
-				})
-			}
-		}
-	}
-
-	return ts, nil
-}
-
-func safetensorsPad(n, s int64) int64 {
-	return 8 + n + s
-}
-
-type safetensor struct {
-	path   string
-	dtype  string
-	offset int64
-	size   int64
-	*tensorBase
-}
-
-func (st safetensor) WriteTo(w io.Writer) (int64, error) {
-	f, err := os.Open(st.path)
-	if err != nil {
-		return 0, err
-	}
-	defer f.Close()
-
-	if _, err = f.Seek(st.offset, io.SeekStart); err != nil {
-		return 0, err
-	}
-
-	var f32s []float32
-	switch st.dtype {
-	case "F32":
-		f32s = make([]float32, st.size/4)
-		if err = binary.Read(f, binary.LittleEndian, f32s); err != nil {
-			return 0, err
-		}
-	case "F16":
-		u16s := make([]uint16, st.size/2)
-		if err = binary.Read(f, binary.LittleEndian, u16s); err != nil {
-			return 0, err
-		}
-
-		for _, b := range u16s {
-			f32s = append(f32s, float16.Frombits(b).Float32())
-		}
-
-	case "BF16":
-		u8s := make([]uint8, st.size)
-		if err = binary.Read(f, binary.LittleEndian, u8s); err != nil {
-			return 0, err
-		}
-
-		f32s = bfloat16.DecodeFloat32(u8s)
-	default:
-		return 0, fmt.Errorf("unknown data type: %s", st.dtype)
-	}
-
-	if st.repacker != nil {
-		f32s, err = st.repacker(st.Name(), f32s, st.Shape())
-		if err != nil {
-			return 0, err
-		}
-	}
-
-	switch st.Kind() {
-	case 0:
-		return 0, binary.Write(w, binary.LittleEndian, f32s)
-	case 1:
-		f16s := make([]uint16, len(f32s))
-		for i := range f32s {
-			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
-		}
-
-		return 0, binary.Write(w, binary.LittleEndian, f16s)
-	default:
-		return 0, fmt.Errorf("unknown storage type: %d", st.Kind())
-	}
-}

convert/reader_torch.go

@@ -1,46 +0,0 @@
-package convert
-
-import (
-	"io"
-
-	"github.com/nlpodyssey/gopickle/pytorch"
-	"github.com/nlpodyssey/gopickle/types"
-)
-
-func parseTorch(ps ...string) ([]Tensor, error) {
-	var ts []Tensor
-	for _, p := range ps {
-		pt, err := pytorch.Load(p)
-		if err != nil {
-			return nil, err
-		}
-
-		for _, k := range pt.(*types.Dict).Keys() {
-			t := pt.(*types.Dict).MustGet(k)
-
-			var shape []uint64
-			for dim := range t.(*pytorch.Tensor).Size {
-				shape = append(shape, uint64(dim))
-			}
-
-			ts = append(ts, torch{
-				storage: t.(*pytorch.Tensor).Source,
-				tensorBase: &tensorBase{
-					name:  k.(string),
-					shape: shape,
-				},
-			})
-		}
-	}
-
-	return ts, nil
-}
-
-type torch struct {
-	storage pytorch.StorageInterface
-	*tensorBase
-}
-
-func (pt torch) WriteTo(w io.Writer) (int64, error) {
-	return 0, nil
-}

convert/safetensors.go

@@ -0,0 +1,309 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"regexp"
+	"slices"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type safetensorWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	filename string
+	dtype    string
+
+	offset, size int64
+	repacker     func(string, []float32, []uint64) ([]float32, error)
+}
+
+type safetensorMetadata struct {
+	Type    string   `json:"dtype"`
+	Shape   []uint64 `json:"shape"`
+	Offsets []int64  `json:"data_offsets"`
+}
+
+type SafetensorFormat struct{}
+
+func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	var tensors []llm.Tensor
+	matches, err := filepath.Glob(filepath.Join(dirpath, "*.safetensors"))
+	if err != nil {
+		return nil, err
+	}
+
+	var offset uint64
+	for _, f := range matches {
+		var t []llm.Tensor
+		var err error
+		t, offset, err = m.readTensors(f, offset, params)
+		if err != nil {
+			return nil, err
+		}
+
+		tensors = append(tensors, t...)
+	}
+	return tensors, nil
+}
+
+func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params) ([]llm.Tensor, uint64, error) {
+	f, err := os.Open(fn)
+	if err != nil {
+		return nil, 0, err
+	}
+	defer f.Close()
+
+	var n int64
+	if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
+		return nil, 0, err
+	}
+
+	b := bytes.NewBuffer(make([]byte, 0, n))
+	if _, err = io.CopyN(b, f, n); err != nil {
+		return nil, 0, err
+	}
+
+	var headers map[string]safetensorMetadata
+	if err := json.NewDecoder(b).Decode(&headers); err != nil {
+		return nil, 0, err
+	}
+
+	var keys []string
+	for key := range headers {
+		if !strings.HasSuffix(key, "self_attn.rotary_embd.inv_freq") {
+			keys = append(keys, key)
+		}
+	}
+
+	slices.Sort(keys)
+
+	var tensors []llm.Tensor
+	for _, key := range keys {
+		value := headers[key]
+
+		var kind uint32
+		switch len(value.Shape) {
+		case 0:
+			// valuedata
+			continue
+		case 2:
+			kind = 1
+		}
+
+		name, err := m.GetLayerName(key)
+		if err != nil {
+			return nil, 0, err
+		}
+
+		shape := make([]uint64, len(value.Shape))
+		copy(shape, value.Shape)
+
+		pad := func(s int64) int64 {
+			return 8 + n + s
+		}
+
+		t := llm.Tensor{
+			Name:   name,
+			Kind:   kind,
+			Offset: offset,
+			Shape:  shape[:],
+		}
+
+		t.WriterTo = safetensorWriterTo{
+			t:        &t,
+			params:   params,
+			bo:       params.ByteOrder,
+			filename: fn,
+			dtype:    value.Type,
+			offset:   pad(value.Offsets[0]),
+			size:     pad(value.Offsets[1]) - pad(value.Offsets[0]),
+		}
+
+		offset += t.Size()
+		tensors = append(tensors, t)
+	}
+
+	return tensors, offset, nil
+}
+
+func (m *SafetensorFormat) GetParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "config.json"))
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var params Params
+
+	if err := json.NewDecoder(f).Decode(&params); err != nil {
+		return nil, err
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return &params, nil
+}
+
+func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
+	directMap := map[string]string{
+		"model.embed_tokens.weight": "token_embd.weight",
+		"lm_head.weight":            "output.weight",
+		"model.norm.weight":         "output_norm.weight",
+	}
+
+	tMap := map[string]string{
+		"model.layers.(\\d+).input_layernorm.weight":                    "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":                      "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":                      "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":                        "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight":           "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":                   "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":                   "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":                   "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":                   "blk.$1.attn_v.weight",
+		"model.layers.(\\d+).block_sparse_moe.gate.weight":              "blk.$1.ffn_gate_inp.weight",
+		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w1.weight": "blk.$1.ffn_gate.$2.weight",
+		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w2.weight": "blk.$1.ffn_down.$2.weight",
+		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w3.weight": "blk.$1.ffn_up.$2.weight",
+	}
+
+	v, ok := directMap[n]
+	if ok {
+		return v, nil
+	}
+
+	// quick hack to rename the layers to gguf format
+	for k, v := range tMap {
+		re := regexp.MustCompile(k)
+		newName := re.ReplaceAllString(n, v)
+		if newName != n {
+			return newName, nil
+		}
+	}
+
+	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
+}
+
+func (r safetensorWriterTo) WriteTo(w io.Writer) (n int64, err error) {
+	f, err := os.Open(r.filename)
+	if err != nil {
+		return 0, err
+	}
+	defer f.Close()
+
+	if _, err = f.Seek(r.offset, io.SeekStart); err != nil {
+		return 0, err
+	}
+
+	var f32s []float32
+	switch r.dtype {
+	case "F32":
+		f32s = make([]float32, r.size/4)
+		if err = binary.Read(f, r.bo, f32s); err != nil {
+			return 0, err
+		}
+	case "F16":
+		u16s := make([]uint16, r.size/2)
+		if err = binary.Read(f, r.bo, u16s); err != nil {
+			return 0, err
+		}
+
+		for _, b := range u16s {
+			f32s = append(f32s, float16.Frombits(b).Float32())
+		}
+
+	case "BF16":
+		u8s := make([]uint8, r.size)
+		if err = binary.Read(f, r.bo, u8s); err != nil {
+			return 0, err
+		}
+
+		f32s = bfloat16.DecodeFloat32(u8s)
+	default:
+		return 0, fmt.Errorf("unknown data type: %s", r.dtype)
+	}
+
+	if r.repacker != nil {
+		f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	switch r.t.Kind {
+	case 0:
+		return 0, binary.Write(w, r.bo, f32s)
+	case 1:
+		f16s := make([]uint16, len(f32s))
+		for i := range f32s {
+			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
+		}
+
+		return 0, binary.Write(w, r.bo, f16s)
+	default:
+		return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
+	}
+}
+
+func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
+	switch len(params.Architectures) {
+	case 0:
+		return nil, fmt.Errorf("No architecture specified to convert")
+	case 1:
+		switch params.Architectures[0] {
+		case "LlamaForCausalLM":
+			return &LlamaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		case "MistralForCausalLM":
+			return &MistralModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		case "MixtralForCausalLM":
+			return &MixtralModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		case "GemmaForCausalLM":
+			return &GemmaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		default:
+			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
+		}
+	}
+
+	return nil, fmt.Errorf("Unknown error")
+}

convert/testdata/Meta-Llama-3-8B-Instruct.json

@@ -1,313 +0,0 @@
-{
-  "general.architecture": "llama",
-  "general.file_type": "1",
-  "general.quantization_version": "2",
-  "llama.block_count": "32",
-  "llama.context_length": "8192",
-  "llama.embedding_length": "4096",
-  "llama.feed_forward_length": "14336",
-  "llama.rope.dimension_count": "128",
-  "llama.rope.freq_base": "500000",
-  "llama.vocab_size": "128256",
-  "llama.attention.head_count": "32",
-  "llama.attention.head_count_kv": "8",
-  "llama.attention.layer_norm_rms_epsilon": "1e-05",
-  "tokenizer.ggml.model": "gpt2",
-  "tokenizer.ggml.pre": "llama-bpe",
-  "tokenizer.ggml.bos_token_id": "128000",
-  "tokenizer.ggml.eos_token_id": "128009",
-  "tokenizer.ggml.merges": "d0cbac1fcc9dcf03724b8db5c9bfb593ae1cf68fb9bc72eb1d15274dcbbf618b",
-  "tokenizer.ggml.token_type": "d70a88809fd7da6f1f028622685cd64268a7a922c5d343c96f25b66327358978",
-  "tokenizer.ggml.tokens": "765b529dbcbc42dd202ce657341c63807b51f3b07e09898f6aa6196326865d5a",
-  "token_embd.weight": "b53102a11d9064bbd404833e3464b1b13e08ce73300b442312cccde2f19b2698",
-  "blk.0.attn_norm.weight": "7318df3cca9e8d153ff0a503026a1265e63d20b2a8c1dd7a2769585082b5d1ee",
-  "blk.0.ffn_down.weight": "b950806a1fc722c9fad7fd0b20c3c0a7fb50f14395e1e7663a590bfd62e20900",
-  "blk.0.ffn_gate.weight": "e73e580af6d4f08e060a74a3c25efdf5d3bed99e183d95a5a85ae859014839fd",
-  "blk.0.ffn_up.weight": "c8158af679ef99746da1befb67eebb19489e0bbe6ce7d97e13e348508244e516",
-  "blk.0.ffn_norm.weight": "7ec69c3c31e95e49a3359003b0033f6b9e85561a3e3fd83e7476661ecdd756bb",
-  "blk.0.attn_k.weight": "2732303257bac969b4964e0e32ec08b5a7f5c031bb02bf6ac4467b3ea0ebcf1e",
-  "blk.0.attn_output.weight": "ecda1d43b4ccc91cd5b366d7e7a275353990ac78561a07c83d9c77031aba12dc",
-  "blk.0.attn_q.weight": "569b1f5faf92b6f00910cf7effb2d5862f91038ce5c3b0019fc10e5d79fbd5e1",
-  "blk.0.attn_v.weight": "aa8416c5ef7e32fb54a1f20d6ac651656845d4af240564b397c39bd83e06e3b8",
-  "blk.1.attn_norm.weight": "03327e02862908c2a44b2f52decdb924bf4201f400b46f8037a9cb2e1d7a61ff",
-  "blk.1.ffn_down.weight": "5a83a87603f38c99f8e1e370a2d5f967bb45ac51d881a609304a7811027321e0",
-  "blk.1.ffn_gate.weight": "31da0572c79e655186c721c231376f85e56cdcc6257c28d08c8c5b40d5c22b40",
-  "blk.1.ffn_up.weight": "e0c811d64ca155c8de10a868e72015d43888834804614ee1aa2953129ffbc90f",
-  "blk.1.ffn_norm.weight": "5861f313d6137d6f0f904d423df47fffc6069e224ff746e1b637ac9c7f0af862",
-  "blk.1.attn_k.weight": "5fbbec0acca6457b9416ebdcd90e526885d0224537b7628f6be376a7f275313d",
-  "blk.1.attn_output.weight": "b237c9763fa3f75166a6f70b70f1566e77d0d89dfa164ed1b3137393e90575c3",
-  "blk.1.attn_q.weight": "c0a9cf4a98b4882b16f3eb2b49d933793dcc5357abb246fd3fe3134ed2b12e1c",
-  "blk.1.attn_v.weight": "96867111727200cac1af7865189dd41fd62b47584e5e5f33a91f1d34509cbd40",
-  "blk.2.attn_norm.weight": "f392f8a88ee3a95b1cc19c40dd4ef66317037b0faaa1800f610779e129ee0539",
-  "blk.2.ffn_down.weight": "73823eef46632aedcc8c1cb08a736b6aa97ca97842cd1fdfc5567d8dec459662",
-  "blk.2.ffn_gate.weight": "f4909ae19fc3848b00bb8b9050122e74f8e903b89e22937036f4cc9fea20a718",
-  "blk.2.ffn_up.weight": "16f4904a3d814ea68f00519724fc4943e48444a84c786bda39aa5efc298a7d84",
-  "blk.2.ffn_norm.weight": "e3ccdf56e75cb969f6f69c39caf6daf7c4e70e89e25df0f4d2e4bc60e159aafe",
-  "blk.2.attn_k.weight": "c3beb1e0a11bcf007ef0f0d8f6bdd3082d8b29090cd29597846b5d51e308a8e5",
-  "blk.2.attn_output.weight": "bb9f66c32cff51154fea92933c2cd62549236f8cb1a767f9ef28d3f99809b343",
-  "blk.2.attn_q.weight": "8eba394132eef2a05c5a92d62d2376000f7948448d7a2dc74e6b608203add20d",
-  "blk.2.attn_v.weight": "88f61f77c53567c617db3eef8f30621109a750e679f6784f7911739bd42c2f02",
-  "blk.3.attn_norm.weight": "7b996675b7ca75fa24107b3ebe0788653ede0f49ac83b8659d71ff54d591f81a",
-  "blk.3.ffn_down.weight": "2cb332bc05e4821962fdc9dcbcc7cc12630f32117711b687d18fb53c0bc4fbf4",
-  "blk.3.ffn_gate.weight": "340b387c7f208c8f0a6db904ef8d87c1e84b7d6ad57177abd32d86c8d18b760f",
-  "blk.3.ffn_up.weight": "07484433f8a7ee061c55aa0de2ecc009f769b0617c9c0ec096e9bb2946df9f0e",
-  "blk.3.ffn_norm.weight": "4f1a4ade36b393af341240bc894a2aab09cff7e4d56dc4658445deb107f9371b",
-  "blk.3.attn_k.weight": "483dcd96acb4528df84b9842970994630dbd82b8715ace394aa8b39fcf8d6291",
-  "blk.3.attn_output.weight": "beaff0810687923585642ee11d929cbf3b43dc6f87f30ddb552c222ab57bdbb3",
-  "blk.3.attn_q.weight": "0739355002f6fce520863add697e0ff25fc88215322dc3f993be7bb68dcce7e8",
-  "blk.3.attn_v.weight": "c216d17b6d90ee3e07f82598b8161fae34de2f392dbb0f745b682b578c324767",
-  "blk.4.attn_norm.weight": "91ab405bc4ba15bf63af233f266aa43aaab43789a9e6596e14a357c2ac7df217",
-  "blk.4.ffn_down.weight": "620f34ee75cdc73aecb8949af5fbb0d2437fd81422b6d8eb7acfc52addb9fc68",
-  "blk.4.ffn_gate.weight": "f6feec7bc9acadf35ec22532f8998d8e50f31afedabb19263590dcf8b9a92eee",
-  "blk.4.ffn_up.weight": "4a72af7cd28fd07b038f6cc4406678d120517280236ea85d9e76eff40ab2cc22",
-  "blk.4.ffn_norm.weight": "1805b37b44d5d682bdbd2fadeafb763ee001617d7870848cc487079ee34b21f9",
-  "blk.4.attn_k.weight": "a1e4f9d97cdf4c1b0d177cf00c4e32d1be30c1984a239b3c9bd73f8848888853",
-  "blk.4.attn_output.weight": "a1547e2497c423b0aff0eee71d9300d6fdf4e4986679418b6e637b69a9a6720b",
-  "blk.4.attn_q.weight": "0677483a9264ea6803d03d304d87a54632242cb516e8b76b6e3e8284c2f4de04",
-  "blk.4.attn_v.weight": "02691ba3af344fcc1969428ab0df811ac94aaa2fd91b0dc4ec1ac0a58806980d",
-  "blk.5.attn_norm.weight": "ba9c028335e5c895b87a5bd1448ca429248f9746ed97bdcb8679923206117156",
-  "blk.5.ffn_down.weight": "ccfdc9006acad1940a6bc05042a3947f1066acd671e0bb53b7684e9eea9ef5c9",
-  "blk.5.ffn_gate.weight": "623157679f1e742ccc3807c0b0153ddc8450104de75ec62f1370ec3807c09cf4",
-  "blk.5.ffn_up.weight": "05748804c65091f963729b58b085f58351891cac8a2861f5eae26b06aa60b2a0",
-  "blk.5.ffn_norm.weight": "84bae55af2efc8b8429f09056c8c04990c466dae31cb3f9356038b8957f1b406",
-  "blk.5.attn_k.weight": "8c766180c726b037d587fc52371de6e3307140c52409011609d1225624b6a3eb",
-  "blk.5.attn_output.weight": "490b582b3b1dc151ae55aee8b6743dad6c01fb49e43afefb6e68394b74be3d73",
-  "blk.5.attn_q.weight": "6f7b8ca4d9025ec836a44bbcca46be30c66b471a9fb62943ddff8288b3731409",
-  "blk.5.attn_v.weight": "9f70df3ba00c9e723214b3da83ff435a2163fff5915f75515c9664c05c866c27",
-  "blk.6.attn_norm.weight": "1a4a66613a682df6f061fc7c4d986f9f7e9175b62f0c42fc1ef31db536bd5942",
-  "blk.6.ffn_down.weight": "c56f25e4e49b443dbc82d88311ee63bc1f5002cc67e52f4787fd5f003aedeac1",
-  "blk.6.ffn_gate.weight": "31a5cf1aa9b831a81588d508550f51fc425f9517c43254d4ef7096d38029cf04",
-  "blk.6.ffn_up.weight": "ce135f3a1163e0c9297a615bdbe68a67ead21edce8debbfa9f6e15e6af8d4c94",
-  "blk.6.ffn_norm.weight": "4e328ce0648c94e732bc40501858ef6262ad1161e2e407b0cdcf4813fa9d45d8",
-  "blk.6.attn_k.weight": "1eb1c4c9f9c4c7ff7f5429075e0dc6a7782bed55109fa88df209a817dd8ef960",
-  "blk.6.attn_output.weight": "3d32986b56873b88655ee1edabdd413fdd9ab18b82108c9ce90bdbc2d3a6f3a3",
-  "blk.6.attn_q.weight": "8432f583b3a2809c99c393f9beb077cb0534dd5d247c17108f2986cadc6651f6",
-  "blk.6.attn_v.weight": "5045381513815bb91839dbac8335ffe49bbc7b0008369de7ea97eb676c5e2b36",
-  "blk.7.attn_norm.weight": "3dabd003638ec2499bfc8a48c49eef34276caab4fe76894eb963207848c2fdaf",
-  "blk.7.ffn_down.weight": "194fae858608bdcffd235be59ab119d0b91c8549f864ea06dae69249e099935f",
-  "blk.7.ffn_gate.weight": "00b24c29c30246892bce0791be804a89701d4c1332777e0bcdad5d9d5666604f",
-  "blk.7.ffn_up.weight": "44d7082a5280080c90cef9e19d410391de34f212ca0736377769b8ddd0c82d5e",
-  "blk.7.ffn_norm.weight": "21fe8a7fd6911c64e0d15a788b3b4cb6d71dd6ec51de65f760ee89afbb6ae53e",
-  "blk.7.attn_k.weight": "57a149eec5f6744a9526cd3925ac073f9d12db0fbcb5afe042ef4dc846458c44",
-  "blk.7.attn_output.weight": "0e9c28a3e81a2880251ce5eed77bcb8be8aaa1a51c9cb6de820b47ed83849fc2",
-  "blk.7.attn_q.weight": "15ee75263ee4e2a43eb322bc159ae004bb7d77e3a7e63ee4ddab700430693fff",
-  "blk.7.attn_v.weight": "440aa970bba4bff429fd7b7b1de21f2ad14fb2952b776cfa4acee68d7c6e9b8f",
-  "blk.8.attn_norm.weight": "af5b44825633c42c1ae964c82bb2be6a242d3a751f0a91f1bae4f593e8f5b6ec",
-  "blk.8.ffn_down.weight": "b11c14c76adca94fa200496dd2c10743becb23aab6642443ef1ae6d8710edbc1",
-  "blk.8.ffn_gate.weight": "7bb03d3325bf8637ae2fa1296b0651356515578d46a7c5ca65c7a923d7de27bc",
-  "blk.8.ffn_up.weight": "b956ef0a0669b5a9c9bf3a8da2d1c24f52d331cfb7354f6d7c51bd65be355e30",
-  "blk.8.ffn_norm.weight": "c78c3d748302edfef76f71ea5cb2055c94352122eee8b9b1173779a1814d224e",
-  "blk.8.attn_k.weight": "c0fba6a596ed9c1c32a7055c31a935a8b31e42b77282ee47c1f03ee3bde736b5",
-  "blk.8.attn_output.weight": "83cf9947080c5d8d571f04a842bc3dcfe7bbb0195fb25b346e22635e8649f2d4",
-  "blk.8.attn_q.weight": "47409350a576b333d97b7c877d69f47f46df504f3765102dfc0be9e521c7ecd6",
-  "blk.8.attn_v.weight": "1999dff91404fdcf1ecb34d9eaaaa9244ec7658a74dec8feb7cfd1fddba0347e",
-  "blk.9.attn_norm.weight": "1e6e29d5c3889ab4e1b0a5b9998cba60179b0f1fca133515df49cbc19d092593",
-  "blk.9.ffn_down.weight": "acb898a6490adff592e10b4c62d70edc5941661ee6da44658500e9205357c8e9",
-  "blk.9.ffn_gate.weight": "4cff63013593aadc3ffbaaa6ed70ffdba1224cd43c3644bf6f4162b5ac1ab542",
-  "blk.9.ffn_up.weight": "f985b5a2d6cf4fe32c7256301c3c89b8ad22b59e516342c52da42d8110766a4e",
-  "blk.9.ffn_norm.weight": "0d659c538bc6b21ed0018f107ab674a7424a00a42946c80e07208b479b21918f",
-  "blk.9.attn_k.weight": "f67611d888780d1b38c1c146b361c65310c8183bdf64fd73e2259985c6e8517f",
-  "blk.9.attn_output.weight": "f12ca1fa62a02ddc3f77f798bfb5707e0c50bf18ee0eaa67025521a98355f26b",
-  "blk.9.attn_q.weight": "3865185f4361a645b086ad47b72904c095313fb1c624e511647bf1a7dfc1c476",
-  "blk.9.attn_v.weight": "92125bbfed63544ab56052bd1e4aa453bbf34c795249ee54cde54907c8c6d1d3",
-  "blk.10.attn_norm.weight": "5d6bfbe545bcc2fcb2fc75c68f64b1f4c918badaf53e0156fe2d88aa977b2f94",
-  "blk.10.ffn_down.weight": "1dd9da8b0d2696ab5531fbca8a29c7d67567620a9d3e5fc2a19ec5d7e4c6cc8a",
-  "blk.10.ffn_gate.weight": "6e55e7f014edaebda0ac6819a426221d3b025c27312a2e18cc5806f31e3db226",
-  "blk.10.ffn_up.weight": "d80dde54af5db51241345ee8d64c1972608644f4deeac1e8195dc423bf27474a",
-  "blk.10.ffn_norm.weight": "f6ca65951d58ae3379eee8247bec34ebd0db05674cc9295593573841b8a55df3",
-  "blk.10.attn_k.weight": "b58e350bd6b49aba0fba4e4dd6865de3a2a0651ab865dbf2419b627b53ffc187",
-  "blk.10.attn_output.weight": "6b26a986e12fe66ec286a21d7d5af5eaa1bfe6f2bf502165d270e4497235a54a",
-  "blk.10.attn_q.weight": "3440e0e5b7e0d1e426424ae5a33f4e057be623249e9035ea12e57dbe5d3893c4",
-  "blk.10.attn_v.weight": "ebfadcfe14bcd6dee933053df0a67e12e7a196d5cc45728c1ffb2a2daedd5ca2",
-  "blk.11.attn_norm.weight": "3ed057b9576cd2de84507ef64c7646dc478c651efca4c2024cbe91a4f3fbf0bc",
-  "blk.11.ffn_down.weight": "8ff1c2487d22f5c499761e4eb721418f141f960160d0bab779595a34e4d68898",
-  "blk.11.ffn_gate.weight": "9c74e4507c7e45bf39b7cc7402198cd1dd77e3fff8c625b0413acaeb16efeb9f",
-  "blk.11.ffn_up.weight": "4367158007161d29939e00a322bb6776016e43f648a94f9b08a96a477aae75be",
-  "blk.11.ffn_norm.weight": "1cc0288c1491072121f4c9a0af20be0e13af49895696a3320e4fcac608768de3",
-  "blk.11.attn_k.weight": "066f5b3c144fce1366835e1ebf376f768b333b8ae29f5b478c42d1d0c809c855",
-  "blk.11.attn_output.weight": "e0d9f3d3f2c54aed59c02713ea4fb562799ddbacbe67ca3998dfc887bc44e47b",
-  "blk.11.attn_q.weight": "28d3ecc8a88cb3815e89a7f7a7d043da7a71f702b337a126e4d3a2ac1cd6370f",
-  "blk.11.attn_v.weight": "7c5cdef10ee73bca0a3b9f6ece5f0a0155664e0ce3d8de90ccdccfab5545e5e7",
-  "blk.12.attn_norm.weight": "973b133301a1af760cd7b3a7955371ea0a750808b442deb6adaf7b98482bd0c6",
-  "blk.12.ffn_down.weight": "d6c87b4b4ca03f75546ddd6a9e7fca720585a309188723c1ace8122438d4b200",
-  "blk.12.ffn_gate.weight": "2189a6e0cab1540bd05d6089b922aa8fd694be51255654933c165f302a0c955f",
-  "blk.12.ffn_up.weight": "5affbec19b58d092b9305721e3552481fe2eff51269ea3ed91cda3b9ef84d4df",
-  "blk.12.ffn_norm.weight": "f650fd42a34e950f758b4a130e7b8b1a712b1dcbede0291bb8edde47aaed0ef6",
-  "blk.12.attn_k.weight": "59b1e86f10450a7cc188beefc0856d2dcf44e8d7fdd9cd8859c30ec1ebaf24b6",
-  "blk.12.attn_output.weight": "446b0d36b2f66bd72a2323f4f4e9d85a0f621e9a58872e89a27248d6b1123238",
-  "blk.12.attn_q.weight": "3ed6bfd39f040301ed99fad882d3e569769d594259f9948445bef0e44ec881fb",
-  "blk.12.attn_v.weight": "e73652cd5d0029b1931be3ba9d82508f6696dce5a29d085476a54fb7a2ddbabc",
-  "blk.13.attn_norm.weight": "491b85278c0bd67bd31b9b8a9720902c244bd067e53a4a03641b7c0994782e82",
-  "blk.13.ffn_down.weight": "ad71cc248a85e9ced49307a24a9bfae01d387e979a7689c82ff59998e09741f3",
-  "blk.13.ffn_gate.weight": "0a55984d53971fab97575ee0ef5882013be7fdecfa76e3fbebb5dc85a07a14d4",
-  "blk.13.ffn_up.weight": "378b697b35e2e53c0de98e8e29b73d42ae3ec112ec16129aa5997a9e2f3b5943",
-  "blk.13.ffn_norm.weight": "f8aff2f69ab286210fad45a62b03f8d10b38f96a420d7baadf6b95d7b0b0bcd2",
-  "blk.13.attn_k.weight": "25ceb841afb1034831bea7f4d6a6c578def2ce4d4c412c780ef147dc9a598360",
-  "blk.13.attn_output.weight": "a242b322889c6bdaa14b67a7bab593db39df8eea3721638ef639abbb74d482e3",
-  "blk.13.attn_q.weight": "d80be9945a369439e835c55cfb0e97828b8a66bb7ced534d9059c92487bf20a9",
-  "blk.13.attn_v.weight": "ac33274cf9b67979d9ecdc967a55175afe0c9c4aeeff6391433cd9840c818706",
-  "blk.14.attn_norm.weight": "12a1e1091de5b2da12c9e7c0b1c8e6f09ce2a749733cf7d5240445b8e21cd093",
-  "blk.14.ffn_down.weight": "cfd41965c88266e32bc2dcdadda512499c35519e8686fefb9a7f249ab2291eb5",
-  "blk.14.ffn_gate.weight": "8dcfe774f07a095c7c6cf0a901c9df70d938bad7b5ba347fbc8f694e7603c0d1",
-  "blk.14.ffn_up.weight": "c7995577fe4a72ea0fb17c4a7b6b87b959072bbfdd5edacc6c367d43465809ae",
-  "blk.14.ffn_norm.weight": "81c41ebde41739e7016ffec31d2256217b825dc3cae049a935f5f61a60d22003",
-  "blk.14.attn_k.weight": "fb708bdebe4384f5c4b479c110028554f4d122f166b8091eda7d8d65e6780eb8",
-  "blk.14.attn_output.weight": "f5295caf2dfdc60553dcabe17537a80577e8b153c902247daac058df23542514",
-  "blk.14.attn_q.weight": "c12b7a3601c68c63ab5dc9d2599ebf3f3a10abc2c59d3a2126fffd5818f2763b",
-  "blk.14.attn_v.weight": "1ce968d9149bf0d5e237d52cc6d6433565b4bbf03252a736262bb00a2b34a687",
-  "blk.15.attn_norm.weight": "266fd2c36d7dcefc6b6bb7f1c9374c41f2bab5d6c84a063b6f91c4f682dad3c4",
-  "blk.15.ffn_down.weight": "6154886e9ef0a6cc08ab0d264a35f497e6f0987efdac992ed04e87088bea7801",
-  "blk.15.ffn_gate.weight": "183d9fd3c1b5657840099053d2fd3f72ad953b1de523296159b7761f20491a76",
-  "blk.15.ffn_up.weight": "51546d4498842ae2340ee226a0888d5f61e7d2ca4d052dfa06a77b0451242d3d",
-  "blk.15.ffn_norm.weight": "ef7378091a41a25a5f58bf1bf9d3bc64ea562e7f421e1c232b1f177c30fd3500",
-  "blk.15.attn_k.weight": "8d556ab8d9639324141774999b6eed0e91d7ee645bf3e7a3dcd200b2e7a00751",
-  "blk.15.attn_output.weight": "54aa6ba87def7cbe18b0c6ab3aff5c351cb3b6ca4a0d7b2cd5f75a1312991429",
-  "blk.15.attn_q.weight": "10731b0dc031ea8e0ef37bd7f010e0a78518a10a6df05a8bae48e3148b73ef3e",
-  "blk.15.attn_v.weight": "cbbe50c2ed7224866d3cf9b489c599f3ec41a4ea1aa3181e9f4e87e1fa0cefec",
-  "blk.16.attn_norm.weight": "387058eb39d4b28c04cf1368247417f1faeae8ae79d894c9f293457e0eaa00b0",
-  "blk.16.ffn_down.weight": "2cb26ccee585e933401ad5c82ed36ddacb3289efa0b28f8cf91b020ffbd9c333",
-  "blk.16.ffn_gate.weight": "d745985efb5bab42304e5d509024631efe35f92f2b2ec4931ead6db97ca9727e",
-  "blk.16.ffn_up.weight": "7a67bd195e0642828ca36eb7818149bb70c2c25f82de07e2b5807c520daf540e",
-  "blk.16.ffn_norm.weight": "7cefd061c8182482a89272f8a4e88a954b12609a62716923ca1cb3593b1c1651",
-  "blk.16.attn_k.weight": "d7968a2de67e755b4533e061aaad1cb62f8882af92dcad67f99d6d5112513439",
-  "blk.16.attn_output.weight": "9e9ab5788272ca3394ea89eadbce8c86ecc3fd75b7899184d6191c134ad9aae0",
-  "blk.16.attn_q.weight": "ef81c261b536c1a3a093b33f44cf2d42b86e5aa2d821674f07a0c80e992ed925",
-  "blk.16.attn_v.weight": "aef38e7958301b4a437cbdd2fbae6197f677b09269ec1eaf63188cd5da428d25",
-  "blk.17.attn_norm.weight": "28f6b289f1bc3131041e9f791b7a2a3a48baee0dfea27bf7051ebbb7ed364d80",
-  "blk.17.ffn_down.weight": "1a502829aafc6a9bd6bc81f12573bf8632d5c8c659f0dfb13c8b2411f3b1ec05",
-  "blk.17.ffn_gate.weight": "ddfd8aa0eb98846ebc9afe31366249159f46ae9815199dd70161527ed241ac4d",
-  "blk.17.ffn_up.weight": "4211a3cc247071bd361b30de2131d02382f552855062bf3b3e004c17992e5d09",
-  "blk.17.ffn_norm.weight": "647e5fa99a5b0d232af36d15816539f4d27e60a50a341b00aa88bb6e4474f8b9",
-  "blk.17.attn_k.weight": "d9125ff33a19c502c0f8846433ffc24395048582fc2f463d34a0301a82156f02",
-  "blk.17.attn_output.weight": "3d64fbb1cfef04444827f37c35fd9ad3413eb2165094d339ef89f00503f09de4",
-  "blk.17.attn_q.weight": "e5b29424028f578beca385fd82e29f37adedf3037cd51e5889d5a1ffb0428ca7",
-  "blk.17.attn_v.weight": "1809c5aaf2ac04c5d65539097564ad62796e87d24bb8b9ce5b095561a61d908a",
-  "blk.18.attn_norm.weight": "99daca58d001c627523d3adfbca1d95f04e590382a326866544d57989d5f4835",
-  "blk.18.ffn_down.weight": "84f30231ce6ca0f10227541dfc602d6418c1a210386b0c4926ef1656e7d4635c",
-  "blk.18.ffn_gate.weight": "ca5bbe4468b541740e54f69b9e08fcc8e478c344b70551dab21b1206acfbaadb",
-  "blk.18.ffn_up.weight": "0b3067b9dded31686dcfdc1e247eae3974a28a61ac59e9862758dbfaad64e8f7",
-  "blk.18.ffn_norm.weight": "8154a102232dbc0f90ce77ae5c1ff8f26f8b6e4dcf326e9ec1645749669e7960",
-  "blk.18.attn_k.weight": "25abb26021ccc481471a30e0d4cbeb7e1db29828417ec5136edeb93fecf09ac4",
-  "blk.18.attn_output.weight": "d87d481d9b046b68efa06ccdd4ed8cbf61e692d61114b75b7fad5ed75f5d87b2",
-  "blk.18.attn_q.weight": "cc6400379e15766992ff1293be79dc67682c28e9e15155a78109f4b64653b164",
-  "blk.18.attn_v.weight": "45c75cb1dd496aea3173aafe2575b841dd1d02cbe010b3198099731eb98f531c",
-  "blk.19.attn_norm.weight": "65389efc75297684773284ef8e5f8789a4504b636c9f33b8a32e0ee42499fa72",
-  "blk.19.ffn_down.weight": "4eefab7e939f64a17e4a214ca3c77a6fa110d94f677e2d6401086f70fc538b04",
-  "blk.19.ffn_gate.weight": "f1c0a59cafda66f466ab585b0b8b4861b58abe87a67cea1f6a488492242edfdf",
-  "blk.19.ffn_up.weight": "c42d045eef588db4a0e56960a57e110e1ff92eb8041107d19899165fd3b90f17",
-  "blk.19.ffn_norm.weight": "a8f33eda6d5d62ff5f333ad9771783caff556641f4e7df713451385676f441fa",
-  "blk.19.attn_k.weight": "0bab5d9e9083492bfb05a5a3bb23b79c0e7b99ef6a6644817b4d57d5c453b8a5",
-  "blk.19.attn_output.weight": "c99c551d70eafad0f7aea98fb6f9251635897168eb3895f76abf0d4ea3b3aa6f",
-  "blk.19.attn_q.weight": "c98bde95627c3b54c9443813ca50b4e14f518319681db6bbf7b2332ba26e9a60",
-  "blk.19.attn_v.weight": "ff3a490518cf64904db89ce0dc7d6eb89e870f1440e41883c6b55a221f82de84",
-  "blk.20.ffn_gate.weight": "761f0e317229cafe9d3754048ab038a0a84e9a287b196ab65f633139f2d29aba",
-  "blk.20.attn_k.weight": "45d13439b41066d282e8490a726785abf513605f46c79bd0c840f6419d27e790",
-  "blk.20.attn_output.weight": "a3b958d84b4a097844179b7d55c18fd0e4f319cb15e918c6fde33b68de1bcac6",
-  "blk.20.attn_q.weight": "127ab8e7d8c3f882874904196a02712bab42e6744fde45871b67350609d19f5e",
-  "blk.20.attn_v.weight": "5f0ad2d14a8ae42dd3bbeccfb33295687a14055fa92c54bc946249373c1c9f17",
-  "blk.20.attn_norm.weight": "77300b1755edc8c70089e0f45efa646056b9add7d8568b2324d2f3e62b64971a",
-  "blk.20.ffn_down.weight": "ab93d0e075b42e9017b701a070d561e698050d90aac4b4b9919256fbe50c3204",
-  "blk.20.ffn_up.weight": "4fd6628a07acc57a48d1ef83f81b7d7aa0bce569c1160a99d307284f8821322c",
-  "blk.20.ffn_norm.weight": "2a9e46b9e48e8e55215de56592e1f189530037c1c94a1428e3d6f106c7f26fb2",
-  "blk.21.attn_norm.weight": "4b3b5912c7bc61eb9da8e47d4651f896e85d9e59c4ecaa65df7acf3c21737298",
-  "blk.21.ffn_down.weight": "7146f931663d93b8771cd84405cd4802ea6560d0729b0d6d44588203c095bc53",
-  "blk.21.ffn_gate.weight": "b44ec5d64388fa40b90b3e9976d97a8b6800fa3b97584f32e64b03daffb8601f",
-  "blk.21.ffn_up.weight": "0cf3643fd23c685e17062cd11e116e17ce57a405e5e78953bab94cd62fe48789",
-  "blk.21.ffn_norm.weight": "4ef2cdb53da166df70b39f3e6b17af51848cfa5ea3c27ad6a1ae2a1bb1da1ce9",
-  "blk.21.attn_k.weight": "5d40f32a706f670c19972b14176bf660d5b045e3637b110dbf8d7de4ff32101a",
-  "blk.21.attn_output.weight": "18afaa916752ce16c9653ec0ec7e2fe60be55faa2aa5025d147be184adb75cac",
-  "blk.21.attn_q.weight": "2621daa5f858931514a4b2f0fe8d81cf9b96f541e6af99bfa7539e9bde8e34ee",
-  "blk.21.attn_v.weight": "63226dafc54c899bbce4aa49efceeedd8908e94faa613450fdda91f332b62864",
-  "blk.22.attn_norm.weight": "cf3058daab4d2c04387e7d169d1553bb8e7358eea66285ec067703f6ce62043a",
-  "blk.22.ffn_down.weight": "6a58d5fd220abdbac6cee7ba048abab794731af318f04982c2506df59413d0b3",
-  "blk.22.ffn_gate.weight": "d5614535324b03c7b91727a903b2a72f8d07ad17f7aa8b61ea173cf9b895069e",
-  "blk.22.ffn_up.weight": "ec20da3949566e93f66cabb67f8cd7eab399047ec6ebf5d43edfaf3669b82296",
-  "blk.22.ffn_norm.weight": "84c82f38f53a649972a44466fc476bf764e064ce18de870291edc302f3700e28",
-  "blk.22.attn_k.weight": "a3d2ecc37fde7c201176bb8abadf27f0d8ede9679a6034913e03d9db924fda12",
-  "blk.22.attn_output.weight": "5a3b8bb433f43a387df43dd371bdf80ddfac986dfeaf38e9bac1d7a0ec6628de",
-  "blk.22.attn_q.weight": "3a875cec661b4859f30a8fd2c866811184b25b68c9e36fe2663d299caf8b59c6",
-  "blk.22.attn_v.weight": "8717a83b79035058dcfd3ef6f8e5b36e71d77379e5a239e1899eef8766fb7703",
-  "blk.23.attn_norm.weight": "2b4a68a0a2f023dd646e4755c9bef17c2f631901154afd839edac7ac006ec99c",
-  "blk.23.ffn_down.weight": "29499b1586c6fc4883c9b7a9c8cf388035146b5aecf90c5c4c8c8e082c71e7d7",
-  "blk.23.ffn_gate.weight": "7d6554036d21c587b9b556428054f9c15cbef96d24b257f906fcef4ae38bd9c8",
-  "blk.23.ffn_up.weight": "19761ecb288d6ebd44b681c4535661583b1e19dc29e96d0c007333cd8f00aacf",
-  "blk.23.ffn_norm.weight": "37dc35500790a4ca33807b39cf7af65065e535dc25b9e94f3ed2759f61887ac9",
-  "blk.23.attn_k.weight": "717547d00323817b0cb40a72ec5f8cf42ecd1f9e3e42715c2cc5e38f07fffffe",
-  "blk.23.attn_output.weight": "a24786feb6a905fdf166d7500133757cbe494779d4ebcba9eb03046b319557df",
-  "blk.23.attn_q.weight": "6a2c4a98f138b928d22136efa163562691d3b4ed526d52d46a2fa2694a8f3965",
-  "blk.23.attn_v.weight": "c6e6081eb9c38a7fda023085957b460e9ea321e1fff408b38c2b58595c39979c",
-  "blk.24.attn_norm.weight": "5e6283f891e538670425f3e244b08dc6f96f33dfa4aefa913f8eb17212421850",
-  "blk.24.ffn_down.weight": "e09eb170f389deea0a4a1cbfdb52c12490768a2c60491b7bef8a4c445e2a08f5",
-  "blk.24.ffn_gate.weight": "af29d815cf49a38fc2ebd0bf9b2dd9933d023a29f2d766981acb9a1b53f09117",
-  "blk.24.ffn_up.weight": "36ccd9333426666de9d3088bd4dcdf5b624b09dca9e3a83a22fc0383f2d950fa",
-  "blk.24.ffn_norm.weight": "a88e1692318826db6ac42582d182e51a3c698c655d0e21e04fa086318832d07b",
-  "blk.24.attn_k.weight": "f7d61d6d1225289bcc502e3bbb0168b4584add0253218c1b77ac92ccef9a1c2e",
-  "blk.24.attn_output.weight": "85a1363b3ccc87312094c2195022687c16b0dad7fafb9e80bb4ec474d53c29ac",
-  "blk.24.attn_q.weight": "53482a2c008f42f4fad779ca323addc3712040149dfc12f782417756388a72bb",
-  "blk.24.attn_v.weight": "67498272369af7dd10097c73b07f731b565cfc9a559e711cc0d526389e7b44e2",
-  "blk.25.attn_norm.weight": "98dd617def5cb7825ee4833132ca2da2121245921585e1d9e36b93344adc321b",
-  "blk.25.ffn_down.weight": "7fd477d6c50aed5f424a878dd284343379cffbee8a34c0b6e55100c8305fa13f",
-  "blk.25.ffn_gate.weight": "f892c9806c8ec22e8aa746734ac9213428c534921cf161239e1d249fdb5d1ec0",
-  "blk.25.ffn_up.weight": "528bed14c9bf9762f790525ee40412545221f4321d2a2323fa8e73c58b7643c5",
-  "blk.25.ffn_norm.weight": "ca5831966672e7be6a578feeb631ec3570d3b5afe12860819ccb96e896ffc346",
-  "blk.25.attn_k.weight": "610d3068cc9b20401f0c3a0efea39a279dd9f564fde19baf3403b2ec2319e4c4",
-  "blk.25.attn_output.weight": "798aaf702e53b657265ac3b5e6caf3a0ab515bdadfeb1a3a156b4f3bfba76666",
-  "blk.25.attn_q.weight": "8a7fa25248de83029fb97b51d036a01baebe31fcb4be121ab00dd8b7de209b10",
-  "blk.25.attn_v.weight": "2a53d5e9f8a1218c66958c6388d3b37400a9af7956c785024ca44bfbc3c7d371",
-  "blk.26.attn_norm.weight": "5f44fc043481eb0771f3e6d2420bcbcf73140afb9a9feb8eddb6575452acebee",
-  "blk.26.ffn_down.weight": "944a60a409d0d5b6a851e33c69aca152454b691711a8b96f5bcc488772ab2833",
-  "blk.26.ffn_gate.weight": "2a0ca4abb3de5593e6693d8be69b63d6d1a639855ac8332a75f520353f030c62",
-  "blk.26.ffn_up.weight": "0b1df496163f9ac07bf89375d3eb441b51a81d41b47d769a04a61efc18dbe35b",
-  "blk.26.ffn_norm.weight": "56b8dd046e9be6ea71f7efd80dbd14e7fb1aa020d3cd38e063275f3873fd12f8",
-  "blk.26.attn_k.weight": "b1dabfabb970e6971c7ea6e53c63cf7ef56341e6a2edd9cf177785cad9af2f9a",
-  "blk.26.attn_output.weight": "39532c7e836baad164a655fb97ec5114ea4da37ffba9fdea2684f6e4450e6f84",
-  "blk.26.attn_q.weight": "8f48bf6aaa1252bc149e98af2be1777a5c0d2c3274c6d314171ea9344a41b604",
-  "blk.26.attn_v.weight": "02fb145f7fd905133750e90571effacadddfd3f4966552dc59982ac3900ab8c4",
-  "blk.27.attn_norm.weight": "654d168fc3cab716d91261f5719f180b7d697218401633b4878a759f1b5283f2",
-  "blk.27.ffn_down.weight": "2823272bec3a1c12f02cc4cb24aa4031abd7e9dbe0b02676e2305b21671818f0",
-  "blk.27.ffn_gate.weight": "b1a1d40cd02f97182cac17a79971d1934ee0daf3aa0bf11303568c636e208a64",
-  "blk.27.ffn_up.weight": "ed62ec72a020d070e64eb7b50237b32213944727b5b2427f45d989f50df5fb2a",
-  "blk.27.ffn_norm.weight": "c69649ac65d694b306a905dee8b03b89eec1ed188b1eaaf38f8e29d4b12e38a0",
-  "blk.27.attn_k.weight": "cc57bbf413f1fd227128dc66efc8590c73634cbd6f96d01ec4878b5e7ca6a925",
-  "blk.27.attn_output.weight": "cac407ad02361d53207b3c7e25ceab84dcb4347b8087055162e2efe14d11d84a",
-  "blk.27.attn_q.weight": "0af18e07cee12015761c07c94407024f4f4d77d97bdb24163db0e16669e2cef3",
-  "blk.27.attn_v.weight": "a1d08fbdfa40af773c5adcf93bd68b78a44ed144e3fc6bbeb8af02e937527eb6",
-  "blk.28.attn_norm.weight": "f39a51f814512b040a1082143150e4a49ff730f85cef49d7f77fc79d83e91f40",
-  "blk.28.ffn_down.weight": "74f29ed51055d1c1adb8f0660bbe538a27e016c65650f2d67efc6f1c84fa1b45",
-  "blk.28.ffn_gate.weight": "ae48bb16487ded6781c60aafc0bf738fb4ae15729952906f247d216592ce249a",
-  "blk.28.ffn_up.weight": "543009727718ac22f11ee4b17815f68ea6f15ba1f3e7ed5ecdb755cf6417565b",
-  "blk.28.ffn_norm.weight": "b8f9e54c322079ff20a82b88948cdc2916c22c7db40b9a9ed6d3cbe89efb727e",
-  "blk.28.attn_k.weight": "55d055ba653b728d6e784f9e013786fed07115c9fdf23367e3941386d5e77db8",
-  "blk.28.attn_output.weight": "155101c03ddbf18f4fd0694bfc982f33c7bae25c9b087d6f5273c2bfbffcf2c9",
-  "blk.28.attn_q.weight": "1ed19bfdd22e9c14eca014739982492e9516d411515a8585f65cf754d849e53f",
-  "blk.28.attn_v.weight": "11ba854dd575c025d37256eee9041f6d1bd2b549a083d6409a09bfc1542913f3",
-  "blk.29.attn_norm.weight": "02b0bf5e2fcefd11a153cc988c81ba672682e4844fcf6442423e21a0e10d566d",
-  "blk.29.ffn_down.weight": "594bb692ec2779938721ff4748666ca8370e0e4fe85229503f616438b8884f5f",
-  "blk.29.ffn_gate.weight": "8bedcf47e91dcb2cf4093de56b048ee411faab6ff472f89ab2c9c113a08e6967",
-  "blk.29.ffn_up.weight": "e241a547b5fd6dfca8200b8141e21c1c487a96cbc4e5855f181a7ed1be91b642",
-  "blk.29.ffn_norm.weight": "e63eba5e4c6b288bfd9f15e46e236086456c8b7f1f9c732c0b5de84962a2e7cc",
-  "blk.29.attn_k.weight": "afe5979d5bcf211aebb526620f5974bcb0a2c39c8be71e815575c55d6385e3aa",
-  "blk.29.attn_output.weight": "9c944ed44b124b014906fc240afd3b90aed56bbd9567f2eddfd5b7a685b3cb48",
-  "blk.29.attn_q.weight": "e234e08e5c1bd9245a2edc8d63e9933b6b879f97c01392209cad4f55f05f3ada",
-  "blk.29.attn_v.weight": "5cb8e3e5f954e775c5a5e4de7a9a62b17e9c6931bb0ff0e2f82c4126fd3e1a1c",
-  "blk.30.attn_norm.weight": "a65483ee51a0b214144ec8a14f28ea5437586e9e12ebe342a57d1f8627ee12af",
-  "blk.30.ffn_down.weight": "417959da77ceb33ead4271cbb9428b195196173a893c44e52880a7ec61b4856b",
-  "blk.30.ffn_gate.weight": "a0d503ffcbe45dc927600bb98c9f6082487e65cb577ab545add400d666a87638",
-  "blk.30.ffn_up.weight": "f8ab957b82ffcd10b21303cb5e866209b6fe95f827b1b94e9a949207952d12c0",
-  "blk.30.ffn_norm.weight": "210c7ceb0514a9ef27b5d4d1b3aff6dde43f1af0345a050d71097940e0e73e03",
-  "blk.30.attn_k.weight": "16861b9abcf5a3fe73c93d977ca45a1e6daa65be0fd85c2cff53486ce2033afa",
-  "blk.30.attn_output.weight": "ca541fb2e57e2257118c35784845b0c731278af8db3036ac53d71aa1681fdbdc",
-  "blk.30.attn_q.weight": "f7834917748e26bb456b945e230bc926c228e93696bc01fbc2b134bdeeac71a1",
-  "blk.30.attn_v.weight": "9292783171dbe5eb689d17c9bda11e537f0e9b328fced6986c938d61ed590e81",
-  "blk.31.ffn_gate.weight": "e4766a04bcd8f937ba883c6a144101e546747804ca66c35c97281d6ccb47b566",
-  "blk.31.ffn_up.weight": "cc1e666116f7e6b06736db4aa4b81003c583f54f4d9200bfa48842249940e16a",
-  "blk.31.attn_k.weight": "fc80b57557687504efae7d24265cb7dc39b8f826bb3d897a11783012dbedc44f",
-  "blk.31.attn_output.weight": "215617f50a1f5d9b2250b82f3652b35a9e9aa0ad9ef2b485d73965a14b2b872a",
-  "blk.31.attn_q.weight": "274b4f1dfb0bdec28632705677049fb3e327ce6d9e1f3baaad1560439039982f",
-  "blk.31.attn_v.weight": "e641b8b926f9dfcbbf6b6da1c02555525ac4b1c306d96f20cfbba7d6662c4e56",
-  "blk.31.attn_norm.weight": "b3243c361d4041ddb892ce6862dd5091f57d87357e3c67e177451b85d8baf34d",
-  "blk.31.ffn_down.weight": "0a00cd3ecd5e91624a27f9e239b1de425d5ba3cfff82c256a11a4ad434abf3c2",
-  "blk.31.ffn_norm.weight": "2a0d67ea2bb1303975712243f07273c92fce83baa11b1cd6d8e42e74ea3c810b",
-  "output.weight": "768615f077fb797967844571c58b94d7c399d884d115be3ab4b0154504cae892",
-  "output_norm.weight": "7cc5b7ce10e5082000fa00bfa68af8c7c5da218e59e2c41cf2f1499d40ca229e"
-}

convert/testdata/Mistral-7B-Instruct-v0.2.json

@@ -1,313 +0,0 @@
-{
-  "general.architecture": "llama",
-  "general.file_type": "1",
-  "general.quantization_version": "2",
-  "llama.block_count": "32",
-  "llama.context_length": "32768",
-  "llama.embedding_length": "4096",
-  "llama.feed_forward_length": "14336",
-  "llama.attention.head_count": "32",
-  "llama.attention.head_count_kv": "8",
-  "llama.attention.layer_norm_rms_epsilon": "1e-05",
-  "llama.rope.dimension_count": "128",
-  "tokenizer.ggml.model": "llama",
-  "tokenizer.ggml.add_bos_token": "true",
-  "tokenizer.ggml.add_eos_token": "false",
-  "tokenizer.ggml.bos_token_id": "1",
-  "tokenizer.ggml.eos_token_id": "2",
-  "tokenizer.ggml.unknown_token_id": "0",
-  "tokenizer.ggml.scores": "e3d3eea80bb41a1213f2d0aa3e8a38581d1f19323be77dbd779c9c7e3b72e676",
-  "tokenizer.ggml.token_type": "6040635e6bd38d98af06698feb75c1802bad35180ee6ae0a503e38c0f60fd71e",
-  "tokenizer.ggml.tokens": "604ac4bfbd019e430d7b6cdf18c6c0cd5b967900601f0307f714ec7773aa5ca6",
-  "token_embd.weight": "cde834ccac5e94324b25cb81b02d27312cac0c551b55a7e1d555d90bf6cb6e81",
-  "blk.0.attn_k.weight": "458bfdd9715c66e017c2447b1ed3c582963a3111479314e664faad8c914f42be",
-  "blk.0.attn_norm.weight": "e1fd60b95f713bae7b7e3ca933c64ae6c9cd1e8d808000204bbfdc19f0ba635b",
-  "blk.0.attn_output.weight": "df13b6a157d9d4f96c53b012b3b9bcd207d0c94144cbd22ae3ec13bb07d6c373",
-  "blk.0.attn_q.weight": "13b4126b4245bf06c915a93317c42b8174e05053535ec99dc576541e4cec7c25",
-  "blk.0.attn_v.weight": "5b1781d3a341214511b27eb4e268674ea3ea829dbdf8ae5a6bb89b3c0b33fafd",
-  "blk.0.ffn_down.weight": "49186f5d8148d316b07458841d13a2e66587f4af69b776188a809591ed9c070d",
-  "blk.0.ffn_gate.weight": "4397e30ece09136f00f4ff84ff49e5241b765a374deb8c5a12e897e2bf73473e",
-  "blk.0.ffn_norm.weight": "43260589aac3850a779bca3f9649f793bbfbe5db538361cb743b3830217f8287",
-  "blk.0.ffn_up.weight": "fd7ac918240a07566f6967527ffca58fcf433a30b78fdd6d84b2136d4ebd9987",
-  "blk.1.attn_k.weight": "209839566c7d235bdc20565a4766378b6ee8553133a5a3315abe8a85baa80712",
-  "blk.1.attn_norm.weight": "58c52986f7c69784ba327cb7f350923420782bee17fa39b1fbd13839d4005357",
-  "blk.1.attn_output.weight": "5067cc628449682665dfcf59b16e58fe2a9d2a81cb099f0fcd42f4f8670c6740",
-  "blk.1.attn_q.weight": "f410f9f0dd5edc09401af597d02e2a4c727f1502ec3ec3898321617b36c6df6b",
-  "blk.1.attn_v.weight": "d40fa49e07c102c0644e130e7909eaa93ed0d54e2edddc0759e721d58a4e4f5e",
-  "blk.1.ffn_down.weight": "594b1eff6ed4defbdd819fabbe2d48764984f08878a860bdb808511d5a25b8db",
-  "blk.1.ffn_gate.weight": "4cda97541e388a5bb607ce4cc8b3db1da7045830a630e7ba4d17807befcff346",
-  "blk.1.ffn_norm.weight": "66c13d7481be65b97aa474735ddc9674f33d512ddda76fa6fb45c7464b09f1ed",
-  "blk.1.ffn_up.weight": "1adc6de288ba4cc1237833ca8b4eb81107149842e38bc452e18e5cfe284338a2",
-  "blk.2.attn_k.weight": "5420423559f236ab22d85a00849f31e0cc6e9c7dd879de724393d8cd2b379153",
-  "blk.2.attn_norm.weight": "495fe1ab40cc52aa054ddd4f0c2d2790f4326c8d103296b1b38f3b1060db2a24",
-  "blk.2.attn_output.weight": "ccb83e7085381f558bfd65588c525ad2671feddcbc3887afb4038ad9c7aac348",
-  "blk.2.attn_q.weight": "2e8f77478392bc93c2a391f2e0f4a173a952bbab88a7aca099c6ee909726409a",
-  "blk.2.attn_v.weight": "d64512590f3b7ebbb9e77c2eb97fbda90b00d45c944f2b174f03a2cb11007567",
-  "blk.2.ffn_down.weight": "1de5084a05dcaa6b1bd926e83517dbe9ebe7fde79235fe56018b3028b1aa6397",
-  "blk.2.ffn_gate.weight": "cbea526b557f49aad8c976973cf367fcd12175b900f551984f498b9e07e4b7fd",
-  "blk.2.ffn_norm.weight": "530aa49b10c7eae08899d143409240deb95dae4e1d5bf78cea3b26393cff3ba1",
-  "blk.2.ffn_up.weight": "13a5fc19b96b4dcc1e9bd01998c8272ebe52034c1933ed123a506b711fae9a5c",
-  "blk.3.attn_k.weight": "1913b63a73305941d8cdc472e7f101c633d3357a78602eac0a4b49a744261075",
-  "blk.3.attn_norm.weight": "9c11bed5ab41f4adbfdae4ead65b525c8f19443e656a8c61ba412a4e1ad1193b",
-  "blk.3.attn_output.weight": "bb0b42c1d34779c5943272ed71f1dbb31ad8edd75f8bcd5c868f88505ac3a610",
-  "blk.3.attn_q.weight": "3461a1fe4e49f5319ea047cae98ccdb46528a3ec23831183fe87610b48c94948",
-  "blk.3.attn_v.weight": "82aa30be6a61526a41fb79bb28a2617416f5909f0477aa9e95e16be9370fcb38",
-  "blk.3.ffn_down.weight": "68521011ae03f5e3b0966127111afa8ee9f2eaeeef8d3a0b86b633e0332e9fbf",
-  "blk.3.ffn_gate.weight": "1e89e26338fd364bb679695968c65106382f15ad55c95cbb5ec9bdfeb766f432",
-  "blk.3.ffn_norm.weight": "c81932529a5a8c417c27b888dbe95fff8b447c2ea5f6f560444ec5d50b93832c",
-  "blk.3.ffn_up.weight": "305021735afd8669afefd713f56137248d5e817e60471a112ad06b7fa07ffe88",
-  "blk.4.attn_k.weight": "cc26ba5c5c28082a79e6abfe61186029e80b145252ca6a7924c437f0bcf2d51b",
-  "blk.4.attn_norm.weight": "302d251fdcc91f7468cf33f80b49484251d8917d7018ad264ab3a85c8ecf9ddd",
-  "blk.4.attn_output.weight": "a012f5bee3520cd4ce51f0076c132ebc3653309f304032ad051aa308f55f36de",
-  "blk.4.attn_q.weight": "3c8d607e447f5ef21e73af71e3c0d32fae16f91f31faae34ff06912cf9cb68fa",
-  "blk.4.attn_v.weight": "49f6c81a634ce46d71c2350206ecbd231b1732af96e4e4e67693c41a07e007d8",
-  "blk.4.ffn_down.weight": "e89504f311a4a34dc819a67b761022f14d71c43df3ead4f892c87aaa8e9f0adf",
-  "blk.4.ffn_gate.weight": "18b22f079a2fbaefe3572eec61fdcd996fd747724e2f0ff4f08cfcb43eb7bfb6",
-  "blk.4.ffn_norm.weight": "22415a492c168a0878912b05c854a631228b01c3ea8842e1d75989ec46c18a65",
-  "blk.4.ffn_up.weight": "f57379eae2874d8853f14ddf0f0fcc4ff1338574d5ed5d7e88331d5fb84f5642",
-  "blk.5.attn_k.weight": "d627af853c40bddf9762ce3988008c1ff17f2686fa8f73a0b5da38010147c316",
-  "blk.5.attn_norm.weight": "9ce01092c7f7f1c3ef72d6b794da12d77aa1f6a24fb96ba1b9bd5a0bcc3e2443",
-  "blk.5.attn_output.weight": "0388da8064c4b6b795ce2d8079e8a36535e82b2c9cf794e38ce8ae460aae726d",
-  "blk.5.attn_q.weight": "039b7ce1c909761fdf475c06cf14cabe5a90199282c89e4dcf460e95a4b6275d",
-  "blk.5.attn_v.weight": "c47bfd8d2496bdb6e00e03b903e15fd0ee806a515094ec257e43cc433147ab7e",
-  "blk.5.ffn_down.weight": "1d62e6708974bae318cbf00a8bf621d9ba0537e549ce4710a536520a8d14168e",
-  "blk.5.ffn_gate.weight": "8b42b1b11c92db19985094cbb50434e3a7c9cfea71ee6f21ea79eae7c49284a5",
-  "blk.5.ffn_norm.weight": "e0bc520f1505e687ec391d632a381d38d8ebcdec19f614a11a2000ab573e8b7b",
-  "blk.5.ffn_up.weight": "8cdcd17d2ea89bb9ab902dbc6bf3f827fa4ee029c6bf19eecbdefd146d8b6f2f",
-  "blk.6.attn_k.weight": "5dc6bcff89794d1756bf57ec665b58622d9352130d31082a6c66e1a079f99932",
-  "blk.6.attn_norm.weight": "13b26008abe0f119b5104b9d78ebd5e797d3cdd68122b93d73a3b4831a54d085",
-  "blk.6.attn_output.weight": "f5a49917ea70c3fb311ccfffbfafa63ab18416a5d55e5429b70ce8bfba57c075",
-  "blk.6.attn_q.weight": "d9c2f652c87dbd09ec3822e12876648fa32e86553ac25afab723b1cd9f8cef90",
-  "blk.6.attn_v.weight": "5ecc5fe67609a35151011cb526f45c56fc0a999079ae0ff37c755ca03c68c555",
-  "blk.6.ffn_down.weight": "0ec125ae0ecb2d9277fdb1b04f17efee94e37d0ae37311057c212ca2db3fe6d1",
-  "blk.6.ffn_gate.weight": "fa4d6d38355ee8aa3b80b476d65ae7e343c9b7770d7b097fc848ee8a6e091d1f",
-  "blk.6.ffn_norm.weight": "30e8f7defc627532e1739dc76d31223d45767391a431f925b63dabe334b0f392",
-  "blk.6.ffn_up.weight": "6b97cc32b290fa9087806b5d65aa6dc1760737730c8c71394cc4f30c2157f9ab",
-  "blk.7.attn_k.weight": "0231cb127cb7c3714cd72b8f39343891d7715a9bab2237ade9e7bc5f4ed2e68a",
-  "blk.7.attn_norm.weight": "7c3187f07eead7d219d98ab2daf87905e88d5f1ace109b6f5fa55dce3914981f",
-  "blk.7.attn_output.weight": "2f30ad972c284ae7c8eb0482053433495ebe8fe9c5ee2c28b4bc4ed1f33050fe",
-  "blk.7.attn_q.weight": "3a2b4b8d61cc9956d304fa9f82a9e65b4bb9fda2196670b16df7e0d8c43eff2c",
-  "blk.7.attn_v.weight": "d2aab97d0dcf0f61dd2f32848f7a8a99c423a4948a660a660a03a546972b8db8",
-  "blk.7.ffn_down.weight": "2270d520468c5549cd30023ff9c452a277058310104c4239a616373fc5a94387",
-  "blk.7.ffn_gate.weight": "4134a3ef71b3eac8f76b6f1a2e58625b3bae48081f175994bc3ed7d8b0d4f2d0",
-  "blk.7.ffn_norm.weight": "42df4abd4b8769b16f3930068f96960af1b061f1aeb7505384f272233b2badff",
-  "blk.7.ffn_up.weight": "c920549054ec16ff8c73a72f5d837cf4e11885e44db57c1c1c584c18fbd7a9a5",
-  "blk.8.attn_k.weight": "01c609bd3bf31ce65688f1f640ee413740e821330134d4ed1877a3065d1527d5",
-  "blk.8.attn_norm.weight": "48857411f769b00290f4e4f2e593e092781fdc2503f80c1e3eeda1b85a20f74d",
-  "blk.8.attn_output.weight": "90fb273f8df83744554bd59236515c16c5a5a698ca3fbedc17cc89ddcee354ff",
-  "blk.8.attn_q.weight": "ade617ac4653c7f00593dbb51837a468afef20a14eaab3780fb96ac3d6714369",
-  "blk.8.attn_v.weight": "c2c37496494864fee5c527d1fe1f88529d31c73f9cbd02ef9b2e9b23611ea50f",
-  "blk.8.ffn_down.weight": "2da58572e9ad79087c03cbb0c23c9ef69f93ec221fd5fe4ed92fb93871d23ffa",
-  "blk.8.ffn_gate.weight": "4483294e628edaa4901708e73e92c917bdd93b780fa01aa74aed57166f2bbf0a",
-  "blk.8.ffn_norm.weight": "c0cbb7a4f8123b62f0c4652a687f3b394802bc32870dc446eefb709e42043a7f",
-  "blk.8.ffn_up.weight": "9eaf8a2060cb9224cd585997cd671866c4051ad885c2c6d9fdc7056c2a5c0d89",
-  "blk.9.attn_k.weight": "5dd36c45fbc9c50fd35c36cd75576288506971eac5c5311d4f5c16ef60099645",
-  "blk.9.attn_norm.weight": "3c8ca64f2f75ed7c8fc1da010c23be787648139a96ca0ef3ad10be7b14942b8d",
-  "blk.9.attn_output.weight": "6277e1f833024f53c409be919ec76d34464a78b278c8f9dbf79e777746e3b995",
-  "blk.9.attn_q.weight": "87352b70d9e328c2d51d59090cf5ea5a046529864a890d0bc8986447a0a5c006",
-  "blk.9.attn_v.weight": "2efdf01161d7a82a9117cc2d87d37dba5ffefcf730781cb94fcc95130e48ff9e",
-  "blk.9.ffn_down.weight": "e7658a2ca984961c7ace16acb679387bedb1fef656b5330bbbf588db19673a75",
-  "blk.9.ffn_gate.weight": "773cd330d4ff5d64be8af00adf2e2722fae4e33fc26bb9d03549f6f4b3b0fe57",
-  "blk.9.ffn_norm.weight": "c8b86cd5c43b332f72060b807091c33a258e5dac01358ff4733b916cd34c9c97",
-  "blk.9.ffn_up.weight": "d8cc3bcff18bd46124ba2aa7caacc71220b44eeef6fccb993b4c6cb53e8f2c3a",
-  "blk.10.attn_k.weight": "964bdf3b4e77b915a216f750ff7b0f2eb1dd6bfa071358aef21010b90111044d",
-  "blk.10.attn_norm.weight": "59ed411d91d14775764eb514acb0895a75a10cbbfbc1c15d453bc50f8046cb7f",
-  "blk.10.attn_output.weight": "4d35a2a44cfe4ac0a83fd3ab0dcf1f5a0bf54cdb3b7be9fc353ed32c8a3eb81c",
-  "blk.10.attn_q.weight": "defff5339450dd881ac352f5c459293f39e07b9619ebd10ed632d79a3f310278",
-  "blk.10.attn_v.weight": "b9803e8d6a54acea58f662d4c0a5c8ebdf986676de7dfe12d4b288937881ce93",
-  "blk.10.ffn_down.weight": "eba856be64e4be20b92fb4639a783454dd92427250759df92a337e39f1971c08",
-  "blk.10.ffn_gate.weight": "2d5c509b066584db4de3632b01234e86edcde35409c5ebce18957dc80fe465e3",
-  "blk.10.ffn_norm.weight": "ecb9a8679945ff0273856624ce435dd250ffe5a440ea0861a5c84f0e4c44d2c6",
-  "blk.10.ffn_up.weight": "e76ec7e993f399af02958778c643aa78368e3067846714165eb5aba9d5f547f5",
-  "blk.11.attn_k.weight": "29c6d1f34bd3ba2f0904e57b32a5bf8dcb2834d439159a33edf234ce0b775677",
-  "blk.11.attn_norm.weight": "b5817b275149cd2abe18a6a10e19854605fc58fd364666744362ceee8cfe49f4",
-  "blk.11.attn_output.weight": "1e05653220e237cbe0cc770033e183c9a0eed5680510997409b16186c6691950",
-  "blk.11.attn_q.weight": "03db725ae669151e4d536e50285b3b047ad097f52475df208ed3e790e31a44be",
-  "blk.11.attn_v.weight": "27cdf1d4e971326c451a4615a0b79a8c7fe9508f9b76c0d52fa01971fc7eb403",
-  "blk.11.ffn_down.weight": "176938cd7c2966094f614cace8ba568b10532e45a0d438f80eccd19b6c2a7f87",
-  "blk.11.ffn_gate.weight": "9782339915dd6fa70013628a01524ee1d01ad8beab04068da7ac6a5ee7603a60",
-  "blk.11.ffn_norm.weight": "8245f6391e3be97811c0ff27f0d8f484ecc82a468a837c893f059745bfcd95eb",
-  "blk.11.ffn_up.weight": "15616ddde096d0d25e906375c548b6de4bd5576d1f6b68eefdc29f14e183af42",
-  "blk.12.attn_k.weight": "66dd21604993edd1b1fe547bcaa06f5bb7e31c9204902d147a227e4badf7feec",
-  "blk.12.attn_norm.weight": "23a69f85dd8a0904b9839cc5d0afcda299b74e82ae2642106224a1c820f2b761",
-  "blk.12.attn_output.weight": "4a98d132e376beb274a39d4ea9b6a1b870ad5c66625439d7ff6f45c229c3ca04",
-  "blk.12.attn_q.weight": "1c6c309d63afcfde32fe37257e300a78e25d01117e33490801107c0e75d1ea66",
-  "blk.12.attn_v.weight": "723d9e4ebe4e2b1974afa01d8f512b52933698fa36717dd47b37b07760c50a10",
-  "blk.12.ffn_down.weight": "00e0fb09e1f1fbbf3803f1dee373eaae7a93756b6e13063ab77f9927bc6f996a",
-  "blk.12.ffn_gate.weight": "89159f7f97aefb1e100107e3ac2d694e1008ad873f79bb953d60c2c1bb22724d",
-  "blk.12.ffn_norm.weight": "5f70aebd0e43a39d6373d8658cc670c13aadd7818831d3d84f761d5f688442f0",
-  "blk.12.ffn_up.weight": "faec21b446f061eb4dca561a3180712724347b77a71eb312e7afe9be9e89fa04",
-  "blk.13.attn_k.weight": "3d440825d19eac3b1753b34d94fee2b3a3cb6636c10b2703ffcf688d3c1eded3",
-  "blk.13.attn_norm.weight": "47b575e57e410738ad13fd3c74bb49c06b3d31030910834ece509cd1a5c6d9be",
-  "blk.13.attn_output.weight": "05436d8e613f4475741c1798a7c371b53d61b229507fa04fe23c504ba1f0e12a",
-  "blk.13.attn_q.weight": "002b5024ce520da41256e3ded5cdc60e5ae07ad9b202cb19d76ab511efd02b1b",
-  "blk.13.attn_v.weight": "c1f2d6763587c50312cee0d7140fa2c7ee326f5b172bc99b2d8946e08329cabd",
-  "blk.13.ffn_down.weight": "b5c4e0d8a3ff96cd76a135e415b89f02d28c28f7f3c16a36af31ef0ab8773da5",
-  "blk.13.ffn_gate.weight": "ae06e9e3d2e1f64c7ad23a4009dc904c2eccd7241f9f91c4974ab2504f116be0",
-  "blk.13.ffn_norm.weight": "e44a22321bcbcb4a3c345b504e939e8071370f54a8cd702fabdb40b97e0d7683",
-  "blk.13.ffn_up.weight": "7e6f366d538e21ad431264b12c011892d0be9dfe4c4da9f730af677f920641ba",
-  "blk.14.attn_k.weight": "95492d6417952ec24b2cab87bceb750fc7e95ac6b1944fc328a3852d980164be",
-  "blk.14.attn_norm.weight": "6b7b09e1c51addcdbb160ea59edf032531421c520ec5645fe1ff9ca4180cef54",
-  "blk.14.attn_output.weight": "75887474e4d72c218e6ab0f69f1bf3ec3dc414d51b36fc59df00cdb23421bb6a",
-  "blk.14.attn_q.weight": "940e33f76e48c21215d19e8a21234c8246d4d084381a7d9806aecb24b071d5bd",
-  "blk.14.attn_v.weight": "c58601cf5a9833f80f7f9a5b2656e8eab5eb133211446ebd48f8be15fed4ebb9",
-  "blk.14.ffn_down.weight": "f9f886e7f9b2a54d717b08947a25a0a93e8c2a5b8bcd5a907c06817c8ee3ac11",
-  "blk.14.ffn_gate.weight": "727ed0ee68594a3f59d704ed3240b6929f083b9c36650fb848d182315737245c",
-  "blk.14.ffn_norm.weight": "bd2471008ff1b2bae9aa26bea019393fb2bbc5b9493b8cec3ebd2c280fca24ca",
-  "blk.14.ffn_up.weight": "b006446769f51e4f93b503c4727deae897bc1fc7f4fad49f85024b63c4548d38",
-  "blk.15.attn_k.weight": "23bb70f9035356624039547a603e46be7d1e4403616eafc2451cc09c5373d522",
-  "blk.15.attn_norm.weight": "718cb371ca052eeb3bfac6ac506abb887df125271821fd171797a7f2d8dd6313",
-  "blk.15.attn_output.weight": "c76a2695a204b43a8e5acfa5720590b5d449a9ad9e082cbe3e80fab5903ea16a",
-  "blk.15.attn_q.weight": "2b3e4037b9e91bdd26d6e8d904cf39f948192dcf09bb6445cb55ca058d4f4626",
-  "blk.15.attn_v.weight": "7c15e89b6acafc8619e86aa9d412f5893ab17843ff2cfaf40eea9637b24910c6",
-  "blk.15.ffn_down.weight": "e16fd4bdc6d1c1209c6b633454df4992870c8cefb2cb0e8c92a7e489e9fb5d19",
-  "blk.15.ffn_gate.weight": "95a46bea366c260337c537fde06b4cbeaeec52484a69c3390bb1d178eb0525c9",
-  "blk.15.ffn_norm.weight": "37730293f704da265dc6d1896b3be00c39c0a41dab07f573af39dc30a481d623",
-  "blk.15.ffn_up.weight": "ba74a199da2d0875d7410824238c4ffafbda3993568812284a72b8800df91f15",
-  "blk.16.attn_k.weight": "f58f79a2a91c9a763adefce0c53a71eb5ce6bd8442f4af554b04b58083bff27e",
-  "blk.16.attn_norm.weight": "0c16e41b95e81978e0e0e3b338e2afe2d297426578cacee94de15df74e94eaad",
-  "blk.16.attn_output.weight": "ead22fc337514e4add49aee19720008558e52090466866e849671953a1fccba4",
-  "blk.16.attn_q.weight": "ef59c4e8fe8918c1add43d7e9c6fb3ef799dd3e1bdd731ec7b6a4a6f97c86048",
-  "blk.16.attn_v.weight": "902e6b84c2b64241470b13e6f412f859f66b4b223bcfb9c15d5cb1106b07ef3b",
-  "blk.16.ffn_down.weight": "2ad6e9eb4d8372c32a554395d460d17cfb02d6dbcb757cc962b6bfa36db4f5ee",
-  "blk.16.ffn_gate.weight": "825b2d50fcce3dbe6a5d8d8a50a95466f83ca4a10343efe67894c20b4628fb15",
-  "blk.16.ffn_norm.weight": "3bf6ac90befb0e17e077c8ea9454a8485a30f89f2d761ec7751b60c90aed1af9",
-  "blk.16.ffn_up.weight": "9fbdd08739b32411f5ab0252174d386bab19eb0b17884862f760429b7d41d78c",
-  "blk.17.attn_k.weight": "4033398718bf3674830ed1b73071ed8482b6dd4ef27f31a6c5fbb998321b6c07",
-  "blk.17.attn_norm.weight": "714f2e8ac9592966a0f1c02ee979eee8f84586405b992e8ee9543e840199ffa1",
-  "blk.17.attn_output.weight": "b6bbb618597d767b8f535117be68f92911e4a71d4eb4d8b5d943444151445ece",
-  "blk.17.attn_q.weight": "b84a0dc00ceb515faa2628125dcec502eed923077b21cfe900a4ff16c2e5f9ed",
-  "blk.17.attn_v.weight": "4387c7d6a17da9cc7a6bca8f4a75618b20407d570792056283a8e93b6ec65f18",
-  "blk.17.ffn_down.weight": "47db95c6f1e12b399c3eaf9ddba261782dd71173dd163b52af96541cf87b5196",
-  "blk.17.ffn_gate.weight": "59abaded0aedfd12f01df81f7a811e84db6a227f51b60abe9a247ca726e87392",
-  "blk.17.ffn_norm.weight": "b7e86445be5c7b722e01ddb98d5c7527ca86cb827ce0354f2c269e0f2558751e",
-  "blk.17.ffn_up.weight": "8e31c293bac649d2f60da4b3fc4a3acdce1111ec6058d8805eeeb242443011de",
-  "blk.18.attn_k.weight": "5ce762ab7b032511c131df81093b587871718c7097f79d8e07d707571f18a47b",
-  "blk.18.attn_norm.weight": "1f52cdc7af1f4dc1f0ef6ad1ad02e18cda32133654e57cfa9c72ada9c0b1d995",
-  "blk.18.attn_output.weight": "6486957f30bf8a88516e25772c6650f98b13923f490a2865a8752e36439d1cfa",
-  "blk.18.attn_q.weight": "93621c8abf69d2ca29c5207180eb628fb2b544d89de6c4a7fb0699be95534899",
-  "blk.18.attn_v.weight": "11604083b5a74828ac1d226af015ad5dc0215a1fdca44fa7131c2163c02d8156",
-  "blk.18.ffn_down.weight": "8f9997feb94385f106915df810239c9753b31efda2bf14bdf18a9fbbeec8233d",
-  "blk.18.ffn_gate.weight": "427c213b3a4e94af703429daf2f65766f70424d8230c123e7e712a18bceb5ecb",
-  "blk.18.ffn_norm.weight": "c45d305c4ea6a54013ba112f12dafaade064a32cf01317373464a3618d8ba44a",
-  "blk.18.ffn_up.weight": "a2811f2e73ac9eb9cce91a21a454e84e230a155244e2cd73f2c12aad3c9b8cfd",
-  "blk.19.attn_k.weight": "b2daed159925eac58c291e2f1e2000beed21002b03c9e1bc7e7a52e22240666c",
-  "blk.19.attn_norm.weight": "6307306ede2ab5bffa1bcac3f8b139354678c0376b1d9f5530c1fcb4268cfeb4",
-  "blk.19.attn_output.weight": "ebb98218b2a9c84d3fb6baeb02c5df264b7ab80d994d1098ba1cd47aa398effe",
-  "blk.19.attn_q.weight": "4f10df2ad09177e7528e9456039b670d07db22940a49417101b725d239c16724",
-  "blk.19.attn_v.weight": "30f1efc5114badaeaafa91fa466dc7fa14b1616db433c6f563ab851f7333a5dd",
-  "blk.19.ffn_down.weight": "be5ec7fe6b48855cd0015b0e430d1b70c620de87a7ff188c7c1afef546d7b6bd",
-  "blk.19.ffn_gate.weight": "10dffea4213881f8a9b583ee0fd370e033756d32255ed15053f794375b9400e9",
-  "blk.19.ffn_norm.weight": "e75cd24ade45dca78fdb0cbcaaa2d4a17d83a5a73dcc94ce0ec2d68fbdb2a881",
-  "blk.19.ffn_up.weight": "63e81bdb951410ffa81bcfba1b94a679ec9ebae59cd1623ce2651ed5d4c78bfd",
-  "blk.20.attn_k.weight": "c2fc5ad39e9bdd45e73c6e54aecc474388d944c4be1ee1921b7fcd035bad02e0",
-  "blk.20.attn_norm.weight": "aaa9169171937bdce20c1f057e94e9252f221cabacf1ced12e11b9586f23d308",
-  "blk.20.attn_output.weight": "a9f4fb496e4bc053e3f6cf2e72e22d4cd2b545ef6c32f7e782c2ef6ebcc21d4b",
-  "blk.20.attn_q.weight": "5a07ac619ed251494170b213921ef3fcc4c2712839da262516d9d5b8ea1ff185",
-  "blk.20.attn_v.weight": "d6689473105d241eacb17f09f06000ee237336916cf5ec4f48271c5b41bcb8e7",
-  "blk.20.ffn_down.weight": "74be38db51df736f26ede7c6b52ea787e385f181cb66231e2cced4556a25c9b8",
-  "blk.20.ffn_gate.weight": "ea91e06dc3d051c0ba0243b5a8bb40edbf254eadfb54fda7247e05cfdd88cbe2",
-  "blk.20.ffn_norm.weight": "5fbd357b3d6f44a7a91e8a4fc246b24303891b7957e0f3c32818ae5dc16ddd8d",
-  "blk.20.ffn_up.weight": "fe3290333e056af4ed12942ac72aeba97a6b562e2db05e79cd35dd07eab5b101",
-  "blk.21.attn_k.weight": "201ec6ee95f06ea5eb80fe86fd07bd016d3ae9ab6abd25d631834414e14a010e",
-  "blk.21.attn_norm.weight": "ea8154f93e06485828475a00b98cc397ac84768dd70e06ecc0c075b5712d7276",
-  "blk.21.attn_output.weight": "9f8af74d531478fd304723fd8e4e01578db598441b80dc7c960cb801dbbc501e",
-  "blk.21.attn_q.weight": "277de9953a8d3cff894ffd06c15ad0ee1407e319df0c1a693d4f45fa9c74ac7f",
-  "blk.21.attn_v.weight": "6bfdc16cfb898909b7788ddd39dd04b928f31d6732772195d53c558004638dca",
-  "blk.21.ffn_down.weight": "173877146cb94801157796ee9e5eecf3f46acb3b5e797f90b83a3fc22395eb30",
-  "blk.21.ffn_gate.weight": "53146713e2ca1be80496024077a028f6b6d749b02e71003c349e113b436f48f4",
-  "blk.21.ffn_norm.weight": "b28b97e18ab20a5c553ba422f7d7f6014f5902f1d62a69abd20d9fe19a5f9462",
-  "blk.21.ffn_up.weight": "5c39d0ac4d602b8ec8909dade93b2efcd6b6d9d84a19b252d76bb66dcfaab87c",
-  "blk.22.attn_k.weight": "01f26272c82917a87a3ccf922fa1d521a952b05de878241b7efe3525b617ac87",
-  "blk.22.attn_norm.weight": "5ffc96249d8873b506e9eb7158bdfd07fa1429e53c1951430ca7505d25f11c76",
-  "blk.22.attn_output.weight": "9c2201569358f720244b9c9497e4da02585a167b1414c8a506b85ad75ba990d0",
-  "blk.22.attn_q.weight": "906036eb4ddf027f6d920f9356a6a2a5e529b96f4e1231a0496d46b4434a5842",
-  "blk.22.attn_v.weight": "30ede8b0d166003a4b8a81fc99437f557719fc36e5c4dd510c9f161f36a47e73",
-  "blk.22.ffn_down.weight": "d04c164beabab30e1837b843e18852260efccfbb9d96a34ddd816e6fb3ba23c5",
-  "blk.22.ffn_gate.weight": "19c889db6b19179f0a62d5981a1506592c65de83760d67afbe00d202202750a8",
-  "blk.22.ffn_norm.weight": "4885eff2d851b32dbd306bd632c725857e6d164f0fa8b3d5857e572e6ef98ee9",
-  "blk.22.ffn_up.weight": "365594d8db8e95cf87cc33ac23947942dc326110175cc8ec5a07b5c7059089a7",
-  "blk.23.attn_k.weight": "badfea1569da0fc6ab817c5727ca3a69b07d9cfd622fb8be5e66678d5b3f7ae2",
-  "blk.23.attn_norm.weight": "8968f78a379ac3ca5458b4ed4251e8d9112aca6d6dd1ef6440b4bb0b380375a4",
-  "blk.23.attn_output.weight": "93e43393c03956287b1fe31e9735ff1cfe84f4ae56b83dbaebe96275e4e11831",
-  "blk.23.attn_q.weight": "aaff73c725a8700ae66bf26ac8869dfe96738eff23a8ff340de2ab53400a5795",
-  "blk.23.attn_v.weight": "3a86a8dcf14a746ed1411f5a7e634064bc4dfd6511c24cfeccfb2c9ebb6b4101",
-  "blk.23.ffn_down.weight": "d4da6f37bd7ef69bb203f7b0dd59f50bce37432c70627e6cf274ab81548af5cf",
-  "blk.23.ffn_gate.weight": "5b6072936c4a693923bb4e3d1473fd45545cb02fc07799aca458ef0449a04061",
-  "blk.23.ffn_norm.weight": "cd76e37025f84773180298ddb15e0d4ba9cfc7d832e19c791049daa47c6d9c10",
-  "blk.23.ffn_up.weight": "cde43b99b83124a13b2e4753d12674b3a61dfb34c04703007ced3e8e2aee1801",
-  "blk.24.attn_k.weight": "457379edc4cce4cbbe107385079019bc922264fdfc7bd1d1ae84343a81460c66",
-  "blk.24.attn_norm.weight": "0ce0dfab2edeede5da419fa7833db78e36222cf25c358d08f3ec664310f031fb",
-  "blk.24.attn_output.weight": "0cf91c2fd40c204d2fd4b9c85b69281e5ad4ea8442972fcd44b5fc8e835ffdf8",
-  "blk.24.attn_q.weight": "87ede30c09eafec6a4e6285674c1bc4637140b168b2da4ed34f36fdb6e176cc9",
-  "blk.24.attn_v.weight": "4c0b078b2798ca35d6d2c2258fe499820d2bc88700654ba4016e4b028f563590",
-  "blk.24.ffn_down.weight": "cdb8540c32b1ab988f984484928d39f6841f2131c1cebe90ad9456737fccbcaf",
-  "blk.24.ffn_gate.weight": "da2e0e913648b5526bd2bbb344038dd067639343aed3b413662b064b0db7556e",
-  "blk.24.ffn_norm.weight": "8940bd781c610d75eb2be63cfc8d869a3af05e53c963dc7fd4c6f653df5a80ab",
-  "blk.24.ffn_up.weight": "90cbac2a58801abe11ed6c24560aa4acb949f79429f2aa8ff129ac05868bb87d",
-  "blk.25.attn_k.weight": "90607131e36998e990ce718ad05cbecd1bcaed010931401ce6baa3b0d93ebce6",
-  "blk.25.attn_norm.weight": "fbf679c85656c04a6cf8fedd5412c1ace22960e6c2d47f2d43997827811fbb97",
-  "blk.25.attn_output.weight": "08412724ee7a2086514406e6f68fb9f622e10bac25b0c373b294709f4b09bd2b",
-  "blk.25.attn_q.weight": "9c1238e98a2747654a0d4371d3e7ea8b979867f609dc42482544f25591e85c7f",
-  "blk.25.attn_v.weight": "a57796a535c6cb09581cbafd6a91dc14adc8cca2a2465a7ffd0aec546cd84074",
-  "blk.25.ffn_down.weight": "f7e34e8a6391b480da08b52640613ccadce268373934b409759743a1735b74d6",
-  "blk.25.ffn_gate.weight": "b8d0b2f4612678b5ce42bd4a683f8024514b75fb5ebf6b22c600811e95582ee4",
-  "blk.25.ffn_norm.weight": "cde1fdba2369d315f3c6940a997c471ec891924e642505db580d732763bd7b75",
-  "blk.25.ffn_up.weight": "72e700c32ac8b9c47559c2222e45888a480b527ea512075423c5dc01678e2bb3",
-  "blk.26.attn_k.weight": "6ac83b3414ae75bf3a9055c32e49d2c40fe611ab21f8444f03d2f465d18122c9",
-  "blk.26.attn_norm.weight": "55f9d6dc9d75973dc75136ecb9d991b4398097ac133070873fb96ec76a6f60bc",
-  "blk.26.attn_output.weight": "ebc4fcbd15b33263e50ed2ad45740867cce15bc90e1216623babcb1820734509",
-  "blk.26.attn_q.weight": "080f057521073e412936fe3fee64fd574c8128fa4a148b879d3e598fe4954581",
-  "blk.26.attn_v.weight": "0fa2830d6746487ac91b243716e4302361f891e4e008eddd14abec47c7809d5e",
-  "blk.26.ffn_down.weight": "cb2ab8af1653adc57111ada49d2825c6995e338c8208455b92de10e580f60f31",
-  "blk.26.ffn_gate.weight": "231ce30966086bce2dc0e0afd34a22a1958cfda7a57c41b3b8e9444c5dfde8a6",
-  "blk.26.ffn_norm.weight": "35d959d25d17b00617590f5d5831bf705c385c51e46297a14375a700effca6af",
-  "blk.26.ffn_up.weight": "367680c8d332538b467d1ef87cfeb36cc5c6af564c5023c5fb50e728e3438287",
-  "blk.27.attn_k.weight": "0bfcb351c6d17aeac5b55a915074fbdf00f11c4bda98babb196ac8804805746b",
-  "blk.27.attn_norm.weight": "5d598a88c2e75ba59dd7ba4fee940bdec92d72038f1286536d2dfb71d008a09c",
-  "blk.27.attn_output.weight": "23a9da7347336479f6a10ded14cb3f46e06b5bd56dc4b0fbc526c688552ec840",
-  "blk.27.attn_q.weight": "b83319dba9055f069208e9c9d66da08bc6874f23e575288fcd81697d1777aa54",
-  "blk.27.attn_v.weight": "36ed34ccb2f36fdf16b2c2dd225a98ea6b7b0e376e7791191136ccd7bd7a4add",
-  "blk.27.ffn_down.weight": "5488e1d3a58c71b5e9ddda430540b4776b268cfe1457cbc1c2622dedd9e4526e",
-  "blk.27.ffn_gate.weight": "4ff48011ee0bac39af704849d9132a2410392c87a509c684f2062f6b76b498fb",
-  "blk.27.ffn_norm.weight": "32afe99675983da3de2961d1b5ca41c98970a356823597fe29e91f6e86abf0e8",
-  "blk.27.ffn_up.weight": "1eae3088a75629571fdbf6a20f141bc2bb2ed3f5ba2b9fd1d949f80695e442a1",
-  "blk.28.attn_k.weight": "c4e80af714962d6f9040d2c09f316f4a1cbc3a2e994e19902d7c653cf3c73dba",
-  "blk.28.attn_norm.weight": "c1ecf85dedc1c83d5d402bb7c94fb8b9c11f1a3e5f64e7680f80912d4a560794",
-  "blk.28.attn_output.weight": "72ba47c061b21f5ebc5213a455eaf6fc49c8f8e04ff9ce37e6ed4921b629161d",
-  "blk.28.attn_q.weight": "c4abc47234307f44b8ca789aa6668e298158fa4b459b2c1e84bd581806591cc1",
-  "blk.28.attn_v.weight": "aeba950799d4950e491ad0fcbe30334e39b8975177990a2cb339031c45ac153c",
-  "blk.28.ffn_down.weight": "4e84ce382a37b994fb8608df451a60040559e3f4f3241c3b3cb8989a3ed50d83",
-  "blk.28.ffn_gate.weight": "04df157acdc8e8534ad60acc2d2a4dd3a7a6610f6382535ec728994fa6f83f83",
-  "blk.28.ffn_norm.weight": "4d0386dae2bd1c1a9d0f9730718333e3a486c3bc6a5c5d482193c75d39832c80",
-  "blk.28.ffn_up.weight": "fec60bb0a3daf182a14bd8311fe6dd1e3fd020c5fc273e2549cdb1a2d6b79b05",
-  "blk.29.attn_k.weight": "b0532a263aa5a4e2a7a80adc83fc5dec974493bd18da7f953e7ebfc3f3a19aae",
-  "blk.29.attn_norm.weight": "593fc3b4000c35b7a59dace09ca1756c08be0105b2edd354a0e1c16c82898859",
-  "blk.29.attn_output.weight": "315b896f9f0cbacd0ca8937384c3a3a227efa908cb8c3a9125ec00c480e32b9b",
-  "blk.29.attn_q.weight": "d482d45386d4ad3394f08e9dff233ee3a70d0427d65c0b8fa05905da7e25ca53",
-  "blk.29.attn_v.weight": "cd3b5a6e2852da796902930a6a84bc87fc6a7c7bf51f8fc23758d12a39013b36",
-  "blk.29.ffn_down.weight": "5b3dba6f9753bd1b1ebcba65ef5373dd62c38e755c44b7231b95d93d45761f89",
-  "blk.29.ffn_gate.weight": "8610d9d2db15c256243ffcca3ffd31786d0ada0af0e7c7aa3fd20524370ab036",
-  "blk.29.ffn_norm.weight": "1a2ef2d38b7ac3e51190b9ccb8b6552ba83ab290e523356a7f851ddb35dedca2",
-  "blk.29.ffn_up.weight": "a5fdd15811bde16dc27677cf1a4c97daab4c28cb12a9530f1a0e573134fdb69c",
-  "blk.30.attn_k.weight": "1efeb0b5f4b45a85cdf47300f892ac77ac1f38000ec3653565d1303d1fb8c743",
-  "blk.30.attn_norm.weight": "c73934c182c7fe80838ec1d0b92f50a583f75f7a3d78d822f009b58ad2c80e65",
-  "blk.30.attn_output.weight": "3a0fd89de2d274614750345d827a9c886a4f97b343a13cdf680390505df596a3",
-  "blk.30.attn_q.weight": "711e113362bdb067db843c66236704eb1cd3fc5f40e3767143e96d510686ef4e",
-  "blk.30.attn_v.weight": "82b12a9a74fd3d91b73cc2e841e2b3f0a5197ccd2998afa17020995f880d2267",
-  "blk.30.ffn_down.weight": "af9f4b1287c0d824ae22d6e335d19e04a70135b835be7caa2435f1d85e931993",
-  "blk.30.ffn_gate.weight": "e2ab3e6f15f5c50fca66c084cb6a57a2b6b82406d65150e82ea0437b93dd9a46",
-  "blk.30.ffn_norm.weight": "c1b9c325c83f00e177386a4d7e769945f2995e60950c4a576c0a2c4ab9703d04",
-  "blk.30.ffn_up.weight": "9b94a21efd419715d82071b490d3b635cf1e8da080620dcc39e5bde976d7e9a6",
-  "blk.31.attn_k.weight": "0db0d82e3ddcc2c06209f5f013e1d72a84a996c40bf00186be485b909cc268e8",
-  "blk.31.attn_norm.weight": "2b8b7239471f57140c5cdfe06bd224a4f6326282f99736e44fba4c7b120ac101",
-  "blk.31.attn_output.weight": "a310b048840cc3ff2be4b84796340e8e2cdf05ec89d14bd3655c109b2bfa9fcd",
-  "blk.31.attn_q.weight": "f45e0cd95645175ea82813455356d171838539bc3f7676d877c698f2af0a0eda",
-  "blk.31.attn_v.weight": "8bde008e809112aa7e7c23e9c3099087bcc557313b01306c87efa0a4a30805ba",
-  "blk.31.ffn_down.weight": "8266fec7e203fbfad7033120861e44984581ff8b6851d01dfb7b81c5d8fa90ec",
-  "blk.31.ffn_gate.weight": "b73bc0aa5baf006d9ef6403104891b8133671b0992398fe038380b67e0d7e2cf",
-  "blk.31.ffn_norm.weight": "9c62cc27a7b6017c1df8ad49bff249a8245e8895c6754f402cd44623fda83268",
-  "blk.31.ffn_up.weight": "5b970a4694ea3171a0167f6e1636d9f00268bc1c9640430ffc35218494884adb",
-  "output.weight": "74fa0ef08c57a30e633e7117b1e9c805f833e2e5e21434bc79ddf9c92c6d7330",
-  "output_norm.weight": "59b8a59fd3fbf39353506116e43e5e76edd0cbf2a2873d869da4cf27a04997c3"
-}

convert/testdata/Mixtral-8x7B-Instruct-v0.1.json

@@ -1,348 +0,0 @@
-{
-  "general.architecture": "llama",
-  "general.file_type": "1",
-  "general.quantization_version": "2",
-  "llama.block_count": "32",
-  "llama.context_length": "32768",
-  "llama.embedding_length": "4096",
-  "llama.feed_forward_length": "14336",
-  "llama.rope.dimension_count": "128",
-  "llama.rope.freq_base": "1e+06",
-  "llama.attention.head_count": "32",
-  "llama.attention.head_count_kv": "8",
-  "llama.attention.layer_norm_rms_epsilon": "1e-05",
-  "llama.expert_count": "8",
-  "llama.expert_used_count": "2",
-  "tokenizer.ggml.model": "llama",
-  "tokenizer.ggml.add_bos_token": "true",
-  "tokenizer.ggml.add_eos_token": "false",
-  "tokenizer.ggml.bos_token_id": "1",
-  "tokenizer.ggml.eos_token_id": "2",
-  "tokenizer.ggml.unknown_token_id": "0",
-  "tokenizer.ggml.scores": "e3d3eea80bb41a1213f2d0aa3e8a38581d1f19323be77dbd779c9c7e3b72e676",
-  "tokenizer.ggml.token_type": "6040635e6bd38d98af06698feb75c1802bad35180ee6ae0a503e38c0f60fd71e",
-  "tokenizer.ggml.tokens": "604ac4bfbd019e430d7b6cdf18c6c0cd5b967900601f0307f714ec7773aa5ca6",
-  "token_embd.weight": "1d1d1d39a867d5a4bfb32792a47247d2638c10c95a6259391d02843583505cc4",
-  "blk.0.ffn_gate_exps.weight": "2e5cd43ac3f26c44f071926ff6c3f239ecc52a34bc9a5b5906d3d4c1bf2fbbfa",
-  "blk.0.ffn_down_exps.weight": "a4dfc7e7c96e7402eb70279601675b956bb7331da8101e63fe5c0a611b6972e5",
-  "blk.0.ffn_up_exps.weight": "2d5d87b378b2319c344ed2c642598b6f7cb6beeb582a8ea51abc9ae690d473c3",
-  "blk.0.ffn_gate_inp.weight": "a46aaf5aba7401ce6e41f158242b4879d34901661f3ede85496cbd0ce79d6314",
-  "blk.0.attn_norm.weight": "3fe37d913bdd2b65076bcdd6efe64a37b0b03cacbb1b80b9f7089068aa35f38c",
-  "blk.0.ffn_norm.weight": "5e14308a3c894734eb204c8f558bdc817e94bbd5b4e9cb4094e91ba388c8f7f2",
-  "blk.0.attn_k.weight": "73d943dcac0911e87bd771f4aa1c901e1bfe1aed293af06e1a67812159859f67",
-  "blk.0.attn_output.weight": "4c5f754c855e262e8d4c94c6fbbb57af06399dc0e170d7d99a1a17fc9aab9227",
-  "blk.0.attn_q.weight": "d6fd7403c873d49c05f6f03208f30d99ad34cb3b71c9990c47334d502a8e4c7b",
-  "blk.0.attn_v.weight": "cf17cf64b2d683bd9de6cebaf60e5c264df6fdc38fe719dde9d54c80334f6366",
-  "blk.1.ffn_gate_inp.weight": "0d524de81cd915816b4e714bf595ad6946a9130b3de731cd89428b2781230809",
-  "blk.1.attn_k.weight": "2ea47f412992b374c70674730fe84700e0c8cce177086ce9b6635e42408964bd",
-  "blk.1.attn_output.weight": "b4b2520794d54113e86c8ff678eacfc62e35be4395a594a6c8c22b4383ebcc0c",
-  "blk.1.attn_q.weight": "5db930c98c4f91f6eab57eb974c72210b158e366d23d6d2890b2759c053bee33",
-  "blk.1.attn_v.weight": "079bdde09668394bf7af9f8bc175017b4f48f0ab64e6dd855a4d7561d1693c0f",
-  "blk.1.ffn_gate_exps.weight": "146a62de19f9ab093deb101f9640534ffc3dc40d69f508be12fc0475d01b0c7a",
-  "blk.1.ffn_down_exps.weight": "949da94a3c0f375160672a979e85f7def284264b10d48d038238aad5f5ece793",
-  "blk.1.ffn_up_exps.weight": "7016a3f467d9e3f2f4b4019579ed86b757469cd367f2b225483305376b4bb3c1",
-  "blk.1.attn_norm.weight": "1614d1e6ed537737275eb888666c7bac533f4eefbe73dec92b591045ca9e1afd",
-  "blk.1.ffn_norm.weight": "405a455fa7d1ec36894652ceb554bbcb09a07fd6405f42741e66dc4a4665c19c",
-  "blk.2.ffn_gate_exps.weight": "90d5003fc7421f44220c0842d43128955e91488f6f785fe570b62d81b719e964",
-  "blk.2.ffn_down_exps.weight": "ecdc2b5a8b504ef0a7833acff47d69b0c1fa9c22126de1bb120ff5e48c3d6e2c",
-  "blk.2.ffn_up_exps.weight": "2cbd9485a32460d315eb50a2f3b00863fd77245bfe885b7565efac1cdb1f191e",
-  "blk.2.ffn_gate_inp.weight": "0d0a17a1a2c7a61f2cca49ecbb479154dc93a870873257bc4f225e7607f2e2c2",
-  "blk.2.attn_norm.weight": "b2e4c5a977f87a6f880896bd73596234c9b83622fa0d7add5892501e3155913c",
-  "blk.2.ffn_norm.weight": "0ab875b4280afa922376cfc7b9aa3f7071c9432ea1254091ce7de3749df0e8e6",
-  "blk.2.attn_k.weight": "bb884af51fb51550acfef54ccf1b58ce8284e587806e6a2f88c8265e1ad05a5e",
-  "blk.2.attn_output.weight": "0f03099ba1ef342ea61af9cd71d028123bbd8b1dd7d7fd9b509aef77815427d9",
-  "blk.2.attn_q.weight": "8fad0d29eb4c9d24e564774ee3316b9eb7a4c4985e4567111d2c836c830f6cf3",
-  "blk.2.attn_v.weight": "fe04c847ff677632401a94e7b6b6fdca60391ab21cb23bd791533115de6303a1",
-  "blk.3.ffn_gate_inp.weight": "29e3aaa724590c070e614af8288939603d2641b0ef11e8c0f476bebb2776673c",
-  "blk.3.attn_k.weight": "231cc5631def10f7f292d8862d6125ff555164cd70480ac76362149fad204497",
-  "blk.3.attn_output.weight": "86467a605c62852e05fda1a7ef43150df2cf715fe59785dbcba09f1c27cfa086",
-  "blk.3.attn_q.weight": "901822402453922225c2d6ac79616691d48217635d5ff7338daa971d5ddee210",
-  "blk.3.attn_v.weight": "27030784f44375720df2f090933645a31a022d3fb3b14573e5ca0b78f44070c1",
-  "blk.3.ffn_gate_exps.weight": "231ba59cc0b988d125d77bf627aa3f04636684870af88f081f3944b48a160d86",
-  "blk.3.ffn_down_exps.weight": "530c3ab44ae4d66e8afa4d10c153ba5dfcdfb7321989a988e62e9d12e7234625",
-  "blk.3.ffn_up_exps.weight": "b85c2d4d9d11332e702b3c0a6610d4f525f9a93e5d12f5c7c55c592c40755e75",
-  "blk.3.attn_norm.weight": "05dbb6d88cfa6b199f9d705ccbda97c0ef13f9ec875c595398a1a42d009a4555",
-  "blk.3.ffn_norm.weight": "6880b1c27d46969ce36fac049c05dc8b89e4bb47dc89df357e32df7e18fc512e",
-  "blk.4.ffn_gate_exps.weight": "a883b4f225b760c5a2f6605dc5e2167ab85bb398c70bf64ceb539fcbd6128dcd",
-  "blk.4.ffn_down_exps.weight": "d291bb656aae77947d4b525e2819bf4112afece53ff31de9dab999af1f65f9c4",
-  "blk.4.ffn_up_exps.weight": "38592afb8ba3dcfb26970f906174f7d3fa62da44fa4be4fc6912a19030ea9164",
-  "blk.4.ffn_gate_inp.weight": "1596cb74e8fd6c3080b937b06468bb397b0dbb661e6d180a6bcbdc43e8bfd0c6",
-  "blk.4.attn_norm.weight": "f90c83c5ff4366281d283384efc941620542b9cfdea160d678dc54a75e33f758",
-  "blk.4.ffn_norm.weight": "d28d8c49d1746b7cc085562d1074905fd14023844de823dc4fb22202bb280790",
-  "blk.4.attn_k.weight": "792bbf412cc357140fdaba543e547a9b2f7582919e307bbd9a80c7d6d8f5f1f9",
-  "blk.4.attn_output.weight": "d98e4a062d2631d9c315f1990d5f6ca9a88e7e0e46387f611ccb0353f876aa12",
-  "blk.4.attn_q.weight": "1a11a55a91d9f748a72176ff6b1c174844df406e00d1b66b9aa64dc6ee4bcd1d",
-  "blk.4.attn_v.weight": "04cb3c02b12a6313c7ac7044513441083d534fb4c5a3f63bbaa58f7edbd2fadb",
-  "blk.5.ffn_gate_inp.weight": "cbd5cdf015d33a2da6703eb74c22fcb97581fb9175435173b6dc4f9e8364320d",
-  "blk.5.attn_k.weight": "4fdf3405e4d657403f5647b51233521310ee984b4b81bbcd901cb3e6ab76b7ff",
-  "blk.5.attn_output.weight": "4a25662c46979a29600ed77e1907cf81fb16ef30e724c155444e54ccb76af481",
-  "blk.5.attn_q.weight": "e2acb30e30b97300039bb20ad0878f05159d5657fa811748a51d5b6fb35d631e",
-  "blk.5.attn_v.weight": "306504b6a26aa123c63dbbed3f4ced0ed2ee8fb6a30bf0093539b817539f5ece",
-  "blk.5.ffn_gate_exps.weight": "7e34df9b9944dbeea5e8565786d3aa6937314a4b87acd4d0874687877c5a39fd",
-  "blk.5.ffn_down_exps.weight": "c4b7a57a42b5ac0a8ae27dcd5cb2646d7a7cc7123126d44a56ab128e85f60b13",
-  "blk.5.ffn_up_exps.weight": "09d47593b6dd6c664a9155bff02fc2eb7ac4a70219a88162d05c802a01d3c6ba",
-  "blk.5.attn_norm.weight": "58804a036d6ac4c1fe357b8b6a97a5c37cae1c2f06ee0086c041d449c1c6ef6a",
-  "blk.5.ffn_norm.weight": "d872dee6789f0826211aa46ca9d0869e3e96bcace9e77d6559a7b6f3e524f3ca",
-  "blk.6.ffn_gate_inp.weight": "fb1eae732e974d6c1d020a5b4ef98c5f33016f984701bcea656f999a99daad66",
-  "blk.6.attn_k.weight": "55e9c59c5051ab5519b3a7962e1b5fa96a3c0251cb6200dc2f177885ad2de470",
-  "blk.6.attn_output.weight": "f3c834a8d0027370350e2b6294d95434d31432e57be6313b013c15a56303d61c",
-  "blk.6.attn_q.weight": "efaefe5f11c2140dc7cb532b0832c2a0b363a165cbda21f00fadae77efca377b",
-  "blk.6.attn_v.weight": "900bd734d75616d846a90a121c97e081c956a3d1ab012f66dd0bc62c43e1ec3c",
-  "blk.6.ffn_gate_exps.weight": "312a99661b1468fcaed2474621116f1681432755e973f3ee79d01912974fd424",
-  "blk.6.ffn_down_exps.weight": "ac9cd7db67a2ef0d2b5def86873673d05e48d49d147dd944469dbb8e2d4c46f6",
-  "blk.6.ffn_up_exps.weight": "57613e7e09579400a1a09fee4445acfbfe83f2f327fdf317877787d96ada6b84",
-  "blk.6.attn_norm.weight": "0e8801e09885c633bc01a9a5b85d4e878d30158a4eb41a937dc5b760ebd044cb",
-  "blk.6.ffn_norm.weight": "b8c58062ac93072f878446b0e7f958c737aa47fb769fc3a8f593133d12db2dd1",
-  "blk.7.ffn_gate_exps.weight": "1ef611732ff13edfa8d30981ed9dac00c15ceba9fc012ed0b199e9280a849948",
-  "blk.7.ffn_down_exps.weight": "856c6811945c7b0fa461ca17811cfa43436b4cdf5326bad23cbc30883486d7cc",
-  "blk.7.ffn_up_exps.weight": "6725e3e33994302ee13fa5ec163631ce2dcaa08aadde8fc166c2265d4561c5c5",
-  "blk.7.ffn_gate_inp.weight": "36b49d7f80c1003dc392b2c1b9960cd49889dd69e77b26b9e4b13d01f3d0a32a",
-  "blk.7.attn_norm.weight": "7a0ec49acc5e20ee71c6f80ca02f4f1e564c485e0ae0621309e7c2eb0c616cf0",
-  "blk.7.ffn_norm.weight": "eeae035c39ab6e64bc06a4baa1bf6e50d4c8b8797cb0ad8abd48be86974802c0",
-  "blk.7.attn_k.weight": "e8f78c1def01a7a38d2d9bf7becb17755e28fefe4927856f7890fbee52840187",
-  "blk.7.attn_output.weight": "5367f05ac3bb49ef8745ba5902e1bdd4442415a3ebff2c7e1a3918d7be6fe948",
-  "blk.7.attn_q.weight": "37c95fc5acc55a4f6e5f02cab9be60e4fe54c08b65f98f4455741b4aa542ff4e",
-  "blk.7.attn_v.weight": "c89f1343486ba55814233511e94090f7365662a8a4214aa4c278cdadc79196c2",
-  "blk.8.ffn_gate_inp.weight": "4e239afe8c7afb8de3a005757c887cf14b1622ca2d224227591cb0e5301f4c17",
-  "blk.8.attn_k.weight": "2ad0229f30fdcc1e85ce64e00d8f75902238294844a81d5af43e14ba75c02983",
-  "blk.8.attn_output.weight": "2e44a4722acb3b521b81d0b910f8ca2f6c286d874a92ddd02150566454061699",
-  "blk.8.attn_q.weight": "1cd2b09cb2f43e08de776b5f7eac197a5a6d4ffdfd52b21baa36319450147bd0",
-  "blk.8.attn_v.weight": "5a22c57ebfd33ac500cbcfd321d5b5b1783f8728801db6f3f8bed51c7183e4db",
-  "blk.8.ffn_gate_exps.weight": "91063fe56cb4f3ff3b41052bb5046fcf8ef61516a603ee90aab893a9d68c15a7",
-  "blk.8.ffn_down_exps.weight": "d4c3abc8f1d1b462f67f70bd8f404b3fcf45dceeaa8527fa120527254c383c90",
-  "blk.8.ffn_up_exps.weight": "76a1a1f08ec577716a2e7027b45293e9205751126424f1bebe1de89c78f087d5",
-  "blk.8.attn_norm.weight": "f980d774da39eb76c52358afac3e38cb4c81cb323deaabbe5c41822e3f17a98e",
-  "blk.8.ffn_norm.weight": "1c937658cf90f1a85db9a5f26e077730fdd4b694607dbeeb825c5fb2bc407e0b",
-  "blk.9.ffn_gate_exps.weight": "a2532471ecb7896d5c78e5a34e10cfaf4125265e1595166c8d0d0dfbe2a3187f",
-  "blk.9.ffn_down_exps.weight": "b47921a28412d48fee450b8b9d97cee42344a2e69f06d407fd9523d7adf13333",
-  "blk.9.ffn_up_exps.weight": "7c461bd1b2a73b439cff6a10d94afa01e8b06f7e6f09d9a6f28e3876aef48bce",
-  "blk.9.ffn_gate_inp.weight": "1648dfb08b5c06d7953a5a97ecb764995fae9487fb729a1c867023b2538149d0",
-  "blk.9.attn_norm.weight": "8635db0f299882a63b7cfcd1d4259c9e53fab22c31d3d054de36b1001380b31b",
-  "blk.9.ffn_norm.weight": "f9309aa323062d174c463613afef9b0a33501b510bfaa58a8e0e866d12ffef3c",
-  "blk.9.attn_k.weight": "dfe62030441e947a588512d18d9c6e4ed72c2f71c227d622c095e4263b23dadf",
-  "blk.9.attn_output.weight": "1977beb75c6349c50ba7dd3865d7c0a9c5c5ddc854413147b0eec98ac4fda351",
-  "blk.9.attn_q.weight": "eb132596719605cd6bd1782487f121994629e115190edd69240b12af66e734f5",
-  "blk.9.attn_v.weight": "9e708f15d332d7c5187b0693b1a977eb30a2fa10bf7df48ed9d7537c0aa6ed99",
-  "blk.10.ffn_gate_inp.weight": "97503a5d166c1925f9b65c0eed980753d411714d66896f3d0fad5286c7aba702",
-  "blk.10.attn_k.weight": "1ebdd222336bd25b48df1b138cdbe09021c4a5562ea7cb78cadd1255d2be3a39",
-  "blk.10.attn_output.weight": "5e98faa38e9d514b9057e1c8342c509cbe1083defd518e506f6bad89117d1f5a",
-  "blk.10.attn_q.weight": "3323a26c87d936d1dd87c577d0b763459fced726679612c874b3de5fc6d969c5",
-  "blk.10.attn_v.weight": "d5fa73cb56aca388e205f44455e4b4f676fdc12ed7fac4542fbb3b41ecea59ad",
-  "blk.10.ffn_gate_exps.weight": "225021b53782800906cd13b70be3a4161e8b300b97f984a959ccad6a6e8adcbd",
-  "blk.10.ffn_down_exps.weight": "f08eb91526bd22f5fd0402fe925d6141cdbb308a1ced0330858d0c85c71f5ef3",
-  "blk.10.ffn_up_exps.weight": "a9f688350c3b53eaada5103b5848bd9a3d7d6b327a70fa16c24bf28ece933eac",
-  "blk.10.attn_norm.weight": "5ba426c9dfc79805015ccd76cd1068b0ad3bb7a8453e14bb1d35486f122d8f95",
-  "blk.10.ffn_norm.weight": "98891d6acbc3986b2581b7a3af9f5946a392d9188972c6a8b15d4e745a4f2482",
-  "blk.11.ffn_gate_inp.weight": "b2365a60566e7dace892e1cb0e62eb73ce387352601723e847052b34874feaa6",
-  "blk.11.attn_k.weight": "0efbc1d1430505543ff71532a4fcda821aeac616ef6c1dca40e00d4f2ff70bea",
-  "blk.11.attn_output.weight": "3d5bd4d9a41236f30d4293edb9ae27beaa113ffb31b4fbfadff3a4c370dfd3e6",
-  "blk.11.attn_q.weight": "aa11e9db14dd9c77951511443077c2a1a78070753d7bd3d9811038473f69e325",
-  "blk.11.attn_v.weight": "5adc567f377aa11d1763d35f50e53fb2896a8b03b623ac36acc45efa2486d512",
-  "blk.11.ffn_gate_exps.weight": "71d07d982aabfab9eed3c733d49c20f023bf475368fc71db5084d91beadc4b47",
-  "blk.11.ffn_down_exps.weight": "9a06e61461e48b3925a9f7d9cca634d048c8b62163d7bc5c43e35899f959319e",
-  "blk.11.ffn_up_exps.weight": "bc05494d0dcec61021b3ac0c5bc1bf502736cadf48224e213bc139d562699a89",
-  "blk.11.attn_norm.weight": "a5758a10bdd0404ae1470e8e9db903985d4d07f60553c5001a5e7b660d4f7ada",
-  "blk.11.ffn_norm.weight": "814ae037563aad3771787316bec4806c95bf6f5991dd6474b4b1e5cc13dc18ee",
-  "blk.12.ffn_gate_exps.weight": "3a68b831ba1606fb9ef6dffed4732032447ecef23ea563ff4e79317586c7eb49",
-  "blk.12.ffn_down_exps.weight": "268b25e13f4b7beab08686e83705a41b21d15251809ee4784526f78a580da829",
-  "blk.12.ffn_up_exps.weight": "9105751a5b5b42ca2614d0456f24f779d2e2ac8cdff0f96842aa7ae2b70f341e",
-  "blk.12.ffn_gate_inp.weight": "d0de1558cc1d458c5c504f63ddc59785c323df7330474bb0644c346104b40a3a",
-  "blk.12.attn_norm.weight": "859a4c8113678e2e202d10299850e0cfb52eb11ea50bcbf4fe3ff39bdd394154",
-  "blk.12.ffn_norm.weight": "7fbf4c459c1760218877e9ee3f5ad49e960956a4369bcfe96c143f04ff9ddf97",
-  "blk.12.attn_k.weight": "0a7e254fdf3730a57372b6ff421a613eabaea68cdefd64800857941411318374",
-  "blk.12.attn_output.weight": "ceb763fc15d88af149d8fb78e82db2b7dab3aeae584af8cf7611a12356a397e5",
-  "blk.12.attn_q.weight": "a43402d23c46cb2d3cb3c2a98c81b19d10026b7e6742370fed6b2880b6e049b5",
-  "blk.12.attn_v.weight": "3bc24f2c0480ce91ef72993ee8f1cf962f7359e12183424583ffa1246bf3db52",
-  "blk.13.ffn_gate_inp.weight": "a6d68c82bfe66d8bab68f980f5f18268a9e2c0cd6b8832ed39010e0de198ae05",
-  "blk.13.attn_k.weight": "0166c39546b37dc2e01b2b396ba43e183f797dd04eaa51a6d103d8b58ee4bace",
-  "blk.13.attn_output.weight": "2ce5eb198deab9557475a58b69b11e9874b547e05c23f223c6e42fa35ddca069",
-  "blk.13.attn_q.weight": "745c1bbdf434284a7fae98f45e821c076dd9c2a2467dba6a9d8cf0041e419dbc",
-  "blk.13.attn_v.weight": "9ece68d5ac64d1421ea7aa32e1cff9cc1fecf5175f4c4da858dd31d8633e3337",
-  "blk.13.ffn_gate_exps.weight": "ccfdcb4670b131689de12d396a010b5ea737795cf5c15a14a304d720b3c7c899",
-  "blk.13.ffn_down_exps.weight": "8b8fb328664764f1aaa5cbdec336d5654e981e965a02ef622bde5f07ea1c164d",
-  "blk.13.ffn_up_exps.weight": "d2ace0236c2fb3365fdc85499d676a7f65813c48e5085348b1df1799922766ec",
-  "blk.13.attn_norm.weight": "1ed29d7d89ce52d7cb4d57e895ff7115430466e917136c049c385c030ed44e9c",
-  "blk.13.ffn_norm.weight": "a194fc542597a4dcfdfaec5e3cba2a2b2b21b21edfc87c39c0d7f7651355bc4d",
-  "blk.14.ffn_gate_exps.weight": "a625e3574e5e740e7f8e2f9c40390f2f382c720aab5b10534e298002dd8d1fb9",
-  "blk.14.ffn_down_exps.weight": "bc366f015b83c865946afd74c8a884943e0ea2c671314a0b7bb72f21a44d2f78",
-  "blk.14.ffn_up_exps.weight": "ee3199bf2086de77b49f57f487676be8ee70e102a2fb5a5ef8ddbbc28a9eff41",
-  "blk.14.ffn_gate_inp.weight": "2b437870c850fa2e2044d032bb02908af634356e37466fdae260b933e48ee8b4",
-  "blk.14.attn_norm.weight": "cd8344d193a1cbd42bd898e17f4bcb1ca0b2918420fbdafa9249a6f2b7f4ae06",
-  "blk.14.ffn_norm.weight": "70eec40374e558fed5b07257283cf36342b6b0129285a00007deb59c32c9f7c8",
-  "blk.14.attn_k.weight": "4053bdb507e0543d724b632570bac86b31707696d90a0db44c49b2a082e0d599",
-  "blk.14.attn_output.weight": "0182632cb0e06a07241b8293d25d109fbc1862e1e337d435f908e8681e2eb1ab",
-  "blk.14.attn_q.weight": "ffc7794a4c1b6f793c842dba969435330a7a80b9212e457b4b2ac33e68b41241",
-  "blk.14.attn_v.weight": "6411805292d528e61bbaad8f9aab9dd073529a17946c057fb06864fad9cf3211",
-  "blk.15.ffn_gate_inp.weight": "77d0744567c76e6abb67f81ba9c715b2b544841186d5b948309571eff213bafb",
-  "blk.15.attn_k.weight": "1f7957954ea4c6521c257b35a360e868ffa02bdb3de91f146d5e06bb4a545c98",
-  "blk.15.attn_output.weight": "d7809d36bd8d3342240c46fd87bcc7f9821a222f48d9a95e45ae50460265d3cf",
-  "blk.15.attn_q.weight": "25f509313ae4d8401b871904059f472a26f5714e7c791c725de77a1a522c976e",
-  "blk.15.attn_v.weight": "96fedf5a591fc0f020e6de10fd72ff12b3ef9cf70cd21dabaa0d3e7b06f54e73",
-  "blk.15.ffn_gate_exps.weight": "8f950d976b2fd9a3d213b84123cf114c1377efde9352767fb2ddee89e177c8ef",
-  "blk.15.ffn_down_exps.weight": "6fd09d1557bb94b06efbd4f6a1ca4be532a202ba290e9315bc8da3d12a5c4c4a",
-  "blk.15.ffn_up_exps.weight": "cbeb59ae7b0266a928dc7e3a6e70a9330b92f9ee1b17ee1ed91022108204a33c",
-  "blk.15.attn_norm.weight": "2005330911ac2edc7b6d27aca021c67d30d16eb632e49b1a13f30fdb2717aed0",
-  "blk.15.ffn_norm.weight": "0e9198f3b548eb78acc8961f2b3350d238d26cec110933ba753a8cf0035c501c",
-  "blk.16.ffn_gate_inp.weight": "a41d1f99d739c8b150c3945b6949763988d0c6a4c5a2b5855592ca1a48ed23d5",
-  "blk.16.attn_k.weight": "b624e2ec88c2d3047f60530fb87e72cb4a5e655a9663f6f3e9b09e5ad32cddaa",
-  "blk.16.attn_output.weight": "687759ea75e45108526ffc1573d6fdf084728079bfc2dc89b9979e76280f43c4",
-  "blk.16.attn_q.weight": "beff3a45c7e9ec82ffc6d3c701126be28654d10aabd747d03441210491fd31b6",
-  "blk.16.attn_v.weight": "43a349b13f0b9d040cacecd942bcb168c030fef8c75c987d59a4fce6c14e855b",
-  "blk.16.ffn_gate_exps.weight": "793406d6c13d727c82bb7b692ca98d65ca975baee69fc57be5378d77c5a19b62",
-  "blk.16.ffn_down_exps.weight": "9bad3dd150d0230404b7f886ac7ff8803225757e813f195cdb26bad245243b4d",
-  "blk.16.ffn_up_exps.weight": "7449d663023fea3496475bf0a9c1de7272ad0ce9adcb3265e8e424badaa674dc",
-  "blk.16.attn_norm.weight": "a424ce34c195a401df1ce37ac4f2794e8a6720b1ee8acb21428e2b68c65e0125",
-  "blk.16.ffn_norm.weight": "405a68bb8e16e1064df2de55ca3cd9ceddda1d9fc0af007a9bd7cad4b2676248",
-  "blk.17.ffn_gate_exps.weight": "97c6e5321491ca5dc039ee88da0eb0e78f347372785411809af84b3298cb19dd",
-  "blk.17.ffn_down_exps.weight": "1617ac19788a1be19bac69277408761e6bdf5719d63a8c7fea14d41cc27641b5",
-  "blk.17.ffn_up_exps.weight": "4ead1c365f112581c10610ea3f63d2a1474311d2503d2060fed4b458ef337f5d",
-  "blk.17.ffn_gate_inp.weight": "ed4b3393f2523f2b5e0fc7680a1caa2842e605728a529b5af68a7fa8d7abf940",
-  "blk.17.attn_norm.weight": "beac17ef86a7fb2b5840cc72f7a95a5e3d6bd24e7fa698e0b0ebb9bdac45c561",
-  "blk.17.ffn_norm.weight": "81cb58ec6d6dc02a0b4ede10adc336dc865fa76f982d4eab0e4a37b40f5b0fac",
-  "blk.17.attn_k.weight": "eab569e5ea8c8b05e5a6a209fba031129453c2e28181eee3e736b3b04b36bbec",
-  "blk.17.attn_output.weight": "f85b70f01438ce8fe5d10599b113f30bf18dee2bbae0657d3eba295870001db3",
-  "blk.17.attn_q.weight": "887ceebfbf6a2b94b43d2df4439ac3a5bbc29311d4b28addc04d525546032047",
-  "blk.17.attn_v.weight": "2df9414d65014c06a93da22ba3a668be7b83e2e8008e98d7771f7dfebed98298",
-  "blk.18.ffn_gate_inp.weight": "9b07741a0950fc667e5fd25937e33bc22e1f764f80eb4ff3119f005327ae0f6e",
-  "blk.18.attn_k.weight": "8649598dbb63938744c39bcda5ce8c31773e29c573be8d4d2c114f5030f8d3e8",
-  "blk.18.attn_output.weight": "f8e391adb92622298ca834d5d1eda48b69c3b1c51c5a584ef6c54a725c298d75",
-  "blk.18.attn_q.weight": "84bf8708a2eed618f48f69c178ed7dd11fa4c468102376e72e910ebd037d131f",
-  "blk.18.attn_v.weight": "31db3cd773f09548c2c1b1eac2718e46364a7810970fe9c433fad9d8de5397eb",
-  "blk.18.ffn_gate_exps.weight": "be2a2ba378002f1b61f86c273a69eede9b93786d5ce96b4fee1861f730dca4c4",
-  "blk.18.ffn_down_exps.weight": "d35196159e37705db50a5343e3989f7335477f1a4add67ef42ad64a638cd07ae",
-  "blk.18.ffn_up_exps.weight": "c6ceedd86e97913a6dcadc838e7abb762d629fb8dd55f15cf02fd9bd66d2ba78",
-  "blk.18.attn_norm.weight": "41f0b1ad83d6e3cb9fbe0d27878c2e7ad4a351b9f554a6bc9117c01745cdf6e5",
-  "blk.18.ffn_norm.weight": "96646204bd0d82f25dc77faba4dbd86b1332e449313e6684e00122da8be99057",
-  "blk.19.ffn_gate_exps.weight": "c6eb7f61e7938bda0492dbc05e51e8f631c99224fe18e99861fc4fc53ba9e9ff",
-  "blk.19.ffn_down_exps.weight": "4384803da3a3a3d44120d7dd192fe2c9bbd9a1a0cb492dbec1fdd7565230f1e8",
-  "blk.19.ffn_up_exps.weight": "22d73de2fbb8bb0f1bd2caf17fad8a355c47d914143f7f6e6d0128f66f074a60",
-  "blk.19.ffn_gate_inp.weight": "9a0cc4a2301a5634022fbce41189021bf0d1a961792d2d9330fd35556d18e5bd",
-  "blk.19.attn_norm.weight": "c5cc56ec5df9a1f7d5ad71fbda49f1433132e58895d45cb44c73420bd61ebd6b",
-  "blk.19.ffn_norm.weight": "77e17de741742ef2482fc7872fd423c8e3c1454dc4d2be89ee939084b6d78bc0",
-  "blk.19.attn_k.weight": "a92ea36ce2e3569656306aeefb835ccd5d1b03b33a86e0d3d030644cc923b813",
-  "blk.19.attn_output.weight": "5e2a912b37855f84ea964907a1a86d609cbdd79efa0c93c3e8e2fc07caf7c226",
-  "blk.19.attn_q.weight": "4ef3a5913292ac3c1a6fd3e9e53d011021f2b41d0276cf849706d1ca925cf7a7",
-  "blk.19.attn_v.weight": "42981b75b68ae852cee638b5433605c147da4392aaa6d7a06e756115b0171f39",
-  "blk.20.ffn_gate_inp.weight": "71381b9879a7c80b9f7b475abc0aa31b8cd71ccc00856ebe89764a2acb9df2dc",
-  "blk.20.attn_k.weight": "1928b7ebc054eb3967929ed6fb446314d5352f4aaf8b475ce55c6345019f2ea4",
-  "blk.20.attn_output.weight": "6071ecd9ca91af0d2ba93fef4a1a56f3b243dd70f862a21a2d164d56f386043b",
-  "blk.20.attn_q.weight": "002e95042a40f36ceed5829e3d0c8072e5f5e4ee86a089e2902b2348fed24dd5",
-  "blk.20.attn_v.weight": "42f509cdb1c0e298f89f896e349be86952c5168e49b3f83bb17badbcb7596d57",
-  "blk.20.ffn_gate_exps.weight": "a684a3ffe4b0a57c819a5fa9cb3521de223f392732927271e97ce925b6e33765",
-  "blk.20.ffn_down_exps.weight": "e3081a7bc7ba750d8a4886bc8ca4f231b55db4ca082b54b4106c7531964725cb",
-  "blk.20.ffn_up_exps.weight": "fad0fd5eca36ab154788da28be8ec25bb5d6db06c9d133db89e96df358a2f6a2",
-  "blk.20.attn_norm.weight": "c3e3f2429715ae95e884ef1246b0b461b23c5cc0ed08beecf70a14cddd184820",
-  "blk.20.ffn_norm.weight": "ff31f609dda65ca496b0584fabea6550e42edd05ebf229812aa6b7bb5ede15e6",
-  "blk.21.ffn_gate_exps.weight": "366f09ef0ecfb86808eb3296cc9abdb957951d27f6533c03f1422b54061da660",
-  "blk.21.ffn_down_exps.weight": "3fc495947d27fcca7fc0893c8a96e5d48ba27b2c8c58f8fcfb8dcfcd5539741c",
-  "blk.21.ffn_up_exps.weight": "6713ed51410bcc8283cbb001c4ad784098f25701e8021f4fa4f411e186859c4a",
-  "blk.21.ffn_gate_inp.weight": "6d4c92c01ec801647134d907bf1108878156df266a6107abc10526332b328b93",
-  "blk.21.attn_norm.weight": "27605719ae2df24f4f2e85a730927cab20367631612cb501631f6bbf38eb1209",
-  "blk.21.ffn_norm.weight": "ca80ee8177db185b15a4a378c1cb6f7143c76546a7f1726bda23f329323d4ffa",
-  "blk.21.attn_k.weight": "9e49f743d4a5bda9b4bd9c40c2ca37cdae5aec7e54cb193897ac8b4945ada14d",
-  "blk.21.attn_output.weight": "ab923540879753feaed152f5950f69cdd83d8f2413ca873f5f038b63ab0aea12",
-  "blk.21.attn_q.weight": "62617fc3f1c9d2aa672a4d91a121c7a91b92d145b65e75f0b06b4bb7c825dc36",
-  "blk.21.attn_v.weight": "15f8b2e72f8e8e992f2f6b3e93238a9d7be7bd6136f91c9d04b4b4cd0cd60369",
-  "blk.22.ffn_gate_inp.weight": "3ddb1773d9257b68add7a2a4e94dad25ed926803e02707863dd742ab9b2dc179",
-  "blk.22.attn_k.weight": "680e45a9e8d5feddee5266e119dc053bf80718fa9af1cf6803e6f493b265f1eb",
-  "blk.22.attn_output.weight": "0d5fae3402fb2c5aa3a860010e3973fc8e3168d1015f7a76b7b2964681693206",
-  "blk.22.attn_q.weight": "eee7e3d426ab533bd18d62c9aa142eedbde394bed07db58313e0fccc82a23237",
-  "blk.22.attn_v.weight": "26b5be1fe3c2b6824c5a648a3e4bdf17691904526fca158fbc3ebb627b67e2f4",
-  "blk.22.ffn_gate_exps.weight": "32ab7a7735313d60f6a75229b1aeee940b6aee176c9648536bf5921b0dc2929a",
-  "blk.22.ffn_down_exps.weight": "67590808f6a67777d3eb7976c31fe616d388b98fecbb12253b72d1241d70753f",
-  "blk.22.ffn_up_exps.weight": "fc245c0183e6d90829ff5e71a4ec93e4860b3d4c1a17b9dda2fb64f5f5c9ed32",
-  "blk.22.attn_norm.weight": "128e99d206d4d6724758ec97468af767fa0aea592149c324b731659c1e74a1a8",
-  "blk.22.ffn_norm.weight": "e45f498033f0cffa15da0eff2c47b4472e43fcf8921729fc4eeb2e3a6b3c78e2",
-  "blk.23.ffn_gate_inp.weight": "d63e686f5325fbc89fa242c2c52a3b8ff54f867dca914c9ae6eea13e9d6f46e5",
-  "blk.23.attn_k.weight": "f71f5a577f46ea12b1818f3a5ff4b85ddc45f9a2afb0fa2e041d71a3e31c6779",
-  "blk.23.attn_output.weight": "92b13563c1e0eac0d748fb67b235dfd7a64c8f16e2dafb316885744582e23b4b",
-  "blk.23.attn_q.weight": "2f9b9c35dc4f912f3f51c06e2d68f417b51a0de0a84aac530a64f9d3d7b0a2dd",
-  "blk.23.attn_v.weight": "268e40813806e74a5c364b19556d087bf8374e76e7b6fcf55c381eb7da13ccd1",
-  "blk.23.ffn_gate_exps.weight": "12f857e7a7ce228afac34d99b602c8d6fe96984f2a21118f459a58cb767ee65e",
-  "blk.23.ffn_down_exps.weight": "cdb082c16599c3bb36a28066dcc122d9529b54fa91b6cf0153437ec960a5e16d",
-  "blk.23.ffn_up_exps.weight": "f4b99f6f44d7b8b5a305894e88633bf5938fc1f6303a2b2092399da9c8b64d7c",
-  "blk.23.attn_norm.weight": "a691392210383915916b4d3886d5e4d56e7855e27e37e414fbd73bf66b3712e6",
-  "blk.23.ffn_norm.weight": "0c3dc72f667e5ae19b69bfa9f2bd2a01a57681f89ef9527bad4eb0d8c7b70da8",
-  "blk.24.ffn_gate_exps.weight": "86baca2a3157994df7fd8ced5e08436d5c1810dc29c0715637c36de723e0e7d1",
-  "blk.24.ffn_down_exps.weight": "ac5d559562b35c34993e34b071f66d15c65be5907797078c2d2a49aba54e3192",
-  "blk.24.ffn_up_exps.weight": "fce0a099cf09777f44fbab3606ceb75f7fae6f0b80725f9e871654b8cdf9262a",
-  "blk.24.ffn_gate_inp.weight": "e7c6800c0cfc56b565b2d35ad6f1dbfdb70dd0b05b338bc8da2286ffc3678d79",
-  "blk.24.attn_norm.weight": "dc6cc18ec52d102d015153c4a1132f9d7a504e29cbdec81c5edbf3b9e65815e1",
-  "blk.24.ffn_norm.weight": "480d5a1397af5e0e657f1e67d20ec0cdef5724e71246a326843321b87ffabd33",
-  "blk.24.attn_k.weight": "338c0597954a9b95a782545b2fe36469553e73f86ae2d2b5697767b28e1c7daa",
-  "blk.24.attn_output.weight": "a77d23b79933c67e52f1eef7f83a3dff4f767ce0bbcc39572f8cec4acd457643",
-  "blk.24.attn_q.weight": "45c9478593002be1998e96e70668aafa2dd3972380fbc1df12fb05c24ba959e0",
-  "blk.24.attn_v.weight": "515729420885408a6a9614bc27cda393ed907521318d14d21335d39a3eff0b61",
-  "blk.25.ffn_gate_inp.weight": "aae4ac40e9ab3925241f9d784b54b38851d9bc999a6c3bc03fc3f17c9b28a67c",
-  "blk.25.attn_k.weight": "4ab4808d02396c35b00b426f536015673b71c17ae6cd55bbc2e6bfe7a4c59d0c",
-  "blk.25.attn_output.weight": "1990bb982b77e0c947cd1a8ef0b36227ee1259e6dbbc2829e5c136edf88675eb",
-  "blk.25.attn_q.weight": "a1490f3048e8c0ec8784f8550c43adf5cc8d0f2f90131c934713fe4b1b015bd7",
-  "blk.25.attn_v.weight": "f15e53c6d45b3b6f58808fa968425d65e0b26b7f9b268127a77abb1227c67431",
-  "blk.25.ffn_gate_exps.weight": "656662447ff54f56ee80f78a1b9483f7efdc40f7375d0cd8a9c72ccf21f77e7b",
-  "blk.25.ffn_down_exps.weight": "db06f101bccbaef19cced0f6c185166e18202465f4a42cddfd535fbe5cbabb4a",
-  "blk.25.ffn_up_exps.weight": "584a7b02456f27fe1d8d3c7ccd21d426b6ea887795a3ed77f704596a1e3841d7",
-  "blk.25.attn_norm.weight": "8f0f3597982930fd237e9d609776c64f2b909a455b21678f83a7ebd4bbb83e64",
-  "blk.25.ffn_norm.weight": "3e7079c32582afba0c55e032f254adc18d2997705eec860185e9a6dd3d82f07e",
-  "blk.26.ffn_gate_exps.weight": "e70341691b583b86489812b29b77aa41eb658b1865733d6118da54c66e3bfcc6",
-  "blk.26.ffn_down_exps.weight": "5c1b812d11dfb064af816ced5ab6463bf9722eefdfc341b8a93705d5038fd781",
-  "blk.26.ffn_up_exps.weight": "e18118362ae54ef7432781c83884f9fb230a9d934e342aabeda8822ea5f71fb6",
-  "blk.26.ffn_gate_inp.weight": "cd1c5f6710166b9567c6b74c97b2348b191c60aa860958c6bc264ab095261dff",
-  "blk.26.attn_norm.weight": "71d087531af2520bda2e676c489e8529cef5db8aeea1eec0a937a8b4f2fa2e54",
-  "blk.26.ffn_norm.weight": "7f704e936fda28eb5c2cc339f0f6a5f78170b5aa43c01265b21668870d819c82",
-  "blk.26.attn_k.weight": "1cc62a0ce0ae251275d898c52c4a9fba5995fca10955d2011d10dd1a59e1afb8",
-  "blk.26.attn_output.weight": "636e881b1505f9cef656a4be98bec6a4765321d51f9bf1dac8933397cf44b765",
-  "blk.26.attn_q.weight": "89a3c4d202d7d6adebb9e0c1bcfd8b775f6456386f1be25e86e43acc949c1e16",
-  "blk.26.attn_v.weight": "ff2cc963b597cdf1a21703f3e7022af3bb4c65a34a19e19d9309a7c5e198b5bd",
-  "blk.27.ffn_gate_inp.weight": "6150139498fefe380bb99d11e72028da47a15ecb73dfc5b2774f726f4bed8f9e",
-  "blk.27.attn_k.weight": "f286eb9e5c56c7b801a497aedc40158c2a27877d7f9fb59b3fc67834798902d2",
-  "blk.27.attn_output.weight": "5dc3d3a05f9f7729509147fd09c16fb53f85f520cdab5cb69abf4bae3fd460c7",
-  "blk.27.attn_q.weight": "8462e40f86b24251960d6f35a9ea99b8793a01937faf1aec2859f2e5395dbb61",
-  "blk.27.attn_v.weight": "bac1a99e38e25953f8315f7212eb9777dc216cadb09b959977885ae62724ceca",
-  "blk.27.ffn_gate_exps.weight": "6a15eca7f0f6ecfd93db2e55c63875348ec4a78c4ff643ec46df9e958c0101e4",
-  "blk.27.ffn_down_exps.weight": "2e1c91247c4359e2073a8e5f26fd7f6426da7be3ed5bc65dcfff701f0a5022b2",
-  "blk.27.ffn_up_exps.weight": "65d6f5c553c9332085eae4aeadf25090b5d7768212ea7b08ed698102c21b29a1",
-  "blk.27.attn_norm.weight": "7fab8ae63ec8e91ce625cd130ab96d8427dad3a7413bb21b25ec5f408c5b9f5a",
-  "blk.27.ffn_norm.weight": "532720546b0fdcd423a02ca6e3e9d8aacb84b1b3e8269968f88a47fe2a69bab4",
-  "blk.28.ffn_gate_inp.weight": "a305ea58d98962d9dcf0c53ad2389b7acc8936fb35a0e3fc9410e7767cd49dea",
-  "blk.28.attn_k.weight": "8315e8a2e4f78dfdf36d4fc18fffc74bc95fe42c3ae4f9af2b6c874612c0f71b",
-  "blk.28.attn_output.weight": "9b5fdedd32d39ef46a22cca7cd5355d7b93bd07ea305f466a8aad6ca5a4f3778",
-  "blk.28.attn_q.weight": "4e8fb96997c30e231c437130f410d7c91d541a816f6c568b5f3bfdb4b8dece74",
-  "blk.28.attn_v.weight": "1fec739cf3bd7b4913f72ca358d4cf31391c304de44ac0ae31ecb825beaa7cfd",
-  "blk.28.ffn_gate_exps.weight": "9f259789d535e09268266b9a8020f32d6a6779966c909d91d3a10574f06238a2",
-  "blk.28.ffn_down_exps.weight": "516d3f8abaedb01b9916a4b67d4672159769138ef2850158bc1b32c41e31f0e8",
-  "blk.28.ffn_up_exps.weight": "f2f1d88d2c31ed588806fb5ad981d68f5134d7284c4fc022fd018de2eef437fc",
-  "blk.28.attn_norm.weight": "960fd005598deadaebd969996f4367a9dbfad90539a863674fe95730935acc64",
-  "blk.28.ffn_norm.weight": "e1993b37ced93d4049e9af2c47b0d9207d8f7e6f2cc3a52f57bef30bc806d805",
-  "blk.29.ffn_gate_exps.weight": "58927146338f443513337476b3cd30e6341742f096c2beb5890d400f10121298",
-  "blk.29.ffn_down_exps.weight": "03a3386e4f0b75a28c5608e23b2de8f0de25f21954e4aa7fc343431bde9db07e",
-  "blk.29.ffn_up_exps.weight": "6916b7490a7ae7b04a5d81cc1e7ac9b20c483434f3b186b12d87fe176bf1567b",
-  "blk.29.ffn_gate_inp.weight": "98e710e467a3d567abe4ce29d78b8e8dc033148762290c0c5e1ae4d78efd8c78",
-  "blk.29.attn_norm.weight": "4e64cb307d37be20d55f38c94faf7e451d11df5e60df347906cbaf9c5441be71",
-  "blk.29.ffn_norm.weight": "696c23a52f742679bd44440d687a4c44b4302d57f1e9dc5610d23374336187e7",
-  "blk.29.attn_k.weight": "e85253652fd6120c623634ba66b725bf7cd491318b54ccdad2c7df8851d64c0a",
-  "blk.29.attn_output.weight": "4f650a71efb150d1f24cd4d114d4187bf570ac424da3b92ea6455abdf1aea705",
-  "blk.29.attn_q.weight": "69fa7da901026ebcbbbc848455b425458b7e3295007d7fc093acf4b38e2166ea",
-  "blk.29.attn_v.weight": "17e2e7590b317b21f106de546aafd955579703d1e95d6aea044ee72ec3a514c9",
-  "blk.30.ffn_gate_inp.weight": "3a03284b4aa60d59d4a2ec86253469b61fc656372afca427cb77a5332fbcc62c",
-  "blk.30.attn_k.weight": "d518cfd0db9708e769eb1399e87ee49357dc54d5afdbac3d4c0ca46c64e789eb",
-  "blk.30.attn_output.weight": "9b44378714d784c5ef9ab604359091baca4e0ec222afa139b7f840eaefb371fd",
-  "blk.30.attn_q.weight": "cbb95365bbfbcad0c9cd99b4eebb5a5d32de68ce08e4063b5ec3e792b7548044",
-  "blk.30.attn_v.weight": "e7985c04fe1740e35a9598f43b67b0922b4fc2d00b68a92a9f917b82c3248de1",
-  "blk.30.ffn_gate_exps.weight": "8ac4bbd07935d98f895ba94dc174e5ad5046c3c222b53729d60f987c05e7eb70",
-  "blk.30.ffn_down_exps.weight": "dd672cc71e82abf05064a18121b8e55fe1a4f19bc1d7cb9a142f4add54bc336e",
-  "blk.30.ffn_up_exps.weight": "12282f664a2a12aa25e2deac58946108715ebb978bafed5274cef24569107646",
-  "blk.30.attn_norm.weight": "1a33458fee054c6c9c896a4bb0a4e1fbfa0293b2408c7dd2b81d692e966e7273",
-  "blk.30.ffn_norm.weight": "311e33b68051f507f1478ed8f2693fddb846170ddb7285a91be43f795c2ce31e",
-  "blk.31.ffn_gate_exps.weight": "8af43d9867a51cd8392fb48b981b0ceee0ae979c491c07d711b3b56b5162c786",
-  "blk.31.ffn_down_exps.weight": "5579cb7758c1600b19d1f540deffe081b575962e37437b3b2efb2fb0a2924e40",
-  "blk.31.ffn_up_exps.weight": "f2e7c005276b3a001fb40753f027fa10b4d5a346f43cf4b4bbdeec6e74e1cf6a",
-  "blk.31.ffn_gate_inp.weight": "89885dc0e30b6b16a90c0331d7fa3174671e941364e8102d934f02132237e61b",
-  "blk.31.attn_norm.weight": "99e4e9bf86a9edf8c404153a7e8a82324ba79da462622196e2faba161bd95172",
-  "blk.31.ffn_norm.weight": "55335997cf6de781bf332b943de96ff4646966b05d9fee86b76ea897e27b6ca7",
-  "blk.31.attn_k.weight": "cee570762b78da6316b637892cc4b080e40f57af5551ffb1866b9a8e80e96628",
-  "blk.31.attn_output.weight": "fa321ff55ec7819ead7b819fd45215262f39744569765ba2113c989c03588802",
-  "blk.31.attn_q.weight": "9e2c409b878f8a2a1436874abf428fceb1c534b21f9ad4dd6f532b8a469007f0",
-  "blk.31.attn_v.weight": "a845d0be68ba537b4a775bfba4d897faf7c82a811a2612b0b7420cc4f3574cb8",
-  "output.weight": "16101cbb74b54cda9ebc07ca3c762e3263a56efb3cc011156184b95807d7cf13",
-  "output_norm.weight": "d7aa61585baedd60157aafe157930785742c55989c288573566a971b02423564"
-}

convert/testdata/gemma-2b-it.json

@@ -1,188 +0,0 @@
-{
-  "general.architecture": "gemma",
-  "general.file_type": "1",
-  "general.quantization_version": "2",
-  "gemma.block_count": "18",
-  "gemma.context_length": "8192",
-  "gemma.embedding_length": "2048",
-  "gemma.feed_forward_length": "16384",
-  "gemma.attention.head_count": "8",
-  "gemma.attention.head_count_kv": "1",
-  "gemma.attention.key_length": "256",
-  "gemma.attention.value_length": "256",
-  "gemma.attention.layer_norm_rms_epsilon": "1e-06",
-  "tokenizer.ggml.model": "llama",
-  "tokenizer.ggml.add_bos_token": "true",
-  "tokenizer.ggml.add_eos_token": "false",
-  "tokenizer.ggml.bos_token_id": "2",
-  "tokenizer.ggml.eos_token_id": "1",
-  "tokenizer.ggml.padding_token_id": "0",
-  "tokenizer.ggml.unknown_token_id": "3",
-  "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
-  "tokenizer.ggml.token_type": "485e40bf3d715a4764818fc097d6a2a41db872d82ee714bc500872a3437ff48d",
-  "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
-  "token_embd.weight": "17b87ab2c01c80657855a5413d0457b4a041afaeda0cc785080e44e2f04acf07",
-  "blk.0.attn_k.weight": "28ac0da05754ad2714ae95da28a5ad191192140b30b8fd22d108d4700c9d989f",
-  "blk.0.attn_norm.weight": "3f9d5675d1ab0eb8a816719dac9fab81f2e95c52be02c34263339acbc087febb",
-  "blk.0.attn_output.weight": "703295c2c63990ff896778685c678f145298886f680f3ed5dc2a7ad54c293265",
-  "blk.0.attn_q.weight": "69c2d0e4870e9d722a190d356203c9605575a16863466c3d1747966ef1cf5791",
-  "blk.0.attn_v.weight": "95219c9c07b5ffe9a9a01e456d845eef2b11f4fc12c93dbbba479db395444c13",
-  "blk.0.ffn_down.weight": "a2feb5eb3d572c57c5bafbf0ab506862df1160fe40965dcfe4b9fd855c08bed7",
-  "blk.0.ffn_gate.weight": "fcca072c445c31f4dc4d5dfaa785b1bdf7271342442099b74fd17268b5829fbf",
-  "blk.0.ffn_norm.weight": "7621f95dbd245cade6fffd6b08797d69d8e3954e960f0b5551b90d967ab95448",
-  "blk.0.ffn_up.weight": "14a9bcdd451403c67136391e1b6e53b3b1830f00199bd911dbcc56d8749c14f4",
-  "blk.1.attn_k.weight": "c70f73c5df20579cb44d971164b48b5f0d8d5abdb38b381e7a8b880ba12aa406",
-  "blk.1.attn_norm.weight": "88b6b91f93a1ef83425a7c7dc2a2fbd3b22704a04c64a80061df376ac8c33626",
-  "blk.1.attn_output.weight": "f031a537490c452be3b3bb51e6b7949a636405756e160976a1c070a792ea00ee",
-  "blk.1.attn_q.weight": "bdb23214b1cf9cfd30f863a0a5868e52c6809d93b7e8f44df096a94204d9896a",
-  "blk.1.attn_v.weight": "e9bbc0b05f2c872fb1403f8f938cd1612b502229ee401f12593b1164c61acc00",
-  "blk.1.ffn_down.weight": "5ff53811038b661a7b8f2bfdf213bebfb185ec1a6060b662f063714f33584d79",
-  "blk.1.ffn_gate.weight": "205085c8c951a5c7543b1495183cd96028fb49f67464b3e9862a2693a6077a33",
-  "blk.1.ffn_norm.weight": "798f354fc85afce9625f5d10093a585a966831698a0560e6c9b97ce659eb4b22",
-  "blk.1.ffn_up.weight": "db92dc5684cb6e90940e13f4d1da555ed20ba4f8cab1e990ddfd7553e2e91315",
-  "blk.2.attn_k.weight": "ef5ce360c4eed6d00d03ca4761e0f8e4b0af4509978468314be14f3d46621044",
-  "blk.2.attn_norm.weight": "6dadbc05dbd0d3fabb4216affa60a3de1378a82d2859dc90b338cbe70f50d455",
-  "blk.2.attn_output.weight": "6bbf87a966f691bbfd7c8d25629aa4e6710107bd431a667434861febb391edc5",
-  "blk.2.attn_q.weight": "4e575c09ae2de417ce9057ce8b073680e860a24aae13a472b68f101b760752e5",
-  "blk.2.attn_v.weight": "cd33f7f01141e9439afdaf2ea1aaced9feaa335e32a58daa136ebd555d4d96f4",
-  "blk.2.ffn_down.weight": "b970ff1b0b6494165defe2fbfa1d31425766ed71e64de9ec4e66ac3955c8bc5f",
-  "blk.2.ffn_gate.weight": "dbb3e1360402e0e369b101995bb686b73f95d4a7673f061be85d64d15dfb0061",
-  "blk.2.ffn_norm.weight": "bfb7980105d8ac9647710454f57a5cdac50598a0f6f4884e16f1d94b00844687",
-  "blk.2.ffn_up.weight": "50ef89339b275a438b664686f6227dd9b6e43853ed6856ec9e33ef4bbd90bda1",
-  "blk.3.attn_k.weight": "be942ea98151434eebcd2c1da4b00e0146152fe524a530689b1fd491cb833d21",
-  "blk.3.attn_norm.weight": "0df2f218daf609c289fb7c60c5f375fa99c0d4e04381ad5a494a19144edd8e20",
-  "blk.3.attn_output.weight": "c2184aaf86aa2cb8f47be49f60b165834e97205f39c6ee1dfd19fd4411a156ce",
-  "blk.3.attn_q.weight": "4f86e2a0a4221c1c84ff9c409ac89893cb95d7208cf65bf1e98e24e01125f991",
-  "blk.3.attn_v.weight": "abfdb8a60c349dadde641d1afc9542025e24fbf41a3238bfa9675e0b1f1e4b68",
-  "blk.3.ffn_down.weight": "58821a8d87008d47d122427911c6fad5272aca70c448bbae223256a74bacd07e",
-  "blk.3.ffn_gate.weight": "776e051f1a0ddd5c4934e69186683a75ca9a3c8c0f61911bba321fed1dd287d2",
-  "blk.3.ffn_norm.weight": "7f380f29335e28be90bfcfae6f6d69fdf5751211b36d2dd62aa5541ed113e4f2",
-  "blk.3.ffn_up.weight": "fc5ae8d488894cbd4951059675468d227da27871d26e925c9941863841c097ee",
-  "blk.4.attn_k.weight": "14833b078cc4c5137bdd5fdc0538047974ca147a99b0282e1b144440c78bc1db",
-  "blk.4.attn_norm.weight": "0a69957d4a15599fb80ad4753558020804925221457d9a5052926754d3768065",
-  "blk.4.attn_output.weight": "887a49b6130fb6297cf10767207c3dd97191b2cf63723449af9c27bca8dbeda0",
-  "blk.4.attn_q.weight": "51fd577b76764824dd6f0d4891c137ebe4736f591b5ca2793c5fff2be49abbde",
-  "blk.4.attn_v.weight": "1a623c43cf9c509d1b7ea0d1a5c04d0af4809665f9f9e93b7d6dba8c5df178fa",
-  "blk.4.ffn_down.weight": "5d61e8856d8941d2b1fd138116d015f63840d0fa1e31e20e20a5ceca1536ceec",
-  "blk.4.ffn_gate.weight": "06640f7273764f8ca5df7e386547417916b6cd7d565a8343153113239a94b0a1",
-  "blk.4.ffn_norm.weight": "91a6c6c41b894228e361435ecbc5058dca34d4911a23da5b56de219299c964d3",
-  "blk.4.ffn_up.weight": "d016dac1055e36d6a10b6317e57f98a904709ea892ef3194342f4d2f6326561e",
-  "blk.5.attn_k.weight": "987146afe124131500808cc0da33c06d207433656d41df6e6d8c99118a83bac5",
-  "blk.5.attn_norm.weight": "6b354938966f2608a2fb8d0f5b363ed0d8b0967c2ec8d0abd5c625b413042ded",
-  "blk.5.attn_output.weight": "cdcbfe02c6ff79d5326882b017a02099f5af71beedf6b1b3eb4de01e3a844536",
-  "blk.5.attn_q.weight": "b910d0cff781d3efb42eab0a302f46f286b2de717079175680d5b42bf8c309c8",
-  "blk.5.attn_v.weight": "66d3a279f747412f9f4b0e8abad44540c122ab2e811a7ee74c1f33bc36caade9",
-  "blk.5.ffn_down.weight": "c9b0efd2212981f16d956d8571f054b68780ad01f4917033647e359b557a4653",
-  "blk.5.ffn_gate.weight": "fe96b94109ca141c01f6a04788e20783019ca6ec334aa1f3134810bdb499e557",
-  "blk.5.ffn_norm.weight": "aa7b016e832e7055a36c6e20de58ea1936f995f390401fff1c5fc65906064e49",
-  "blk.5.ffn_up.weight": "555ce27c4873d3375394f38ad3b45e3d8848f9d5642dc1602383d0f0a33c2a14",
-  "blk.6.attn_k.weight": "88280d461db324c4f36475ce396793063e61a27283ec64511b0480890fb5b3b4",
-  "blk.6.attn_norm.weight": "af8f460c411f660d33196286d208f1845fd5a2b45f7b56549a4df31e7515447a",
-  "blk.6.attn_output.weight": "dd9996fb0a256e8375ad3917705258a33fce006bcea0f536caae420a77974d8b",
-  "blk.6.attn_q.weight": "7a4841541191e037cfb9b07930c4d8cab451809658b182f0ada6ccde9615c003",
-  "blk.6.attn_v.weight": "ae81e6a592b64d701a9d40233e986039a56cba8d8d24f61aea93c6393cf3078a",
-  "blk.6.ffn_down.weight": "622dd1ce1706355cbc659a8ab2c4509678ffe0f3ad34258e5e25ed2a5d951bcd",
-  "blk.6.ffn_gate.weight": "8389a735c0bd5591010f8ced9805a2a12c749f6df0d3c18ad4d05c2a302e7168",
-  "blk.6.ffn_norm.weight": "621f5346400382474d61358397bd58fb1459b07c53e376e4bca15e08b3f9b3fb",
-  "blk.6.ffn_up.weight": "8d834e4c42f13c251dfee36cf89e12f1bd400680d00d5c2e6cac0459e9ce2f7f",
-  "blk.7.attn_k.weight": "8bd0412de65a3e64901ef8fe6a28c95e116bf39dc9aa22f0126b9d36688e5ea7",
-  "blk.7.attn_norm.weight": "056d8e56be4e87d6dc6f900762f0dc6fde07bfdc50dd85bfc510415e2bba3f3d",
-  "blk.7.attn_output.weight": "27972eda51da53d416ff95aed78149a2c5a287b47d2cd46f2f544ca692ecb3bb",
-  "blk.7.attn_q.weight": "41eca977b9371f7932800c11a9c45b931310196919e2a0651b847703b180fc7f",
-  "blk.7.attn_v.weight": "13c74fd7e07f08883a09fb070a1fe5bbdd2341b4cb8d1cac07c4b637049b5774",
-  "blk.7.ffn_down.weight": "9e75db42468800849a9a7da603d0072c5e86c8ed2b4d8b20a312a51fb86a7a10",
-  "blk.7.ffn_gate.weight": "db6bdc3117f910088aaf7db51f2da63ea5bd933de36af5599c215bfb26f7db2b",
-  "blk.7.ffn_norm.weight": "48bb82b49bfc8679a1e77f282ee182d952db7a3c11be7ef9a102ee2ddd8011e2",
-  "blk.7.ffn_up.weight": "feebea87175817a0f3585ec0af09dc873d94c203581ae97a712eb356d3b49efe",
-  "blk.8.attn_k.weight": "d5640ad71b6af68d88e17bf8e7fc26c907d2262605457a84247dd9afc2884d69",
-  "blk.8.attn_norm.weight": "75b850c481a69083ae09d0207ba7317b37c735a39fcf5fef5400e6c84fb1257f",
-  "blk.8.attn_output.weight": "cbd669dbdea2bdd90f9f0cc97566b3dffff3c56cecb4f47290ceef30da83b2d6",
-  "blk.8.attn_q.weight": "9edcb63087a431bac361822497e6ecdaa06d9ea4a1a754e36da7ba9f8db81c7c",
-  "blk.8.attn_v.weight": "3fb72c2c4f95a83626aa3e30062f9450b09ab37c7871e229f18bbc5cf744633c",
-  "blk.8.ffn_down.weight": "bd69d2c9172974fff154441b237b4787fb53b2d185325442d5048130ef5bc4ef",
-  "blk.8.ffn_gate.weight": "d04689c80553edd011d1cbaa5d570fffa7fa91e88b66cf1352d89ab60b72f908",
-  "blk.8.ffn_norm.weight": "e49984183b735b7f2c4e4730c289eed9394056d2e283a00fd83ea0915df31a73",
-  "blk.8.ffn_up.weight": "8fe62a1ce8e847e567add6c6f6bf2922bc467495b5eb4c116b3cb85b85b3b211",
-  "blk.9.attn_k.weight": "d90904959e5004cf0d6e729c6bff18cc33c094798b802473c1ec55ab8d276183",
-  "blk.9.attn_norm.weight": "79277f290cc07411115d8fa138045edf4a17b3416ab2145409cbe8ab829fd4ee",
-  "blk.9.attn_output.weight": "5a21bf2e1f09a81405025f96d4153ffb630158e17269cff8ffff935c38ceb1a7",
-  "blk.9.attn_q.weight": "51b1d0febc3b350945be4504f55afa4347517bde0f710e1a4b88e6b17e71e7c7",
-  "blk.9.attn_v.weight": "aab7e1db0a8b50a03036356791ffce736ab010d15674c96eaef8049d80076054",
-  "blk.9.ffn_down.weight": "cbf43ec84becb40c9359a181ab0e641fd7faae7d34b549501f7cfb7afdc3d764",
-  "blk.9.ffn_gate.weight": "dce0e8661c778327bed7f03b6790d26710764188aed9dc746e6e05863891fa57",
-  "blk.9.ffn_norm.weight": "6d41642104f995c77bf31122b13237caebda3e7fcccb1367ce91db36b015e923",
-  "blk.9.ffn_up.weight": "82fe4c67bf24e7b2d6f6e05f7b1234c2bf90c3932951091a9066211b8e15ecbb",
-  "blk.10.attn_k.weight": "f6a9ed8fd8d3229b5d03175c413ffc56a07f2ce7236271986361dd3d8993f9aa",
-  "blk.10.attn_norm.weight": "cebbef89f0326ca8e02df3867a571e4d61c20c2a12f295f98ae590d62bc86010",
-  "blk.10.attn_output.weight": "34f5efb86accb4f06347d83a32558ea8eab3039d128969161a741ebacbb656ff",
-  "blk.10.attn_q.weight": "1e0efe27df2d5d50f7157253ba2cfd436d6781c3dc78ca176d0c16a210b5b763",
-  "blk.10.attn_v.weight": "8f085bf50a2b0f83cd6cdda3c8ef5a9e204a36348ed95871aac725d1f68640cf",
-  "blk.10.ffn_down.weight": "bf3b3cb4cace435809ac7b4cc933f20853af12f1f272d3dcefe7f19c0f203b8b",
-  "blk.10.ffn_gate.weight": "d3df7a1413b1c5adf1a1dcda9e5225a15c89874bae53bb6137ad1ea42fca2d34",
-  "blk.10.ffn_norm.weight": "a1da603b0480471b5ed8e862148cecd5fed918f8304d6933ab0bdb25b8d2fb8f",
-  "blk.10.ffn_up.weight": "bffbba605922e972dc47dda88a0b4659aa52236c76e5fe861a949e6d9a367492",
-  "blk.11.attn_k.weight": "9f31c63d66cd32c29b1eb8bb829d0c8525ce2ae936e0eefdaab6335a2d12a3df",
-  "blk.11.attn_norm.weight": "0bde1a266d8b2e8f202bb7e2e88b19147ca83021901f6d3cae77a4df5548c754",
-  "blk.11.attn_output.weight": "e10725c7cf746ed4a7e472cf7aea6cb564e5db6a1d5197adc980d650a387ccea",
-  "blk.11.attn_q.weight": "05ee758a7d065802630f8c65dca424364c1c8825e389aa33f9405c45e8a50cce",
-  "blk.11.attn_v.weight": "0c3ae7090f11775d24c51120db6e305db6aff706493e7ee123dcab74485ba789",
-  "blk.11.ffn_down.weight": "7ba40b8e12c09c5fb2006b77a771cb01ce894e88a3b3e1877f927a5b89c91709",
-  "blk.11.ffn_gate.weight": "db76388a023b98097972d354ba1c6a5e26efdeb1c596b9c28bf2cd8f6596975e",
-  "blk.11.ffn_norm.weight": "a38c3ae1b89a68ddc7b72c99c5b28be7fe3787c4fad9904d0c43d64eaf00c474",
-  "blk.11.ffn_up.weight": "13c8142f9cf1eddc658babf978daf3515c4ccc45f849f3e7e3930aa18a8480a0",
-  "blk.12.attn_k.weight": "f03241c36ac87cb57429a2ef22186b8d7d0b590a8b173beb01fa13d93772f3b1",
-  "blk.12.attn_norm.weight": "4568f654e6d65104d586e7c16ba960c83428698ce103022b7e0be15e2884e13b",
-  "blk.12.attn_output.weight": "04867603f82f91e41306e09b33ecda0104b3ee4834061f2c0bbdc8da33c72509",
-  "blk.12.attn_q.weight": "70fe04b9a8e08b6100cc8d6b58bf4cbbad15ca1de82d63baca5d352ba6c4cbae",
-  "blk.12.attn_v.weight": "15cb28db61a86c98687991d7e611bc92a1fcc6007f3432149cfb5fe518a4f65e",
-  "blk.12.ffn_down.weight": "6d10c790a4e3dc44c2dc36d96251ae97cdf30a4fa04d4c43e31bfbd038e6a7b7",
-  "blk.12.ffn_gate.weight": "3462a2d8f6b4743b25e24da51b90018ac2858d05ac7e582bcb69063cfdac1104",
-  "blk.12.ffn_norm.weight": "1f96392c1faa34e34ae5dea55a6a86c5aa4c79758952075d53d28de89dd88456",
-  "blk.12.ffn_up.weight": "d22eacc612a7411953d948483c5fb201e11722955ee0754da866e7bec578ac6d",
-  "blk.13.attn_k.weight": "5864977e6b733ea942647d6feed5c76156c48c200649c22e4e11b9e5860e57f3",
-  "blk.13.attn_norm.weight": "87e053535144723db4145aa5402acc54331b7696752d852bb9fc542ff33f0fb5",
-  "blk.13.attn_output.weight": "078145f5ad83f8b14f97a869346f7fd1583b24d1e3edadaa95d3da4242973f8f",
-  "blk.13.attn_q.weight": "3b8caf35504cbc4d1a7dd6e011a95760703b7f71e2218b030b1254f811362dd7",
-  "blk.13.attn_v.weight": "4fdf8365a603e043e5b40c4a21c84ac167f9be62794178f9d8a608dfe5653bf9",
-  "blk.13.ffn_down.weight": "a07d3abbfcacf48ba028df2cab895be32cc15022d23389a745286e79c1b1d1fd",
-  "blk.13.ffn_gate.weight": "1d2ab39666aa2909acc96787432a3ed13b19d25170f74665fadff9b17bbaffb1",
-  "blk.13.ffn_norm.weight": "4f2e809fda5f3eadf52578ee50e0ba36e53be91e55dce418c12dfe595f5f18e7",
-  "blk.13.ffn_up.weight": "8783d2720c2c37ca176a5801e0b3ef1f9cc9cf3ef1cd37af423aaf6b2a27e2bd",
-  "blk.14.attn_k.weight": "ce9428e2b55d43ae0c6690dbd56182f99adc427694ba8236b405cc8ea5035e86",
-  "blk.14.attn_norm.weight": "6abb35f9db8251d6ae954bda147c6ada2371b0574d11702e828f3c6ac99b7cc0",
-  "blk.14.attn_output.weight": "fe3880916d0ceb5bff672c88bbefb7060a545be609bf049beb2024b38221836d",
-  "blk.14.attn_q.weight": "7c8ad81be6f4a350931fd108b5f7c9e366e8c26ef62d1d85ffef5dca8fd893f8",
-  "blk.14.attn_v.weight": "e4bdedffacbebe38567a0734dfd67db90e911d9a9669fcde9a7c4ad8a0066c52",
-  "blk.14.ffn_down.weight": "ef6694dff1e05820aac0cd2b22f39ac7788b4967afc9250775575554c66aab2c",
-  "blk.14.ffn_gate.weight": "db63c4179e2db704bc505e2b4696e055b593e295a1b7c4c586fc793bdd5aab19",
-  "blk.14.ffn_norm.weight": "2796a62d832a9710148f95d533320492a33e712b2e5218659c548705bd11684d",
-  "blk.14.ffn_up.weight": "3f78c78d8c2d54df45f799d4ff902316628af296834afe4ceed63d4a324ff03e",
-  "blk.15.attn_k.weight": "6e810ee3859e07695645ee0c9a5efc7962668984a5f0a9325f47e462743b447c",
-  "blk.15.attn_norm.weight": "0956b576ae96db0b28cb09f761f801cfd9281432284664f0fe181c8d9c55d1ec",
-  "blk.15.attn_output.weight": "03a17f7e94208177aace5cc41b7f54670ba57873b7274ff6e23caf58cce110ca",
-  "blk.15.attn_q.weight": "b8edafe7d2216a6f8b4ae4905a906475490e6ea418f6e1d3cec563dbdc6fab91",
-  "blk.15.attn_v.weight": "f8ae8cae0f4cfa34a459824eba57350c3c248104ba5607e7d9dc7d7c39aaf4a6",
-  "blk.15.ffn_down.weight": "8d02eb439da852246d2ca67e9b7b6de0b090b80744355e64728a23e41926505b",
-  "blk.15.ffn_gate.weight": "ed5bf361c67db8731f186b775826f21c33bdb521111fd2d922539719a770239f",
-  "blk.15.ffn_norm.weight": "5942ca3c73209ac9a0c8bfd9b4aab7f7be7aee9aa12d9c35833493b44af76767",
-  "blk.15.ffn_up.weight": "f4bebf4ad99ec5f911327dec347be6c595814885309c7bc5647ce28c7f4d1cf5",
-  "blk.16.attn_k.weight": "756a534c19364448e0958b8948fe33891c6ccda0fbb4dfa2024e1f532a87804b",
-  "blk.16.attn_norm.weight": "386b7b9e4e6509f6af9c022d942b6c6c6cc136aeed8751ecb037c74d7c4bfb93",
-  "blk.16.attn_output.weight": "3ba1a766a25830b84d7c22178203635f9c5624caad290bc5e5d73da5d5e7a2ec",
-  "blk.16.attn_q.weight": "d39b0c91e1fda7685d50a0f7cc8d18c44b5bdc90a142c7fda0bc329cca1afa74",
-  "blk.16.attn_v.weight": "98b33fcb0ee3483cff1b06ecb44d7b7ffb4d34c268248e4d73dfdf82b2065b2f",
-  "blk.16.ffn_down.weight": "14006f5e4acb2f9416271ae562e299359cd2585739c7fc77ccbca54495563948",
-  "blk.16.ffn_gate.weight": "12f8abae2d301d8f88bedb6af98b1daecc7b0b8d05148594f931f30958d77aca",
-  "blk.16.ffn_norm.weight": "129a15a046ee96d06de288bd43c80f77a6b0fb3a159c7367154c6e4aaf362672",
-  "blk.16.ffn_up.weight": "b4a5911a45f3871ef1d4efb7dc7108645a564b70f818eccf45beebef2e844ee9",
-  "blk.17.attn_k.weight": "5e1bfcff0146ebdde3817b656952892eb671e14e75afc92fa53f84f8eecbec4c",
-  "blk.17.attn_norm.weight": "60bc988fab7c4b29ee9de599df41a8de00caa94fcd74677da011fac82f60f465",
-  "blk.17.attn_output.weight": "ba49b40d6a0b5685f749c24b0edbed3adc44dbe13b5d5e5fa1e56169fc746555",
-  "blk.17.attn_q.weight": "82bb415d24efcd14d03ace03f907bb70db6a204c76a0bdd1892e0fba165db87d",
-  "blk.17.attn_v.weight": "73dbe54beb91a899884e275ea81ffc5187a20cb7d5b68d5c299b783096999d94",
-  "blk.17.ffn_down.weight": "7c086166241e0664f8963fd1ca4ed74c737abfb2525ec20f8435821ff50158f3",
-  "blk.17.ffn_gate.weight": "51a32f78244d42a539f619c5ce661db9e6cf41636280a826d439b5444edcd28c",
-  "blk.17.ffn_norm.weight": "c4bb247fccd1ecc84875028af63dd20aaf5cbd17eb94a9bc36679c09285dccab",
-  "blk.17.ffn_up.weight": "b5886182790bc6fbadd63de9bc4ffee416f3b69a66280d197ab8c18edf769abf",
-  "output_norm.weight": "481f3097d0a20412e35b3a739b1b958487bcd41ff67744baa3c9acbddd2ee4d4"
-}

convert/tokenizer.go

@@ -3,148 +3,19 @@ package convert
 import (
 	"cmp"
 	"crypto/sha256"
-	"encoding/hex"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"log/slog"
 	"os"
-	"path/filepath"
 	"slices"
-)
 
-const (
-	_ int32 = iota
-	tokenTypeNormal
-	tokenTypeUnknown
-	tokenTypeControl
-	tokenTypeUserDefined
-	tokenTypeUnused
-	tokenTypeByte
+	"golang.org/x/exp/maps"
 )
 
 type Tokenizer struct {
-	*Vocabulary
-	SpecialVocabulary []*SpecialVocabulary
-	Merges            []string
-
-	Pre      string
-	Template string
-}
-
-func parseTokenizer(d string, specialTypes []string) (*Tokenizer, error) {
-	v, err := parseVocabulary(d)
-	if err != nil {
-		return nil, err
-	}
-
-	t := &Tokenizer{
-		Vocabulary: v,
-		Pre:        "default",
-	}
-
-	addedTokens := make(map[string]token)
-	if f, err := os.Open(filepath.Join(d, "tokenizer.json")); errors.Is(err, os.ErrNotExist) {
-	} else if err != nil {
-		return nil, err
-	} else {
-		defer f.Close()
-
-		var tt tokenizer
-		if err := json.NewDecoder(f).Decode(&tt); err != nil {
-			return nil, err
-		}
-
-		for _, t := range tt.AddedTokens {
-			addedTokens[t.Content] = t
-		}
-
-		t.Merges = tt.Model.Merges
-
-		sha256sum := sha256.New()
-		for _, pt := range tt.PreTokenizer.PreTokenizers {
-			switch pt.Type {
-			case "Split":
-				if pt.Pattern.Regex != "" {
-					sha256sum.Write([]byte(pt.Pattern.Regex))
-				}
-			}
-		}
-
-		switch digest := hex.EncodeToString(sha256sum.Sum(nil)); digest {
-		case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
-			t.Pre = "llama-bpe"
-		case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
-			t.Pre = "deepseek-llm"
-		case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
-			t.Pre = "deepseek-coder"
-		case "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855":
-			// noop, empty pretokenizer
-		default:
-			slog.Warn("unknown pretokenizer, using default", "digest", digest)
-		}
-	}
-
-	if f, err := os.Open(filepath.Join(d, "tokenizer_config.json")); errors.Is(err, os.ErrNotExist) {
-	} else if err != nil {
-		return nil, err
-	} else {
-		defer f.Close()
-
-		var p map[string]json.RawMessage
-		if err := json.NewDecoder(f).Decode(&p); err != nil {
-			return nil, err
-		}
-
-		if template, ok := p["chat_template"]; ok {
-			if err := json.Unmarshal(template, &t.Template); err != nil {
-				return nil, err
-			}
-		}
-
-		for _, st := range specialTypes {
-			sv := SpecialVocabulary{Type: st}
-			if bts, ok := p[fmt.Sprintf("add_%s_token", st)]; ok {
-				if err := json.Unmarshal(bts, &sv.AddToken); err != nil {
-					return nil, err
-				}
-			}
-
-			if bts, ok := p[fmt.Sprintf("%s_token", st)]; ok {
-				var content string
-				if err := json.Unmarshal(bts, &content); err != nil {
-					var mm map[string]any
-					if err := json.Unmarshal(bts, &mm); err != nil {
-						continue
-					}
-
-					content, ok = mm["content"].(string)
-					if !ok {
-						continue
-					}
-				}
-
-				sv.Content = content
-			}
-
-			if id, ok := addedTokens[sv.Content]; ok {
-				sv.ID = id.ID
-				t.SpecialVocabulary = append(t.SpecialVocabulary, &sv)
-			}
-		}
-	}
-
-	return t, nil
-}
-
-type tokenizer struct {
-	Version     string  `json:"version"`
-	AddedTokens []token `json:"added_tokens"`
-	Model       struct {
-		Type   string         `json:"type"`
-		Vocab  map[string]int `json:"vocab"`
-		Merges []string       `json:"merges"`
-	} `json:"model"`
+	Version     string         `json:"version"`
+	AddedTokens []Token        `json:"added_tokens"`
+	Model       TokenizerModel `json:"model"`
 
 	PreTokenizer struct {
 		PreTokenizers []struct {
@@ -156,106 +27,83 @@ type tokenizer struct {
 	} `json:"pre_tokenizer"`
 }
 
-type token struct {
+type TokenizerModel struct {
+	Type   string         `json:"type"`
+	Vocab  map[string]int `json:"vocab"`
+	Merges []string       `json:"merges"`
+	Tokens []Token
+}
+
+type Token struct {
 	ID          int    `json:"id"`
 	Content     string `json:"content"`
 	Special     bool   `json:"special"`
 	UserDefined bool
 }
 
-type Vocabulary struct {
-	Model  string
-	Tokens []string
-	Scores []float32
-	Types  []int32
+func (t *Token) Type() int32 {
+	switch {
+	case t.Special:
+		return tokenTypeControl
+	case t.UserDefined:
+		return tokenTypeUserDefined
+	default:
+		return tokenTypeNormal
+	}
 }
 
-func parseVocabularyFromTokenizer(p string) (*Vocabulary, error) {
-	f, err := os.Open(filepath.Join(p, "tokenizer.json"))
+func (t *Tokenizer) maxID() int {
+	return max(
+		slices.Max(maps.Values(t.Model.Vocab)),
+		slices.MaxFunc(t.AddedTokens, func(a, b Token) int {
+			return cmp.Compare(a.ID, b.ID)
+		}).ID,
+	)
+}
+
+func parseTokens(dirpath string) (pre string, tokens []Token, merges []string, err error) {
+	f, err := os.Open(dirpath)
 	if err != nil {
-		return nil, err
+		panic(err)
 	}
 	defer f.Close()
 
-	var t tokenizer
+	var t Tokenizer
 	if err := json.NewDecoder(f).Decode(&t); err != nil {
-		return nil, err
+		return "", nil, nil, err
 	}
 
-	var tokens []token
+	tokens = make([]Token, t.maxID()+1)
 	for k, v := range t.Model.Vocab {
-		tokens = append(tokens, token{
-			ID:      v,
-			Content: k,
-		})
+		tokens[v] = Token{ID: v, Content: k, Special: false, UserDefined: false}
 	}
 
-	for _, t := range t.AddedTokens {
-		t.UserDefined = true
-		tokens = append(tokens, t)
+	for _, v := range t.AddedTokens {
+		v.UserDefined = true
+		tokens[v.ID] = v
 	}
 
-	slices.SortFunc(tokens, func(i, j token) int {
-		return cmp.Compare(i.ID, j.ID)
-	})
-
-	v := Vocabulary{Model: "gpt2"}
-	for _, t := range tokens {
-		v.Tokens = append(v.Tokens, t.Content)
-		v.Scores = append(v.Scores, float32(t.ID))
-
-		switch {
-		case t.Special:
-			v.Types = append(v.Types, tokenTypeControl)
-		case t.UserDefined:
-			v.Types = append(v.Types, tokenTypeUserDefined)
-		default:
-			v.Types = append(v.Types, tokenTypeNormal)
+	sha256sum := sha256.New()
+	for _, pt := range t.PreTokenizer.PreTokenizers {
+		switch pt.Type {
+		case "Split":
+			if pt.Pattern.Regex != "" {
+				sha256sum.Write([]byte(pt.Pattern.Regex))
+			}
 		}
 	}
 
-	return &v, nil
-}
-
-func parseVocabulary(d string) (*Vocabulary, error) {
-	patterns := map[string]func(string) (*Vocabulary, error){
-		"tokenizer.model": parseSentencePiece,
-		"tokenizer.json":  parseVocabularyFromTokenizer,
+	switch digest := fmt.Sprintf("%x", sha256sum.Sum(nil)); digest {
+	case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
+		pre = "llama-bpe"
+	case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
+		pre = "deepseek-llm"
+	case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
+		pre = "deepseek-coder"
+	default:
+		slog.Warn("unknown pretokenizer, using default", "digest", digest)
+		pre = "default"
 	}
 
-	for pattern, parseFn := range patterns {
-		matches, err := filepath.Glob(filepath.Join(d, pattern))
-		if err != nil {
-			return nil, err
-		}
-
-		if len(matches) > 0 {
-			return parseFn(d)
-		}
-	}
-
-	return nil, errors.New("unknown tensor format")
-}
-
-type SpecialVocabulary struct {
-	Type     string
-	ID       int
-	Content  string
-	AddToken bool
-}
-
-func (sv SpecialVocabulary) Key() string {
-	switch t := sv.Type; t {
-	case "bos", "eos", "cls", "mask":
-		return t
-	case "unk":
-		return "unknown"
-	case "sep":
-		//nolint:misspell // this is an upstream typo
-		return "seperator"
-	case "pad":
-		return "padding"
-	}
-
-	panic("unknown special vocabulary type")
+	return pre, tokens, t.Model.Merges, nil
 }

convert/tokenizer_spm.go

@@ -1,83 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"encoding/json"
-	"errors"
-	"fmt"
-	"os"
-	"path/filepath"
-	"slices"
-
-	"google.golang.org/protobuf/proto"
-
-	"github.com/ollama/ollama/convert/sentencepiece"
-)
-
-func parseSentencePiece(d string) (*Vocabulary, error) {
-	bts, err := os.ReadFile(filepath.Join(d, "tokenizer.model"))
-	if err != nil {
-		return nil, err
-	}
-
-	var spm sentencepiece.ModelProto
-	if err := proto.Unmarshal(bts, &spm); err != nil {
-		return nil, err
-	}
-
-	v := Vocabulary{Model: "llama"}
-	for _, piece := range spm.GetPieces() {
-		v.Tokens = append(v.Tokens, piece.GetPiece())
-		v.Scores = append(v.Scores, piece.GetScore())
-
-		switch t := piece.GetType(); t {
-		case sentencepiece.ModelProto_SentencePiece_UNKNOWN,
-			sentencepiece.ModelProto_SentencePiece_CONTROL,
-			sentencepiece.ModelProto_SentencePiece_UNUSED,
-			sentencepiece.ModelProto_SentencePiece_BYTE:
-			v.Types = append(v.Types, int32(t))
-		default:
-			v.Types = append(v.Types, int32(sentencepiece.ModelProto_SentencePiece_NORMAL))
-		}
-	}
-
-	f, err := os.Open(filepath.Join(d, "added_tokens.json"))
-	if errors.Is(err, os.ErrNotExist) {
-		return &v, nil
-	} else if err != nil {
-		return nil, err
-	}
-	defer f.Close()
-
-	var atm map[string]int
-	if err := json.NewDecoder(f).Decode(&atm); err != nil {
-		return nil, err
-	}
-
-	type t struct {
-		id      int
-		content string
-	}
-
-	var ts []t
-	for content, id := range atm {
-		ts = append(ts, t{id, content})
-	}
-
-	slices.SortFunc(ts, func(i, j t) int {
-		return cmp.Compare(i.id, j.id)
-	})
-
-	n := len(v.Tokens)
-	for i, t := range ts {
-		if t.id != i+n {
-			return nil, fmt.Errorf("invalid token id: %d", t.id)
-		}
-
-		v.Tokens = append(v.Tokens, t.content)
-		v.Scores = append(v.Scores, -1000.0)
-		v.Types = append(v.Types, tokenTypeUserDefined)
-	}
-
-	return &v, nil
-}

convert/torch.go

@@ -0,0 +1,288 @@
+package convert
+
+import (
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/nlpodyssey/gopickle/types"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type torchWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	storage  pytorch.StorageInterface
+	repacker func(string, []float32, []uint64) ([]float32, error)
+}
+
+type TorchFormat struct{}
+
+func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	slog.Debug("getting torch tensors")
+
+	var files []string
+	if pt, _ := filepath.Glob(filepath.Join(dirpath, "consolidated*.pth")); len(pt) > 0 {
+		files = append(files, pt...)
+	} else if pt, _ := filepath.Glob(filepath.Join(dirpath, "pytorch_model*.pth")); len(pt) > 0 {
+		files = append(files, pt...)
+	}
+
+	var offset uint64
+	var tensors []llm.Tensor
+	for _, fn := range files {
+		m, err := pytorch.Load(fn)
+		if err != nil {
+			slog.Error(fmt.Sprintf("error unpickling: %q", err))
+			return []llm.Tensor{}, err
+		}
+
+		for _, k := range m.(*types.Dict).Keys() {
+			if strings.HasSuffix(k.(string), "self_attn.rotary_emb.inv_freq") {
+				continue
+			}
+
+			t, _ := m.(*types.Dict).Get(k)
+			tshape := t.(*pytorch.Tensor).Size
+
+			var size uint64
+			var kind uint32
+			switch len(tshape) {
+			case 0:
+				continue
+			case 1:
+				// convert to float32
+				kind = 0
+				size = uint64(tshape[0] * 4)
+			case 2:
+				// convert to float16
+				kind = 1
+				size = uint64(tshape[0] * tshape[1] * 2)
+			}
+
+			ggufName, err := tf.GetLayerName(k.(string))
+			if err != nil {
+				slog.Error(err.Error())
+				return nil, err
+			}
+			slog.Debug(fmt.Sprintf("'%35s': '%30s' %10d [%#v]", k.(string), ggufName, size, tshape))
+
+			shape := []uint64{0, 0, 0, 0}
+			for i := range tshape {
+				shape[i] = uint64(tshape[i])
+			}
+
+			tensor := llm.Tensor{
+				Name:   ggufName,
+				Kind:   kind,
+				Offset: offset, // calculate the offset
+				Shape:  shape[:],
+			}
+
+			tensor.WriterTo = torchWriterTo{
+				t:       &tensor,
+				params:  params,
+				bo:      params.ByteOrder,
+				storage: t.(*pytorch.Tensor).Source,
+			}
+
+			tensors = append(tensors, tensor)
+			offset += size
+		}
+	}
+
+	return tensors, nil
+
+}
+
+func getAltParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "params.json"))
+	if err != nil {
+		slog.Error("no params.json")
+		return nil, err
+	}
+	defer f.Close()
+
+	type TorchParams struct {
+		HiddenSize     int     `json:"dim"`
+		AttentionHeads int     `json:"n_heads"`
+		KeyValHeads    int     `json:"n_kv_heads"`
+		HiddenLayers   int     `json:"n_layers"`
+		RopeTheta      float64 `json:"rope_theta"`
+		NormEPS        float64 `json:"norm_eps"`
+	}
+
+	var tparams TorchParams
+
+	d := json.NewDecoder(f)
+	err = d.Decode(&tparams)
+	if err != nil {
+		return nil, err
+	}
+
+	params := &Params{
+		Architectures:  []string{"LlamaForCausalLM"},
+		HiddenSize:     tparams.HiddenSize,
+		AttentionHeads: tparams.AttentionHeads,
+		KeyValHeads:    tparams.KeyValHeads,
+		HiddenLayers:   tparams.HiddenLayers,
+		NormEPS:        tparams.NormEPS,
+	}
+
+	switch {
+	case tparams.RopeTheta == 1000000:
+		// Codellama
+		params.ContextSize = 16384
+	case tparams.NormEPS == 1e-06:
+		// llama2
+		slog.Debug("Found llama2 - setting context size to 4096")
+		params.ContextSize = 4096
+	default:
+		params.ContextSize = 2048
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return params, nil
+}
+
+func (m *TorchFormat) GetParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "config.json"))
+	if err != nil {
+		if os.IsNotExist(err) {
+			// try params.json instead
+			return getAltParams(dirpath)
+		} else {
+			return nil, err
+		}
+	}
+
+	var params Params
+	d := json.NewDecoder(f)
+	err = d.Decode(&params)
+	if err != nil {
+		return nil, err
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return &params, nil
+}
+
+func (m *TorchFormat) GetLayerName(n string) (string, error) {
+	directMap := map[string]string{
+		"tok_embeddings.weight":     "token_embd.weight",
+		"output.weight":             "output.weight",
+		"norm.weight":               "output_norm.weight",
+		"rope.freqs":                "rope_freqs.weight",
+		"model.embed_tokens.weight": "token_embd.weight",
+		"lm_head.weight":            "output.weight",
+		"model.norm.weight":         "output_norm.weight",
+	}
+
+	lMap := map[string]string{
+		"layers.(\\d+).attention_norm.weight":                 "blk.$1.attn_norm.weight",
+		"layers.(\\d+).attention_output_norm.weight":          "blk.$1.attn_norm.weight",
+		"layers.(\\d+).feed_forward.w2.weight":                "blk.$1.ffn_down.weight",
+		"layers.(\\d+).feed_forward.w1.weight":                "blk.$1.ffn_gate.weight",
+		"layers.(\\d+).feed_forward.w3.weight":                "blk.$1.ffn_up.weight",
+		"layers.(\\d+).ffn_norm.weight":                       "blk.$1.ffn_norm.weight",
+		"layers.(\\d+).attention.wk.weight":                   "blk.$1.attn_k.weight",
+		"layers.(\\d+).attention.wo.weight":                   "blk.$1.attn_output.weight",
+		"layers.(\\d+).attention.wq.weight":                   "blk.$1.attn_q.weight",
+		"layers.(\\d+).attention.wv.weight":                   "blk.$1.attn_v.weight",
+		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
+	}
+
+	v, ok := directMap[n]
+	if ok {
+		return v, nil
+	}
+
+	// quick hack to rename the layers to gguf format
+	for k, v := range lMap {
+		re := regexp.MustCompile(k)
+		newName := re.ReplaceAllString(n, v)
+		if newName != n {
+			return newName, nil
+		}
+	}
+
+	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
+}
+
+func (r torchWriterTo) WriteTo(w io.Writer) (n int64, err error) {
+	var f32s []float32
+	switch s := r.storage.(type) {
+	case *pytorch.FloatStorage:
+		f32s = s.Data
+	case *pytorch.HalfStorage:
+		f32s = s.Data
+	case *pytorch.BFloat16Storage:
+		f32s = s.Data
+	default:
+		return 0, fmt.Errorf("unknown data type: %T", s)
+	}
+
+	if r.repacker != nil {
+		f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	switch r.t.Kind {
+	case 0:
+		return 0, binary.Write(w, r.bo, f32s)
+	case 1:
+		f16s := make([]uint16, len(f32s))
+		for i := range f32s {
+			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
+		}
+
+		return 0, binary.Write(w, r.bo, f16s)
+	default:
+		return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
+	}
+}
+
+func (m *TorchFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
+	switch len(params.Architectures) {
+	case 0:
+		return nil, fmt.Errorf("No architecture specified to convert")
+	case 1:
+		switch params.Architectures[0] {
+		case "LlamaForCausalLM":
+			return &LlamaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		default:
+			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
+		}
+	}
+
+	return nil, fmt.Errorf("Unknown error")
+}

docs/api.md

@@ -12,7 +12,6 @@
 - [Pull a Model](#pull-a-model)
 - [Push a Model](#push-a-model)
 - [Generate Embeddings](#generate-embeddings)
-- [List Running Models](#list-running-models)
 
 ## Conventions
 
@@ -26,7 +25,7 @@ All durations are returned in nanoseconds.
 
 ### Streaming responses
 
-Certain endpoints stream responses as JSON objects. Streaming can be disabled by providing `{"stream": false}` for these endpoints.
+Certain endpoints stream responses as JSON objects and can optional return non-streamed responses.
 
 ## Generate a completion
 
@@ -250,7 +249,7 @@ curl http://localhost:11434/api/generate -d '{
 
 #### Request (Reproducible outputs)
 
-For reproducible outputs, set `seed` to a number:
+For reproducible outputs, set `temperature` to 0 and `seed` to a number:
 
 ##### Request
 
@@ -259,7 +258,8 @@ curl http://localhost:11434/api/generate -d '{
   "model": "mistral",
   "prompt": "Why is the sky blue?",
   "options": {
-    "seed": 123
+    "seed": 123,
+    "temperature": 0
   }
 }'
 ```
@@ -777,12 +777,11 @@ A single JSON object will be returned.
 POST /api/show
 ```
 
-Show information about a model including details, modelfile, template, parameters, license, system prompt.
+Show information about a model including details, modelfile, template, parameters, license, and system prompt.
 
 ### Parameters
 
 - `name`: name of the model to show
-- `verbose`: (optional) if set to `true`, returns full data for verbose response fields
 
 ### Examples
 
@@ -799,40 +798,14 @@ curl http://localhost:11434/api/show -d '{
 ```json
 {
   "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
-  "parameters": "num_keep                       24\nstop                           \"<|start_header_id|>\"\nstop                           \"<|end_header_id|>\"\nstop                           \"<|eot_id|>\"",
-  "template": "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>",
+  "parameters": "num_ctx                        4096\nstop                           \u003c/s\u003e\nstop                           USER:\nstop                           ASSISTANT:",
+  "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: ",
   "details": {
-    "parent_model": "",
     "format": "gguf",
     "family": "llama",
-    "families": [
-      "llama"
-    ],
-    "parameter_size": "8.0B",
+    "families": ["llama", "clip"],
+    "parameter_size": "7B",
     "quantization_level": "Q4_0"
-  },
-  "model_info": {
-    "general.architecture": "llama",
-    "general.file_type": 2,
-    "general.parameter_count": 8030261248,
-    "general.quantization_version": 2,
-    "llama.attention.head_count": 32,
-    "llama.attention.head_count_kv": 8,
-    "llama.attention.layer_norm_rms_epsilon": 0.00001,
-    "llama.block_count": 32,
-    "llama.context_length": 8192,
-    "llama.embedding_length": 4096,
-    "llama.feed_forward_length": 14336,
-    "llama.rope.dimension_count": 128,
-    "llama.rope.freq_base": 500000,
-    "llama.vocab_size": 128256,
-    "tokenizer.ggml.bos_token_id": 128000,
-    "tokenizer.ggml.eos_token_id": 128009,
-    "tokenizer.ggml.merges": [],            // populates if `verbose=true`
-    "tokenizer.ggml.model": "gpt2",
-    "tokenizer.ggml.pre": "llama-bpe",
-    "tokenizer.ggml.token_type": [],        // populates if `verbose=true`
-    "tokenizer.ggml.tokens": []             // populates if `verbose=true`
   }
 }
 ```
@@ -1062,47 +1035,3 @@ curl http://localhost:11434/api/embeddings -d '{
   ]
 }
 ```
-
-## List Running Models
-```shell
-GET /api/ps
-```
-
-List models that are currently loaded into memory.
-
-#### Examples
-
-### Request
-
-```shell
-curl http://localhost:11434/api/ps
-```
-
-#### Response
-
-A single JSON object will be returned.
-
-```json
-{
-  "models": [
-    {
-      "name": "mistral:latest",
-      "model": "mistral:latest",
-      "size": 5137025024,
-      "digest": "2ae6f6dd7a3dd734790bbbf58b8909a606e0e7e97e94b7604e0aa7ae4490e6d8",
-      "details": {
-        "parent_model": "",
-        "format": "gguf",
-        "family": "llama",
-        "families": [
-          "llama"
-        ],
-        "parameter_size": "7.2B",
-        "quantization_level": "Q4_0"
-      },
-      "expires_at": "2024-06-04T14:38:31.83753-07:00",
-      "size_vram": 5137025024
-    }
-  ]
-}
-```

docs/development.md

@@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:
 
 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
+OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
 go build .
 ```
 
@@ -114,18 +114,15 @@ If you have Docker available, you can build linux binaries with `./scripts/build
 
 ### Windows
 
-Note: The Windows build for Ollama is still under development.
+Note: The windows build for Ollama is still under development.
 
-First, install required tools:
+Install required tools:
 
 - MSVC toolchain - C/C++ and cmake as minimal requirements
 - Go version 1.22 or higher
 - MinGW (pick one variant) with GCC.
   - [MinGW-w64](https://www.mingw-w64.org/)
   - [MSYS2](https://www.msys2.org/)
-- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser`
-
-Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"

docs/faq.md

@@ -257,19 +257,3 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP
 ## How do I manage the maximum number of requests the Ollama server can queue?
 
 If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded.  You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
-
-## How does Ollama handle concurrent requests?
-
-Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.
-
-If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded.  As prior models become idle, one or more will be unloaded to make room for the new model.  Queued requests will be processed in order.  When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.
-
-Parallel request processing for a given model results in increasing the context size by the number of parallel requests.  For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation.
-
-The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
-
-- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
-- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
-- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
-
-Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.

docs/gpu.md

@@ -8,7 +8,7 @@ Check your compute compatibility to see if your card is supported:
 | Compute Capability | Family              | Cards                                                                                                       |
 | ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
 | 9.0                | NVIDIA              | `H100`                                                                                                      |
-| 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
+| 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080` `RTX 4070 Ti` `RTX 4060 Ti`                                                           |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
 | 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
 |                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
@@ -18,7 +18,7 @@ Check your compute compatibility to see if your card is supported:
 |                    | Quadro              | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000`                                                                 |
 | 7.0                | NVIDIA              | `TITAN V` `V100` `Quadro GV100`                                                                             |
 | 6.1                | NVIDIA TITAN        | `TITAN Xp` `TITAN X`                                                                                        |
-|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050 Ti` `GTX 1050`                       |
+|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050`                                     |
 |                    | Quadro              | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` |
 |                    | Tesla               | `P40` `P4`                                                                                                  |
 | 6.0                | NVIDIA              | `Tesla P100` `Quadro GP100`                                                                                 |

docs/import.md

@@ -1,88 +1,170 @@
-# Import
+# Import a model
 
-GGUF models and select Safetensors models can be imported directly into Ollama.
+This guide walks through importing a GGUF, PyTorch or Safetensors model.
 
-## Import GGUF
+## Importing (GGUF)
 
-A binary GGUF file can be imported directly into Ollama through a Modelfile.
+### Step 1: Write a `Modelfile`
 
-```dockerfile
-FROM /path/to/file.gguf
+Start by creating a `Modelfile`. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more.
+
+```
+FROM ./mistral-7b-v0.1.Q4_0.gguf
 ```
 
-## Import Safetensors
+(Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`:
 
-If the model being imported is one of these architectures, it can be imported directly into Ollama through a Modelfile:
-
- - LlamaForCausalLM
- - MistralForCausalLM
- - GemmaForCausalLM
-
-```dockerfile
-FROM /path/to/safetensors/directory
+```
+FROM ./mistral-7b-v0.1.Q4_0.gguf
+TEMPLATE "[INST] {{ .Prompt }} [/INST]"
 ```
 
-For architectures not directly convertable by Ollama, see llama.cpp's [guide](https://github.com/ggerganov/llama.cpp/blob/master/README.md#prepare-and-quantize) on conversion. After conversion, see [Import GGUF](#import-gguf).
+### Step 2: Create the Ollama model
 
-## Automatic Quantization
+Finally, create a model from your `Modelfile`:
 
-> [!NOTE]
-> Automatic quantization requires v0.1.35 or higher.
-
-Ollama is capable of quantizing FP16 or FP32 models to any of the supported quantizations with the `-q/--quantize` flag in `ollama create`.
-
-```dockerfile
-FROM /path/to/my/gemma/f16/model
 ```
+ollama create example -f Modelfile
+```
+
+### Step 3: Run your model
+
+Next, test the model with `ollama run`:
+
+```
+ollama run example "What is your favourite condiment?"
+```
+
+## Importing (PyTorch & Safetensors)
+
+> Importing from PyTorch and Safetensors is a longer process than importing from GGUF. Improvements that make it easier are a work in progress.
+
+### Setup
+
+First, clone the `ollama/ollama` repo:
+
+```
+git clone git@github.com:ollama/ollama.git ollama
+cd ollama
+```
+
+and then fetch its `llama.cpp` submodule:
 
 ```shell
-$ ollama create -q Q4_K_M mymodel
-transferring model data
-quantizing F16 model to Q4_K_M
-creating new layer sha256:735e246cc1abfd06e9cdcf95504d6789a6cd1ad7577108a70d9902fef503c1bd
-creating new layer sha256:0853f0ad24e5865173bbf9ffcc7b0f5d56b66fd690ab1009867e45e7d2c4db0f
-writing manifest
-success
+git submodule init
+git submodule update llm/llama.cpp
 ```
 
-### Supported Quantizations
+Next, install the Python dependencies:
 
-- `Q4_0`
-- `Q4_1`
-- `Q5_0`
-- `Q5_1`
-- `Q8_0`
-
-#### K-means Quantizations
-
-- `Q3_K_S`
-- `Q3_K_M`
-- `Q3_K_L`
-- `Q4_K_S`
-- `Q4_K_M`
-- `Q5_K_S`
-- `Q5_K_M`
-- `Q6_K`
-
-## Template Detection
-
-> [!NOTE]
-> Template detection requires v0.1.42 or higher.
-
-Ollama uses model metadata, specifically `tokenizer.chat_template`, to automatically create a template appropriate for the model you're importing.
-
-```dockerfile
-FROM /path/to/my/gemma/model
+```
+python3 -m venv llm/llama.cpp/.venv
+source llm/llama.cpp/.venv/bin/activate
+pip install -r llm/llama.cpp/requirements.txt
 ```
 
-```shell
-$ ollama create mymodel
-transferring model data
-using autodetected template gemma-instruct
-creating new layer sha256:baa2a0edc27d19cc6b7537578a9a7ba1a4e3214dc185ed5ae43692b319af7b84
-creating new layer sha256:ba66c3309914dbef07e5149a648fd1877f030d337a4f240d444ea335008943cb
-writing manifest
-success
+Then build the `quantize` tool:
+
+```
+make -C llm/llama.cpp quantize
 ```
 
-Defining a template in the Modelfile will disable this feature which may be useful if you want to use a different template than the autodetected one.
+### Clone the HuggingFace repository (optional)
+
+If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model.
+
+Install [Git LFS](https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage), verify it's installed, and then clone the model's repository:
+
+```
+git lfs install
+git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1 model
+```
+
+### Convert the model
+
+> Note: some model architectures require using specific convert scripts. For example, Qwen models require running `convert-hf-to-gguf.py` instead of `convert.py`
+
+```
+python llm/llama.cpp/convert.py ./model --outtype f16 --outfile converted.bin
+```
+
+### Quantize the model
+
+```
+llm/llama.cpp/quantize converted.bin quantized.bin q4_0
+```
+
+### Step 3: Write a `Modelfile`
+
+Next, create a `Modelfile` for your model:
+
+```
+FROM quantized.bin
+TEMPLATE "[INST] {{ .Prompt }} [/INST]"
+```
+
+### Step 4: Create the Ollama model
+
+Finally, create a model from your `Modelfile`:
+
+```
+ollama create example -f Modelfile
+```
+
+### Step 5: Run your model
+
+Next, test the model with `ollama run`:
+
+```
+ollama run example "What is your favourite condiment?"
+```
+
+## Publishing your model (optional – early alpha)
+
+Publishing models is in early alpha. If you'd like to publish your model to share with others, follow these steps:
+
+1. Create [an account](https://ollama.com/signup)
+2. Copy your Ollama public key:
+  - macOS: `cat ~/.ollama/id_ed25519.pub | pbcopy`
+  - Windows: `type %USERPROFILE%\.ollama\id_ed25519.pub`
+  - Linux: `cat /usr/share/ollama/.ollama/id_ed25519.pub`
+3. Add your public key to your [Ollama account](https://ollama.com/settings/keys)
+
+Next, copy your model to your username's namespace:
+
+```
+ollama cp example <your username>/example
+```
+
+> Note: model names may only contain lowercase letters, digits, and the characters `.`, `-`, and `_`.
+
+Then push the model:
+
+```
+ollama push <your username>/example
+```
+
+After publishing, your model will be available at `https://ollama.com/<your username>/example`.
+
+## Quantization reference
+
+The quantization options are as follow (from highest highest to lowest levels of quantization). Note: some architectures such as Falcon do not support K quants.
+
+- `q2_K`
+- `q3_K`
+- `q3_K_S`
+- `q3_K_M`
+- `q3_K_L`
+- `q4_0` (recommended)
+- `q4_1`
+- `q4_K`
+- `q4_K_S`
+- `q4_K_M`
+- `q5_0`
+- `q5_1`
+- `q5_K`
+- `q5_K_S`
+- `q5_K_M`
+- `q6_K`
+- `q8_0`
+- `f16`

docs/linux.md

@@ -100,16 +100,6 @@ sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
 sudo chmod +x /usr/bin/ollama
 ```
 
-## Installing specific versions
-
-Use `OLLAMA_VERSION` environment variable with the install script to install a specific version of Ollama, including pre-releases. You can find the version numbers in the [releases page](https://github.com/ollama/ollama/releases). 
-
-For example:
-
-```
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.1.32 sh
-```
-
 ## Viewing logs
 
 To view logs of Ollama running as a startup service, run:

docs/openai.md

@@ -65,7 +65,6 @@ curl http://localhost:11434/v1/chat/completions \
             }
         ]
     }'
-
 ```
 
 ## Endpoints
@@ -105,6 +104,8 @@ curl http://localhost:11434/v1/chat/completions \
 
 #### Notes
 
+- Setting `seed` will always set `temperature` to `0`
+- `finish_reason` will always be `stop`
 - `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached
 
 ## Models

docs/troubleshooting.md

@@ -22,7 +22,7 @@ docker logs <container-name>
 If manually running `ollama serve` in a terminal, the logs will be on that terminal.
 
 When you run Ollama on **Windows**, there are a few different locations. You can view them in the explorer window by hitting `<cmd>+R` and type in:
-- `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
+- `explorer %LOCALAPPDATA%\Ollama` to view logs
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
 - `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
@@ -70,18 +70,13 @@ curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 
 If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
 
-## NVIDIA GPU Discovery
+## Container fails to run on NVIDIA GPU
 
-When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available.  Sometimes this discovery can fail to find your GPUs.  In general, running the latest driver will yield the best results.
+Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
 
-### Linux NVIDIA Troubleshooting
+Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
 
-If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
-
-Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
-
-- If you are using a container, is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
-- Is the uvm driver loaded? `sudo nvidia-modprobe -u`
+- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
 - Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
 - Try rebooting
 - Make sure you're running the latest nvidia drivers
@@ -89,8 +84,3 @@ Sometimes the Ollama can have difficulties initializing the GPU. When you check
 If none of those resolve the problem, gather additional information and file an issue:
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
-
-
-## Windows Terminal Errors
-
-Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.

docs/tutorials/langchainpy.md

@@ -45,7 +45,7 @@ all_splits = text_splitter.split_documents(data)
 ```
 
 It's split up, but we have to find the relevant splits and then submit those to the model. We can do this by creating embeddings and storing them in a vector database. We can use Ollama directly to instantiate an embedding model. We will use ChromaDB in this example for a vector database. `pip install chromadb`
-We also need to pull embedding model: `ollama pull nomic-embed-text`
+
 ```python
 from langchain.embeddings import OllamaEmbeddings
 from langchain.vectorstores import Chroma
@@ -68,8 +68,7 @@ The next thing is to send the question and the relevant parts of the docs to the
 ```python
 from langchain.chains import RetrievalQA
 qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
-res = qachain.invoke({"query": question})
-print(res['result'])
+qachain.invoke({"query": question})
 ```
 
 The answer received from this chain was:

docs/windows.md

@@ -19,7 +19,7 @@ Logs will often be helpful in diagnosing the problem (see
 
 ## System Requirements
 
-* Windows 10 22H2 or newer, Home or Pro
+* Windows 10 or newer, Home or Pro
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
 * AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
 
@@ -39,8 +39,8 @@ server.
 Ollama on Windows stores files in a few different locations.  You can view them in
 the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
-    - *app.log* contains most resent logs from the GUI application
-    - *server.log* contains the most recent server logs
+    - *app.log* contains logs from the GUI application
+    - *server.log* contains the server logs
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration

envconfig/config.go

@@ -1,31 +1,15 @@
 package envconfig
 
 import (
-	"errors"
 	"fmt"
 	"log/slog"
-	"math"
-	"net"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
-	"time"
 )
 
-type OllamaHost struct {
-	Scheme string
-	Host   string
-	Port   string
-}
-
-func (o OllamaHost) String() string {
-	return fmt.Sprintf("%s://%s:%s", o.Scheme, o.Host, o.Port)
-}
-
-var ErrInvalidHostPort = errors.New("invalid port specified in OLLAMA_HOST")
-
 var (
 	// Set via OLLAMA_ORIGINS in the environment
 	AllowOrigins []string
@@ -33,10 +17,8 @@ var (
 	Debug bool
 	// Experimental flash attention
 	FlashAttention bool
-	// Set via OLLAMA_HOST in the environment
-	Host *OllamaHost
 	// Set via OLLAMA_KEEP_ALIVE in the environment
-	KeepAlive time.Duration
+	KeepAlive string
 	// Set via OLLAMA_LLM_LIBRARY in the environment
 	LLMLibrary string
 	// Set via OLLAMA_MAX_LOADED_MODELS in the environment
@@ -45,8 +27,6 @@ var (
 	MaxQueuedRequests int
 	// Set via OLLAMA_MAX_VRAM in the environment
 	MaxVRAM uint64
-	// Set via OLLAMA_MODELS in the environment
-	ModelsDir string
 	// Set via OLLAMA_NOHISTORY in the environment
 	NoHistory bool
 	// Set via OLLAMA_NOPRUNE in the environment
@@ -55,23 +35,8 @@ var (
 	NumParallel int
 	// Set via OLLAMA_RUNNERS_DIR in the environment
 	RunnersDir string
-	// Set via OLLAMA_SCHED_SPREAD in the environment
-	SchedSpread bool
 	// Set via OLLAMA_TMPDIR in the environment
 	TmpDir string
-	// Set via OLLAMA_INTEL_GPU in the environment
-	IntelGpu bool
-
-	// Set via CUDA_VISIBLE_DEVICES in the environment
-	CudaVisibleDevices string
-	// Set via HIP_VISIBLE_DEVICES in the environment
-	HipVisibleDevices string
-	// Set via ROCR_VISIBLE_DEVICES in the environment
-	RocrVisibleDevices string
-	// Set via GPU_DEVICE_ORDINAL in the environment
-	GpuDeviceOrdinal string
-	// Set via HSA_OVERRIDE_GFX_VERSION in the environment
-	HsaOverrideGfxVersion string
 )
 
 type EnvVar struct {
@@ -81,33 +46,23 @@ type EnvVar struct {
 }
 
 func AsMap() map[string]EnvVar {
-	ret := map[string]EnvVar{
+	return map[string]EnvVar{
 		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug, "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
 		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention, "Enabled flash attention"},
-		"OLLAMA_HOST":              {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"},
+		"OLLAMA_HOST":              {"OLLAMA_HOST", "", "IP Address for the ollama server (default 127.0.0.1:11434)"},
 		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
-		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
-		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"},
+		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_ORIGINS", LLMLibrary, ""},
+		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"},
 		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
-		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
-		"OLLAMA_MODELS":            {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"},
+		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, ""},
+		"OLLAMA_MODELS":            {"OLLAMA_MODELS", "", "The path to the models directory"},
 		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
-		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"},
+		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"},
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
-		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"},
-		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"},
+		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, ""},
 		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir, "Location for temporary files"},
 	}
-	if runtime.GOOS != "darwin" {
-		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices, "Set which NVIDIA devices are visible"}
-		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices, "Set which AMD devices are visible"}
-		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices, "Set which AMD devices are visible"}
-		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal, "Set which AMD devices are visible"}
-		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion, "Override the gfx used for all detected AMD GPUs"}
-		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGpu, "Enable experimental Intel GPU detection"}
-	}
-	return ret
 }
 
 func Values() map[string]string {
@@ -131,10 +86,9 @@ func clean(key string) string {
 
 func init() {
 	// default values
-	NumParallel = 0 // Autoselect
-	MaxRunners = 0  // Autoselect
+	NumParallel = 1
+	MaxRunners = 1
 	MaxQueuedRequests = 512
-	KeepAlive = 5 * time.Minute
 
 	LoadConfig()
 }
@@ -172,7 +126,7 @@ func LoadConfig() {
 		var paths []string
 		for _, root := range []string{filepath.Dir(appExe), cwd} {
 			paths = append(paths,
-				root,
+				filepath.Join(root),
 				filepath.Join(root, "windows-"+runtime.GOARCH),
 				filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
 			)
@@ -208,8 +162,8 @@ func LoadConfig() {
 
 	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {
 		val, err := strconv.Atoi(onp)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_NUM_PARALLEL", onp, "error", err)
+		if err != nil || val <= 0 {
+			slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
 		} else {
 			NumParallel = val
 		}
@@ -219,15 +173,6 @@ func LoadConfig() {
 		NoHistory = true
 	}
 
-	if spread := clean("OLLAMA_SCHED_SPREAD"); spread != "" {
-		s, err := strconv.ParseBool(spread)
-		if err == nil {
-			SchedSpread = s
-		} else {
-			SchedSpread = true
-		}
-	}
-
 	if noprune := clean("OLLAMA_NOPRUNE"); noprune != "" {
 		NoPrune = true
 	}
@@ -239,22 +184,16 @@ func LoadConfig() {
 		AllowOrigins = append(AllowOrigins,
 			fmt.Sprintf("http://%s", allowOrigin),
 			fmt.Sprintf("https://%s", allowOrigin),
-			fmt.Sprintf("http://%s", net.JoinHostPort(allowOrigin, "*")),
-			fmt.Sprintf("https://%s", net.JoinHostPort(allowOrigin, "*")),
+			fmt.Sprintf("http://%s:*", allowOrigin),
+			fmt.Sprintf("https://%s:*", allowOrigin),
 		)
 	}
 
-	AllowOrigins = append(AllowOrigins,
-		"app://*",
-		"file://*",
-		"tauri://*",
-	)
-
 	maxRunners := clean("OLLAMA_MAX_LOADED_MODELS")
 	if maxRunners != "" {
 		m, err := strconv.Atoi(maxRunners)
 		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
+			slog.Error("invalid setting", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
 		} else {
 			MaxRunners = m
 		}
@@ -263,111 +202,11 @@ func LoadConfig() {
 	if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" {
 		p, err := strconv.Atoi(onp)
 		if err != nil || p <= 0 {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_QUEUE", onp, "error", err)
+			slog.Error("invalid setting", "OLLAMA_MAX_QUEUE", onp, "error", err)
 		} else {
 			MaxQueuedRequests = p
 		}
 	}
 
-	ka := clean("OLLAMA_KEEP_ALIVE")
-	if ka != "" {
-		loadKeepAlive(ka)
-	}
-
-	var err error
-	ModelsDir, err = getModelsDir()
-	if err != nil {
-		slog.Error("invalid setting", "OLLAMA_MODELS", ModelsDir, "error", err)
-	}
-
-	Host, err = getOllamaHost()
-	if err != nil {
-		slog.Error("invalid setting", "OLLAMA_HOST", Host, "error", err, "using default port", Host.Port)
-	}
-
-	if set, err := strconv.ParseBool(clean("OLLAMA_INTEL_GPU")); err == nil {
-		IntelGpu = set
-	}
-
-	CudaVisibleDevices = clean("CUDA_VISIBLE_DEVICES")
-	HipVisibleDevices = clean("HIP_VISIBLE_DEVICES")
-	RocrVisibleDevices = clean("ROCR_VISIBLE_DEVICES")
-	GpuDeviceOrdinal = clean("GPU_DEVICE_ORDINAL")
-	HsaOverrideGfxVersion = clean("HSA_OVERRIDE_GFX_VERSION")
-}
-
-func getModelsDir() (string, error) {
-	if models, exists := os.LookupEnv("OLLAMA_MODELS"); exists {
-		return models, nil
-	}
-	home, err := os.UserHomeDir()
-	if err != nil {
-		return "", err
-	}
-	return filepath.Join(home, ".ollama", "models"), nil
-}
-
-func getOllamaHost() (*OllamaHost, error) {
-	defaultPort := "11434"
-
-	hostVar := os.Getenv("OLLAMA_HOST")
-	hostVar = strings.TrimSpace(strings.Trim(strings.TrimSpace(hostVar), "\"'"))
-
-	scheme, hostport, ok := strings.Cut(hostVar, "://")
-	switch {
-	case !ok:
-		scheme, hostport = "http", hostVar
-	case scheme == "http":
-		defaultPort = "80"
-	case scheme == "https":
-		defaultPort = "443"
-	}
-
-	// trim trailing slashes
-	hostport = strings.TrimRight(hostport, "/")
-
-	host, port, err := net.SplitHostPort(hostport)
-	if err != nil {
-		host, port = "127.0.0.1", defaultPort
-		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
-			host = ip.String()
-		} else if hostport != "" {
-			host = hostport
-		}
-	}
-
-	if portNum, err := strconv.ParseInt(port, 10, 32); err != nil || portNum > 65535 || portNum < 0 {
-		return &OllamaHost{
-			Scheme: scheme,
-			Host:   host,
-			Port:   defaultPort,
-		}, ErrInvalidHostPort
-	}
-
-	return &OllamaHost{
-		Scheme: scheme,
-		Host:   host,
-		Port:   port,
-	}, nil
-}
-
-func loadKeepAlive(ka string) {
-	v, err := strconv.Atoi(ka)
-	if err != nil {
-		d, err := time.ParseDuration(ka)
-		if err == nil {
-			if d < 0 {
-				KeepAlive = time.Duration(math.MaxInt64)
-			} else {
-				KeepAlive = d
-			}
-		}
-	} else {
-		d := time.Duration(v) * time.Second
-		if d < 0 {
-			KeepAlive = time.Duration(math.MaxInt64)
-		} else {
-			KeepAlive = d
-		}
-	}
+	KeepAlive = clean("OLLAMA_KEEP_ALIVE")
 }

envconfig/config_test.go

@@ -1,13 +1,8 @@
 package envconfig
 
 import (
-	"fmt"
-	"math"
-	"net"
 	"testing"
-	"time"
 
-	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
@@ -25,64 +20,4 @@ func TestConfig(t *testing.T) {
 	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
 	LoadConfig()
 	require.True(t, FlashAttention)
-	t.Setenv("OLLAMA_KEEP_ALIVE", "")
-	LoadConfig()
-	require.Equal(t, 5*time.Minute, KeepAlive)
-	t.Setenv("OLLAMA_KEEP_ALIVE", "3")
-	LoadConfig()
-	require.Equal(t, 3*time.Second, KeepAlive)
-	t.Setenv("OLLAMA_KEEP_ALIVE", "1h")
-	LoadConfig()
-	require.Equal(t, 1*time.Hour, KeepAlive)
-	t.Setenv("OLLAMA_KEEP_ALIVE", "-1s")
-	LoadConfig()
-	require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
-	t.Setenv("OLLAMA_KEEP_ALIVE", "-1")
-	LoadConfig()
-	require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
-}
-
-func TestClientFromEnvironment(t *testing.T) {
-	type testCase struct {
-		value  string
-		expect string
-		err    error
-	}
-
-	hostTestCases := map[string]*testCase{
-		"empty":               {value: "", expect: "127.0.0.1:11434"},
-		"only address":        {value: "1.2.3.4", expect: "1.2.3.4:11434"},
-		"only port":           {value: ":1234", expect: ":1234"},
-		"address and port":    {value: "1.2.3.4:1234", expect: "1.2.3.4:1234"},
-		"hostname":            {value: "example.com", expect: "example.com:11434"},
-		"hostname and port":   {value: "example.com:1234", expect: "example.com:1234"},
-		"zero port":           {value: ":0", expect: ":0"},
-		"too large port":      {value: ":66000", err: ErrInvalidHostPort},
-		"too small port":      {value: ":-1", err: ErrInvalidHostPort},
-		"ipv6 localhost":      {value: "[::1]", expect: "[::1]:11434"},
-		"ipv6 world open":     {value: "[::]", expect: "[::]:11434"},
-		"ipv6 no brackets":    {value: "::1", expect: "[::1]:11434"},
-		"ipv6 + port":         {value: "[::1]:1337", expect: "[::1]:1337"},
-		"extra space":         {value: " 1.2.3.4 ", expect: "1.2.3.4:11434"},
-		"extra quotes":        {value: "\"1.2.3.4\"", expect: "1.2.3.4:11434"},
-		"extra space+quotes":  {value: " \" 1.2.3.4 \" ", expect: "1.2.3.4:11434"},
-		"extra single quotes": {value: "'1.2.3.4'", expect: "1.2.3.4:11434"},
-	}
-
-	for k, v := range hostTestCases {
-		t.Run(k, func(t *testing.T) {
-			t.Setenv("OLLAMA_HOST", v.value)
-			LoadConfig()
-
-			oh, err := getOllamaHost()
-			if err != v.err {
-				t.Fatalf("expected %s, got %s", v.err, err)
-			}
-
-			if err == nil {
-				host := net.JoinHostPort(oh.Host, oh.Port)
-				assert.Equal(t, v.expect, host, fmt.Sprintf("%s: expected %s, got %s", k, v.expect, host))
-			}
-		})
-	}
 }

examples/langchain-python-rag-privategpt/ingest.py

@@ -77,21 +77,13 @@ LOADER_MAPPING = {
 
 
 def load_single_document(file_path: str) -> List[Document]:
-    if os.path.getsize(file_path) != 0:
-        filename, ext = os.path.splitext(file_path)
-        if ext in LOADER_MAPPING:
-            loader_class, loader_args = LOADER_MAPPING[ext]
-            try:
-                loader = loader_class(file_path, **loader_args)
-                if loader:
-                    return loader.load()
-            except:
-                print(f"Corrupted file {file_path}. Ignoring it.")
-        else:
-            print(f"Unsupported file {file_path}. Ignoring it.")
-    else:
-        print(f"Empty file {file_path}. Ignoring it.")
+    ext = "." + file_path.rsplit(".", 1)[-1]
+    if ext in LOADER_MAPPING:
+        loader_class, loader_args = LOADER_MAPPING[ext]
+        loader = loader_class(file_path, **loader_args)
+        return loader.load()
 
+    raise ValueError(f"Unsupported file extension '{ext}'")
 
 def load_documents(source_dir: str, ignored_files: List[str] = []) -> List[Document]:
     """
@@ -108,8 +100,7 @@ def load_documents(source_dir: str, ignored_files: List[str] = []) -> List[Docum
         results = []
         with tqdm(total=len(filtered_files), desc='Loading new documents', ncols=80) as pbar:
             for i, docs in enumerate(pool.imap_unordered(load_single_document, filtered_files)):
-                if docs:
-                    results.extend(docs)
+                results.extend(docs)
                 pbar.update()
 
     return results

examples/langchain-python-rag-privategpt/requirements.txt

@@ -11,5 +11,4 @@ tabulate==0.9.0
 pandoc==2.3
 pypandoc==1.11
 tqdm==4.66.1
-sentence_transformers==2.2.2
-numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability
+sentence_transformers==2.2.2

format/format_test.go

@@ -5,6 +5,7 @@ import (
 )
 
 func TestHumanNumber(t *testing.T) {
+
 	type testCase struct {
 		input    uint64
 		expected string

go.mod

@@ -16,12 +16,10 @@ require (
 )
 
 require (
-	github.com/agnivade/levenshtein v1.1.1
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
-	github.com/sbinet/npyio v0.9.0
 )
 
 require (
@@ -72,7 +70,7 @@ require (
 	golang.org/x/net v0.25.0 // indirect
 	golang.org/x/sys v0.20.0
 	golang.org/x/term v0.20.0
-	golang.org/x/text v0.15.0
+	golang.org/x/text v0.15.0 // indirect
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -4,14 +4,10 @@ dmitri.shuralyov.com/gpu/mtl v0.0.0-20190408044501-666a987793e9/go.mod h1:H6x//7
 gioui.org v0.0.0-20210308172011-57750fc8a0a6/go.mod h1:RSH6KIUZ0p2xy5zHDxgAM4zumjgTw83q2ge/PI+yyw8=
 github.com/BurntSushi/toml v0.3.1/go.mod h1:xHWCNGjB5oqiDr8zfno3MHue2Ht5sIBksp03qcyfWMU=
 github.com/BurntSushi/xgb v0.0.0-20160522181843-27f122750802/go.mod h1:IVnqGOEym/WlBOVXweHU+Q+/VP0lqqI8lqeDx9IjBqo=
-github.com/agnivade/levenshtein v1.1.1 h1:QY8M92nrzkmr798gCo3kmMyqXFzdQVpxLlGPRBij0P8=
-github.com/agnivade/levenshtein v1.1.1/go.mod h1:veldBMzWxcCG2ZvUTKD2kJNRdCk5hVbJomOvKkmgYbo=
 github.com/ajstarks/svgo v0.0.0-20180226025133-644b8db467af/go.mod h1:K08gAheRH3/J6wwsYMMT4xOr94bZjxIelGM0+d/wbFw=
 github.com/antihax/optional v1.0.0/go.mod h1:uupD/76wgC+ih3iEmQUL+0Ugr19nfwCT1kdvxnR2qWY=
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 h1:q4dksr6ICHXqG5hm0ZW5IHyeEJXoIJSOZeBLmWPNeIQ=
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40/go.mod h1:Q7yQnSMnLvcXlZ8RV+jwz/6y1rQTqbX6C82SndT52Zs=
-github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0 h1:jfIu9sQUG6Ig+0+Ap1h4unLjW6YQJpKZVmUzxsD4E/Q=
-github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0/go.mod h1:t2tdKJDJF9BV14lnkjHmOQgcvEKgtqs5a1N3LNdJhGE=
 github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
 github.com/bytedance/sonic v1.11.6 h1:oUp34TzMlL+OY1OUWxHqsdkgC/Zfc85zGqw9siXjrc0=
 github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1ZaiQtds4=
@@ -40,8 +36,6 @@ github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1/go.mod h1:uw2gLc
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
-github.com/dgryski/trifles v0.0.0-20200323201526-dd97f9abfb48 h1:fRzb/w+pyskVMQ+UbP35JkH8yB7MYb4q/qhBarqZE6g=
-github.com/dgryski/trifles v0.0.0-20200323201526-dd97f9abfb48/go.mod h1:if7Fbed8SFyPtHLHbg49SI7NAdJiC5WIA09pe59rfAA=
 github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
 github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
@@ -171,8 +165,6 @@ github.com/rogpeppe/go-internal v1.8.0 h1:FCbCCtXNOY3UtUuHUYaghJg4y7Fd14rXifAYUA
 github.com/rogpeppe/go-internal v1.8.0/go.mod h1:WmiCO8CzOY8rg0OYDC4/i/2WRWAB6poM+XZ2dLUbcbE=
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
 github.com/ruudk/golang-pdf417 v0.0.0-20181029194003-1af4ab5afa58/go.mod h1:6lfFZQK844Gfx8o5WFuvpxWRwnSoipWe/p622j1v06w=
-github.com/sbinet/npyio v0.9.0 h1:A7h8OyYsOsc+NPRtynRMSf70xSgATZNpamNp8nQ8Tjc=
-github.com/sbinet/npyio v0.9.0/go.mod h1:vgjQEMRTS9aMS9GdXhr+5jounCmGqjDO2JI+IpSokns=
 github.com/spf13/cobra v1.7.0 h1:hyqWnYt1ZQShIddO5kBpj3vu05/++x6tJ6dg8EC572I=
 github.com/spf13/cobra v1.7.0/go.mod h1:uLxZILRyS/50WlhOIKD7W6V5bgeIt+4sICxh6uRMrb0=
 github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA=

gpu/amd_linux.go

@@ -13,7 +13,6 @@ import (
 	"strconv"
 	"strings"
 
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )
 
@@ -26,16 +25,7 @@ const (
 
 	// Prefix with the node dir
 	GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
-
-	// Direct Rendering Manager sysfs location
-	DRMDeviceDirGlob   = "/sys/class/drm/card*/device"
-	DRMTotalMemoryFile = "mem_info_vram_total"
-	DRMUsedMemoryFile  = "mem_info_vram_used"
-
-	// In hex; properties file is in decimal
-	DRMUniqueIDFile = "unique_id"
-	DRMVendorFile   = "vendor"
-	DRMDeviceFile   = "device"
+	GPUUsedMemoryFileGlob  = "mem_banks/*/used_memory"
 )
 
 var (
@@ -45,8 +35,8 @@ var (
 )
 
 // Gather GPU information from the amdgpu driver if any supported GPUs are detected
-func AMDGetGPUInfo() []RocmGPUInfo {
-	resp := []RocmGPUInfo{}
+func AMDGetGPUInfo() []GpuInfo {
+	resp := []GpuInfo{}
 	if !AMDDetected() {
 		return resp
 	}
@@ -60,9 +50,9 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 
 	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
 	var visibleDevices []string
-	hipVD := envconfig.HipVisibleDevices   // zero based index only
-	rocrVD := envconfig.RocrVisibleDevices // zero based index or UUID, but consumer cards seem to not support UUID
-	gpuDO := envconfig.GpuDeviceOrdinal    // zero based index
+	hipVD := os.Getenv("HIP_VISIBLE_DEVICES")   // zero based index only
+	rocrVD := os.Getenv("ROCR_VISIBLE_DEVICES") // zero based index or UUID, but consumer cards seem to not support UUID
+	gpuDO := os.Getenv("GPU_DEVICE_ORDINAL")    // zero based index
 	switch {
 	// TODO is this priorty order right?
 	case hipVD != "":
@@ -75,7 +65,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		visibleDevices = strings.Split(gpuDO, ",")
 	}
 
-	gfxOverride := envconfig.HsaOverrideGfxVersion
+	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
 	var supported []string
 	libDir := ""
 
@@ -100,7 +90,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		scanner := bufio.NewScanner(fp)
 		isCPU := false
 		var major, minor, patch uint64
-		var vendor, device, uniqueID uint64
+		var vendor, device uint64
 		for scanner.Scan() {
 			line := strings.TrimSpace(scanner.Text())
 			// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
@@ -131,43 +121,30 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			} else if strings.HasPrefix(line, "vendor_id") {
 				ver := strings.Fields(line)
 				if len(ver) != 2 {
-					slog.Debug("malformed", "vendor_id", line)
+					slog.Debug("malformed vendor_id", "vendor_id", line)
 					continue
 				}
-				vendor, err = strconv.ParseUint(ver[1], 10, 64)
+				vendor, err = strconv.ParseUint(ver[1], 10, 32)
 				if err != nil {
-					slog.Debug("malformed", "vendor_id", line, "error", err)
+					slog.Debug("malformed vendor_id" + line)
 				}
 			} else if strings.HasPrefix(line, "device_id") {
 				ver := strings.Fields(line)
 				if len(ver) != 2 {
-					slog.Debug("malformed", "device_id", line)
+					slog.Debug("malformed device_id", "device_id", line)
 					continue
 				}
-				device, err = strconv.ParseUint(ver[1], 10, 64)
+				device, err = strconv.ParseUint(ver[1], 10, 32)
 				if err != nil {
-					slog.Debug("malformed", "device_id", line, "error", err)
-				}
-			} else if strings.HasPrefix(line, "unique_id") {
-				ver := strings.Fields(line)
-				if len(ver) != 2 {
-					slog.Debug("malformed", "unique_id", line)
-					continue
-				}
-				uniqueID, err = strconv.ParseUint(ver[1], 10, 64)
-				if err != nil {
-					slog.Debug("malformed", "unique_id", line, "error", err)
+					slog.Debug("malformed device_id" + line)
 				}
 			}
+
 			// TODO - any other properties we want to extract and record?
 			// vendor_id + device_id -> pci lookup for "Name"
 			// Other metrics that may help us understand relative performance between multiple GPUs
 		}
 
-		// Note: while ./mem_banks/*/used_memory exists, it doesn't appear to take other VRAM consumers
-		// into consideration, so we instead map the device over to the DRM driver sysfs nodes which
-		// do reliably report VRAM usage.
-
 		if isCPU {
 			cpuCount++
 			continue
@@ -179,7 +156,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		// Shouldn't happen, but just in case...
 		if gpuID < 0 {
 			slog.Error("unexpected amdgpu sysfs data resulted in negative GPU ID, please set OLLAMA_DEBUG=1 and report an issue")
-			return nil
+			return []GpuInfo{}
 		}
 
 		if int(major) < RocmComputeMin {
@@ -190,68 +167,65 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		// Look up the memory for the current node
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
-		var usedFile string
-		mapping := []struct {
-			id       uint64
-			filename string
-		}{
-			{vendor, DRMVendorFile},
-			{device, DRMDeviceFile},
-			{uniqueID, DRMUniqueIDFile}, // Not all devices will report this
+		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUTotalMemoryFileGlob)
+		propFiles, err := filepath.Glob(propGlob)
+		if err != nil {
+			slog.Warn("error looking up total GPU memory", "glob", propGlob, "error", err)
 		}
-		slog.Debug("mapping amdgpu to drm sysfs nodes", "amdgpu", match, "vendor", vendor, "device", device, "unique_id", uniqueID)
-		// Map over to DRM location to find the total/free memory
-		drmMatches, _ := filepath.Glob(DRMDeviceDirGlob)
-		for _, devDir := range drmMatches {
-			matched := true
-			for _, m := range mapping {
-				if m.id == 0 {
-					// Null ID means it didn't populate, so we can't use it to match
-					continue
-				}
-				filename := filepath.Join(devDir, m.filename)
-				buf, err := os.ReadFile(filename)
-				if err != nil {
-					slog.Debug("failed to read sysfs node", "file", filename, "error", err)
-					matched = false
-					break
-				}
-				// values here are in hex, strip off the lead 0x and parse so we can compare the numeric (decimal) values in amdgpu
-				cmp, err := strconv.ParseUint(strings.TrimPrefix(strings.TrimSpace(string(buf)), "0x"), 16, 64)
-				if err != nil {
-					slog.Debug("failed to parse sysfs node", "file", filename, "error", err)
-					matched = false
-					break
-				}
-				if cmp != m.id {
-					matched = false
-					break
-				}
-			}
-			if !matched {
+		// 1 or more memory banks - sum the values of all of them
+		for _, propFile := range propFiles {
+			fp, err := os.Open(propFile)
+			if err != nil {
+				slog.Warn("failed to open sysfs node", "file", propFile, "erroir", err)
 				continue
 			}
-
-			// Found the matching DRM directory
-			slog.Debug("matched", "amdgpu", match, "drm", devDir)
-			totalFile := filepath.Join(devDir, DRMTotalMemoryFile)
-			buf, err := os.ReadFile(totalFile)
-			if err != nil {
-				slog.Debug("failed to read sysfs node", "file", totalFile, "error", err)
-				break
+			defer fp.Close()
+			scanner := bufio.NewScanner(fp)
+			for scanner.Scan() {
+				line := strings.TrimSpace(scanner.Text())
+				if strings.HasPrefix(line, "size_in_bytes") {
+					ver := strings.Fields(line)
+					if len(ver) != 2 {
+						slog.Warn("malformed " + line)
+						continue
+					}
+					bankSizeInBytes, err := strconv.ParseUint(ver[1], 10, 64)
+					if err != nil {
+						slog.Warn("malformed int " + line)
+						continue
+					}
+					totalMemory += bankSizeInBytes
+				}
 			}
-			totalMemory, err = strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
+		}
+		if totalMemory == 0 {
+			slog.Warn("amdgpu reports zero total memory", "gpu", gpuID)
+			continue
+		}
+		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUUsedMemoryFileGlob)
+		usedFiles, err := filepath.Glob(usedGlob)
+		if err != nil {
+			slog.Warn("error looking up used GPU memory", "glob", usedGlob, "error", err)
+			continue
+		}
+		for _, usedFile := range usedFiles {
+			fp, err := os.Open(usedFile)
 			if err != nil {
-				slog.Debug("failed to parse sysfs node", "file", totalFile, "error", err)
-				break
+				slog.Warn("failed to open sysfs node", "file", usedFile, "error", err)
+				continue
 			}
-
-			usedFile = filepath.Join(devDir, DRMUsedMemoryFile)
-			usedMemory, err = getFreeMemory(usedFile)
+			defer fp.Close()
+			data, err := io.ReadAll(fp)
 			if err != nil {
-				slog.Debug("failed to update used memory", "error", err)
+				slog.Warn("failed to read sysfs node", "file", usedFile, "error", err)
+				continue
 			}
-			break
+			used, err := strconv.ParseUint(strings.TrimSpace(string(data)), 10, 64)
+			if err != nil {
+				slog.Warn("malformed used memory", "data", string(data), "error", err)
+				continue
+			}
+			usedMemory += used
 		}
 
 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
@@ -267,21 +241,18 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 
 		slog.Debug("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
-		gpuInfo := RocmGPUInfo{
-			GpuInfo: GpuInfo{
-				Library: "rocm",
-				memInfo: memInfo{
-					TotalMemory: totalMemory,
-					FreeMemory:  (totalMemory - usedMemory),
-				},
-				ID:            strconv.Itoa(gpuID),
-				Name:          name,
-				Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
-				MinimumMemory: rocmMinimumMemory,
-				DriverMajor:   driverMajor,
-				DriverMinor:   driverMinor,
+		gpuInfo := GpuInfo{
+			Library: "rocm",
+			memInfo: memInfo{
+				TotalMemory: totalMemory,
+				FreeMemory:  (totalMemory - usedMemory),
 			},
-			usedFilepath: usedFile,
+			ID:            fmt.Sprintf("%d", gpuID),
+			Name:          name,
+			Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
+			MinimumMemory: rocmMinimumMemory,
+			DriverMajor:   driverMajor,
+			DriverMinor:   driverMinor,
 		}
 
 		// If the user wants to filter to a subset of devices, filter out if we aren't a match
@@ -305,7 +276,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			libDir, err = AMDValidateLibDir()
 			if err != nil {
 				slog.Warn("unable to verify rocm library, will use cpu", "error", err)
-				return nil
+				return []GpuInfo{}
 			}
 		}
 		gpuInfo.DependencyPath = libDir
@@ -316,7 +287,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 				supported, err = GetSupportedGFX(libDir)
 				if err != nil {
 					slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
-					return nil
+					return []GpuInfo{}
 				}
 				slog.Debug("rocm supported GPUs", "types", supported)
 			}
@@ -333,11 +304,6 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
 		}
 
-		// Check for env var workarounds
-		if name == "1002:687f" { // Vega RX 56
-			gpuInfo.EnvWorkarounds = append(gpuInfo.EnvWorkarounds, [2]string{"HSA_ENABLE_SDMA", "0"})
-		}
-
 		// The GPU has passed all the verification steps and is supported
 		resp = append(resp, gpuInfo)
 	}
@@ -412,31 +378,3 @@ func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	}
 	return driverMajor, driverMinor, nil
 }
-
-func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
-	if len(gpus) == 0 {
-		return nil
-	}
-	for i := range gpus {
-		usedMemory, err := getFreeMemory(gpus[i].usedFilepath)
-		if err != nil {
-			return err
-		}
-		slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(gpus[i].TotalMemory-usedMemory))
-		gpus[i].FreeMemory = gpus[i].TotalMemory - usedMemory
-	}
-	return nil
-}
-
-func getFreeMemory(usedFile string) (uint64, error) {
-	buf, err := os.ReadFile(usedFile)
-	if err != nil {
-		return 0, fmt.Errorf("failed to read sysfs node %s %w", usedFile, err)
-	}
-	usedMemory, err := strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
-	if err != nil {
-		slog.Debug("failed to parse sysfs node", "file", usedFile, "error", err)
-		return 0, fmt.Errorf("failed to parse sysfs node %s %w", usedFile, err)
-	}
-	return usedMemory, nil
-}

gpu/amd_windows.go

@@ -7,10 +7,8 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
-	"strconv"
 	"strings"
 
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )
 
@@ -26,8 +24,8 @@ var (
 	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\5.7\\bin"} // TODO glob?
 )
 
-func AMDGetGPUInfo() []RocmGPUInfo {
-	resp := []RocmGPUInfo{}
+func AMDGetGPUInfo() []GpuInfo {
+	resp := []GpuInfo{}
 	hl, err := NewHipLib()
 	if err != nil {
 		slog.Debug(err.Error())
@@ -54,7 +52,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	}
 
 	var supported []string
-	gfxOverride := envconfig.HsaOverrideGfxVersion
+	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
 	if gfxOverride == "" {
 		supported, err = GetSupportedGFX(libDir)
 		if err != nil {
@@ -67,7 +65,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 
 	slog.Debug("detected hip devices", "count", count)
 	// TODO how to determine the underlying device ID when visible devices is causing this to subset?
-	for i := range count {
+	for i := 0; i < count; i++ {
 		err = hl.HipSetDevice(i)
 		if err != nil {
 			slog.Warn("set device", "id", i, "error", err)
@@ -115,29 +113,25 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			continue
 		}
 
+		// TODO revisit this once ROCm v6 is available on windows.
+		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
 		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
-		gpuInfo := RocmGPUInfo{
-			GpuInfo: GpuInfo{
-				Library: "rocm",
-				memInfo: memInfo{
-					TotalMemory: totalMemory,
-					FreeMemory:  freeMemory,
-				},
-				// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
-				UnreliableFreeMemory: true,
-
-				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				DependencyPath: libDir,
-				MinimumMemory:  rocmMinimumMemory,
-				Name:           name,
-				Compute:        gfx,
-
-				// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
-				// DriverMajor:    driverMajor,
-				// DriverMinor:    driverMinor,
+		gpuInfo := GpuInfo{
+			Library: "rocm",
+			memInfo: memInfo{
+				TotalMemory: totalMemory,
+				FreeMemory:  freeMemory,
 			},
-			index: i,
+			ID:             fmt.Sprintf("%d", i), // TODO this is probably wrong if we specify visible devices
+			DependencyPath: libDir,
+			MinimumMemory:  rocmMinimumMemory,
+			Name:           name,
+			Compute:        gfx,
+
+			// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
+			// DriverMajor:    driverMajor,
+			// DriverMinor:    driverMinor,
 		}
 
 		resp = append(resp, gpuInfo)
@@ -165,30 +159,3 @@ func AMDValidateLibDir() (string, error) {
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
 	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
 }
-
-func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
-	if len(gpus) == 0 {
-		return nil
-	}
-	hl, err := NewHipLib()
-	if err != nil {
-		slog.Debug(err.Error())
-		return nil
-	}
-	defer hl.Release()
-
-	for i := range gpus {
-		err := hl.HipSetDevice(gpus[i].index)
-		if err != nil {
-			return err
-		}
-		freeMemory, _, err := hl.HipMemGetInfo()
-		if err != nil {
-			slog.Warn("get mem info", "id", i, "error", err)
-			continue
-		}
-		slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(freeMemory))
-		gpus[i].FreeMemory = freeMemory
-	}
-	return nil
-}

gpu/assets.go

@@ -77,27 +77,20 @@ func cleanupTmpDirs() {
 			continue
 		}
 		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
+		if err == nil {
+			pid, err := strconv.Atoi(string(raw))
+			if err == nil {
+				if proc, err := os.FindProcess(int(pid)); err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+					// Another running ollama, ignore this tmpdir
+					continue
+				}
+			}
+		} else {
+			slog.Debug("failed to open ollama.pid", "path", d, "error", err)
+		}
+		err = os.RemoveAll(d)
 		if err != nil {
-			slog.Warn("failed to read ollama.pid", "path", d, "error", err)
-			// No pid, ignore this tmpdir
-			continue
-		}
-
-		pid, err := strconv.Atoi(string(raw))
-		if err != nil {
-			slog.Warn("failed to parse pid", "path", d, "error", err)
-			continue
-		}
-
-		proc, err := os.FindProcess(pid)
-		if err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
-			slog.Warn("found running ollama", "pid", pid, "path", d)
-			// Another running ollama, ignore this tmpdir
-			continue
-		}
-
-		if err := os.Remove(d); err != nil {
-			slog.Warn("unable to cleanup stale tmpdir", "path", d, "error", err)
+			slog.Debug("unable to cleanup stale tmpdir", "path", d, "error", err)
 		}
 	}
 }

gpu/cpu_common.go

@@ -1,16 +1,21 @@
 package gpu
 
 import (
+	"log/slog"
+
 	"golang.org/x/sys/cpu"
 )
 
-func GetCPUCapability() CPUCapability {
+func GetCPUVariant() string {
 	if cpu.X86.HasAVX2 {
-		return CPUCapabilityAVX2
+		slog.Debug("CPU has AVX2")
+		return "avx2"
 	}
 	if cpu.X86.HasAVX {
-		return CPUCapabilityAVX
+		slog.Debug("CPU has AVX")
+		return "avx"
 	}
+	slog.Debug("CPU does not have vector extensions")
 	// else LCD
-	return CPUCapabilityNone
+	return ""
 }

gpu/cuda_common.go

@@ -18,4 +18,5 @@ func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 		ids = append(ids, info.ID)
 	}
 	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
+
 }

gpu/gpu.go

@@ -16,45 +16,28 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"sync"
 	"unsafe"
 
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/envconfig"
 )
 
-type cudaHandles struct {
+type handles struct {
 	deviceCount int
 	cudart      *C.cudart_handle_t
 	nvcuda      *C.nvcuda_handle_t
-	nvml        *C.nvml_handle_t
-}
-
-type oneapiHandles struct {
 	oneapi      *C.oneapi_handle_t
-	deviceCount int
 }
 
 const (
 	cudaMinimumMemory = 457 * format.MebiByte
 	rocmMinimumMemory = 457 * format.MebiByte
-	// TODO OneAPI minimum memory
 )
 
-var (
-	gpuMutex      sync.Mutex
-	bootstrapped  bool
-	cpuCapability CPUCapability
-	cpus          []CPUInfo
-	cudaGPUs      []CudaGPUInfo
-	nvcudaLibPath string
-	cudartLibPath string
-	oneapiLibPath string
-	nvmlLibPath   string
-	rocmGPUs      []RocmGPUInfo
-	oneapiGPUs    []OneapiGPUInfo
-)
+var gpuMutex sync.Mutex
 
 // With our current CUDA compile flags, older than 5.0 will not work properly
 var CudaComputeMin = [2]C.int{5, 0}
@@ -64,113 +47,130 @@ var RocmComputeMin = 9
 // TODO find a better way to detect iGPU instead of minimum memory
 const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
 
+var CudartLinuxGlobs = []string{
+	"/usr/local/cuda/lib64/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
+	"/usr/lib/wsl/lib/libcudart.so*",
+	"/usr/lib/wsl/drivers/*/libcudart.so*",
+	"/opt/cuda/lib64/libcudart.so*",
+	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
+	"/usr/local/cuda/lib*/libcudart.so*",
+	"/usr/lib*/libcudart.so*",
+	"/usr/local/lib*/libcudart.so*",
+}
+
+var CudartWindowsGlobs = []string{
+	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
+}
+
+var NvcudaLinuxGlobs = []string{
+	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
+	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
+	"/usr/lib/*-linux-gnu/libcuda.so*",
+	"/usr/lib/wsl/lib/libcuda.so*",
+	"/usr/lib/wsl/drivers/*/libcuda.so*",
+	"/opt/cuda/lib*/libcuda.so*",
+	"/usr/local/cuda/lib*/libcuda.so*",
+	"/usr/lib*/libcuda.so*",
+	"/usr/local/lib*/libcuda.so*",
+}
+
+var NvcudaWindowsGlobs = []string{
+	"c:\\windows\\system*\\nvcuda.dll",
+}
+
+var OneapiWindowsGlobs = []string{
+	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
+}
+
+var OneapiLinuxGlobs = []string{
+	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
+	"/usr/lib*/libze_intel_gpu.so*",
+}
+
 // Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
 // Note: gpuMutex must already be held
-func initCudaHandles() *cudaHandles {
+func initGPUHandles() *handles {
 
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 
-	cHandles := &cudaHandles{}
-	// Short Circuit if we already know which library to use
-	if nvmlLibPath != "" {
-		cHandles.nvml, _ = LoadNVMLMgmt([]string{nvmlLibPath})
-		return cHandles
-	}
-	if nvcudaLibPath != "" {
-		cHandles.deviceCount, cHandles.nvcuda, _ = LoadNVCUDAMgmt([]string{nvcudaLibPath})
-		return cHandles
-	}
-	if cudartLibPath != "" {
-		cHandles.deviceCount, cHandles.cudart, _ = LoadCUDARTMgmt([]string{cudartLibPath})
-		return cHandles
-	}
-
-	slog.Debug("searching for GPU discovery libraries for NVIDIA")
+	gpuHandles := &handles{}
+	var cudartMgmtName string
 	var cudartMgmtPatterns []string
+	var nvcudaMgmtName string
+	var nvcudaMgmtPatterns []string
+	var oneapiMgmtName string
+	var oneapiMgmtPatterns []string
 
-	// Aligned with driver, we can't carry as payloads
-	nvcudaMgmtPatterns := NvcudaGlobs
-
-	if runtime.GOOS == "windows" {
-		localAppData := os.Getenv("LOCALAPPDATA")
-		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
-	}
 	tmpDir, _ := PayloadsDir()
-	if tmpDir != "" {
-		// TODO - add "payloads" for subprocess
-		cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", CudartMgmtName)}
-	}
-	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)
-
-	if len(NvmlGlobs) > 0 {
-		nvmlLibPaths := FindGPULibs(NvmlMgmtName, NvmlGlobs)
-		if len(nvmlLibPaths) > 0 {
-			nvml, libPath := LoadNVMLMgmt(nvmlLibPaths)
-			if nvml != nil {
-				slog.Debug("nvidia-ml loaded", "library", libPath)
-				cHandles.nvml = nvml
-				nvmlLibPath = libPath
-			}
+	switch runtime.GOOS {
+	case "windows":
+		cudartMgmtName = "cudart64_*.dll"
+		localAppData := os.Getenv("LOCALAPPDATA")
+		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", cudartMgmtName)}
+		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartWindowsGlobs...)
+		// Aligned with driver, we can't carry as payloads
+		nvcudaMgmtName = "nvcuda.dll"
+		nvcudaMgmtPatterns = NvcudaWindowsGlobs
+		oneapiMgmtName = "ze_intel_gpu64.dll"
+		oneapiMgmtPatterns = OneapiWindowsGlobs
+	case "linux":
+		cudartMgmtName = "libcudart.so*"
+		if tmpDir != "" {
+			// TODO - add "payloads" for subprocess
+			cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", cudartMgmtName)}
 		}
+		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartLinuxGlobs...)
+		// Aligned with driver, we can't carry as payloads
+		nvcudaMgmtName = "libcuda.so*"
+		nvcudaMgmtPatterns = NvcudaLinuxGlobs
+		oneapiMgmtName = "libze_intel_gpu.so"
+		oneapiMgmtPatterns = OneapiLinuxGlobs
+	default:
+		return gpuHandles
 	}
 
-	nvcudaLibPaths := FindGPULibs(NvcudaMgmtName, nvcudaMgmtPatterns)
+	slog.Debug("Detecting GPUs")
+	nvcudaLibPaths := FindGPULibs(nvcudaMgmtName, nvcudaMgmtPatterns)
 	if len(nvcudaLibPaths) > 0 {
 		deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
 		if nvcuda != nil {
 			slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
-			cHandles.nvcuda = nvcuda
-			cHandles.deviceCount = deviceCount
-			nvcudaLibPath = libPath
-			return cHandles
+			gpuHandles.nvcuda = nvcuda
+			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
 		}
 	}
 
-	cudartLibPaths := FindGPULibs(CudartMgmtName, cudartMgmtPatterns)
+	cudartLibPaths := FindGPULibs(cudartMgmtName, cudartMgmtPatterns)
 	if len(cudartLibPaths) > 0 {
 		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
 		if cudart != nil {
 			slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
-			cHandles.cudart = cudart
-			cHandles.deviceCount = deviceCount
-			cudartLibPath = libPath
-			return cHandles
+			gpuHandles.cudart = cudart
+			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
 		}
 	}
 
-	return cHandles
-}
-
-// Note: gpuMutex must already be held
-func initOneAPIHandles() *oneapiHandles {
-	oHandles := &oneapiHandles{}
-
-	// Short Circuit if we already know which library to use
-	if oneapiLibPath != "" {
-		oHandles.deviceCount, oHandles.oneapi, _ = LoadOneapiMgmt([]string{oneapiLibPath})
-		return oHandles
-	}
-
-	oneapiLibPaths := FindGPULibs(OneapiMgmtName, OneapiGlobs)
+	oneapiLibPaths := FindGPULibs(oneapiMgmtName, oneapiMgmtPatterns)
 	if len(oneapiLibPaths) > 0 {
-		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath = LoadOneapiMgmt(oneapiLibPaths)
+		deviceCount, oneapi, libPath := LoadOneapiMgmt(oneapiLibPaths)
+		if oneapi != nil {
+			slog.Debug("detected Intel GPUs", "library", libPath, "count", deviceCount)
+			gpuHandles.oneapi = oneapi
+			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
+		}
 	}
 
-	return oHandles
-}
-
-func GetCPUInfo() GpuInfoList {
-	gpuMutex.Lock()
-	if !bootstrapped {
-		gpuMutex.Unlock()
-		GetGPUInfo()
-	} else {
-		gpuMutex.Unlock()
-	}
-	return GpuInfoList{cpus[0].GpuInfo}
+	return gpuHandles
 }
 
 func GetGPUInfo() GpuInfoList {
@@ -178,258 +178,124 @@ func GetGPUInfo() GpuInfoList {
 	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
 	gpuMutex.Lock()
 	defer gpuMutex.Unlock()
-	needRefresh := true
-	var cHandles *cudaHandles
-	var oHandles *oneapiHandles
+
+	gpuHandles := initGPUHandles()
 	defer func() {
-		if cHandles != nil {
-			if cHandles.cudart != nil {
-				C.cudart_release(*cHandles.cudart)
-			}
-			if cHandles.nvcuda != nil {
-				C.nvcuda_release(*cHandles.nvcuda)
-			}
-			if cHandles.nvml != nil {
-				C.nvml_release(*cHandles.nvml)
-			}
+		if gpuHandles.cudart != nil {
+			C.cudart_release(*gpuHandles.cudart)
 		}
-		if oHandles != nil {
-			if oHandles.oneapi != nil {
-				// TODO - is this needed?
-				C.oneapi_release(*oHandles.oneapi)
-			}
+		if gpuHandles.nvcuda != nil {
+			C.nvcuda_release(*gpuHandles.nvcuda)
 		}
 	}()
 
-	if !bootstrapped {
-		slog.Info("looking for compatible GPUs")
-		needRefresh = false
-		cpuCapability = GetCPUCapability()
-		var memInfo C.mem_info_t
-
-		mem, err := GetCPUMem()
-		if err != nil {
-			slog.Warn("error looking up system memory", "error", err)
-		}
-		cpus = []CPUInfo{CPUInfo{
-			GpuInfo: GpuInfo{
-				memInfo: mem,
-				Library: "cpu",
-				Variant: cpuCapability,
-				ID:      "0",
-			},
-		}}
-
-		// Fallback to CPU mode if we're lacking required vector extensions on x86
-		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
-			slog.Warn("CPU does not have minimum vector extensions, GPU inference disabled", "required", GPURunnerCPUCapability, "detected", cpuCapability)
-			bootstrapped = true
-			// No need to do any GPU discovery, since we can't run on them
-			return GpuInfoList{cpus[0].GpuInfo}
-		}
-
-		// On windows we bundle the nvidia library one level above the runner dir
-		depPath := ""
-		if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
-			depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir), "cuda")
-		}
-
-		// Load ALL libraries
-		cHandles = initCudaHandles()
-
-		// NVIDIA
-		for i := range cHandles.deviceCount {
-			if cHandles.cudart != nil || cHandles.nvcuda != nil {
-				gpuInfo := CudaGPUInfo{
-					GpuInfo: GpuInfo{
-						Library: "cuda",
-					},
-					index: i,
-				}
-				var driverMajor int
-				var driverMinor int
-				if cHandles.cudart != nil {
-					C.cudart_bootstrap(*cHandles.cudart, C.int(i), &memInfo)
-				} else {
-					C.nvcuda_bootstrap(*cHandles.nvcuda, C.int(i), &memInfo)
-					driverMajor = int(cHandles.nvcuda.driver_major)
-					driverMinor = int(cHandles.nvcuda.driver_minor)
-				}
-				if memInfo.err != nil {
-					slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
-					C.free(unsafe.Pointer(memInfo.err))
-					continue
-				}
-				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
-					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
-					continue
-				}
-				gpuInfo.TotalMemory = uint64(memInfo.total)
-				gpuInfo.FreeMemory = uint64(memInfo.free)
-				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-				gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
-				gpuInfo.MinimumMemory = cudaMinimumMemory
-				gpuInfo.DependencyPath = depPath
-				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-				gpuInfo.DriverMajor = driverMajor
-				gpuInfo.DriverMinor = driverMinor
-
-				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
-				cudaGPUs = append(cudaGPUs, gpuInfo)
-			}
-		}
-
-		// Intel
-		if envconfig.IntelGpu {
-			oHandles = initOneAPIHandles()
-			// On windows we bundle the oneapi library one level above the runner dir
-			depPath = ""
-			if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
-				depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir), "oneapi")
-			}
-
-			for d := range oHandles.oneapi.num_drivers {
-				if oHandles.oneapi == nil {
-					// shouldn't happen
-					slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
-					continue
-				}
-				devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
-				for i := range devCount {
-					gpuInfo := OneapiGPUInfo{
-						GpuInfo: GpuInfo{
-							Library: "oneapi",
-						},
-						driverIndex: int(d),
-						gpuIndex:    int(i),
-					}
-					// TODO - split bootstrapping from updating free memory
-					C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
-					// TODO - convert this to MinimumMemory based on testing...
-					var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
-					memInfo.free = C.uint64_t(totalFreeMem)
-					gpuInfo.TotalMemory = uint64(memInfo.total)
-					gpuInfo.FreeMemory = uint64(memInfo.free)
-					gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-					gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-					gpuInfo.DependencyPath = depPath
-					oneapiGPUs = append(oneapiGPUs, gpuInfo)
-				}
-			}
-		}
-
-		rocmGPUs = AMDGetGPUInfo()
-		bootstrapped = true
-		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
-			slog.Info("no compatible GPUs were discovered")
-		}
+	// All our GPU builds on x86 have AVX enabled, so fallback to CPU if we don't detect at least AVX
+	cpuVariant := GetCPUVariant()
+	if cpuVariant == "" && runtime.GOARCH == "amd64" {
+		slog.Warn("CPU does not have AVX or AVX2, disabling GPU support.")
 	}
 
-	// For detected GPUs, load library if not loaded
+	// On windows we bundle the nvidia library one level above the runner dir
+	depPath := ""
+	if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
+		depPath = filepath.Dir(envconfig.RunnersDir)
+	}
 
-	// Refresh free memory usage
-	if needRefresh {
-		mem, err := GetCPUMem()
-		if err != nil {
-			slog.Warn("error looking up system memory", "error", err)
-		} else {
-			slog.Debug("updating system memory data",
-				slog.Group(
-					"before",
-					"total", format.HumanBytes2(cpus[0].TotalMemory),
-					"free", format.HumanBytes2(cpus[0].FreeMemory),
-				),
-				slog.Group(
-					"now",
-					"total", format.HumanBytes2(mem.TotalMemory),
-					"free", format.HumanBytes2(mem.FreeMemory),
-				),
-			)
-			cpus[0].FreeMemory = mem.FreeMemory
-		}
+	var memInfo C.mem_info_t
+	resp := []GpuInfo{}
 
-		var memInfo C.mem_info_t
-		if cHandles == nil && len(cudaGPUs) > 0 {
-			cHandles = initCudaHandles()
+	// NVIDIA first
+	for i := 0; i < gpuHandles.deviceCount; i++ {
+		// TODO once we support CPU compilation variants of GPU libraries refine this...
+		if cpuVariant == "" && runtime.GOARCH == "amd64" {
+			continue
 		}
-		for i, gpu := range cudaGPUs {
-			if cHandles.nvml != nil {
-				C.nvml_get_free(*cHandles.nvml, C.int(gpu.index), &memInfo.free, &memInfo.total, &memInfo.used)
-			} else if cHandles.cudart != nil {
-				C.cudart_bootstrap(*cHandles.cudart, C.int(gpu.index), &memInfo)
-			} else if cHandles.nvcuda != nil {
-				C.nvcuda_get_free(*cHandles.nvcuda, C.int(gpu.index), &memInfo.free, &memInfo.total)
-				memInfo.used = memInfo.total - memInfo.free
+		if gpuHandles.cudart != nil || gpuHandles.nvcuda != nil {
+			gpuInfo := GpuInfo{
+				Library: "cuda",
+			}
+			var driverMajor int
+			var driverMinor int
+			if gpuHandles.cudart != nil {
+				C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
 			} else {
-				// shouldn't happen
-				slog.Warn("no valid cuda library loaded to refresh vram usage")
-				break
+				C.nvcuda_check_vram(*gpuHandles.nvcuda, C.int(i), &memInfo)
+				driverMajor = int(gpuHandles.nvcuda.driver_major)
+				driverMinor = int(gpuHandles.nvcuda.driver_minor)
 			}
 			if memInfo.err != nil {
-				slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+				slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
 				C.free(unsafe.Pointer(memInfo.err))
 				continue
 			}
-			if memInfo.free == 0 {
-				slog.Warn("error looking up nvidia GPU memory")
+			if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+				slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
 				continue
 			}
-			slog.Debug("updating cuda memory data",
-				"gpu", gpu.ID,
-				"name", gpu.Name,
-				slog.Group(
-					"before",
-					"total", format.HumanBytes2(gpu.TotalMemory),
-					"free", format.HumanBytes2(gpu.FreeMemory),
-				),
-				slog.Group(
-					"now",
-					"total", format.HumanBytes2(uint64(memInfo.total)),
-					"free", format.HumanBytes2(uint64(memInfo.free)),
-					"used", format.HumanBytes2(uint64(memInfo.used)),
-				),
-			)
-			cudaGPUs[i].FreeMemory = uint64(memInfo.free)
-		}
+			gpuInfo.TotalMemory = uint64(memInfo.total)
+			gpuInfo.FreeMemory = uint64(memInfo.free)
+			gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+			gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
+			gpuInfo.MinimumMemory = cudaMinimumMemory
+			gpuInfo.DependencyPath = depPath
+			gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+			gpuInfo.DriverMajor = int(driverMajor)
+			gpuInfo.DriverMinor = int(driverMinor)
 
-		if oHandles == nil && len(oneapiGPUs) > 0 {
-			oHandles = initOneAPIHandles()
+			// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
+			resp = append(resp, gpuInfo)
 		}
-		for i, gpu := range oneapiGPUs {
-			if oHandles.oneapi == nil {
-				// shouldn't happen
-				slog.Warn("nil oneapi handle with device count", "count", oHandles.deviceCount)
-				continue
+		if gpuHandles.oneapi != nil {
+			gpuInfo := GpuInfo{
+				Library: "oneapi",
 			}
-			C.oneapi_check_vram(*oHandles.oneapi, C.int(gpu.driverIndex), C.int(gpu.gpuIndex), &memInfo)
-			// TODO - convert this to MinimumMemory based on testing...
+			C.oneapi_check_vram(*gpuHandles.oneapi, &memInfo)
 			var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
 			memInfo.free = C.uint64_t(totalFreeMem)
-			oneapiGPUs[i].FreeMemory = uint64(memInfo.free)
-		}
-
-		err = RocmGPUInfoList(rocmGPUs).RefreshFreeMemory()
-		if err != nil {
-			slog.Debug("problem refreshing ROCm free memory", "error", err)
+			gpuInfo.TotalMemory = uint64(memInfo.total)
+			gpuInfo.FreeMemory = uint64(memInfo.free)
+			gpuInfo.ID = strconv.Itoa(i)
+			resp = append(resp, gpuInfo)
 		}
 	}
 
-	resp := []GpuInfo{}
-	for _, gpu := range cudaGPUs {
-		resp = append(resp, gpu.GpuInfo)
-	}
-	for _, gpu := range rocmGPUs {
-		resp = append(resp, gpu.GpuInfo)
-	}
-	for _, gpu := range oneapiGPUs {
-		resp = append(resp, gpu.GpuInfo)
-	}
+	// Then AMD
+	resp = append(resp, AMDGetGPUInfo()...)
+
 	if len(resp) == 0 {
-		resp = append(resp, cpus[0].GpuInfo)
+		C.cpu_check_ram(&memInfo)
+		if memInfo.err != nil {
+			slog.Info("error looking up CPU memory", "error", C.GoString(memInfo.err))
+			C.free(unsafe.Pointer(memInfo.err))
+			return resp
+		}
+		gpuInfo := GpuInfo{
+			Library: "cpu",
+			Variant: cpuVariant,
+		}
+		gpuInfo.TotalMemory = uint64(memInfo.total)
+		gpuInfo.FreeMemory = uint64(memInfo.free)
+		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+
+		resp = append(resp, gpuInfo)
 	}
+
 	return resp
 }
 
+func GetCPUMem() (memInfo, error) {
+	var ret memInfo
+	var info C.mem_info_t
+	C.cpu_check_ram(&info)
+	if info.err != nil {
+		defer C.free(unsafe.Pointer(info.err))
+		return ret, fmt.Errorf(C.GoString(info.err))
+	}
+	ret.FreeMemory = uint64(info.free)
+	ret.TotalMemory = uint64(info.total)
+	return ret, nil
+}
+
 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
@@ -460,7 +326,6 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 		// Nvidia PhysX known to return bogus results
 		if strings.Contains(pattern, "PhysX") {
 			slog.Debug("skipping PhysX cuda library path", "path", pattern)
-			continue
 		}
 		// Ignore glob discovery errors
 		matches, _ := filepath.Glob(pattern)
@@ -517,23 +382,7 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
-			// Decide what log level based on the type of error message to help users understand why
-			msg := C.GoString(resp.err)
-			switch resp.cudaErr {
-			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
-				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
-			case C.CUDA_ERROR_NO_DEVICE:
-				slog.Info("no nvidia devices detected", "library", libPath)
-			case C.CUDA_ERROR_UNKNOWN:
-				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
-				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
-			default:
-				if strings.Contains(msg, "wrong ELF class") {
-					slog.Debug("skipping 32bit library", "library", libPath)
-				} else {
-					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
-				}
-			}
+			slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
 			return int(resp.num_devices), &resp.ch, libPath
@@ -542,26 +391,8 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 	return 0, nil, ""
 }
 
-func LoadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string) {
-	var resp C.nvml_init_resp_t
-	resp.ch.verbose = getVerboseState()
-	for _, libPath := range nvmlLibPaths {
-		lib := C.CString(libPath)
-		defer C.free(unsafe.Pointer(lib))
-		C.nvml_init(lib, &resp)
-		if resp.err != nil {
-			slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
-			C.free(unsafe.Pointer(resp.err))
-		} else {
-			return &resp.ch, libPath
-		}
-	}
-	return nil, ""
-}
-
 func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
 	var resp C.oneapi_init_resp_t
-	num_devices := 0
 	resp.oh.verbose = getVerboseState()
 	for _, libPath := range oneapiLibPaths {
 		lib := C.CString(libPath)
@@ -571,10 +402,7 @@ func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
 			slog.Debug("Unable to load oneAPI management library", "library", libPath, "error", C.GoString(resp.err))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			for i := range resp.oh.num_drivers {
-				num_devices += int(C.oneapi_get_device_count(resp.oh, C.int(i)))
-			}
-			return num_devices, &resp.oh, libPath
+			return int(resp.num_devices), &resp.oh, libPath
 		}
 	}
 	return 0, nil, ""

gpu/gpu_darwin.go

@@ -24,7 +24,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
-				Variant: GetCPUCapability(),
+				Variant: GetCPUVariant(),
 				memInfo: mem,
 			},
 		}
@@ -42,21 +42,10 @@ func GetGPUInfo() GpuInfoList {
 	return []GpuInfo{info}
 }
 
-func GetCPUInfo() GpuInfoList {
-	mem, _ := GetCPUMem()
-	return []GpuInfo{
-		{
-			Library: "cpu",
-			Variant: GetCPUCapability(),
-			memInfo: mem,
-		},
-	}
-}
-
 func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
-		FreeMemory:  uint64(C.getFreeMemory()),
+		FreeMemory:  0,
 	}, nil
 }
 

gpu/gpu_info.h

@@ -47,7 +47,6 @@ typedef struct mem_info {
   char gpu_name[GPU_NAME_LEN];
   uint64_t total;
   uint64_t free;
-  uint64_t used;
 
   // Compute Capability
   int major; 
@@ -63,7 +62,6 @@ void cpu_check_ram(mem_info_t *resp);
 
 #include "gpu_info_cudart.h"
 #include "gpu_info_nvcuda.h"
-#include "gpu_info_nvml.h"
 #include "gpu_info_oneapi.h"
 
 #endif  // __GPU_INFO_H__

gpu/gpu_info_cpu.c

@@ -0,0 +1,45 @@
+#include "gpu_info.h"
+// Fallbacks for CPU mode
+
+#ifdef _WIN32
+#include <sysinfoapi.h>
+void cpu_check_ram(mem_info_t *resp) {
+  resp->err = NULL;
+  MEMORYSTATUSEX info;
+  info.dwLength = sizeof(info);
+  if (GlobalMemoryStatusEx(&info) != 0) {
+    resp->total = info.ullTotalPhys;
+    resp->free = info.ullAvailPhys;
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
+  } else {
+    resp->err = LOAD_ERR();
+  }
+  return;
+}
+
+#elif __linux__
+#include <errno.h>
+#include <string.h>
+#include <sys/sysinfo.h>
+void cpu_check_ram(mem_info_t *resp) {
+  struct sysinfo info;
+  resp->err = NULL;
+  if (sysinfo(&info) != 0) {
+    resp->err = strdup(strerror(errno));
+  } else {
+    resp->total = info.totalram * info.mem_unit;
+    resp->free = info.freeram * info.mem_unit;
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
+  }
+  return;
+}
+
+#elif __APPLE__
+// TODO consider an Apple implementation that does something useful
+// mem_info_t cpu_check_ram() {
+//   mem_info_t resp = {0, 0, NULL};
+//   return resp;
+// }
+#else
+#error "Unsupported platform"
+#endif

gpu/gpu_info_cudart.c

@@ -40,7 +40,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
 
   for (i = 0; l[i].s != NULL; i++) {
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*(l[i].p)) {
+    if (!l[i].p) {
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);
       UNLOAD_LIBRARY(resp->ch.handle);
@@ -94,7 +94,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
 }
 
 
-void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
+void cudart_check_vram(cudart_handle_t h, int i, mem_info_t *resp) {
   resp->err = NULL;
   cudartMemory_t memInfo = {0,0,0};
   cudartReturn_t ret;
@@ -166,11 +166,9 @@ void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
 
   resp->total = memInfo.total;
   resp->free = memInfo.free;
-  resp->used = memInfo.used;
 
   LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
   LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
-  LOG(h.verbose, "[%s] CUDA usedMem %lu\n", resp->gpu_id, resp->used);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 }
 

gpu/gpu_info_cudart.h

@@ -140,8 +140,7 @@ typedef struct cudart_init_resp {
 } cudart_init_resp_t;
 
 void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
-void cudart_bootstrap(cudart_handle_t ch, int device_id, mem_info_t *resp);
-// TODO - if we keep this library longer term, add cudart_get_free
+void cudart_check_vram(cudart_handle_t ch, int device_id, mem_info_t *resp);
 void cudart_release(cudart_handle_t ch);
 
 #endif  // __GPU_INFO_CUDART_H__

gpu/gpu_info_darwin.h

@@ -2,4 +2,3 @@
 #include <stdint.h>
 uint64_t getRecommendedMaxVRAM();
 uint64_t getPhysicalMemory();
-uint64_t getFreeMemory();

gpu/gpu_info_darwin.m

@@ -1,5 +1,4 @@
-#import <Foundation/Foundation.h>
-#import <mach/mach.h>
+// go:build darwin
 #include "gpu_info_darwin.h"
 
 uint64_t getRecommendedMaxVRAM() {
@@ -9,27 +8,6 @@ uint64_t getRecommendedMaxVRAM() {
   return result;
 }
 
-// getPhysicalMemory returns the total physical memory in bytes
 uint64_t getPhysicalMemory() {
-  return [NSProcessInfo processInfo].physicalMemory;
-}
-
-// getFreeMemory returns the total free memory in bytes, including inactive
-// memory that can be reclaimed by the system.
-uint64_t getFreeMemory() {
-  mach_port_t host_port = mach_host_self();
-  mach_msg_type_number_t host_size = sizeof(vm_statistics64_data_t) / sizeof(integer_t);
-  vm_size_t pagesize;
-  vm_statistics64_data_t vm_stat;
-
-  host_page_size(host_port, &pagesize);
-  if (host_statistics64(host_port, HOST_VM_INFO64, (host_info64_t)&vm_stat, &host_size) != KERN_SUCCESS) {
-    return 0;
-  }
-
-  uint64_t free_memory = (uint64_t)vm_stat.free_count * pagesize;
-  free_memory += (uint64_t)vm_stat.speculative_count * pagesize;
-  free_memory += (uint64_t)vm_stat.inactive_count * pagesize;
-
-  return free_memory;
+  return [[NSProcessInfo processInfo] physicalMemory];
 }

gpu/gpu_info_nvcuda.c

@@ -7,7 +7,6 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
   CUresult ret;
   resp->err = NULL;
   resp->num_devices = 0;
-  resp->cudaErr = CUDA_SUCCESS;
   const int buflen = 256;
   char buf[buflen + 1];
   int i;
@@ -39,13 +38,12 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
             nvcuda_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
-    resp->cudaErr = -1;
     return;
   }
 
   for (i = 0; l[i].s != NULL; i++) {
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*(l[i].p)) {
+    if (!*l[i].p) {
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);
       UNLOAD_LIBRARY(resp->ch.handle);
@@ -54,7 +52,6 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
               msg);
       free(msg);
       resp->err = strdup(buf);
-      resp->cudaErr = -1;
       return;
     }
   }
@@ -64,9 +61,12 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
+    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
+      return;
+    }
+    snprintf(buf, buflen, "nvcuda init failure: %d", ret);
     resp->err = strdup(buf);
-    resp->cudaErr = ret;
     return;
   }
 
@@ -91,13 +91,12 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     resp->ch.handle = NULL;
     snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
-    resp->cudaErr = ret;
     return;
   }
 }
 
 const int buflen = 256;
-void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
+void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
   resp->err = NULL;
   nvcudaMemory_t memInfo = {0,0};
   CUresult ret;
@@ -107,13 +106,13 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 
   if (h.handle == NULL) {
-    resp->err = strdup("cuda driver library handle isn't initialized");
+    resp->err = strdup("nvcuda handle isn't initialized");
     return;
   }
 
   ret = (*h.cuDeviceGet)(&device, i);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "cuda driver library device failed to initialize");
+    snprintf(buf, buflen, "nvcuda device failed to initialize");
     resp->err = strdup(buf);
     return;
   }
@@ -169,14 +168,14 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
+    snprintf(buf, buflen, "nvcuda failed to get primary device context %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
   ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
+    snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
     resp->err = strdup(buf);
     // Best effort on failure...
     (*h.cuCtxDestroy)(ctx);
@@ -194,47 +193,12 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
 
   ret = (*h.cuCtxDestroy)(ctx);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library failed to release device context %d", ret);
-  }
-}
-
-void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total) {
-  CUresult ret;
-  CUcontext ctx = NULL;
-  CUdevice device = -1;
-  *free = 0;
-  *total = 0;
-
-  ret = (*h.cuDeviceGet)(&device, i);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library device failed to initialize");
-    return;
-  }
-
-
-  // To get memory we have to set (and release) a context
-  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library failed to get device context %d", ret);
-    return;
-  }
-
-  ret = (*h.cuMemGetInfo_v2)(free, total);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library device memory info lookup failure %d", ret);
-    // Best effort on failure...
-    (*h.cuCtxDestroy)(ctx);
-    return;
-  }
-
-  ret = (*h.cuCtxDestroy)(ctx);
-  if (ret != CUDA_SUCCESS) {
-    LOG(1, "cuda driver library failed to release device context %d", ret);
+    LOG(1, "nvcuda failed to release primary device context %d", ret);
   }
 }
 
 void nvcuda_release(nvcuda_handle_t h) {
-  LOG(h.verbose, "releasing cuda driver library\n");
+  LOG(h.verbose, "releasing nvcuda library\n");
   UNLOAD_LIBRARY(h.handle);
   // TODO and other context release logic?
   h.handle = NULL;

gpu/gpu_info_nvcuda.h

@@ -7,12 +7,9 @@
 typedef enum cudaError_enum {
   CUDA_SUCCESS = 0,
   CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_OUT_OF_MEMORY = 2,
+  CUDA_ERROR_MEMORY_ALLOCATION = 2,
   CUDA_ERROR_NOT_INITIALIZED = 3,
   CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
-  CUDA_ERROR_NO_DEVICE = 100,
-  CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
-  CUDA_ERROR_UNKNOWN = 999,
   // Other values omitted for now...
 } CUresult;
 
@@ -67,12 +64,10 @@ typedef struct nvcuda_init_resp {
   char *err;  // If err is non-null handle is invalid
   nvcuda_handle_t ch;
   int num_devices;
-  CUresult cudaErr;
 } nvcuda_init_resp_t;
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
-void nvcuda_bootstrap(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
-void nvcuda_get_free(nvcuda_handle_t ch,  int device_id, uint64_t *free, uint64_t *total);
+void nvcuda_check_vram(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
 void nvcuda_release(nvcuda_handle_t ch);
 
 #endif  // __GPU_INFO_NVCUDA_H__

gpu/gpu_info_nvml.c

@@ -1,104 +0,0 @@
-#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
-
-#include <string.h>
-
-#include "gpu_info_nvml.h"
-
-void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
-  nvmlReturn_t ret;
-  resp->err = NULL;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
-      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
-      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.nvmlDeviceGetHandleByIndex},
-      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
-      {NULL, NULL},
-  };
-
-  resp->ch.handle = LOAD_LIBRARY(nvml_lib_path, RTLD_LAZY);
-  if (!resp->ch.handle) {
-    char *msg = LOAD_ERR();
-    LOG(resp->ch.verbose, "library %s load err: %s\n", nvml_lib_path, msg);
-    snprintf(buf, buflen,
-             "Unable to load %s library to query for Nvidia GPUs: %s",
-             nvml_lib_path, msg);
-    free(msg);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  // TODO once we've squashed the remaining corner cases remove this log
-  // LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", nvml_lib_path);
-  
-  for (i = 0; l[i].s != NULL; i++) {
-    // TODO once we've squashed the remaining corner cases remove this log
-    // LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
-
-    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*(l[i].p)) {
-      resp->ch.handle = NULL;
-      char *msg = LOAD_ERR();
-      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
-      UNLOAD_LIBRARY(resp->ch.handle);
-      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
-               msg);
-      free(msg);
-      resp->err = strdup(buf);
-      return;
-    }
-  }
-
-  ret = (*resp->ch.nvmlInit_v2)();
-  if (ret != NVML_SUCCESS) {
-    LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-}
-
-
-void nvml_get_free(nvml_handle_t h, int device_id, uint64_t *free, uint64_t *total, uint64_t *used) {
-    nvmlDevice_t device;
-    nvmlMemory_t memInfo = {0};
-    nvmlReturn_t ret;
-    ret = (*h.nvmlDeviceGetHandleByIndex)(device_id, &device);
-    if (ret != NVML_SUCCESS) {
-        LOG(1, "unable to get device handle %d: %d", device_id, ret);
-        *free = 0;
-        return;
-    }
-
-    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
-    if (ret != NVML_SUCCESS) {
-        LOG(1, "device memory info lookup failure %d: %d", device_id, ret);
-        *free = 0;
-        return;
-    }
-    *free = memInfo.free;
-    *total = memInfo.total;
-    *used = memInfo.used;
-}
-
-
-void nvml_release(nvml_handle_t h) {
-  LOG(h.verbose, "releasing nvml library\n");
-  nvmlReturn_t ret;
-  ret = (*h.nvmlShutdown)();
-  if (ret != NVML_SUCCESS) {
-    LOG(1, "error during nvmlShutdown %d", ret);
-  }
-  UNLOAD_LIBRARY(h.handle);
-  h.handle = NULL;
-}
-
-#endif  // __APPLE__

gpu/gpu_info_nvml.h

@@ -1,48 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_NVML_H__
-#define __GPU_INFO_NVML_H__
-#include "gpu_info.h"
-
-// Just enough typedef's to dlopen/dlsym for memory information
-typedef enum nvmlReturn_enum {
-  NVML_SUCCESS = 0,
-  // Other values omitted for now...
-} nvmlReturn_t;
-typedef void *nvmlDevice_t;  // Opaque is sufficient
-typedef struct nvmlMemory_st {
-  unsigned long long total;
-  unsigned long long free;
-  unsigned long long used;
-} nvmlMemory_t;
-
-typedef enum nvmlBrandType_enum
-{
-    NVML_BRAND_UNKNOWN          = 0,
-} nvmlBrandType_t;
-
-typedef struct nvml_handle {
-  void *handle;
-  uint16_t verbose;
-  nvmlReturn_t (*nvmlInit_v2)(void);
-  nvmlReturn_t (*nvmlShutdown)(void);
-  nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(unsigned int, nvmlDevice_t *);
-  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
-} nvml_handle_t;
-
-typedef struct nvml_init_resp {
-  char *err;  // If err is non-null handle is invalid
-  nvml_handle_t ch;
-} nvml_init_resp_t;
-
-typedef struct nvml_compute_capability {
-  char *err;
-  int major;
-  int minor;
-} nvml_compute_capability_t;
-
-void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp);
-void nvml_get_free(nvml_handle_t ch,  int device_id, uint64_t *free, uint64_t *total, uint64_t *used);
-void nvml_release(nvml_handle_t ch);
-
-#endif  // __GPU_INFO_NVML_H__
-#endif  // __APPLE__

gpu/gpu_info_oneapi.c

@@ -4,17 +4,15 @@
 
 #include <string.h>
 
-void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
+void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp)
+{
   ze_result_t ret;
   resp->err = NULL;
-  resp->oh.devices = NULL;
-  resp->oh.num_devices = NULL;
-  resp->oh.drivers = NULL;
-  resp->oh.num_drivers = 0;
   const int buflen = 256;
   char buf[buflen + 1];
-  int i, d;
-  struct lookup {
+  int i;
+  struct lookup
+  {
     char *s;
     void **p;
   } l[] = {
@@ -30,7 +28,8 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
   };
 
   resp->oh.handle = LOAD_LIBRARY(oneapi_lib_path, RTLD_LAZY);
-  if (!resp->oh.handle) {
+  if (!resp->oh.handle)
+  {
     char *msg = LOAD_ERR();
     snprintf(buf, buflen,
              "Unable to load %s library to query for Intel GPUs: %s\n",
@@ -45,12 +44,14 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
       "wiring Level-Zero management library functions in %s\n",
       oneapi_lib_path);
 
-  for (i = 0; l[i].s != NULL; i++) {
+  for (i = 0; l[i].s != NULL; i++)
+  {
     // TODO once we've squashed the remaining corner cases remove this log
     LOG(resp->oh.verbose, "dlsym: %s\n", l[i].s);
 
     *l[i].p = LOAD_SYMBOL(resp->oh.handle, l[i].s);
-    if (!*(l[i].p)) {
+    if (!l[i].p)
+    {
       resp->oh.handle = NULL;
       char *msg = LOAD_ERR();
       LOG(resp->oh.verbose, "dlerr: %s\n", msg);
@@ -62,70 +63,23 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
     }
   }
 
-  LOG(resp->oh.verbose, "calling zesInit\n");
-
   ret = (*resp->oh.zesInit)(0);
-  if (ret != ZE_RESULT_SUCCESS) {
-    LOG(resp->oh.verbose, "zesInit err: %x\n", ret);
-    snprintf(buf, buflen, "oneapi vram init failure: %x", ret);
+  if (ret != ZE_RESULT_SUCCESS)
+  {
+    LOG(resp->oh.verbose, "zesInit err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->oh.handle);
+    resp->oh.handle = NULL;
+    snprintf(buf, buflen, "oneapi vram init failure: %d", ret);
     resp->err = strdup(buf);
-    oneapi_release(resp->oh);
-    return;
   }
 
-  LOG(resp->oh.verbose, "calling zesDriverGet\n");
-  ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, NULL);
-  if (ret != ZE_RESULT_SUCCESS) {
-    LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
-    snprintf(buf, buflen, "unable to get driver count: %x", ret);
-    resp->err = strdup(buf);
-    oneapi_release(resp->oh);
-    return;
-  }
-  LOG(resp->oh.verbose, "oneapi driver count: %d\n", resp->oh.num_drivers);
-  resp->oh.drivers = malloc(resp->oh.num_drivers * sizeof(zes_driver_handle_t));
-  resp->oh.num_devices = malloc(resp->oh.num_drivers * sizeof(uint32_t));
-  memset(&resp->oh.num_devices[0], 0, resp->oh.num_drivers * sizeof(uint32_t));
-  resp->oh.devices =
-      malloc(resp->oh.num_drivers * sizeof(zes_device_handle_t *));
-  ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, &resp->oh.drivers[0]);
-  if (ret != ZE_RESULT_SUCCESS) {
-    LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
-    snprintf(buf, buflen, "unable to get driver count: %x", ret);
-    resp->err = strdup(buf);
-    oneapi_release(resp->oh);
-    return;
-  }
-
-  for (d = 0; d < resp->oh.num_drivers; d++) {
-    LOG(resp->oh.verbose, "calling zesDeviceGet count %d: %p\n", d, resp->oh.drivers[d]);
-    ret = (*resp->oh.zesDeviceGet)(resp->oh.drivers[d],
-                                   &resp->oh.num_devices[d], NULL);
-    if (ret != ZE_RESULT_SUCCESS) {
-      LOG(resp->oh.verbose, "zesDeviceGet err: %x\n", ret);
-      snprintf(buf, buflen, "unable to get device count: %x", ret);
-      resp->err = strdup(buf);
-      oneapi_release(resp->oh);
-      return;
-    }
-    resp->oh.devices[d] =
-        malloc(resp->oh.num_devices[d] * sizeof(zes_device_handle_t));
-    ret = (*resp->oh.zesDeviceGet)(
-        resp->oh.drivers[d], &resp->oh.num_devices[d], resp->oh.devices[d]);
-    if (ret != ZE_RESULT_SUCCESS) {
-      LOG(resp->oh.verbose, "zesDeviceGet err: %x\n", ret);
-      snprintf(buf, buflen, "unable to get device count: %x", ret);
-      resp->err = strdup(buf);
-      oneapi_release(resp->oh);
-      return;
-    }
-  }
+  (*resp->oh.zesDriverGet)(&resp->num_devices, NULL);
 
   return;
 }
 
-void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
-                       mem_info_t *resp) {
+void oneapi_check_vram(oneapi_handle_t h, mem_info_t *resp)
+{
   ze_result_t ret;
   resp->err = NULL;
   uint64_t totalMem = 0;
@@ -134,126 +88,127 @@ void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
   char buf[buflen + 1];
   int i, d, m;
 
-  if (h.handle == NULL) {
+  if (h.handle == NULL)
+  {
     resp->err = strdup("Level-Zero handle not initialized");
     return;
   }
 
-  if (driver > h.num_drivers || device > h.num_devices[driver]) {
-    resp->err = strdup("driver of device index out of bounds");
+  uint32_t driversCount = 0;
+  ret = (*h.zesDriverGet)(&driversCount, NULL);
+  if (ret != ZE_RESULT_SUCCESS)
+  {
+    snprintf(buf, buflen, "unable to get driver count: %d", ret);
+    resp->err = strdup(buf);
     return;
   }
+  LOG(h.verbose, "discovered %d Level-Zero drivers\n", driversCount);
+
+  zes_driver_handle_t *allDrivers =
+      malloc(driversCount * sizeof(zes_driver_handle_t));
+  (*h.zesDriverGet)(&driversCount, allDrivers);
 
   resp->total = 0;
   resp->free = 0;
 
-  zes_device_ext_properties_t ext_props;
-  ext_props.stype = ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES;
-  ext_props.pNext = NULL;
-
-  zes_device_properties_t props;
-  props.stype = ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES;
-  props.pNext = &ext_props;
-
-  ret = (*h.zesDeviceGetProperties)(h.devices[driver][device], &props);
-  if (ret != ZE_RESULT_SUCCESS) {
-    snprintf(buf, buflen, "unable to get device properties: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  snprintf(&resp->gpu_name[0], GPU_NAME_LEN, "%s", props.modelName);
-
-  // TODO this needs to map to ONEAPI_DEVICE_SELECTOR syntax
-  // (this is probably wrong...)
-  // TODO - the driver isn't included - what if there are multiple drivers?
-  snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", device);
-
-  if (h.verbose) {
-    // When in verbose mode, report more information about
-    // the card we discover.
-    LOG(h.verbose, "[%d:%d] oneAPI device name: %s\n", driver, device,
-        props.modelName);
-    LOG(h.verbose, "[%d:%d] oneAPI brand: %s\n", driver, device,
-        props.brandName);
-    LOG(h.verbose, "[%d:%d] oneAPI vendor: %s\n", driver, device,
-        props.vendorName);
-    LOG(h.verbose, "[%d:%d] oneAPI S/N: %s\n", driver, device,
-        props.serialNumber);
-    LOG(h.verbose, "[%d:%d] oneAPI board number: %s\n", driver, device,
-        props.boardNumber);
-  }
-
-  // TODO
-  // Compute Capability equivalent in resp->major, resp->minor, resp->patch
-
-  uint32_t memCount = 0;
-  ret = (*h.zesDeviceEnumMemoryModules)(h.devices[driver][device], &memCount,
-                                        NULL);
-  if (ret != ZE_RESULT_SUCCESS) {
-    snprintf(buf, buflen, "unable to enumerate Level-Zero memory modules: %x",
-             ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  LOG(h.verbose, "discovered %d Level-Zero memory modules\n", memCount);
-
-  zes_mem_handle_t *mems = malloc(memCount * sizeof(zes_mem_handle_t));
-  (*h.zesDeviceEnumMemoryModules)(h.devices[driver][device], &memCount, mems);
-
-  for (m = 0; m < memCount; m++) {
-    zes_mem_state_t state;
-    state.stype = ZES_STRUCTURE_TYPE_MEM_STATE;
-    state.pNext = NULL;
-    ret = (*h.zesMemoryGetState)(mems[m], &state);
-    if (ret != ZE_RESULT_SUCCESS) {
-      snprintf(buf, buflen, "unable to get memory state: %x", ret);
+  for (d = 0; d < driversCount; d++)
+  {
+    uint32_t deviceCount = 0;
+    ret = (*h.zesDeviceGet)(allDrivers[d], &deviceCount, NULL);
+    if (ret != ZE_RESULT_SUCCESS)
+    {
+      snprintf(buf, buflen, "unable to get device count: %d", ret);
       resp->err = strdup(buf);
-      free(mems);
+      free(allDrivers);
       return;
     }
 
-    resp->total += state.size;
-    resp->free += state.free;
-  }
+    LOG(h.verbose, "discovered %d Level-Zero devices\n", deviceCount);
 
-  free(mems);
-}
+    zes_device_handle_t *devices =
+        malloc(deviceCount * sizeof(zes_device_handle_t));
+    (*h.zesDeviceGet)(allDrivers[d], &deviceCount, devices);
 
-void oneapi_release(oneapi_handle_t h) {
-  int d;
-  LOG(h.verbose, "releasing oneapi library\n");
-  for (d = 0; d < h.num_drivers; d++) {
-    if (h.devices != NULL && h.devices[d] != NULL) {
-      free(h.devices[d]);
+    for (i = 0; i < deviceCount; i++)
+    {
+      zes_device_ext_properties_t ext_props;
+      ext_props.stype = ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES;
+      ext_props.pNext = NULL;
+
+      zes_device_properties_t props;
+      props.stype = ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+      props.pNext = &ext_props;
+
+      ret = (*h.zesDeviceGetProperties)(devices[i], &props);
+      if (ret != ZE_RESULT_SUCCESS)
+      {
+        snprintf(buf, buflen, "unable to get device properties: %d", ret);
+        resp->err = strdup(buf);
+        free(allDrivers);
+        free(devices);
+        return;
+      }
+
+      if (h.verbose)
+      {
+        // When in verbose mode, report more information about
+        // the card we discover.
+        LOG(h.verbose, "[%d] oneAPI device name: %s\n", i,
+            props.modelName);
+        LOG(h.verbose, "[%d] oneAPI brand: %s\n", i,
+            props.brandName);
+        LOG(h.verbose, "[%d] oneAPI vendor: %s\n", i,
+            props.vendorName);
+        LOG(h.verbose, "[%d] oneAPI S/N: %s\n", i,
+            props.serialNumber);
+        LOG(h.verbose, "[%d] oneAPI board number: %s\n", i,
+            props.boardNumber);
+      }
+
+      uint32_t memCount = 0;
+      ret = (*h.zesDeviceEnumMemoryModules)(devices[i], &memCount, NULL);
+      if (ret != ZE_RESULT_SUCCESS)
+      {
+        snprintf(buf, buflen,
+                 "unable to enumerate Level-Zero memory modules: %d", ret);
+        resp->err = strdup(buf);
+        free(allDrivers);
+        free(devices);
+        return;
+      }
+
+      LOG(h.verbose, "discovered %d Level-Zero memory modules\n", memCount);
+
+      zes_mem_handle_t *mems = malloc(memCount * sizeof(zes_mem_handle_t));
+      (*h.zesDeviceEnumMemoryModules)(devices[i], &memCount, mems);
+
+      for (m = 0; m < memCount; m++)
+      {
+        zes_mem_state_t state;
+        state.stype = ZES_STRUCTURE_TYPE_MEM_STATE;
+        state.pNext = NULL;
+        ret = (*h.zesMemoryGetState)(mems[m], &state);
+        if (ret != ZE_RESULT_SUCCESS)
+        {
+          snprintf(buf, buflen, "unable to get memory state: %d", ret);
+          resp->err = strdup(buf);
+          free(allDrivers);
+          free(devices);
+          free(mems);
+          return;
+        }
+
+        resp->total += state.size;
+        resp->free += state.free;
+      }
+
+      free(mems);
     }
-  }
-  if (h.devices != NULL) {
-    free(h.devices);
-    h.devices = NULL;
-  }
-  if (h.num_devices != NULL) {
-    free(h.num_devices);
-    h.num_devices = NULL;
-  }
-  if (h.drivers != NULL) {
-    free(h.drivers);
-    h.drivers = NULL;
-  }
-  h.num_drivers = 0;
-  UNLOAD_LIBRARY(h.handle);
-  h.handle = NULL;
-}
 
-int oneapi_get_device_count(oneapi_handle_t h, int driver) {
-  if (h.handle == NULL || h.num_devices == NULL) {
-    return 0;
+    free(devices);
   }
-  if (driver > h.num_drivers) {
-    return 0;
-  }
-  return (int)h.num_devices[driver];
+
+  free(allDrivers);
 }
 
 #endif // __APPLE__

gpu/gpu_info_oneapi.h

@@ -9,7 +9,8 @@
 #define ZE_BIT(_i) (1 << _i)
 
 // Just enough typedef's to dlopen/dlsym for memory information
-typedef enum ze_result_t {
+typedef enum ze_result_t
+{
   ZE_RESULT_SUCCESS = 0,
   // Other values omitted for now...
 } ze_result_t;
@@ -19,11 +20,13 @@ typedef struct _zes_driver_handle_t *zes_driver_handle_t;
 typedef struct _zes_device_handle_t *zes_device_handle_t;
 typedef struct _zes_mem_handle_t *zes_mem_handle_t;
 
-typedef enum _ze_structure_type_t {
+typedef enum _ze_structure_type_t
+{
   ZE_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
 } ze_structure_type_t;
 
-typedef enum _zes_structure_type_t {
+typedef enum _zes_structure_type_t
+{
   ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES = 0x1,
   ZES_STRUCTURE_TYPE_MEM_PROPERTIES = 0xb,
   ZES_STRUCTURE_TYPE_MEM_STATE = 0x1e,
@@ -31,29 +34,35 @@ typedef enum _zes_structure_type_t {
   ZES_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
 } zes_structure_type_t;
 
-typedef enum _zes_mem_type_t {
+typedef enum _zes_mem_type_t
+{
   ZES_MEM_TYPE_FORCE_UINT32 = 0x7fffffff
 } zes_mem_type_t;
 
-typedef enum _zes_mem_loc_t {
+typedef enum _zes_mem_loc_t
+{
   ZES_MEM_LOC_SYSTEM = 0,
   ZES_MEM_LOC_DEVICE = 1,
   ZES_MEM_LOC_FORCE_UINT32 = 0x7fffffff
 } zes_mem_loc_t;
 
-typedef enum _zes_mem_health_t {
+typedef enum _zes_mem_health_t
+{
   ZES_MEM_HEALTH_FORCE_UINT32 = 0x7fffffff
 } zes_mem_health_t;
 
-typedef struct _ze_device_uuid_t {
+typedef struct _ze_device_uuid_t
+{
   uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
 } ze_device_uuid_t;
 
-typedef struct _zes_uuid_t {
+typedef struct _zes_uuid_t
+{
   uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
 } zes_uuid_t;
 
-typedef enum _ze_device_type_t {
+typedef enum _ze_device_type_t
+{
   ZE_DEVICE_TYPE_GPU = 1,
   ZE_DEVICE_TYPE_CPU = 2,
   ZE_DEVICE_TYPE_FPGA = 3,
@@ -62,7 +71,8 @@ typedef enum _ze_device_type_t {
   ZE_DEVICE_TYPE_FORCE_UINT32 = 0x7fffffff
 } ze_device_type_t;
 
-typedef enum _zes_device_type_t {
+typedef enum _zes_device_type_t
+{
   ZES_DEVICE_TYPE_GPU = 1,
   ZES_DEVICE_TYPE_CPU = 2,
   ZES_DEVICE_TYPE_FPGA = 3,
@@ -72,7 +82,8 @@ typedef enum _zes_device_type_t {
 } zes_device_type_t;
 
 typedef uint32_t ze_device_property_flags_t;
-typedef enum _ze_device_property_flag_t {
+typedef enum _ze_device_property_flag_t
+{
   ZE_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
   ZE_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
   ZE_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
@@ -81,7 +92,8 @@ typedef enum _ze_device_property_flag_t {
 } ze_device_property_flag_t;
 
 typedef uint32_t zes_device_property_flags_t;
-typedef enum _zes_device_property_flag_t {
+typedef enum _zes_device_property_flag_t
+{
   ZES_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
   ZES_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
   ZES_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
@@ -89,7 +101,8 @@ typedef enum _zes_device_property_flag_t {
   ZES_DEVICE_PROPERTY_FLAG_FORCE_UINT32 = 0x7fffffff
 } zes_device_property_flag_t;
 
-typedef struct _ze_device_properties_t {
+typedef struct _ze_device_properties_t
+{
   ze_structure_type_t stype;
   void *pNext;
   ze_device_type_t type;
@@ -113,7 +126,8 @@ typedef struct _ze_device_properties_t {
   char name[ZE_MAX_DEVICE_NAME];
 } ze_device_properties_t;
 
-typedef struct _zes_device_properties_t {
+typedef struct _zes_device_properties_t
+{
   zes_structure_type_t stype;
   void *pNext;
   ze_device_properties_t core;
@@ -126,7 +140,8 @@ typedef struct _zes_device_properties_t {
   char driverVersion[ZES_STRING_PROPERTY_SIZE];
 } zes_device_properties_t;
 
-typedef struct _zes_device_ext_properties_t {
+typedef struct _zes_device_ext_properties_t
+{
   zes_structure_type_t stype;
   void *pNext;
   zes_uuid_t uuid;
@@ -134,7 +149,8 @@ typedef struct _zes_device_ext_properties_t {
   zes_device_property_flags_t flags;
 } zes_device_ext_properties_t;
 
-typedef struct _zes_mem_properties_t {
+typedef struct _zes_mem_properties_t
+{
   zes_structure_type_t stype;
   void *pNext;
   zes_mem_type_t type;
@@ -146,7 +162,8 @@ typedef struct _zes_mem_properties_t {
   int32_t numChannels;
 } zes_mem_properties_t;
 
-typedef struct _zes_mem_state_t {
+typedef struct _zes_mem_state_t
+{
   zes_structure_type_t stype;
   const void *pNext;
   zes_mem_health_t health;
@@ -154,19 +171,10 @@ typedef struct _zes_mem_state_t {
   uint64_t size;
 } zes_mem_state_t;
 
-typedef struct oneapi_handle {
+typedef struct oneapi_handle
+{
   void *handle;
   uint16_t verbose;
-
-  uint32_t num_drivers;
-  zes_driver_handle_t *drivers;
-  uint32_t *num_devices;
-  zes_device_handle_t **devices;
-
-  // TODO Driver major, minor information
-  // int driver_major;
-  // int driver_minor;
-
   ze_result_t (*zesInit)(int);
   ze_result_t (*zesDriverGet)(uint32_t *pCount, zes_driver_handle_t *phDrivers);
   ze_result_t (*zesDeviceGet)(zes_driver_handle_t hDriver, uint32_t *pCount,
@@ -183,21 +191,21 @@ typedef struct oneapi_handle {
 
 } oneapi_handle_t;
 
-typedef struct oneapi_init_resp {
+typedef struct oneapi_init_resp
+{
   char *err; // If err is non-null handle is invalid
+  int num_devices;
   oneapi_handle_t oh;
 } oneapi_init_resp_t;
 
-typedef struct oneapi_version_resp {
+typedef struct oneapi_version_resp
+{
   ze_result_t status;
   char *str; // Contains version or error string if status != 0
 } oneapi_version_resp_t;
 
 void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp);
-void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
-                       mem_info_t *resp);
-void oneapi_release(oneapi_handle_t h);
-int oneapi_get_device_count(oneapi_handle_t h, int driver);
+void oneapi_check_vram(oneapi_handle_t rh, mem_info_t *resp);
 
 #endif // __GPU_INFO_INTEL_H__
 #endif // __APPLE__

gpu/gpu_linux.go

@@ -1,89 +0,0 @@
-package gpu
-
-import (
-	"bufio"
-	"fmt"
-	"os"
-	"strings"
-
-	"github.com/ollama/ollama/format"
-)
-
-var CudartGlobs = []string{
-	"/usr/local/cuda/lib64/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
-	"/usr/lib/wsl/lib/libcudart.so*",
-	"/usr/lib/wsl/drivers/*/libcudart.so*",
-	"/opt/cuda/lib64/libcudart.so*",
-	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
-	"/usr/local/cuda/lib*/libcudart.so*",
-	"/usr/lib*/libcudart.so*",
-	"/usr/local/lib*/libcudart.so*",
-}
-
-var NvmlGlobs = []string{}
-
-var NvcudaGlobs = []string{
-	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
-	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
-	"/usr/lib/*-linux-gnu/libcuda.so*",
-	"/usr/lib/wsl/lib/libcuda.so*",
-	"/usr/lib/wsl/drivers/*/libcuda.so*",
-	"/opt/cuda/lib*/libcuda.so*",
-	"/usr/local/cuda/lib*/libcuda.so*",
-	"/usr/lib*/libcuda.so*",
-	"/usr/local/lib*/libcuda.so*",
-}
-
-var OneapiGlobs = []string{
-	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
-	"/usr/lib*/libze_intel_gpu.so*",
-}
-
-var CudartMgmtName = "libcudart.so*"
-var NvcudaMgmtName = "libcuda.so*"
-var NvmlMgmtName = "" // not currently wired on linux
-var OneapiMgmtName = "libze_intel_gpu.so"
-
-func GetCPUMem() (memInfo, error) {
-	var mem memInfo
-	var total, available, free, buffers, cached uint64
-	f, err := os.Open("/proc/meminfo")
-	if err != nil {
-		return mem, err
-	}
-	defer f.Close()
-	s := bufio.NewScanner(f)
-	for s.Scan() {
-		line := s.Text()
-		switch {
-		case strings.HasPrefix(line, "MemTotal:"):
-			_, err = fmt.Sscanf(line, "MemTotal:%d", &total)
-		case strings.HasPrefix(line, "MemAvailable:"):
-			_, err = fmt.Sscanf(line, "MemAvailable:%d", &available)
-		case strings.HasPrefix(line, "MemFree:"):
-			_, err = fmt.Sscanf(line, "MemFree:%d", &free)
-		case strings.HasPrefix(line, "Buffers:"):
-			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
-		case strings.HasPrefix(line, "Cached:"):
-			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
-		default:
-			continue
-		}
-		if err != nil {
-			return mem, err
-		}
-
-		if total > 0 && available > 0 {
-			mem.TotalMemory = total * format.KibiByte
-			mem.FreeMemory = available * format.KibiByte
-			return mem, nil
-		}
-	}
-	mem.TotalMemory = total * format.KibiByte
-	mem.FreeMemory = (free + buffers + cached) * format.KibiByte
-	return mem, nil
-}

gpu/gpu_test.go

@@ -5,12 +5,11 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
 )
 
 func TestBasicGetGPUInfo(t *testing.T) {
 	info := GetGPUInfo()
-	assert.NotEmpty(t, len(info))
+	assert.Greater(t, len(info), 0)
 	assert.Contains(t, "cuda rocm cpu metal", info[0].Library)
 	if info[0].Library != "cpu" {
 		assert.Greater(t, info[0].TotalMemory, uint64(0))
@@ -20,7 +19,7 @@ func TestBasicGetGPUInfo(t *testing.T) {
 
 func TestCPUMemInfo(t *testing.T) {
 	info, err := GetCPUMem()
-	require.NoError(t, err)
+	assert.NoError(t, err)
 	switch runtime.GOOS {
 	case "darwin":
 		t.Skip("CPU memory not populated on darwin")

gpu/gpu_windows.go

@@ -1,55 +0,0 @@
-package gpu
-
-import (
-	"fmt"
-	"syscall"
-	"unsafe"
-)
-
-type MEMORYSTATUSEX struct {
-	length               uint32
-	MemoryLoad           uint32
-	TotalPhys            uint64
-	AvailPhys            uint64
-	TotalPageFile        uint64
-	AvailPageFile        uint64
-	TotalVirtual         uint64
-	AvailVirtual         uint64
-	AvailExtendedVirtual uint64
-}
-
-var (
-	k32                      = syscall.NewLazyDLL("kernel32.dll")
-	globalMemoryStatusExProc = k32.NewProc("GlobalMemoryStatusEx")
-	sizeofMemoryStatusEx     = uint32(unsafe.Sizeof(MEMORYSTATUSEX{}))
-)
-
-var CudartGlobs = []string{
-	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
-}
-
-var NvmlGlobs = []string{
-	"c:\\Windows\\System32\\nvml.dll",
-}
-
-var NvcudaGlobs = []string{
-	"c:\\windows\\system*\\nvcuda.dll",
-}
-
-var OneapiGlobs = []string{
-	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
-}
-
-var CudartMgmtName = "cudart64_*.dll"
-var NvcudaMgmtName = "nvcuda.dll"
-var NvmlMgmtName = "nvml.dll"
-var OneapiMgmtName = "ze_intel_gpu64.dll"
-
-func GetCPUMem() (memInfo, error) {
-	memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}
-	r1, _, err := globalMemoryStatusExProc.Call(uintptr(unsafe.Pointer(&memStatus)))
-	if r1 == 0 {
-		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
-	}
-	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys}, nil
-}

gpu/types.go

@@ -18,7 +18,7 @@ type GpuInfo struct {
 	Library string `json:"library,omitempty"`
 
 	// Optional variant to select (e.g. versions, cpu feature flags)
-	Variant CPUCapability `json:"variant"`
+	Variant string `json:"variant,omitempty"`
 
 	// MinimumMemory represents the minimum memory required to use the GPU
 	MinimumMemory uint64 `json:"-"`
@@ -26,14 +26,6 @@ type GpuInfo struct {
 	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
 	DependencyPath string `json:"lib_path,omitempty"`
 
-	// Extra environment variables specific to the GPU as list of [key,value]
-	EnvWorkarounds [][2]string `json:"envs,omitempty"`
-
-	// Set to true if we can NOT reliably discover FreeMemory.  A value of true indicates
-	// the FreeMemory is best effort, and may over or under report actual memory usage
-	// False indicates FreeMemory can generally be trusted on this GPU
-	UnreliableFreeMemory bool
-
 	// GPU information
 	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
 	Name    string `json:"name"`    // user friendly name if available
@@ -46,30 +38,6 @@ type GpuInfo struct {
 	// TODO other performance capability info to help in scheduling decisions
 }
 
-type CPUInfo struct {
-	GpuInfo
-}
-
-type CudaGPUInfo struct {
-	GpuInfo
-	index int //nolint:unused,nolintlint
-}
-type CudaGPUInfoList []CudaGPUInfo
-
-type RocmGPUInfo struct {
-	GpuInfo
-	usedFilepath string //nolint:unused,nolintlint
-	index        int    //nolint:unused,nolintlint
-}
-type RocmGPUInfoList []RocmGPUInfo
-
-type OneapiGPUInfo struct {
-	GpuInfo
-	driverIndex int //nolint:unused,nolintlint
-	gpuIndex    int //nolint:unused,nolintlint
-}
-type OneapiGPUInfoList []OneapiGPUInfo
-
 type GpuInfoList []GpuInfo
 
 // Split up the set of gpu info's by Library and variant
@@ -79,8 +47,8 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != CPUCapabilityNone {
-			requested += "_" + info.Variant.String()
+		if info.Variant != "" {
+			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {
 			if lib == requested {
@@ -118,26 +86,3 @@ type ByFreeMemory []GpuInfo
 func (a ByFreeMemory) Len() int           { return len(a) }
 func (a ByFreeMemory) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }
-
-type CPUCapability uint32
-
-// Override at build time when building base GPU runners
-var GPURunnerCPUCapability = CPUCapabilityAVX
-
-const (
-	CPUCapabilityNone CPUCapability = iota
-	CPUCapabilityAVX
-	CPUCapabilityAVX2
-	// TODO AVX512
-)
-
-func (c CPUCapability) String() string {
-	switch c {
-	case CPUCapabilityAVX:
-		return "avx"
-	case CPUCapabilityAVX2:
-		return "avx2"
-	default:
-		return "no vector extensions"
-	}
-}

integration/concurrency_test.go

@@ -19,19 +19,17 @@ func TestMultiModelConcurrency(t *testing.T) {
 	var (
 		req = [2]api.GenerateRequest{
 			{
-				Model:     "orca-mini",
-				Prompt:    "why is the ocean blue?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "orca-mini",
+				Prompt: "why is the ocean blue?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "tinydolphin",
-				Prompt:    "what is the origin of the us thanksgiving holiday?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "tinydolphin",
+				Prompt: "what is the origin of the us thanksgiving holiday?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
@@ -40,64 +38,42 @@ func TestMultiModelConcurrency(t *testing.T) {
 		}
 		resp = [2][]string{
 			[]string{"sunlight"},
-			[]string{"england", "english", "massachusetts", "pilgrims", "british"},
+			[]string{"england", "english", "massachusetts", "pilgrims"},
 		}
 	)
 	var wg sync.WaitGroup
 	wg.Add(len(req))
-	ctx, cancel := context.WithTimeout(context.Background(), time.Second*240)
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
 	defer cancel()
-
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	for i := 0; i < len(req); i++ {
-		require.NoError(t, PullIfMissing(ctx, client, req[i].Model))
-	}
-
 	for i := 0; i < len(req); i++ {
 		go func(i int) {
 			defer wg.Done()
-			DoGenerate(ctx, t, client, req[i], resp[i], 60*time.Second, 10*time.Second)
+			GenerateTestHelper(ctx, t, req[i], resp[i])
 		}(i)
 	}
 	wg.Wait()
 }
 
 func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
-	req, resp := GenerateRequests()
-	reqLimit := len(req)
-	iterLimit := 5
-
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
-	if vram != "" {
-		max, err := strconv.ParseUint(vram, 10, 64)
-		require.NoError(t, err)
-		// Don't hammer on small VRAM cards...
-		if max < 4*1024*1024*1024 {
-			reqLimit = min(reqLimit, 2)
-			iterLimit = 2
-		}
-	}
-
-	ctx, cancel := context.WithTimeout(context.Background(), 9*time.Minute)
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute) // GTX 750 2G card takes ~9 minutes
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
 
+	req, resp := GenerateRequests()
 	// Get the server running (if applicable) warm the model up with a single initial request
-	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 5*time.Second)
 
 	var wg sync.WaitGroup
-	wg.Add(reqLimit)
-	for i := 0; i < reqLimit; i++ {
+	wg.Add(len(req))
+	for i := 0; i < len(req); i++ {
 		go func(i int) {
 			defer wg.Done()
-			for j := 0; j < iterLimit; j++ {
+			for j := 0; j < 5; j++ {
 				slog.Info("Starting", "req", i, "iter", j)
-				// On slower GPUs it can take a while to process the concurrent requests
+				// On slower GPUs it can take a while to process the 4 concurrent requests
 				// so we allow a much longer initial timeout
-				DoGenerate(ctx, t, client, req[i], resp[i], 120*time.Second, 20*time.Second)
+				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
 			}
 		}(i)
 	}
@@ -245,23 +221,5 @@ func TestMultiModelStress(t *testing.T) {
 			}
 		}(i)
 	}
-	go func() {
-		for {
-			time.Sleep(2 * time.Second)
-			select {
-			case <-ctx.Done():
-				return
-			default:
-				models, err := client.ListRunning(ctx)
-				if err != nil {
-					slog.Warn("failed to list running models", "error", err)
-					continue
-				}
-				for _, m := range models.Models {
-					slog.Info("loaded model snapshot", "model", m)
-				}
-			}
-		}
-	}()
 	wg.Wait()
 }

integration/context_test.go

@@ -11,8 +11,7 @@ import (
 )
 
 func TestContextExhaustion(t *testing.T) {
-	// Longer needed for small footprint GPUs
-	ctx, cancel := context.WithTimeout(context.Background(), 6*time.Minute)
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute) // TODO maybe shorter?
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{

integration/llm_image_test.go

@@ -32,11 +32,7 @@ func TestIntegrationMultimodal(t *testing.T) {
 	resp := "the ollam"
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
-	// llava models on CPU can be quite slow to start,
-	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
+	GenerateTestHelper(ctx, t, req, []string{resp})
 }
 
 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb

integration/utils_test.go

@@ -140,7 +140,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	showCtx, cancel := context.WithDeadlineCause(
 		ctx,
-		time.Now().Add(10*time.Second),
+		time.Now().Add(5*time.Second),
 		fmt.Errorf("show for existing model %s took too long", modelName),
 	)
 	defer cancel()
@@ -287,46 +287,41 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 	return []api.GenerateRequest{
 			{
-				Model:     "orca-mini",
-				Prompt:    "why is the ocean blue?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "orca-mini",
+				Prompt: "why is the ocean blue?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
-				Prompt:    "why is the color of dirt brown?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "orca-mini",
+				Prompt: "why is the color of dirt brown?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
-				Prompt:    "what is the origin of the us thanksgiving holiday?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "orca-mini",
+				Prompt: "what is the origin of the us thanksgiving holiday?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
-				Prompt:    "what is the origin of independence day?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "orca-mini",
+				Prompt: "what is the origin of independence day?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
-				Prompt:    "what is the composition of air?",
-				Stream:    &stream,
-				KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				Model:  "orca-mini",
+				Prompt: "what is the composition of air?",
+				Stream: &stream,
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
@@ -336,7 +331,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 		[][]string{
 			[]string{"sunlight"},
 			[]string{"soil", "organic", "earth", "black", "tan"},
-			[]string{"england", "english", "massachusetts", "pilgrims", "british"},
+			[]string{"england", "english", "massachusetts", "pilgrims"},
 			[]string{"fourth", "july", "declaration", "independence"},
 			[]string{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}

llm/ext_server/CMakeLists.txt

@@ -1,13 +1,14 @@
-set(TARGET ollama_llama_server)
-option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
-install(TARGETS ${TARGET} RUNTIME)
-target_compile_definitions(${TARGET} PRIVATE
-    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
-)
-target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
-if (WIN32)
-    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
-endif()
+
+set(TARGET ollama_llama_server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

llm/ext_server/server.cpp

@@ -56,6 +56,7 @@ struct server_params {
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
+    std::string chat_template = "";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
@@ -139,6 +140,7 @@ struct server_slot {
     std::vector<llama_token> cache_tokens;
     std::vector<completion_token_output> generated_token_probs;
 
+    bool infill = false;
     bool embedding = false;
     bool has_next_token = true;
     bool truncated = false;
@@ -185,6 +187,7 @@ struct server_slot {
         n_past                 = 0;
         n_sent_text            = 0;
         n_sent_token_probs     = 0;
+        infill                 = false;
         ga_i                   = 0;
         n_past_se              = 0;
 
@@ -358,6 +361,7 @@ struct llama_server_context
 
     // slots / clients
     std::vector<server_slot> slots;
+    json default_generation_settings_for_props;
 
     llama_server_queue    queue_tasks;
     llama_server_response queue_results;
@@ -426,6 +430,16 @@ struct llama_server_context
         return true;
     }
 
+    void validate_model_chat_template(server_params & sparams) {
+        llama_chat_message chat[] = {{"user", "test"}};
+        std::vector<char> buf(1);
+        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
+        if (res < 0) {
+            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
+            sparams.chat_template = "chatml";
+        }
+    }
+
     void initialize() {
         // create slots
         all_slots_are_idle = true;
@@ -471,6 +485,9 @@ struct llama_server_context
             slots.push_back(slot);
         }
 
+        default_generation_settings_for_props = get_formated_generation(slots.front());
+        default_generation_settings_for_props["seed"] = -1;
+
         batch = llama_batch_init(n_ctx, 0, params.n_parallel);
     }
 
@@ -569,7 +586,7 @@ struct llama_server_context
         slot->sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
         slot->sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
         slot->params.n_keep             = json_value(data, "n_keep",            slot->params.n_keep);
-        slot->sparams.seed              = json_value(data, "seed",              default_params.seed);
+        slot->params.seed               = json_value(data, "seed",              default_params.seed);
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
         slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
@@ -583,6 +600,16 @@ struct llama_server_context
             slot->params.n_predict = slot->n_predict;
         }
 
+        // infill
+        if (data.count("input_prefix") != 0)
+        {
+            slot->params.input_prefix = data["input_prefix"];
+        }
+        else
+        {
+            slot->params.input_prefix = "";
+        }
+
         if (data.count("input_suffix") != 0)
         {
             slot->params.input_suffix = data["input_suffix"];
@@ -796,6 +823,7 @@ struct llama_server_context
             llama_sampling_free(slot->ctx_sampling);
         }
         slot->ctx_sampling = llama_sampling_init(slot->sparams);
+        llama_set_rng_seed(ctx, slot->params.seed);
         slot->command = LOAD_PROMPT;
 
         all_slots_are_idle = false;
@@ -819,7 +847,7 @@ struct llama_server_context
         system_tokens.clear();
 
         if (!system_prompt.empty()) {
-            system_tokens = ::llama_tokenize(ctx, system_prompt, true);
+            system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
 
             llama_batch_clear(batch);
 
@@ -869,6 +897,15 @@ struct llama_server_context
         system_need_update = true;
     }
 
+    void system_prompt_process(const json &sys_props) {
+        system_prompt  = sys_props.value("prompt", "");
+        name_user      = sys_props.value("anti_prompt", "");
+        name_assistant = sys_props.value("assistant_name", "");
+
+
+        system_prompt_notify();
+    }
+
     static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
                                         const stop_type type, server_slot &slot)
     {
@@ -1226,12 +1263,13 @@ struct llama_server_context
         queue_results.send(res);
     }
 
-    void request_completion(int task_id, json data, bool embedding, int multitask_id)
+    void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id)
     {
         task_server task;
         task.id = task_id;
         task.target_id = 0;
         task.data = std::move(data);
+        task.infill_mode = infill;
         task.embedding_mode = embedding;
         task.type = TASK_TYPE_COMPLETION;
         task.multitask_id = multitask_id;
@@ -1377,55 +1415,17 @@ struct llama_server_context
             json subtask_data = multiprompt_task.data;
             subtask_data["prompt"] = subtask_data["prompt"][i];
 
-            // subtasks inherit everything else (embedding mode, etc.)
-            request_completion(subtask_ids[i], subtask_data, multiprompt_task.embedding_mode, multitask_id);
+            // subtasks inherit everything else (infill mode, embedding mode, etc.)
+            request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id);
         }
     }
 
-    std::string common_prefix(const std::string& str1, const std::string& str2) {
-        auto mismatch_pair = std::mismatch(str1.begin(), str1.end(), str2.begin());
-        return std::string(str1.begin(), mismatch_pair.first);
-    }
-
-    // Find the slot that has the greatest common prefix
-    server_slot *prefix_slot(const json &prompt) {
-        if (!prompt.is_string()) {
-            return nullptr;
-        }
-
-        std::string prompt_str = prompt.get<std::string>();
-        server_slot *slot = nullptr;
-        size_t longest = 0;
-
-        for (server_slot &s : slots) {
-            if (s.available() && s.prompt.is_string()) {
-                std::string s_prompt = s.prompt.get<std::string>();
-                std::string prefix = common_prefix(s_prompt, prompt_str);
-
-                if (prefix.size() > longest) {
-                    slot = &s;
-                    longest = prefix.size();
-                }
-            }
-        }
-
-        if (!slot) {
-            return get_slot(-1);
-        }
-
-        LOG_DEBUG("slot with common prefix found", {{
-            "slot_id", slot->id,
-            "characters", longest
-        }});
-        return slot;
-    }
-
     void process_single_task(task_server& task)
     {
         switch (task.type)
         {
             case TASK_TYPE_COMPLETION: {
-                server_slot *slot = prefix_slot(task.data["prompt"]);
+                server_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
@@ -1434,8 +1434,26 @@ struct llama_server_context
                     break;
                 }
 
+                if (task.data.contains("system_prompt"))
+                {
+                    if (!all_slots_are_idle) {
+                        send_error(task, "system prompt can only be updated when all slots are idle");
+                        break;
+                    }
+                    system_prompt_process(task.data["system_prompt"]);
+
+                    // reset cache_tokens for all slots
+                    for (server_slot &slot : slots)
+                    {
+                        slot.cache_tokens.clear();
+                        slot.n_past    = 0;
+                        slot.n_past_se = 0;
+                    }
+                }
+
                 slot->reset();
 
+                slot->infill       = task.infill_mode;
                 slot->embedding    = task.embedding_mode;
                 slot->task_id      = task.id;
                 slot->multitask_id = task.multitask_id;
@@ -1661,7 +1679,8 @@ struct llama_server_context
                 const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
 
                 // empty prompt passed -> release the slot and send empty response
-                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
+                // note: infill mode allows empty prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
                 {
                     slot.release();
                     slot.print_timings();
@@ -1678,7 +1697,33 @@ struct llama_server_context
                     slot.t_start_process_prompt = ggml_time_us();
                     slot.t_start_genereration = 0;
 
-                    prompt_tokens = tokenize(slot.prompt, system_prompt.empty());  // add BOS if there isn't system prompt
+                    if (slot.infill)
+                    {
+                        bool suff_rm_leading_spc = true;
+                        if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1)
+                        {
+                            params.input_suffix.erase(0, 1);
+                            suff_rm_leading_spc = false;
+                        }
+                        auto prefix_tokens = tokenize(slot.params.input_prefix, false);
+                        auto suffix_tokens = tokenize(slot.params.input_suffix, false);
+
+                        const int space_token = 29871; // TODO: this should not be hardcoded
+                        if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) {
+                            suffix_tokens.erase(suffix_tokens.begin());
+                        }
+
+                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
+                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
+                        prefix_tokens.insert(prefix_tokens.end(),   llama_token_suffix(model));
+                        prefix_tokens.insert(prefix_tokens.end(),   suffix_tokens.begin(), suffix_tokens.end());
+                        prefix_tokens.push_back(llama_token_middle(model));
+                        prompt_tokens = prefix_tokens;
+                    }
+                    else
+                    {
+                        prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token);  // add BOS if there isn't system prompt
+                    }
 
                     slot.n_prompt_tokens = prompt_tokens.size();
 
@@ -1692,23 +1737,22 @@ struct llama_server_context
                     if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx)
                     {
                         const int n_left = slot.n_ctx - slot.params.n_keep;
-                        const int n_shift = n_left / 2;
-                        const int n_erase = slot.n_prompt_tokens - slot.params.n_keep - n_shift;
+                        const int n_block_size = n_left / 2;
+                        const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
 
                         std::vector<llama_token> new_tokens(
                             prompt_tokens.begin(),
                             prompt_tokens.begin() + slot.params.n_keep);
                         new_tokens.insert(
                             new_tokens.end(),
-                            prompt_tokens.begin() + slot.params.n_keep + n_erase,
+                            prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size,
                             prompt_tokens.end());
 
-                        LOG_INFO("input truncated", {
-                            {"n_ctx",        slot.n_ctx},
-                            {"n_keep",       slot.params.n_keep},
-                            {"n_left",       n_left},
-                            {"n_shift",      n_shift},
-                            {"n_erase",      n_erase},
+                        LOG_VERBOSE("input truncated", {
+                            {"n_ctx",      slot.n_ctx},
+                            {"n_keep",     slot.params.n_keep},
+                            {"n_left",     n_left},
+                            {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
                         });
                         slot.truncated = true;
                         prompt_tokens = new_tokens;
@@ -1743,7 +1787,7 @@ struct llama_server_context
                             slot.n_past -= 1;
                         }
 
-                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens;
+                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens - slot.n_past;
 
                         if (slot.ga_n != 1)
                         {
@@ -2086,7 +2130,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("\n");
 }
 
-static void server_params_parse(int argc, char **argv, server_params &sparams, gpt_params &params)
+static void server_params_parse(int argc, char **argv, server_params &sparams,
+                                gpt_params &params, llama_server_context& llama)
 {
     gpt_params default_params;
     server_params default_sparams;
@@ -2363,9 +2408,9 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-#ifndef GGML_USE_CUDA
-            fprintf(stderr, "warning: llama.cpp was compiled without CUDA. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUDA
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS
         }
         else if (arg == "--tensor-split" || arg == "-ts")
         {
@@ -2374,7 +2419,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-#if defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL)
             std::string arg_next = argv[i];
 
             // split string by , and /
@@ -2395,8 +2440,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 }
             }
 #else
-            LOG_WARNING("llama.cpp was compiled without CUDA. It is not possible to set a tensor split.\n", {});
-#endif // GGML_USE_CUDA
+            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
+#endif // GGML_USE_CUBLAS
         }
         else if (arg == "--main-gpu" || arg == "-mg")
         {
@@ -2405,7 +2450,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-#if defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL)
             params.main_gpu = std::stoi(argv[i]);
 #else
             LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {});
@@ -2501,6 +2546,27 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
             }
             params.n_predict = std::stoi(argv[i]);
         }
+        else if (arg == "-spf" || arg == "--system-prompt-file")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::string systm_content;
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(systm_content)
+            );
+            llama.system_prompt_process(json::parse(systm_content));
+        }
         else if (arg == "-ctk" || arg == "--cache-type-k") {
             params.cache_type_k = argv[++i];
         }
@@ -2563,6 +2629,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
+            sparams.chat_template = argv[i];
         }
         else if (arg == "--override-kv")
         {
@@ -2751,7 +2818,7 @@ int main(int argc, char **argv) {
     // struct that contains llama context and inference
     llama_server_context llama;
 
-    server_params_parse(argc, argv, sparams, params);
+    server_params_parse(argc, argv, sparams, params, llama);
 
     if (params.model_alias == "unknown")
     {
@@ -3035,6 +3102,11 @@ int main(int argc, char **argv) {
     }
     const auto model_meta = llama.model_meta();
 
+    if (sparams.chat_template.empty()) { // custom chat template is not supplied
+        // check if the template comes with the model is supported by us
+        llama.validate_model_chat_template(sparams);
+    }
+
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
@@ -3078,7 +3150,7 @@ int main(int argc, char **argv) {
                 json data = json::parse(req.body);
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
-                llama.request_completion(task_id, data, false, -1);
+                llama.request_completion(task_id, data, false, false, -1);
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.queue_results.recv(task_id);
@@ -3200,7 +3272,7 @@ int main(int argc, char **argv) {
                 // create and queue the task
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
-                llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, true, -1);
+                llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1);
 
                 // get the result
                 task_result result = llama.queue_results.recv(task_id);

llm/generate/gen_darwin.sh

@@ -18,77 +18,74 @@ sign() {
     fi
 }
 
-COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
+COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on"
 
 case "${GOARCH}" in
 "amd64")
-    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
 
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_BLAS=off -DGGML_ACCELERATE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
 
-    if [ -z "$OLLAMA_SKIP_CPU_GENERATE" ]; then
-        #
-        # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
-        #
-        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu"
-        echo "Building LCD CPU"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
 
-        #
-        # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
-        # Approximately 400% faster than LCD on same CPU
-        #
-        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
-        echo "Building AVX CPU"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
+    #
+    # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
+    #
+    init_vars
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}/cpu"
+    echo "Building LCD CPU"
+    build
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
 
-        #
-        # ~2013 CPU Dynamic library
-        # Approximately 10% faster than AVX on same CPU
-        #
-        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
-        echo "Building AVX2 CPU"
-        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
-    fi
+    #
+    # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
+    # Approximately 400% faster than LCD on same CPU
+    #
+    init_vars
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
+    echo "Building AVX CPU"
+    build
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
+
+    #
+    # ~2013 CPU Dynamic library
+    # Approximately 10% faster than AVX on same CPU
+    #
+    init_vars
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
+    echo "Building AVX2 CPU"
+    EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
+    build
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
     ;;
 "arm64")
 
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
 
-    if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
-        init_vars
-        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/metal"
-        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
-        build
-        sign ${BUILD_DIR}/bin/ollama_llama_server
-        compress
-    fi
+    init_vars
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}/metal"
+    EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
+    build
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
     ;;
 *)
     echo "GOARCH must be set"

llm/generate/gen_linux.sh

@@ -51,7 +51,7 @@ if [ -z "${CUDACXX}" ]; then
         export CUDACXX=$(command -v nvcc)
     fi
 fi
-COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
+COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@@ -64,7 +64,7 @@ if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ];
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}_static"
     echo "Building static library"
     build
@@ -77,29 +77,29 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
         init_vars
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
         compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
-        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
         # Note: the following seem to yield slower results than AVX2 - ymmv
-        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
+        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
+        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
         if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
             #
             # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
             #
             init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
             BUILD_DIR="../build/linux/${ARCH}/cpu"
             echo "Building LCD CPU"
             build
@@ -116,7 +116,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 400% faster than LCD on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
                 echo "Building AVX CPU"
                 build
@@ -129,7 +129,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 10% faster than AVX on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
                 echo "Building AVX2 CPU"
                 build
@@ -170,15 +170,15 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         #
         # CUDA compute < 6.0 lacks proper FP16 support on ARM.
         # Disabling has minimal performance effect while maintaining compatibility.
-        ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
+        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
     fi
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DGGML_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} -DCMAKE_LIBRARY_PATH=/usr/local/cuda/compat"
+        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
     fi
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
@@ -211,12 +211,12 @@ if [ -z "${ONEAPI_ROOT}" ]; then
     ONEAPI_ROOT=/opt/intel/oneapi
 fi
 
-if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
+if [ -d "${ONEAPI_ROOT}" ]; then
     echo "OneAPI libraries detected - building dynamic OneAPI library"
     init_vars
     source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
     CC=icx
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
     BUILD_DIR="../build/linux/${ARCH}/oneapi"
     EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
     DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@@ -254,7 +254,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""

llm/generate/gen_windows.ps1

@@ -39,8 +39,7 @@ function init_vars {
     }
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
-        "-DGGML_NATIVE=off",
-        "-DGGML_OPENMP=off"
+        "-DLLAMA_NATIVE=off"
         )
     $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
     $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
@@ -123,13 +122,8 @@ function build {
     & cmake --version
     & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    if ($cmakeDefs -contains "-G") {
-        $extra=@("-j8")
-    } else {
-        $extra= @("--", "/p:CL_MPcount=8")
-    }
-    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
-    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
+    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ })"
+    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ })
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     # Rearrange output to be consistent between different generators
     if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
@@ -182,9 +176,9 @@ function cleanup {
 }
 
 
-# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
 
 function build_static() {
@@ -204,13 +198,12 @@ function build_static() {
             "-DCMAKE_C_COMPILER=gcc.exe",
             "-DCMAKE_CXX_COMPILER=g++.exe",
             "-DBUILD_SHARED_LIBS=off",
-            "-DGGML_NATIVE=off",
-            "-DGGML_AVX=off",
-            "-DGGML_AVX2=off",
-            "-DGGML_AVX512=off",
-            "-DGGML_F16C=off",
-            "-DGGML_FMA=off",
-            "-DGGML_OPENMP=off")
+            "-DLLAMA_NATIVE=off",
+            "-DLLAMA_AVX=off",
+            "-DLLAMA_AVX2=off",
+            "-DLLAMA_AVX512=off",
+            "-DLLAMA_F16C=off",
+            "-DLLAMA_FMA=off")
         $script:buildDir="../build/windows/${script:ARCH}_static"
         write-host "Building static library"
         build
@@ -224,7 +217,7 @@ function build_cpu($gen_arch) {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
         # remaining llama.cpp builds use MSVC 
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu"
         $script:distDir="$script:DIST_BASE\cpu"
         write-host "Building LCD CPU"
@@ -239,7 +232,7 @@ function build_cpu($gen_arch) {
 function build_cpu_avx() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
         $script:distDir="$script:DIST_BASE\cpu_avx"
         write-host "Building AVX CPU"
@@ -254,7 +247,7 @@ function build_cpu_avx() {
 function build_cpu_avx2() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
         $script:distDir="$script:DIST_BASE\cpu_avx2"
         write-host "Building AVX2 CPU"
@@ -277,15 +270,7 @@ function build_cuda() {
         init_vars
         $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
         $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
-        $script:cmakeDefs += @(
-            "-A", "x64",
-            "-DGGML_CUDA=ON",
-            "-DGGML_AVX=on",
-            "-DGGML_AVX2=off",
-            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
-            "-DCMAKE_CUDA_FLAGS=-t8",
-            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
-            )
+        $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUDA=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
         if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
             write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
             $script:cmakeDefs +=@("${env:OLLAMA_CUSTOM_CUDA_DEFS}")
@@ -295,19 +280,17 @@ function build_cuda() {
         sign
         install
 
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
     } else {
         write-host "Skipping CUDA generation step"
     }
 }
 
 function build_oneapi() {
-  if ((-not "${env:OLLAMA_SKIP_ONEAPI_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
+  if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
     # Get oneAPI version
     $script:ONEAPI_VERSION = icpx --version
     $script:ONEAPI_VERSION = [regex]::Match($script:ONEAPI_VERSION, '(?<=oneAPI DPC\+\+/C\+\+ Compiler )(?<version>\d+\.\d+\.\d+)').Value
@@ -319,7 +302,7 @@ function build_oneapi() {
     $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
     $script:cmakeDefs += @(
       "-G", "MinGW Makefiles",
-      "-DGGML_SYCL=ON",
+      "-DLLAMA_SYCL=ON",
       "-DCMAKE_C_COMPILER=icx",
       "-DCMAKE_CXX_COMPILER=icx",
       "-DCMAKE_BUILD_TYPE=Release"
@@ -334,18 +317,16 @@ function build_oneapi() {
     sign
     install
 
-    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:distDir}"
   } else {
     Write-Host "Skipping oneAPI generation step"
   }
@@ -365,10 +346,10 @@ function build_rocm() {
             "-G", "Ninja", 
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DGGML_HIPBLAS=on",
+            "-DLLAMA_HIPBLAS=on",
             "-DHIP_PLATFORM=amd",
-            "-DGGML_AVX=on",
-            "-DGGML_AVX2=off",
+            "-DLLAMA_AVX=on",
+            "-DLLAMA_AVX2=off",
             "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
             "-DAMDGPU_TARGETS=$(amdGPUs)",
             "-DGPU_TARGETS=$(amdGPUs)"

llm/ggla.go

@@ -1,12 +1,9 @@
 package llm
 
 import (
-	"bytes"
 	"encoding/binary"
 	"errors"
-	"fmt"
 	"io"
-	"log/slog"
 	"slices"
 )
 
@@ -19,7 +16,6 @@ func (c *containerGGLA) Name() string {
 }
 
 func (c *containerGGLA) Decode(rs io.ReadSeeker) (model, error) {
-	slog.Info("decoding ggla")
 	if err := binary.Read(rs, binary.LittleEndian, &c.version); err != nil {
 		return nil, err
 	}
@@ -40,8 +36,6 @@ type ggla struct {
 
 	kv      KV
 	tensors []*Tensor
-
-	tensorOffset uint64
 }
 
 func newGGLA(container *containerGGLA) *ggla {
@@ -56,10 +50,7 @@ func (llm *ggla) KV() KV {
 }
 
 func (llm *ggla) Tensors() Tensors {
-	return Tensors{
-		Items:  llm.tensors,
-		Offset: llm.tensorOffset,
-	}
+	return llm.tensors
 }
 
 func (llm *ggla) decode(rs io.ReadSeeker) error {
@@ -75,19 +66,9 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 	}
 	llm.kv["alpha"] = alpha
 
-	offset, err := rs.Seek(0, io.SeekCurrent)
-	if err != nil {
-		return err
-	}
-
-	llm.tensorOffset = uint64(offset)
-
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
-			if errors.Is(err, io.EOF) {
-				break
-			}
 			return err
 		}
 
@@ -121,7 +102,6 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 		}
 
 		t.Name = string(name)
-		slog.Info(fmt.Sprintf("%s: [%d, %d] k=%d", t.Name, t.Shape[0], t.Shape[1], t.Kind))
 
 		offset, err := rs.Seek(0, io.SeekCurrent)
 		if err != nil {
@@ -145,87 +125,4 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 
 		llm.tensors = append(llm.tensors, &t)
 	}
-	return nil
-}
-
-func WriteGGLA(ws io.WriteSeeker, kv KV, ts []*Tensor) error {
-	slog.Debug("writing ggla")
-	if err := binary.Write(ws, binary.LittleEndian, []byte("algg")); err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, uint32(1)); err != nil {
-		return err
-	}
-
-	var r uint32
-	var alpha uint32
-	var ok bool
-
-	if r, ok = kv["r"].(uint32); !ok {
-		r = 8
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, r); err != nil {
-		return err
-	}
-
-	if alpha, ok = kv["alpha"].(uint32); !ok {
-		alpha = 16
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, alpha); err != nil {
-		return err
-	}
-
-	for _, t := range ts {
-		dims := 0
-		for cnt := range len(t.Shape) {
-			if t.Shape[cnt] > 0 {
-				dims++
-			}
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, uint32(dims)); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, uint32(len(t.Name))); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, t.Kind); err != nil {
-			return err
-		}
-
-		for cnt := range dims {
-			if err := binary.Write(ws, binary.LittleEndian, uint32(t.Shape[dims-1-cnt])); err != nil {
-				return err
-			}
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, []byte(t.Name)); err != nil {
-			return err
-		}
-
-		offset, err := ws.Seek(0, io.SeekCurrent)
-		if err != nil {
-			return err
-		}
-
-		var alignment int32 = 32
-		pad := gglaPadding(int32(offset), alignment)
-		if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(pad))); err != nil {
-			return err
-		}
-
-		if _, err := t.WriteTo(ws); err != nil {
-			return err
-		}
-	}
-	return nil
-}
-
-func gglaPadding(offset, align int32) int32 {
-	return (align - offset%align) % align
 }

llm/ggml.go

@@ -6,8 +6,6 @@ import (
 	"fmt"
 	"io"
 	"strings"
-
-	"github.com/ollama/ollama/util/bufioutil"
 )
 
 type GGML struct {
@@ -71,30 +69,6 @@ func (kv KV) HeadCountKV() uint64 {
 	return 1
 }
 
-func (kv KV) EmbeddingHeadCount() uint64 {
-	if heads := kv.HeadCount(); heads > 0 {
-		return kv.EmbeddingLength() / kv.HeadCount()
-	}
-
-	return 0
-}
-
-func (kv KV) EmbeddingHeadCountK() uint64 {
-	if k := kv.u64(fmt.Sprintf("%s.attention.key_length", kv.Architecture())); k > 0 {
-		return k
-	}
-
-	return kv.EmbeddingHeadCount()
-}
-
-func (kv KV) EmbeddingHeadCountV() uint64 {
-	if v := kv.u64(fmt.Sprintf("%s.attention.value_length", kv.Architecture())); v > 0 {
-		return v
-	}
-
-	return kv.EmbeddingHeadCount()
-}
-
 func (kv KV) GQA() uint64 {
 	return kv.HeadCount() / kv.HeadCountKV()
 }
@@ -107,19 +81,11 @@ func (kv KV) ContextLength() uint64 {
 	return kv.u64(fmt.Sprintf("%s.context_length", kv.Architecture()))
 }
 
-func (kv KV) ChatTemplate() string {
-	s, _ := kv["tokenizer.chat_template"].(string)
-	return s
-}
-
-type Tensors struct {
-	Items  []*Tensor
-	Offset uint64
-}
+type Tensors []*Tensor
 
 func (ts Tensors) Layers() map[string]Layer {
 	layers := make(map[string]Layer)
-	for _, t := range ts.Items {
+	for _, t := range ts {
 		parts := strings.Split(t.Name, ".")
 		if parts[0] == "blk" {
 			// join first and second part, e.g. blk.%d
@@ -283,18 +249,7 @@ func DetectGGMLType(b []byte) string {
 	}
 }
 
-// DecodeGGML decodes a GGML model from the given reader.
-//
-// It collects array values for arrays with a size less than or equal to
-// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
-// the maxArraySize is negative, all arrays are collected.
-func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
-	if maxArraySize == 0 {
-		maxArraySize = 1024
-	}
-
-	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
-
+func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
 		return nil, 0, err
@@ -307,15 +262,17 @@ func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	case FILE_MAGIC_GGLA:
 		c = &containerGGLA{}
 	case FILE_MAGIC_GGUF_LE:
-		c = &containerGGUF{ByteOrder: binary.LittleEndian, maxArraySize: maxArraySize}
+		c = &containerGGUF{ByteOrder: binary.LittleEndian}
 	case FILE_MAGIC_GGUF_BE:
-		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
+		c = &containerGGUF{ByteOrder: binary.BigEndian}
 	default:
 		return nil, 0, errors.New("invalid file magic")
 	}
 
 	model, err := c.Decode(rs)
-	if err != nil {
+	if errors.Is(err, io.EOF) {
+		// noop
+	} else if err != nil {
 		return nil, 0, err
 	}
 
@@ -335,10 +292,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 	embedding := llm.KV().EmbeddingLength()
 	heads := llm.KV().HeadCount()
 	headsKV := llm.KV().HeadCountKV()
-	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)
-
-	embeddingHeads := llm.KV().EmbeddingHeadCount()
-	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
+	vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
 
 	layers := llm.Tensors().Layers()
 
@@ -348,8 +302,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 
 		partialOffload = 4 * batch * embedding
 		partialOffload += max(
-			// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
-			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
+			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
 
@@ -357,30 +310,21 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			// mixtral 8x22b
 			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
-				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
-				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
+				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
 			)
 		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
 			// mixtral 8x7b
 			ffnGateWeight1 := ffnGateWeight.Shape[1]
 			fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
 			partialOffload = max(
-				4*batch*(3+embeddingHeads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
+				4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
 				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
 			)
 		}
-	case "gemma", "gemma2":
-		fullOffload = max(
-			4*batch*(embedding+vocab),
-			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
-		)
-
-		partialOffload = max(
-			4*embedding*batch+embedding*vocab*105/128+4*vocab*batch,
-			4*batch*(2*embedding+1+2*embeddingHeadsK*heads+context+context*heads)+
-				4*embeddingHeadsK*context*8+
-				embedding*embeddingHeadsK*heads*9/16,
-		)
+	case "gemma":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
@@ -417,16 +361,6 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(vocab+2*embedding),
 			fullOffload,
 		)
-	case "deepseek2":
-		fullOffload = max(
-			4*batch*(3*embedding+vocab),
-			4*batch*(3*embedding+2+context*(1+headsKV)+2*embeddingHeadsK*headsKV),
-		)
-
-		partialOffload = max(
-			4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
-			4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
-		)
 	}
 
 	return

llm/ggml_test.go

@@ -1 +0,0 @@
-package llm

llm/gguf.go

@@ -2,16 +2,12 @@ package llm
 
 import (
 	"bytes"
-	"cmp"
 	"encoding/binary"
-	"encoding/json"
 	"fmt"
 	"io"
-	"log/slog"
-	"slices"
 	"strings"
 
-	"golang.org/x/exp/maps"
+	"log/slog"
 )
 
 type containerGGUF struct {
@@ -33,12 +29,6 @@ type containerGGUF struct {
 		NumTensor uint64
 		NumKV     uint64
 	}
-
-	maxArraySize int
-}
-
-func (c *containerGGUF) canCollectArray(size int) bool {
-	return c.maxArraySize < 0 || size <= c.maxArraySize
 }
 
 func (c *containerGGUF) Name() string {
@@ -64,6 +54,7 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 	}
 
 	model := newGGUF(c)
+	slog.Debug(fmt.Sprintf("model = %#v", model))
 	if err := model.Decode(rs); err != nil {
 		return nil, err
 	}
@@ -94,9 +85,6 @@ type gguf struct {
 	tensors []*Tensor
 
 	parameters uint64
-	tensorOffset uint64
-
-	scratch [16 << 10]byte
 }
 
 func newGGUF(container *containerGGUF) *gguf {
@@ -106,15 +94,16 @@ func newGGUF(container *containerGGUF) *gguf {
 	}
 }
 
+func NewGGUFV3(bo binary.ByteOrder) *gguf {
+	return newGGUF(&containerGGUF{ByteOrder: bo, Version: 3})
+}
+
 func (llm *gguf) KV() KV {
 	return llm.kv
 }
 
 func (llm *gguf) Tensors() Tensors {
-	return Tensors{
-		Items:  llm.tensors,
-		Offset: llm.tensorOffset,
-	}
+	return llm.tensors
 }
 
 func (llm *gguf) numTensor() uint64 {
@@ -192,34 +181,34 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	// decode tensors
-	for range llm.numTensor() {
+	for i := 0; uint64(i) < llm.numTensor(); i++ {
 		name, err := readGGUFString(llm, rs)
 		if err != nil {
-			return fmt.Errorf("failed to read tensor name: %w", err)
+			return err
 		}
 
 		// dims is the number of dimensions in the tensor
 		dims, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return fmt.Errorf("failed to read tensor dimensions: %w", err)
+			return err
 		}
 
-		shape := make([]uint64, dims)
+		shape := [4]uint64{1, 1, 1, 1}
 		for i := 0; uint32(i) < dims; i++ {
 			shape[i], err = readGGUF[uint64](llm, rs)
 			if err != nil {
-				return fmt.Errorf("failed to read tensor shape: %w", err)
+				return err
 			}
 		}
 
 		kind, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return fmt.Errorf("failed to read tensor kind: %w", err)
+			return err
 		}
 
 		offset, err := readGGUF[uint64](llm, rs)
 		if err != nil {
-			return fmt.Errorf("failed to read tensor offset: %w", err)
+			return err
 		}
 
 		tensor := Tensor{
@@ -246,22 +235,19 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		return err
 	}
 
-	padding := ggufPadding(offset, int64(alignment))
-	llm.tensorOffset = uint64(offset + padding)
+	padding := llm.padding(offset, int64(alignment))
+	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
+		return err
+	}
 
 	for _, tensor := range llm.tensors {
-		offset, err := rs.Seek(0, io.SeekCurrent)
-		if err != nil {
-			return fmt.Errorf("failed to get current offset: %w", err)
-		}
-
-		padding := ggufPadding(offset, int64(alignment))
-		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-			return fmt.Errorf("failed to seek to init padding: %w", err)
-		}
-
 		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
-			return fmt.Errorf("failed to seek to tensor: %w", err)
+			return err
+		}
+
+		padding := llm.padding(int64(tensor.Size()), int64(alignment))
+		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
+			return err
 		}
 	}
 
@@ -274,12 +260,12 @@ func readGGUF[T any](llm *gguf, r io.Reader) (T, error) {
 	return t, err
 }
 
-func writeGGUF[V any](w io.Writer, t uint32, v V) error {
-	if err := binary.Write(w, binary.LittleEndian, t); err != nil {
+func writeGGUF[V any](llm *gguf, w io.Writer, t uint32, v V) error {
+	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
 		return err
 	}
 
-	return binary.Write(w, binary.LittleEndian, v)
+	return binary.Write(w, llm.ByteOrder, v)
 }
 
 func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
@@ -299,56 +285,30 @@ func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
 	return b.String(), nil
 }
 
-func discardGGUFString(llm *gguf, r io.Reader) error {
-	buf := llm.scratch[:8]
-	_, err := io.ReadFull(r, buf)
-	if err != nil {
-		return err
-	}
-
-	size := int(llm.ByteOrder.Uint64(buf))
-	for size > 0 {
-		n, err := r.Read(llm.scratch[:min(size, cap(llm.scratch))])
-		if err != nil {
-			return err
-		}
-		size -= n
-	}
-	return nil
-}
-
 func readGGUFString(llm *gguf, r io.Reader) (string, error) {
 	if llm.Version == 1 {
 		return readGGUFV1String(llm, r)
 	}
 
-	buf := llm.scratch[:8]
-	_, err := io.ReadFull(r, buf)
-	if err != nil {
+	var length uint64
+	if err := binary.Read(r, llm.ByteOrder, &length); err != nil {
 		return "", err
 	}
 
-	length := int(llm.ByteOrder.Uint64(buf))
-	if length > len(llm.scratch) {
-		buf = make([]byte, length)
-	} else {
-		buf = llm.scratch[:length]
-	}
-	clear(buf)
-
-	_, err = io.ReadFull(r, buf)
-	if err != nil {
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(length)); err != nil {
 		return "", err
 	}
-	return string(buf), nil
+
+	return b.String(), nil
 }
 
-func writeGGUFString(w io.Writer, s string) error {
-	if err := binary.Write(w, binary.LittleEndian, ggufTypeString); err != nil {
+func writeGGUFString(llm *gguf, w io.Writer, s string) error {
+	if err := binary.Write(w, llm.ByteOrder, ggufTypeString); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, binary.LittleEndian, uint64(len(s))); err != nil {
+	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
 		return err
 	}
 
@@ -356,16 +316,7 @@ func writeGGUFString(w io.Writer, s string) error {
 	return err
 }
 
-type array struct {
-	size   int
-	values []any
-}
-
-func (a *array) MarshalJSON() ([]byte, error) {
-	return json.Marshal(a.values)
-}
-
-func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
+func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 	t, err := readGGUF[uint32](llm, r)
 	if err != nil {
 		return nil, err
@@ -376,12 +327,7 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
 		return nil, err
 	}
 
-	a := &array{size: int(n)}
-	if llm.canCollectArray(int(n)) {
-		a.values = make([]any, 0, int(n))
-	}
-
-	for i := range n {
+	for i := 0; uint32(i) < n; i++ {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -415,15 +361,13 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
 			return nil, err
 		}
 
-		if a.values != nil {
-			a.values[i] = e
-		}
+		a = append(a, e)
 	}
 
-	return a, nil
+	return
 }
 
-func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
+func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 	if llm.Version == 1 {
 		return readGGUFV1Array(llm, r)
 	}
@@ -438,12 +382,7 @@ func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 		return nil, err
 	}
 
-	a := &array{size: int(n)}
-	if llm.canCollectArray(int(n)) {
-		a.values = make([]any, int(n))
-	}
-
-	for i := range n {
+	for i := 0; uint64(i) < n; i++ {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -469,11 +408,7 @@ func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 		case ggufTypeBool:
 			e, err = readGGUF[bool](llm, r)
 		case ggufTypeString:
-			if a.values != nil {
-				e, err = readGGUFString(llm, r)
-			} else {
-				err = discardGGUFString(llm, r)
-			}
+			e, err = readGGUFString(llm, r)
 		default:
 			return nil, fmt.Errorf("invalid array type: %d", t)
 		}
@@ -481,29 +416,27 @@ func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 			return nil, err
 		}
 
-		if a.values != nil {
-			a.values[i] = e
-		}
+		a = append(a, e)
 	}
 
-	return a, nil
+	return
 }
 
-func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
-	if err := binary.Write(w, binary.LittleEndian, ggufTypeArray); err != nil {
+func writeGGUFArray[S ~[]E, E any](llm *gguf, w io.Writer, t uint32, s S) error {
+	if err := binary.Write(w, llm.ByteOrder, ggufTypeArray); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, binary.LittleEndian, t); err != nil {
+	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, binary.LittleEndian, uint64(len(s))); err != nil {
+	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
 		return err
 	}
 
 	for _, e := range s {
-		if err := binary.Write(w, binary.LittleEndian, e); err != nil {
+		if err := binary.Write(w, llm.ByteOrder, e); err != nil {
 			return err
 		}
 	}
@@ -511,159 +444,210 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 	return nil
 }
 
-func WriteGGUF(ws io.WriteSeeker, kv KV, ts []*Tensor) error {
-	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, uint32(3)); err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(ts))); err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(kv))); err != nil {
-		return err
-	}
-
-	keys := maps.Keys(kv)
-	slices.Sort(keys)
-
-	for _, key := range keys {
-		if err := ggufWriteKV(ws, key, kv[key]); err != nil {
-			return err
-		}
-	}
-
-	slices.SortFunc(ts, func(a, b *Tensor) int {
-		var i, j int
-		if n, err := fmt.Sscanf(a.Name, "blk.%d", &i); err != nil || n != 1 {
-			return cmp.Compare(a.Name, b.Name)
-		} else if n, err := fmt.Sscanf(b.Name, "blk.%d", &j); err != nil || n != 1 {
-			return cmp.Compare(a.Name, b.Name)
-		}
-
-		return cmp.Compare(i, j)
-	})
-
-	var s uint64
-	for _, t := range ts {
-		t.Offset = s
-		if err := ggufWriteTensorInfo(ws, t); err != nil {
-			return err
-		}
-		s += t.Size()
-	}
-
-	var alignment int64 = 32
-	for _, t := range ts {
-		if err := ggufWriteTensor(ws, t, alignment); err != nil {
-			return err
-		}
-	}
-
-	return nil
+var ggufKVOrder = map[string][]string{
+	"llama": {
+		"general.architecture",
+		"general.name",
+		"llama.vocab_size",
+		"llama.context_length",
+		"llama.embedding_length",
+		"llama.block_count",
+		"llama.feed_forward_length",
+		"llama.attention.head_count",
+		"llama.attention.head_count_kv",
+		"llama.attention.layer_norm_rms_epsilon",
+		"llama.rope.freq_base",
+		"llama.rope.dimension_count",
+		"llama.expert_count",
+		"llama.expert_used_count",
+		"gemma.context_length",
+		"gemma.embedding_length",
+		"gemma.block_count",
+		"gemma.feed_forward_length",
+		"gemma.attention.head_count",
+		"gemma.attention.head_count_kv",
+		"gemma.attention.layer_norm_rms_epsilon",
+		"gemma.attention.key_length",
+		"gemma.attention.value_length",
+		"general.file_type",
+		"tokenizer.ggml.pre",
+		"tokenizer.ggml.model",
+		"tokenizer.ggml.tokens",
+		"tokenizer.ggml.scores",
+		"tokenizer.ggml.merges",
+		"tokenizer.ggml.token_type",
+		"tokenizer.ggml.bos_token_id",
+		"tokenizer.ggml.eos_token_id",
+		"tokenizer.ggml.unknown_token_id",
+		"tokenizer.ggml.padding_token_id",
+		"tokenizer.ggml.add_bos_token",
+		"tokenizer.ggml.add_eos_token",
+		"tokenizer.chat_template",
+	},
 }
 
-func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
-	slog.Debug(k, "type", fmt.Sprintf("%T", v))
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(k))); err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, []byte(k)); err != nil {
-		return err
-	}
-
-	var err error
-	switch v := v.(type) {
-	case uint32:
-		err = writeGGUF(ws, ggufTypeUint32, v)
-	case float32:
-		err = writeGGUF(ws, ggufTypeFloat32, v)
-	case bool:
-		err = writeGGUF(ws, ggufTypeBool, v)
-	case string:
-		err = writeGGUFString(ws, v)
-	case []int32:
-		err = writeGGUFArray(ws, ggufTypeInt32, v)
-	case []uint32:
-		err = writeGGUFArray(ws, ggufTypeUint32, v)
-	case []float32:
-		err = writeGGUFArray(ws, ggufTypeFloat32, v)
-	case []string:
-		if err := binary.Write(ws, binary.LittleEndian, ggufTypeArray); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, ggufTypeString); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, uint64(len(v))); err != nil {
-			return err
-		}
-
-		for _, e := range v {
-			if err := binary.Write(ws, binary.LittleEndian, uint64(len(e))); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, binary.LittleEndian, []byte(e)); err != nil {
-				return err
-			}
-		}
+func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
+	switch llm.Version {
+	case 3:
+		llm.V3.NumTensor = uint64(len(tensors))
+		llm.V3.NumKV = uint64(len(kv))
 	default:
-		return fmt.Errorf("improper type for '%s'", k)
+		return fmt.Errorf("not implemented: ggufv%d", llm.Version)
 	}
 
-	return err
-}
-
-func ggufWriteTensorInfo(ws io.WriteSeeker, t *Tensor) error {
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(t.Name))); err != nil {
+	if err := binary.Write(ws, llm.ByteOrder, []byte("GGUF")); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, []byte(t.Name)); err != nil {
+	if err := binary.Write(ws, llm.ByteOrder, llm.Version); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint32(len(t.Shape))); err != nil {
+	if err := binary.Write(ws, llm.ByteOrder, llm.numTensor()); err != nil {
 		return err
 	}
 
-	for i := range len(t.Shape) {
-		if err := binary.Write(ws, binary.LittleEndian, t.Shape[len(t.Shape)-i-1]); err != nil {
+	if err := binary.Write(ws, llm.ByteOrder, llm.numKV()); err != nil {
+		return err
+	}
+
+	kvCheck := make(map[string]bool)
+	for k := range kv {
+		kvCheck[k] = false
+	}
+
+	for _, k := range ggufKVOrder["llama"] {
+		v, ok := kv[k]
+		if !ok {
+			continue
+		}
+		kvCheck[k] = true
+
+		if err := binary.Write(ws, llm.ByteOrder, uint64(len(k))); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, []byte(k)); err != nil {
+			return err
+		}
+
+		var err error
+		switch v := v.(type) {
+		case uint32:
+			err = writeGGUF(llm, ws, ggufTypeUint32, v)
+		case float32:
+			err = writeGGUF(llm, ws, ggufTypeFloat32, v)
+		case bool:
+			err = writeGGUF(llm, ws, ggufTypeBool, v)
+		case string:
+			err = writeGGUFString(llm, ws, v)
+		case []int32:
+			err = writeGGUFArray(llm, ws, ggufTypeInt32, v)
+		case []uint32:
+			err = writeGGUFArray(llm, ws, ggufTypeUint32, v)
+		case []float32:
+			err = writeGGUFArray(llm, ws, ggufTypeFloat32, v)
+		case []string:
+			if err := binary.Write(ws, llm.ByteOrder, ggufTypeArray); err != nil {
+				return err
+			}
+
+			if err := binary.Write(ws, llm.ByteOrder, ggufTypeString); err != nil {
+				return err
+			}
+
+			if err := binary.Write(ws, llm.ByteOrder, uint64(len(v))); err != nil {
+				return err
+			}
+
+			for _, e := range v {
+				if err := binary.Write(ws, llm.ByteOrder, uint64(len(e))); err != nil {
+					return err
+				}
+
+				if err := binary.Write(ws, llm.ByteOrder, []byte(e)); err != nil {
+					return err
+				}
+			}
+		default:
+			return fmt.Errorf("improper type for '%s'", k)
+		}
+		if err != nil {
 			return err
 		}
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, t.Kind); err != nil {
-		return err
+	for k, v := range kvCheck {
+		if !v {
+			return fmt.Errorf("Didn't know how to write kv %s", k)
+		}
 	}
 
-	return binary.Write(ws, binary.LittleEndian, t.Offset)
-}
+	for _, tensor := range tensors {
+		if err := binary.Write(ws, llm.ByteOrder, uint64(len(tensor.Name))); err != nil {
+			return err
+		}
 
-func ggufWriteTensor(ws io.WriteSeeker, t *Tensor, alignment int64) error {
-	slog.Debug(t.Name, "kind", t.Kind, "shape", t.Shape, "offset", t.Offset)
+		if err := binary.Write(ws, llm.ByteOrder, []byte(tensor.Name)); err != nil {
+			return err
+		}
+
+		dims := 0
+		for cnt := 0; cnt < len(tensor.Shape); cnt++ {
+			if tensor.Shape[cnt] > 0 {
+				dims++
+			}
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, uint32(dims)); err != nil {
+			return err
+		}
+
+		for i := 0; i < dims; i++ {
+			if err := binary.Write(ws, llm.ByteOrder, uint64(tensor.Shape[dims-1-i])); err != nil {
+				return err
+			}
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, tensor.Kind); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, tensor.Offset); err != nil {
+			return err
+		}
+	}
 
 	offset, err := ws.Seek(0, io.SeekCurrent)
 	if err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(ggufPadding(offset, alignment)))); err != nil {
+	var alignment int64 = 32
+	padding := llm.padding(offset, alignment)
+	if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
 		return err
 	}
 
-	_, err = t.WriteTo(ws)
-	return err
+	for _, tensor := range tensors {
+		if _, err := tensor.WriteTo(ws); err != nil {
+			return err
+		}
+
+		offset, err := ws.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return err
+		}
+
+		padding := llm.padding(offset, alignment)
+		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
+			return err
+		}
+	}
+
+	return nil
 }
 
-func ggufPadding(offset, align int64) int64 {
+func (gguf) padding(offset, align int64) int64 {
 	return (align - offset%align) % align
 }