update llama.cpp submodule to 8960fe8

move Ollama static build to its own flag

.github/ISSUE_TEMPLATE/10_bug_report.yml

@@ -0,0 +1,60 @@
+name: Bug report
+labels: [bug]
+description: Something isn't working right.
+body:
+  - type: textarea
+    id: description
+    attributes:
+      label: What is the issue?
+      description: What happened? What did you expect to happen?
+    validations:
+      required: true
+  - type: dropdown
+    id: os
+    attributes:
+      label: OS
+      description: Which operating system are you using?
+      multiple: true
+      options:
+        - Linux
+        - macOS
+        - Windows
+        - Docker
+        - WSL2
+    validations:
+      required: false
+  - type: dropdown
+    id: gpu
+    attributes:
+      label: GPU
+      description: Which GPU are you using?
+      multiple: true
+      options:
+        - Nvidia
+        - AMD
+        - Intel
+        - Apple
+        - Other
+    validations:
+      required: false
+  - type: dropdown
+    id: cpu
+    attributes:
+      label: CPU
+      description: Which CPU are you using?
+      multiple: true
+      options:
+        - Intel
+        - AMD
+        - Apple
+        - Other
+    validations:
+      required: false
+  - type: input
+    id: version
+    attributes:
+      label: Ollama version
+      description: What version of Ollama are you using? (`ollama --version`)
+      placeholder: e.g., 0.1.32
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/10_model_request.yml

@@ -1,18 +0,0 @@
-name: Model request
-description: Request a new model for the library
-labels: [mr]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Please check if your Model request is [already available](https://ollama.com/search) or that you cannot [import it](https://github.com/ollama/ollama/blob/main/docs/import.md#import-a-model) yourself.
-        Tell us about which Model you'd like to see in the library!
-  - type: textarea
-    id: problem
-    attributes:
-      label: What model would you like?
-      description: Please provide a link to the model.
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for filing a model request!

.github/ISSUE_TEMPLATE/20_feature_request.md

@@ -0,0 +1,6 @@
+---
+name: Feature request
+about: Request a new feature
+labels: feature request
+---
+

.github/ISSUE_TEMPLATE/20_feature_request.yml

@@ -1,41 +0,0 @@
-name: Feature request
-description: Propose a new feature
-labels: [needs-triage, fr]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Please check if your feature request is [already filed](https://github.com/ollama/ollama/issues).
-        Tell us about your idea!
-  - type: textarea
-    id: problem
-    attributes:
-      label: What are you trying to do?
-      description: Tell us about the problem you're trying to solve.
-    validations:
-      required: false
-  - type: textarea
-    id: solution
-    attributes:
-      label: How should we solve this?
-      description: If you have an idea of how you'd like to see this feature work, let us know.
-    validations:
-      required: false
-  - type: textarea
-    id: alternative
-    attributes:
-      label: What is the impact of not solving this?
-      description: (How) Are you currently working around the issue?
-    validations:
-      required: false
-  - type: textarea
-    id: context
-    attributes:
-      label: Anything else?
-      description: Any additional context to share, e.g., links
-    validations:
-      required: false
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for filing a feature request!

.github/ISSUE_TEMPLATE/30_model_request.md

@@ -0,0 +1,5 @@
+---
+name: Model request
+about: Request support for a new model to be added to Ollama
+labels: model request
+---

.github/ISSUE_TEMPLATE/90_bug_report.yml

@@ -1,125 +0,0 @@
-name: Bug report
-description: File a bug report. If you need help, please join our Discord server.
-labels: [needs-triage, bug]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Please check if your bug is [already filed](https://github.com/ollama/ollama/issues) before filing a new one.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What is the issue?
-      description: What happened? What did you expect to happen?
-    validations:
-      required: true
-  - type: textarea
-    id: what-was-expected
-    attributes:
-      label: What did you expect to see?
-      description: What did you expect to see/happen instead?
-    validations:
-      required: false
-  - type: textarea
-    id: steps
-    attributes:
-      label: Steps to reproduce
-      description: What are the steps you took that hit this issue?
-    validations:
-      required: false
-  - type: textarea
-    id: changes
-    attributes:
-      label: Are there any recent changes that introduced the issue?
-      description: If so, what are those changes?
-    validations:
-      required: false
-  - type: dropdown
-    id: os
-    attributes:
-      label: OS
-      description: What OS are you using? You may select more than one.
-      multiple: true
-      options:
-        - Linux
-        - macOS
-        - Windows
-        - Other
-    validations:
-      required: false
-  - type: dropdown
-    id: architecture
-    attributes:
-      label: Architecture
-      description: What architecture are you using? You may select more than one.
-      multiple: true
-      options:
-        - arm64
-        - amd64
-        - x86
-        - Other
-  - type: dropdown
-    id: platform
-    attributes:
-      label: Platform
-      description: What platform are you using? You may select more than one.
-      multiple: true
-      options:
-        - Docker
-        - WSL
-        - WSL2
-    validations:
-      required: false
-  - type: input
-    id: ollama-version
-    attributes:
-      label: Ollama version
-      description: What Ollama version are you using? (`ollama --version`)
-      placeholder: e.g., 1.14.4
-    validations:
-      required: false
-  - type: dropdown
-    id: gpu
-    attributes:
-      label: GPU
-      description: What GPU, if any, are you using? You may select more than one.
-      multiple: true
-      options:
-        - Nvidia
-        - AMD
-        - Intel
-        - Apple
-        - Other
-    validations:
-      required: false
-  - type: textarea
-    id: gpu-info
-    attributes:
-      label: GPU info
-      description: What GPU info do you have? (`nvidia-smi`, `rocminfo`, `system_profiler SPDisplaysDataType`, etc.)
-    validations:
-      required: false
-  - type: dropdown
-    id: cpu
-    attributes:
-      label: CPU
-      description: What CPU are you using? You may select more than one.
-      multiple: true
-      options:
-        - Intel
-        - AMD
-        - Apple
-        - Other
-    validations:
-      required: false
-  - type: textarea
-    id: other-software
-    attributes:
-      label: Other software
-      description: What other software are you using that might be related to this issue?
-    validations:
-      required: false
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for filing a bug report!

Dockerfile

@@ -18,7 +18,7 @@ ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 
 FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
 ARG CMAKE_VERSION
@@ -28,7 +28,7 @@ ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
 ARG CMAKE_VERSION
@@ -40,7 +40,7 @@ COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 ARG AMDGPU_TARGETS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 RUN mkdir /tmp/scratch && \
     for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
         cp ${dep} /tmp/scratch/ || exit 1 ; \
@@ -64,11 +64,11 @@ WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
 RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
+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_CPU_TARGET="cpu_avx2" sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
 
 FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64
 ARG CMAKE_VERSION
@@ -84,7 +84,7 @@ WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
 RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
 FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
-RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
 
 
 # Intermediate stage used for ./scripts/build_linux.sh

README.md

@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 2](https://ollama.com/library/llama2):
+To run and chat with [Llama 3](https://ollama.com/library/llama3):
 
 ```
-ollama run llama2
+ollama run llama3
 ```
 
 ## Model library
@@ -49,7 +49,8 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 2            | 7B         | 3.8GB | `ollama run llama2`            |
+| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
+| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
 | Dolphin Phi        | 2.7B       | 1.6GB | `ollama run dolphin-phi`       |
 | Phi-2              | 2.7B       | 1.7GB | `ollama run phi`               |
@@ -60,7 +61,6 @@ Here are some example models that can be downloaded:
 | Llama 2 13B        | 13B        | 7.3GB | `ollama run llama2:13b`        |
 | Llama 2 70B        | 70B        | 39GB  | `ollama run llama2:70b`        |
 | Orca Mini          | 3B         | 1.9GB | `ollama run orca-mini`         |
-| Vicuna             | 7B         | 3.8GB | `ollama run vicuna`            |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 | Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
 | Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
@@ -98,16 +98,16 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama2` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3` model:
 
 ```
-ollama pull llama2
+ollama pull llama3
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama2
+FROM llama3
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -142,7 +142,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama2
+ollama pull llama3
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -150,13 +150,13 @@ ollama pull llama2
 ### Remove a model
 
 ```
-ollama rm llama2
+ollama rm llama3
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama2 my-llama2
+ollama cp llama3 my-model
 ```
 
 ### Multiline input
@@ -180,7 +180,7 @@ The image features a yellow smiley face, which is likely the central focus of th
 ### Pass in prompt as arguments
 
 ```
-$ ollama run llama2 "Summarize this file: $(cat README.md)"
+$ 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.
 ```
 
@@ -227,7 +227,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama2
+./ollama run llama3
 ```
 
 ## REST API
@@ -238,7 +238,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama2",
+  "model": "llama3",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -247,7 +247,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "mistral",
+  "model": "llama3",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -378,3 +378,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
+
+### Supported backends 
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 

api/types.go

@@ -2,6 +2,7 @@ package api
 
 import (
 	"encoding/json"
+	"errors"
 	"fmt"
 	"math"
 	"os"
@@ -307,7 +308,7 @@ func (m *Metrics) Summary() {
 	}
 }
 
-var ErrInvalidOpts = fmt.Errorf("invalid options")
+var ErrInvalidOpts = errors.New("invalid options")
 
 func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct

docs/api.md

@@ -90,7 +90,7 @@ The final response in the stream also includes additional data about the generat
 - `load_duration`: time spent in nanoseconds loading the model
 - `prompt_eval_count`: number of tokens in the prompt
 - `prompt_eval_duration`: time spent in nanoseconds evaluating the prompt
-- `eval_count`: number of tokens the response
+- `eval_count`: number of tokens in the response
 - `eval_duration`: time in nanoseconds spent generating the response
 - `context`: an encoding of the conversation used in this response, this can be sent in the next request to keep a conversational memory
 - `response`: empty if the response was streamed, if not streamed, this will contain the full response

docs/tutorials/nvidia-jetson.md

@@ -1,38 +1,15 @@
 # Running Ollama on NVIDIA Jetson Devices
 
-With some minor configuration, Ollama runs well on [NVIDIA Jetson Devices](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/). The following has been tested on [JetPack 5.1.2](https://developer.nvidia.com/embedded/jetpack).
+Ollama runs well on [NVIDIA Jetson Devices](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/) and should run out of the box with the standard installation instructions. 
 
-NVIDIA Jetson devices are Linux-based embedded AI computers that are purpose-built for AI applications.
-
-Jetsons have an integrated GPU that is wired directly to the memory controller of the machine. For this reason, the `nvidia-smi` command is unrecognized, and Ollama proceeds to operate in "CPU only"
-mode. This can be verified by using a monitoring tool like jtop.
-
-In order to address this, we simply pass the path to the Jetson's pre-installed CUDA libraries into `ollama serve` (while in a tmux session). We then hardcode the num_gpu parameters into a cloned
-version of our target model.
-
-Prerequisites:
-
-- curl
-- tmux
-
-Here are the steps:
+The following has been tested on [JetPack 5.1.2](https://developer.nvidia.com/embedded/jetpack), but should also work on JetPack 6.0.
 
 - Install Ollama via standard Linux command (ignore the 404 error): `curl https://ollama.com/install.sh | sh`
-- Stop the Ollama service: `sudo systemctl stop ollama`
-- Start Ollama serve in a tmux session called ollama_jetson and reference the CUDA libraries path: `tmux has-session -t ollama_jetson 2>/dev/null || tmux new-session -d -s ollama_jetson 
-'LD_LIBRARY_PATH=/usr/local/cuda/lib64 ollama serve'`
 - Pull the model you want to use (e.g. mistral): `ollama pull mistral`
-- Create a new Modelfile specifically for enabling GPU support on the Jetson: `touch ModelfileMistralJetson`
-- In the ModelfileMistralJetson file, specify the FROM model and the num_gpu PARAMETER as shown below:
-
-```
-FROM mistral
-PARAMETER num_gpu 999
-```
-
-- Create a new model from your Modelfile: `ollama create mistral-jetson -f ./ModelfileMistralJetson`
-- Run the new model: `ollama run mistral-jetson`
-
-If you run a monitoring tool like jtop you should now see that Ollama is using the Jetson's integrated GPU.
+- Start an interactive session: `ollama run mistral`
 
 And that's it!
+
+# Running Ollama in Docker
+
+When running GPU accelerated applications in Docker, it is highly recommended to use [dusty-nv jetson-containers repo](https://github.com/dusty-nv/jetson-containers).

gpu/gpu_darwin.go

@@ -55,6 +55,6 @@ func getCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
 		FreeMemory:  0,
-		DeviceCount: 0,
+		DeviceCount: 1,
 	}, nil
 }

llm/ext_server/server.cpp

@@ -39,6 +39,10 @@
 #include "httplib.h"
 #include "json.hpp"
 
+#if defined(_WIN32)
+#include <windows.h>
+#endif
+
 #include <cstddef>
 #include <thread>
 #include <chrono>
@@ -2770,8 +2774,28 @@ inline void signal_handler(int signal) {
     shutdown_handler(signal);
 }
 
-int main(int argc, char **argv)
-{
+#if defined(_WIN32)
+char* wchar_to_char(const wchar_t* wstr) {
+    if (wstr == nullptr) return nullptr;
+
+    // Determine the number of bytes needed for the UTF-8 string
+    int bytes = WideCharToMultiByte(CP_UTF8, 0, wstr, -1, nullptr, 0, nullptr, nullptr);
+    char* str = new char[bytes];
+
+    // Convert the wide-character string to a UTF-8 string
+    WideCharToMultiByte(CP_UTF8, 0, wstr, -1, str, bytes, nullptr, nullptr);
+    return str;
+}
+
+int wmain(int argc, wchar_t **wargv) {
+    char** argv = new char*[argc];
+    for (int i = 0; i < argc; ++i) {
+        argv[i] = wchar_to_char(wargv[i]);
+    }
+#else
+int main(int argc, char **argv) {
+#endif
+
 #if SERVER_VERBOSE != 1
     log_disable();
 #endif
@@ -3282,6 +3306,11 @@ int main(int argc, char **argv)
         return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
     };
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+
+    for (int i = 0; i < argc; ++i) {
+        delete[] argv[i];
+    }
+    delete[] argv;
 #endif
     llama.queue_tasks.start_loop();
     svr.stop();

llm/generate/gen_linux.sh

@@ -57,21 +57,21 @@ init_vars
 git_module_setup
 apply_patches
 
+init_vars
+if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
+    # Builds by default, allows skipping, forces build if OLLAMA_CPU_TARGET="static"
+    # Enables optimized Dockerfile builds using a blanket skip and targeted overrides
+    # Static build for linking into the Go binary
+    init_vars
+    CMAKE_TARGETS="--target llama --target ggml"
+    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
+fi
 
 init_vars
 if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
-
-    if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
-        # Static build for linking into the Go binary
-        init_vars
-        CMAKE_TARGETS="--target llama --target ggml"
-        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
-    fi
-
-
     # Users building from source can tune the exact flags we pass to cmake for configuring
     # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then

llm/ggml.go

@@ -164,7 +164,8 @@ func (ts Tensors) Layers() map[string]Layer {
 	for _, t := range ts {
 		parts := strings.Split(t.Name, ".")
 		if parts[0] == "blk" {
-			parts = parts[1:]
+			// join first and second part, e.g. blk.%d
+			parts = append([]string{fmt.Sprintf("%s.%s", parts[0], parts[1])}, parts[2:]...)
 		}
 
 		if _, ok := layers[parts[0]]; !ok {
@@ -380,6 +381,12 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 		)
 
 		partialOffload = 4*batch*(2*embedding+vocab) + embedding*vocab*105/128
+	case "stablelm":
+		fullOffload = 4 * batch * (context*(1+heads) + 3*embedding + 2)
+		partialOffload = max(
+			4*batch*(vocab+2*embedding),
+			fullOffload,
+		)
 	}
 
 	return

llm/gguf.go

@@ -248,13 +248,17 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	padding := llm.padding(offset, int64(alignment))
-	if _, err := rs.Seek(padding-offset, io.SeekCurrent); err != nil {
+	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
 		return err
 	}
 
 	for _, tensor := range llm.tensors {
-		padded := (int64(tensor.size()) + int64(alignment) - 1) & ^(int64(alignment) - 1)
-		if _, err := rs.Seek(padded, io.SeekCurrent); err != nil {
+		if _, err := rs.Seek(int64(tensor.size()), io.SeekCurrent); err != nil {
+			return err
+		}
+
+		padding := llm.padding(int64(tensor.size()), int64(alignment))
+		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
 			return err
 		}
 	}
@@ -623,8 +627,9 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 		return err
 	}
 
-	padding := llm.padding(offset, 32)
-	if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding-offset))); 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
 	}
 
@@ -638,8 +643,8 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 			return err
 		}
 
-		padding := llm.padding(offset, 32)
-		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding-offset))); err != nil {
+		padding := llm.padding(offset, alignment)
+		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
 			return err
 		}
 	}
@@ -648,5 +653,5 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 }
 
 func (gguf) padding(offset, align int64) int64 {
-	return (offset + align - 1) / align * align
+	return (align - offset%align) % align
 }

llm/payload.go

@@ -17,7 +17,7 @@ import (
 	"github.com/ollama/ollama/gpu"
 )
 
-var errPayloadMissing = fmt.Errorf("expected payloads not included in this build of ollama")
+var errPayloadMissing = errors.New("expected payloads not included in this build of ollama")
 
 func Init() error {
 	payloadsDir, err := gpu.PayloadsDir()

llm/server.go

@@ -79,6 +79,9 @@ func NewLlamaServer(model string, adapters, projectors []string, opts api.Option
 		graphFullOffload = graphPartialOffload
 	}
 
+	graphFullOffload *= uint64(info.DeviceCount)
+	graphPartialOffload *= uint64(info.DeviceCount)
+
 	// memoryRequiredTotal represents the memory required for full GPU offloading (all layers)
 	memoryRequiredTotal := memoryMinimum + graphFullOffload
 
@@ -94,7 +97,7 @@ func NewLlamaServer(model string, adapters, projectors []string, opts api.Option
 	var layerCount int
 	layers := ggml.Tensors().Layers()
 	for i := 0; i < int(ggml.KV().BlockCount()); i++ {
-		memoryLayer := layers[fmt.Sprintf("%d", i)].size()
+		memoryLayer := layers[fmt.Sprintf("blk.%d", i)].size()
 
 		// KV is proportional to the number of layers
 		memoryLayer += kv / ggml.KV().BlockCount()
@@ -106,7 +109,13 @@ func NewLlamaServer(model string, adapters, projectors []string, opts api.Option
 		}
 	}
 
-	memoryLayerOutput := layers["output"].size()
+	var memoryLayerOutput uint64
+	for k, v := range layers {
+		if !strings.HasPrefix(k, "blk.") {
+			memoryLayerOutput += v.size()
+		}
+	}
+
 	memoryRequiredTotal += memoryLayerOutput
 
 	if info.Library == "metal" && memoryRequiredTotal > info.TotalMemory {
@@ -121,16 +130,47 @@ func NewLlamaServer(model string, adapters, projectors []string, opts api.Option
 		opts.NumGPU = layerCount
 	}
 
+	memoryWeights := memoryRequiredTotal - memoryMinimum - graphFullOffload - kv
+
 	slog.Info(
 		"offload to gpu",
-		"reallayers", opts.NumGPU,
-		"layers", layerCount,
-		"required", format.HumanBytes2(memoryRequiredTotal),
-		"used", format.HumanBytes2(memoryRequiredPartial),
-		"available", format.HumanBytes2(memoryAvailable),
-		"kv", format.HumanBytes2(kv),
-		"fulloffload", format.HumanBytes2(graphFullOffload),
-		"partialoffload", format.HumanBytes2(graphPartialOffload),
+		slog.Group(
+			"layers",
+			// actual number of layers offloaded
+			"real", opts.NumGPU,
+			// estimated number of layers that can be offloaded
+			"estimate", layerCount,
+		),
+		slog.Group(
+			"memory",
+			// memory available for offloading
+			"available", format.HumanBytes2(memoryAvailable),
+			slog.Group(
+				"required",
+				// memory required for full offloading
+				"full", format.HumanBytes2(memoryRequiredTotal),
+				// memory required to offload layers.estimate layers
+				"partial", format.HumanBytes2(memoryRequiredPartial),
+				// memory of KV cache
+				"kv", format.HumanBytes2(kv),
+			),
+			slog.Group(
+				"weights",
+				// memory of the weights
+				"total", format.HumanBytes2(memoryWeights),
+				// memory of repeating layers
+				"repeating", format.HumanBytes2(memoryWeights-memoryLayerOutput),
+				// memory of non-repeating layers
+				"nonrepeating", format.HumanBytes2(memoryLayerOutput),
+			),
+			slog.Group(
+				"graph",
+				// memory of graph when fully offloaded
+				"full", format.HumanBytes2(graphFullOffload),
+				// memory of graph when not fully offloaded
+				"partial", format.HumanBytes2(graphPartialOffload),
+			),
+		),
 	)
 
 	if len(adapters) > 1 {

server/images.go

@@ -1137,7 +1137,7 @@ func GetSHA256Digest(r io.Reader) (string, int64) {
 	return fmt.Sprintf("sha256:%x", h.Sum(nil)), n
 }
 
-var errUnauthorized = fmt.Errorf("unauthorized")
+var errUnauthorized = errors.New("unauthorized")
 
 func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.URL, headers http.Header, body io.ReadSeeker, regOpts *registryOptions) (*http.Response, error) {
 	for i := 0; i < 2; i++ {
@@ -1255,7 +1255,7 @@ func parseRegistryChallenge(authStr string) registryChallenge {
 	}
 }
 
-var errDigestMismatch = fmt.Errorf("digest mismatch, file must be downloaded again")
+var errDigestMismatch = errors.New("digest mismatch, file must be downloaded again")
 
 func verifyBlob(digest string) error {
 	fp, err := GetBlobsPath(digest)

types/model/digest.go

@@ -1,6 +1,7 @@
 package model
 
 import (
+	"fmt"
 	"log/slog"
 	"strings"
 	"unicode"
@@ -47,8 +48,11 @@ var (
 // Digest.
 func ParseDigest(s string) Digest {
 	typ, digest, ok := strings.Cut(s, "-")
+	if !ok {
+		typ, digest, ok = strings.Cut(s, ":")
+	}
 	if ok && isValidDigestType(typ) && isValidHex(digest) {
-		return Digest{s: s}
+		return Digest{s: fmt.Sprintf("%s-%s", typ, digest)}
 	}
 	return Digest{}
 }

types/model/name.go

@@ -156,7 +156,7 @@ func ParseName(s, fill string) Name {
 			r = Name{}
 			return false
 		}
-		if kind == PartExtraneous || !isValidPart(kind, part) {
+		if kind == PartExtraneous || !IsValidNamePart(kind, part) {
 			r = Name{}
 			return false
 		}
@@ -176,7 +176,7 @@ func parseMask(s string) Name {
 			// mask part; treat as empty but valid
 			return true
 		}
-		if !isValidPart(kind, part) {
+		if !IsValidNamePart(kind, part) {
 			panic(fmt.Errorf("invalid mask part %s: %q", kind, part))
 		}
 		r.parts[kind] = part
@@ -185,8 +185,8 @@ func parseMask(s string) Name {
 	return r
 }
 
-func MustParseName(s, defaults string) Name {
-	r := ParseName(s, "")
+func MustParseName(s, fill string) Name {
+	r := ParseName(s, fill)
 	if !r.IsValid() {
 		panic("invalid Name: " + s)
 	}
@@ -521,6 +521,8 @@ func parts(s string) iter_Seq2[PartKind, string] {
 						return
 					}
 					state, j, partLen = PartModel, i, 0
+				case PartHost:
+					// noop: support for host:port
 				default:
 					yield(PartExtraneous, s[i+1:j])
 					return
@@ -606,7 +608,7 @@ func ParseNameFromFilepath(s, fill string) Name {
 	var r Name
 	for i := range PartBuild + 1 {
 		part, rest, _ := strings.Cut(s, string(filepath.Separator))
-		if !isValidPart(i, part) {
+		if !IsValidNamePart(i, part) {
 			return Name{}
 		}
 		r.parts[i] = part
@@ -643,9 +645,21 @@ func (r Name) Filepath() string {
 	return filepath.Join(r.parts[:]...)
 }
 
-// isValidPart reports if s contains all valid characters for the given
-// part kind.
-func isValidPart(kind PartKind, s string) bool {
+// FilepathNoBuild returns a complete, canonicalized, relative file path using
+// the parts of a complete Name, but without the build part.
+func (r Name) FilepathNoBuild() string {
+	for i := range PartBuild {
+		r.parts[i] = strings.ToLower(r.parts[i])
+	}
+	return filepath.Join(r.parts[:PartBuild]...)
+}
+
+// IsValidNamePart reports if s contains all valid characters for the given
+// part kind and is under MaxNamePartLen bytes.
+func IsValidNamePart(kind PartKind, s string) bool {
+	if len(s) > MaxNamePartLen {
+		return false
+	}
 	if s == "" {
 		return false
 	}
@@ -669,6 +683,9 @@ func isValidByteFor(kind PartKind, c byte) bool {
 	if kind == PartNamespace && c == '.' {
 		return false
 	}
+	if kind == PartHost && c == ':' {
+		return true
+	}
 	if c == '.' || c == '-' {
 		return true
 	}

types/model/name_test.go

@@ -40,6 +40,7 @@ var testNames = map[string]fields{
 	"user/model":                     {namespace: "user", model: "model"},
 	"example.com/ns/mistral:7b+Q4_0": {host: "example.com", namespace: "ns", model: "mistral", tag: "7b", build: "Q4_0"},
 	"example.com/ns/mistral:7b+X":    {host: "example.com", namespace: "ns", model: "mistral", tag: "7b", build: "X"},
+	"localhost:5000/ns/mistral":      {host: "localhost:5000", namespace: "ns", model: "mistral"},
 
 	// invalid digest
 	"mistral:latest@invalid256-": {},
@@ -104,6 +105,12 @@ var testNames = map[string]fields{
 	strings.Repeat("a", MaxNamePartLen+1): {},
 }
 
+func TestIsValidNameLen(t *testing.T) {
+	if IsValidNamePart(PartNamespace, strings.Repeat("a", MaxNamePartLen+1)) {
+		t.Errorf("unexpectedly valid long name")
+	}
+}
+
 // TestConsecutiveDots tests that consecutive dots are not allowed in any
 // part, to avoid path traversal. There also are some tests in testNames, but
 // this test is more exhaustive and exists to emphasize the importance of
@@ -318,6 +325,13 @@ func TestNameGoString(t *testing.T) {
 	}
 }
 
+func TestDisplayLongest(t *testing.T) {
+	g := ParseName("example.com/library/mistral:latest+Q4_0", FillNothing).DisplayLongest()
+	if g != "example.com/library/mistral:latest" {
+		t.Errorf("got = %q; want %q", g, "example.com/library/mistral:latest")
+	}
+}
+
 func TestDisplayShortest(t *testing.T) {
 	cases := []struct {
 		in        string
@@ -478,30 +492,36 @@ func TestNamePath(t *testing.T) {
 	}
 }
 
-func TestNameFromFilepath(t *testing.T) {
+func TestNameFilepath(t *testing.T) {
 	cases := []struct {
-		in   string
-		want string
+		in          string
+		want        string
+		wantNoBuild string
 	}{
 		{
-			in:   "example.com/library/mistral:latest+Q4_0",
-			want: "example.com/library/mistral/latest/Q4_0",
+			in:          "example.com/library/mistral:latest+Q4_0",
+			want:        "example.com/library/mistral/latest/Q4_0",
+			wantNoBuild: "example.com/library/mistral/latest",
 		},
 		{
-			in:   "Example.Com/Library/Mistral:Latest+Q4_0",
-			want: "example.com/library/mistral/latest/Q4_0",
+			in:          "Example.Com/Library/Mistral:Latest+Q4_0",
+			want:        "example.com/library/mistral/latest/Q4_0",
+			wantNoBuild: "example.com/library/mistral/latest",
 		},
 		{
-			in:   "Example.Com/Library/Mistral:Latest+Q4_0",
-			want: "example.com/library/mistral/latest/Q4_0",
+			in:          "Example.Com/Library/Mistral:Latest+Q4_0",
+			want:        "example.com/library/mistral/latest/Q4_0",
+			wantNoBuild: "example.com/library/mistral/latest",
 		},
 		{
-			in:   "example.com/library/mistral:latest",
-			want: "example.com/library/mistral/latest",
+			in:          "example.com/library/mistral:latest",
+			want:        "example.com/library/mistral/latest",
+			wantNoBuild: "example.com/library/mistral/latest",
 		},
 		{
-			in:   "",
-			want: "",
+			in:          "",
+			want:        "",
+			wantNoBuild: "",
 		},
 	}
 	for _, tt := range cases {
@@ -513,6 +533,11 @@ func TestNameFromFilepath(t *testing.T) {
 			if g != tt.want {
 				t.Errorf("got = %q; want %q", g, tt.want)
 			}
+			g = p.FilepathNoBuild()
+			g = filepath.ToSlash(g)
+			if g != tt.wantNoBuild {
+				t.Errorf("got = %q; want %q", g, tt.wantNoBuild)
+			}
 		})
 	}
 }