Linux uninstall instructions (#894)

Separate GGUF and PyTorch guides

README.md

@@ -233,3 +233,4 @@ curl -X POST http://localhost:11434/api/generate -d '{
 - [oterm](https://github.com/ggozad/oterm)
 - [Ellama Emacs client](https://github.com/s-kostyaev/ellama)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
+- [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)

docs/import.md

@@ -1,8 +1,43 @@
 # Import a model
 
-This guide walks through importing a PyTorch, Safetensors or GGUF model.
+This guide walks through importing a GGUF, PyTorch or Safetensors model.
 
-## Supported models
+## Importing (GGUF)
+
+### Step 1: Write a `Modelfile`
+
+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
+```
+
+(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`:
+
+```
+FROM ./q4_0.bin
+TEMPLATE "[INST] {{ .Prompt }} [/INST]"
+```
+
+### Step 2: Create the Ollama model
+
+Finally, create a model from your `Modelfile`:
+
+```
+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)
+
+### Supported models
 
 Ollama supports a set of model architectures, with support for more coming soon:
 
@@ -13,8 +48,6 @@ Ollama supports a set of model architectures, with support for more coming soon:
 
 To view a model's architecture, check the `config.json` file in its HuggingFace repo. You should see an entry under `architectures` (e.g. `LlamaForCausalLM`).
 
-## Importing
-
 ### Step 1: Clone the HuggingFace repository (optional)
 
 If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model.
@@ -44,7 +77,7 @@ This will output two files into the directory:
 
 ### Step 3: Write a `Modelfile`
 
-Next, create a `Modelfile` for your model. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more.
+Next, create a `Modelfile` for your model:
 
 ```
 FROM ./q4_0.bin
@@ -65,13 +98,15 @@ 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?"
 ```
 
-### Step 5: Publish your model (optional – early alpha)
+## 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:
 

docs/linux.md

@@ -80,3 +80,23 @@ To view logs of Ollama running as a startup service, run:
 ```bash
 journalctl -u ollama
 ```
+
+## Uninstall
+
+Remove the ollama service:
+```bash
+systemctl stop ollama
+systemctl disable ollama
+rm /etc/systemd/system/ollama.service
+```
+
+Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr/bin`, or `/bin`):
+```bash
+rm /usr/local/bin/ollama
+```
+
+Remove the downloaded models and Ollama service user:
+```bash
+rm /usr/share/ollama
+userdel ollama
+```

llm/ggml.go

@@ -175,7 +175,8 @@ const (
 	// Magic constant for `ggla` files (LoRA adapter).
 	FILE_MAGIC_GGLA = 0x67676C61
 	// Magic constant for `gguf` files (versioned, gguf)
-	FILE_MAGIC_GGUF = 0x46554747
+	FILE_MAGIC_GGUF_LE = 0x46554747
+	FILE_MAGIC_GGUF_BE = 0x47475546
 )
 
 func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
@@ -191,8 +192,10 @@ func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
 		ggml.container = &containerGGJT{}
 	case FILE_MAGIC_GGLA:
 		ggml.container = &containerLORA{}
-	case FILE_MAGIC_GGUF:
-		ggml.container = &containerGGUF{}
+	case FILE_MAGIC_GGUF_LE:
+		ggml.container = &containerGGUF{bo: binary.LittleEndian}
+	case FILE_MAGIC_GGUF_BE:
+		ggml.container = &containerGGUF{bo: binary.BigEndian}
 	default:
 		return nil, errors.New("invalid file magic")
 	}

llm/gguf.go

@@ -3,12 +3,13 @@ package llm
 import (
 	"bytes"
 	"encoding/binary"
-	"errors"
 	"fmt"
 	"io"
 )
 
 type containerGGUF struct {
+	bo binary.ByteOrder
+
 	Version uint32
 
 	V1 struct {
@@ -27,15 +28,13 @@ func (c *containerGGUF) Name() string {
 }
 
 func (c *containerGGUF) Decode(r io.Reader) (model, error) {
-	binary.Read(r, binary.LittleEndian, &c.Version)
+	binary.Read(r, c.bo, &c.Version)
 
 	switch c.Version {
 	case 1:
-		binary.Read(r, binary.LittleEndian, &c.V1)
-	case 2:
-		binary.Read(r, binary.LittleEndian, &c.V2)
+		binary.Read(r, c.bo, &c.V1)
 	default:
-		return nil, errors.New("invalid version")
+		binary.Read(r, c.bo, &c.V2)
 	}
 
 	model := newGGUFModel(c)
@@ -209,75 +208,75 @@ func (llm *ggufModel) NumLayers() int64 {
 	return int64(v)
 }
 
-func (ggufModel) readU8(r io.Reader) uint8 {
+func (llm ggufModel) readU8(r io.Reader) uint8 {
 	var u8 uint8
-	binary.Read(r, binary.LittleEndian, &u8)
+	binary.Read(r, llm.bo, &u8)
 	return u8
 }
 
-func (ggufModel) readI8(r io.Reader) int8 {
+func (llm ggufModel) readI8(r io.Reader) int8 {
 	var i8 int8
-	binary.Read(r, binary.LittleEndian, &i8)
+	binary.Read(r, llm.bo, &i8)
 	return i8
 }
 
-func (ggufModel) readU16(r io.Reader) uint16 {
+func (llm ggufModel) readU16(r io.Reader) uint16 {
 	var u16 uint16
-	binary.Read(r, binary.LittleEndian, &u16)
+	binary.Read(r, llm.bo, &u16)
 	return u16
 }
 
-func (ggufModel) readI16(r io.Reader) int16 {
+func (llm ggufModel) readI16(r io.Reader) int16 {
 	var i16 int16
-	binary.Read(r, binary.LittleEndian, &i16)
+	binary.Read(r, llm.bo, &i16)
 	return i16
 }
 
-func (ggufModel) readU32(r io.Reader) uint32 {
+func (llm ggufModel) readU32(r io.Reader) uint32 {
 	var u32 uint32
-	binary.Read(r, binary.LittleEndian, &u32)
+	binary.Read(r, llm.bo, &u32)
 	return u32
 }
 
-func (ggufModel) readI32(r io.Reader) int32 {
+func (llm ggufModel) readI32(r io.Reader) int32 {
 	var i32 int32
-	binary.Read(r, binary.LittleEndian, &i32)
+	binary.Read(r, llm.bo, &i32)
 	return i32
 }
 
-func (ggufModel) readU64(r io.Reader) uint64 {
+func (llm ggufModel) readU64(r io.Reader) uint64 {
 	var u64 uint64
-	binary.Read(r, binary.LittleEndian, &u64)
+	binary.Read(r, llm.bo, &u64)
 	return u64
 }
 
-func (ggufModel) readI64(r io.Reader) int64 {
+func (llm ggufModel) readI64(r io.Reader) int64 {
 	var i64 int64
-	binary.Read(r, binary.LittleEndian, &i64)
+	binary.Read(r, llm.bo, &i64)
 	return i64
 }
 
-func (ggufModel) readF32(r io.Reader) float32 {
+func (llm ggufModel) readF32(r io.Reader) float32 {
 	var f32 float32
-	binary.Read(r, binary.LittleEndian, &f32)
+	binary.Read(r, llm.bo, &f32)
 	return f32
 }
 
-func (ggufModel) readF64(r io.Reader) float64 {
+func (llm ggufModel) readF64(r io.Reader) float64 {
 	var f64 float64
-	binary.Read(r, binary.LittleEndian, &f64)
+	binary.Read(r, llm.bo, &f64)
 	return f64
 }
 
-func (ggufModel) readBool(r io.Reader) bool {
+func (llm ggufModel) readBool(r io.Reader) bool {
 	var b bool
-	binary.Read(r, binary.LittleEndian, &b)
+	binary.Read(r, llm.bo, &b)
 	return b
 }
 
-func (ggufModel) readStringV1(r io.Reader) (string, error) {
+func (llm ggufModel) readStringV1(r io.Reader) (string, error) {
 	var nameLength uint32
-	binary.Read(r, binary.LittleEndian, &nameLength)
+	binary.Read(r, llm.bo, &nameLength)
 
 	var b bytes.Buffer
 	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {
@@ -292,7 +291,7 @@ func (ggufModel) readStringV1(r io.Reader) (string, error) {
 
 func (llm ggufModel) readString(r io.Reader) (string, error) {
 	var nameLength uint64
-	binary.Read(r, binary.LittleEndian, &nameLength)
+	binary.Read(r, llm.bo, &nameLength)
 
 	var b bytes.Buffer
 	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {

llm/llama.cpp/generate_darwin_amd64.go

@@ -12,7 +12,8 @@ package llm
 //go:generate mv ggml/build/cpu/bin/server ggml/build/cpu/bin/ollama-runner
 
 //go:generate git submodule update --force gguf
-//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
+//go:generate git -C gguf apply ../patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner

llm/llama.cpp/generate_darwin_arm64.go

@@ -12,7 +12,8 @@ package llm
 //go:generate mv ggml/build/metal/bin/server ggml/build/metal/bin/ollama-runner
 
 //go:generate git submodule update --force gguf
-//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
+//go:generate git -C gguf apply ../patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch
 //go:generate cmake -S gguf -B gguf/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/metal --target server --config Release
 //go:generate mv gguf/build/metal/bin/server gguf/build/metal/bin/ollama-runner

llm/llama.cpp/generate_linux.go

@@ -13,7 +13,7 @@ package llm
 
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-copy-cuda-runtime-libraries.patch
-//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner

llm/llama.cpp/generate_windows.go

@@ -10,7 +10,7 @@ package llm
 //go:generate cmd /c move ggml\build\cpu\bin\Release\server.exe ggml\build\cpu\bin\Release\ollama-runner.exe
 
 //go:generate git submodule update --force gguf
-//go:generate git -C gguf apply ../patches/0001-remove-warm-up-logging.patch
+//go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate cmd /c move gguf\build\cpu\bin\Release\server.exe gguf\build\cpu\bin\Release\ollama-runner.exe

llm/llama.cpp/patches/0001-metal-handle-ggml_scale-for-n-4-0-close-3754.patch

@@ -0,0 +1,91 @@
+From 469c9addef75893e6be12edda852d12e840bf064 Mon Sep 17 00:00:00 2001
+From: Georgi Gerganov <ggerganov@gmail.com>
+Date: Tue, 24 Oct 2023 09:46:50 +0300
+Subject: [PATCH 1/2] metal : handle ggml_scale for n%4 != 0 (close #3754)
+
+ggml-ci
+---
+ ggml-metal.m     | 18 +++++++++++++-----
+ ggml-metal.metal | 10 +++++++++-
+ 2 files changed, 22 insertions(+), 6 deletions(-)
+
+diff --git a/ggml-metal.m b/ggml-metal.m
+index c908106..c1901dc 100644
+--- a/ggml-metal.m
++++ b/ggml-metal.m
+@@ -62,6 +62,7 @@
+     GGML_METAL_DECL_KERNEL(mul);
+     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+     GGML_METAL_DECL_KERNEL(scale);
++    GGML_METAL_DECL_KERNEL(scale_4);
+     GGML_METAL_DECL_KERNEL(silu);
+     GGML_METAL_DECL_KERNEL(relu);
+     GGML_METAL_DECL_KERNEL(gelu);
+@@ -249,6 +250,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){
+         GGML_METAL_ADD_KERNEL(mul);
+         GGML_METAL_ADD_KERNEL(mul_row);
+         GGML_METAL_ADD_KERNEL(scale);
++        GGML_METAL_ADD_KERNEL(scale_4);
+         GGML_METAL_ADD_KERNEL(silu);
+         GGML_METAL_ADD_KERNEL(relu);
+         GGML_METAL_ADD_KERNEL(gelu);
+@@ -347,6 +349,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
+     GGML_METAL_DEL_KERNEL(mul);
+     GGML_METAL_DEL_KERNEL(mul_row);
+     GGML_METAL_DEL_KERNEL(scale);
++    GGML_METAL_DEL_KERNEL(scale_4);
+     GGML_METAL_DEL_KERNEL(silu);
+     GGML_METAL_DEL_KERNEL(relu);
+     GGML_METAL_DEL_KERNEL(gelu);
+@@ -923,15 +926,20 @@ void ggml_metal_graph_compute(
+ 
+                             const float scale = *(const float *) src1->data;
+ 
+-                            [encoder setComputePipelineState:ctx->pipeline_scale];
++                            int64_t n = ggml_nelements(dst);
++
++                            if (n % 4 == 0) {
++                                n /= 4;
++                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
++                            } else {
++                                [encoder setComputePipelineState:ctx->pipeline_scale];
++                            }
++
+                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+ 
+-                            const int64_t n = ggml_nelements(dst);
+-                            GGML_ASSERT(n % 4 == 0);
+-
+-                            [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
++                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                         } break;
+                     case GGML_OP_UNARY:
+                         switch (ggml_get_unary_op(gf->nodes[i])) {
+diff --git a/ggml-metal.metal b/ggml-metal.metal
+index 69fc713..f4b4605 100644
+--- a/ggml-metal.metal
++++ b/ggml-metal.metal
+@@ -125,9 +125,17 @@ kernel void kernel_mul_row(
+ }
+ 
+ kernel void kernel_scale(
++        device const float * src0,
++        device       float * dst,
++        constant     float & scale,
++        uint tpig[[thread_position_in_grid]]) {
++    dst[tpig] = src0[tpig] * scale;
++}
++
++kernel void kernel_scale_4(
+         device const float4 * src0,
+         device       float4 * dst,
+-        constant     float & scale,
++        constant     float  & scale,
+         uint tpig[[thread_position_in_grid]]) {
+     dst[tpig] = src0[tpig] * scale;
+ }
+-- 
+2.39.3 (Apple Git-145)
+

llm/llama.cpp/patches/0001-remove-warm-up-logging.patch

@@ -1,25 +0,0 @@
-From 8dbb5449db259a9c24796e7927d89bee98b6c8f5 Mon Sep 17 00:00:00 2001
-From: Bruce MacDonald <brucewmacdonald@gmail.com>
-Date: Thu, 5 Oct 2023 11:21:12 -0400
-Subject: [PATCH] remove warm up logging
-
----
- common/common.cpp | 2 --
- 1 file changed, 2 deletions(-)
-
-diff --git a/common/common.cpp b/common/common.cpp
-index 7370017..c4433fe 100644
---- a/common/common.cpp
-+++ b/common/common.cpp
-@@ -839,8 +839,6 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
-     }
- 
-     {
--        LOG("warming up the model with an empty run\n");
--
-         std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
-         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-         llama_kv_cache_tokens_rm(lctx, -1, -1);
--- 
-2.39.2 (Apple Git-143)
-

llm/llama.cpp/patches/0001-update-default-log-target.patch

@@ -0,0 +1,25 @@
+From 6465fec6290f0a7f5d4d0fbe6bcf634e4810dde6 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 23 Oct 2023 10:39:34 -0700
+Subject: [PATCH] default log stderr
+
+---
+ common/log.h | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/common/log.h b/common/log.h
+index b8953fd..25522cd 100644
+--- a/common/log.h
++++ b/common/log.h
+@@ -90,7 +90,7 @@
+ //  }
+ //
+ #ifndef LOG_TARGET
+-    #define LOG_TARGET log_handler()
++    #define LOG_TARGET nullptr
+ #endif
+ 
+ #ifndef LOG_TEE_TARGET
+-- 
+2.42.0
+

scripts/install.sh

@@ -74,6 +74,9 @@ configure_systemd() {
         $SUDO useradd -r -s /bin/false -m -d /usr/share/ollama ollama
     fi
 
+    status "Adding current user to ollama group..."
+    $SUDO usermod -a -G ollama $(whoami)
+
     status "Creating ollama systemd service..."
     cat <<EOF | $SUDO tee /etc/systemd/system/ollama.service >/dev/null
 [Unit]