llm: update llama.cpp submodule to 7c26775

2024-06-17 13:46:02 -04:00
58 changed files with 850 additions and 1930 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -437,7 +437,6 @@ jobs:
    env:
      OLLAMA_SKIP_IMAGE_BUILD: '1'
      PUSH: '1'
-      GH_TOKEN: ${{ github.token }}
    steps:
      - uses: actions/checkout@v4
      - name: Set Version
@@ -461,20 +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
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -73,12 +73,12 @@ jobs:
          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
          $env:PATH="$gopath;$gccpath;$env:PATH"
          echo $env:PATH
-          $env:GOARCH=""; $env:OLLAMA_BUILD_TARGET_ARCH="${{ matrix.arch }}"; go generate -x ./...
+          go generate -x ./...
        if: ${{ startsWith(matrix.os, 'windows-') }}
-        name: 'Windows Generate'
-      - run: GOARCH= OLLAMA_BUILD_TARGET_ARCH=${{ matrix.arch }} go generate -x ./...
+        name: 'Windows Go Generate'
+      - run: go generate -x ./...
        if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        name: 'Unix Generate'
+        name: 'Unix Go Generate'
      - uses: actions/upload-artifact@v4
        with:
          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
@@ -184,7 +184,7 @@ jobs:
          $env:OLLAMA_SKIP_CPU_GENERATE="1"
          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
          go generate -x ./...
-        name: go generate -x ./...
+        name: go generate
        env:
          OLLAMA_SKIP_CPU_GENERATE: '1'
      # TODO - do we need any artifacts?
@@ -217,7 +217,7 @@ jobs:
      - name: 'Verify CUDA'
        run: nvcc -V
      - run: go get ./...
-      - name: go generate -x ./...
+      - name: go generate
        run: |
          $gopath=(get-command go).source | split-path -parent
          $cudabin=(get-command nvcc).source | split-path
@@ -312,10 +312,7 @@ jobs:
          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
        if: ${{ startsWith(matrix.os, 'macos-') }}
        shell: bash
-      - run: $env:GOARCH=""; $env:OLLAMA_BUILD_TARGET_ARCH="${{ matrix.arch }}"; go generate -x ./...
-        if: ${{ startsWith(matrix.os, 'windows-') }}
-      - run: GOARCH= OLLAMA_BUILD_TARGET_ARCH=${{ matrix.arch }} go generate -x ./...
-        if: ${{ ! startsWith(matrix.os, 'windows-') }}
+      - run: go generate ./...
      - run: go build
      - run: go test -v ./...
      - uses: actions/upload-artifact@v4
--- a/README.md
+++ b/README.md
@@ -1,12 +1,12 @@
 <div align="center">
-  <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+ <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>

 # Ollama

 [![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)

-Get up and running with large language models locally.
+Get up and running with large language models.

 ### macOS

@@ -51,15 +51,17 @@ Here are some example models that can be downloaded:
 | ------------------ | ---------- | ----- | ------------------------------ |
 | Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
 | Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
-| Phi-3              | 3,8B       | 2.3GB | `ollama run phi3`              |
+| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
+| Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
+| 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`       |
 | Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
-| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
-| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Solar              | 10.7B      | 6.1GB | `ollama run solar`             |

 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -173,7 +175,7 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 The image features a yellow smiley face, which is likely the central focus of the picture.
 ```

-### Pass in prompt as arguments
+### Pass the prompt as an argument

 ```
 $ ollama run llama3 "Summarize this file: $(cat README.md)"
@@ -192,19 +194,7 @@ ollama list

 ## Building

-Install `cmake` and `go`:
-
-```
-brew install cmake go
-```
-
-Then build the binary:
-
-```
-go run build.go
-```
-
-More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
+See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)

 ### Running local builds

@@ -252,6 +242,7 @@ See the [API documentation](./docs/api.md) for all endpoints.

 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
+- [Hollama](https://github.com/fmaclen/hollama)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
 - [LibreChat](https://github.com/danny-avila/LibreChat)
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
@@ -278,17 +269,23 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [OllamaGUI](https://github.com/enoch1118/ollamaGUI)
 - [OpenAOE](https://github.com/InternLM/OpenAOE)
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
- [LLM-X: Progressive Web App](https://github.com/mrdjohnson/llm-x)
+- [LLM-X](https://github.com/mrdjohnson/llm-x) (Progressive Web App)
 - [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
 - [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
- [QA-Pilot: Chat with Code Repository](https://github.com/reid41/QA-Pilot)
- [ChatOllama: Open Source Chatbot based on Ollama with Knowledge Bases](https://github.com/sugarforever/chat-ollama)
- [CRAG Ollama Chat: Simple Web Search with Corrective RAG](https://github.com/Nagi-ovo/CRAG-Ollama-Chat)
- [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
- [chat: chat web app for teams](https://github.com/swuecho/chat)
+- [QA-Pilot](https://github.com/reid41/QA-Pilot) (Chat with Code Repository)
+- [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
+- [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
+- [RAGFlow](https://github.com/infiniflow/ragflow) (Open-source Retrieval-Augmented Generation engine based on deep document understanding)
+- [StreamDeploy](https://github.com/StreamDeploy-DevRel/streamdeploy-llm-app-scaffold) (LLM Application Scaffold)
+- [chat](https://github.com/swuecho/chat) (chat web app for teams)
 - [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
- [Ollama RAG Chatbot: Local Chat with multiples PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)
+- [Ollama RAG Chatbot](https://github.com/datvodinh/rag-chatbot.git) (Local Chat with multiple PDFs using Ollama and RAG)
+- [BrainSoup](https://www.nurgo-software.com/products/brainsoup) (Flexible native client with RAG & multi-agent automation)
+- [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)

 ### Terminal

@@ -311,6 +308,7 @@ 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

@@ -321,17 +319,20 @@ See the [API documentation](./docs/api.md) for all endpoints.

 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
+- [Guix channel](https://codeberg.org/tusharhero/ollama-guix)

 ### Libraries

 - [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)
@@ -342,9 +343,13 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Haystack](https://github.com/deepset-ai/haystack-integrations/blob/main/integrations/ollama.md)
 - [Elixir LangChain](https://github.com/brainlid/langchain)
 - [Ollama for R - rollama](https://github.com/JBGruber/rollama)
+- [Ollama for R - ollama-r](https://github.com/hauselin/ollama-r)
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
 - [Testcontainers](https://testcontainers.com/modules/ollama/)
+- [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

@@ -364,18 +369,23 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama Telegram Bot](https://github.com/ruecat/ollama-telegram)
 - [Hass Ollama Conversation](https://github.com/ej52/hass-ollama-conversation)
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
- [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
 - [Cliobot](https://github.com/herval/cliobot) (Telegram bot with Ollama support)
 - [Copilot for Obsidian plugin](https://github.com/logancyang/obsidian-copilot)
 - [Obsidian Local GPT plugin](https://github.com/pfrankov/obsidian-local-gpt)
 - [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
+- [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
+- [Ollama Copilot](https://github.com/bernardo-bruning/ollama-copilot) (Proxy that allows you to use ollama as a copilot like Github copilot)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
 - [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)
 - [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
+
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.

-### Supported backends 
- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 
--- a/api/types.go
+++ b/api/types.go
@@ -159,49 +159,18 @@ 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   TriState `json:"use_mmap,omitempty"`
-	UseMLock  bool     `json:"use_mlock,omitempty"`
-	NumThread int      `json:"num_thread,omitempty"`
-}
-
-type TriState int
-
-const (
-	TriStateUndefined TriState = -1
-	TriStateFalse     TriState = 0
-	TriStateTrue      TriState = 1
-)
-
-func (b *TriState) UnmarshalJSON(data []byte) error {
-	var v bool
-	if err := json.Unmarshal(data, &v); err != nil {
-		return err
-	}
-	if v {
-		*b = TriStateTrue
-	}
-	*b = TriStateFalse
-	return nil
-}
-
-func (b *TriState) MarshalJSON() ([]byte, error) {
-	if *b == TriStateUndefined {
-		return nil, nil
-	}
-	var v bool
-	if *b == TriStateTrue {
-		v = true
-	}
-	return json.Marshal(v)
+	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].
@@ -253,7 +222,6 @@ 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"`

@@ -263,16 +231,14 @@ 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"`
+	ModifiedAt time.Time    `json:"modified_at,omitempty"`
 }

 // CopyRequest is the request passed to [Client.Copy].
@@ -437,19 +403,6 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 				continue
 			}

-			if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
-				val, ok := val.(bool)
-				if !ok {
-					return fmt.Errorf("option %q must be of type boolean", key)
-				}
-				if val {
-					field.SetInt(int64(TriStateTrue))
-				} else {
-					field.SetInt(int64(TriStateFalse))
-				}
-				continue
-			}
-
 			switch field.Kind() {
 			case reflect.Int:
 				switch t := val.(type) {
@@ -538,7 +491,7 @@ func DefaultOptions() Options {
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   TriStateUndefined,
+			UseMMap:   true,
 			UseNUMA:   false,
 		},
 	}
@@ -608,19 +561,6 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 		} else {
 			field := valueOpts.FieldByName(opt.Name)
 			if field.IsValid() && field.CanSet() {
-				if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
-					boolVal, err := strconv.ParseBool(vals[0])
-					if err != nil {
-						return nil, fmt.Errorf("invalid bool value %s", vals)
-					}
-					if boolVal {
-						out[key] = TriStateTrue
-					} else {
-						out[key] = TriStateFalse
-					}
-					continue
-				}
-
 				switch field.Kind() {
 				case reflect.Float32:
 					floatVal, err := strconv.ParseFloat(vals[0], 32)
--- a/api/types_test.go
+++ b/api/types_test.go
@@ -2,7 +2,6 @@ package api

 import (
 	"encoding/json"
-	"fmt"
 	"math"
 	"testing"
 	"time"
@@ -106,101 +105,3 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
 		})
 	}
 }
-
-func TestUseMmapParsingFromJSON(t *testing.T) {
-	tests := []struct {
-		name string
-		req  string
-		exp  TriState
-	}{
-		{
-			name: "Undefined",
-			req:  `{ }`,
-			exp:  TriStateUndefined,
-		},
-		{
-			name: "True",
-			req:  `{ "use_mmap": true }`,
-			exp:  TriStateTrue,
-		},
-		{
-			name: "False",
-			req:  `{ "use_mmap": false }`,
-			exp:  TriStateFalse,
-		},
-	}
-
-	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) {
-	tests := []struct {
-		name string
-		req  map[string][]string
-		exp  TriState
-		err  error
-	}{
-		{
-			name: "True",
-			req: map[string][]string{
-				"use_mmap": []string{"true"},
-			},
-			exp: TriStateTrue,
-			err: nil,
-		},
-		{
-			name: "False",
-			req: map[string][]string{
-				"use_mmap": []string{"false"},
-			},
-			exp: TriStateFalse,
-			err: nil,
-		},
-		{
-			name: "Numeric True",
-			req: map[string][]string{
-				"use_mmap": []string{"1"},
-			},
-			exp: TriStateTrue,
-			err: nil,
-		},
-		{
-			name: "Numeric False",
-			req: map[string][]string{
-				"use_mmap": []string{"0"},
-			},
-			exp: TriStateFalse,
-			err: nil,
-		},
-		{
-			name: "invalid string",
-			req: map[string][]string{
-				"use_mmap": []string{"foo"},
-			},
-			exp: TriStateUndefined,
-			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, err, test.err)
-			respVal, ok := resp["use_mmap"]
-			if test.exp != TriStateUndefined {
-				assert.True(t, ok, "resp: %v", resp)
-				assert.Equal(t, test.exp, respVal)
-			}
-		})
-	}
-}
--- a/app/lifecycle/logging.go
+++ b/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)
-			}
-		}
-	}
-}
--- a/app/lifecycle/logging_test.go
+++ b/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"))
-	}
-}
--- a/app/lifecycle/paths.go
+++ b/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() {
--- a/app/lifecycle/server.go
+++ b/app/lifecycle/server.go
@@ -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)
--- a/app/ollama.iss
+++ b/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
--- a/build.go
+++ b/build.go
@@ -1,199 +0,0 @@
-//go:build ignore
-
-package main
-
-import (
-	"cmp"
-	"errors"
-	"flag"
-	"log"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"runtime"
-)
-
-// Flags
-var (
-	flagRegenerateDestroy = flag.Bool("d", false, "force regenerate the dependencies (destructive)")
-	flagRegenerateGently  = flag.Bool("g", false, "regenerate the dependencies (non-destructive)")
-	flagSkipBuild         = flag.Bool("s", false, "generate dependencies only (e.g. skip 'go build .')")
-
-	// Flags to set GOARCH explicitly for cross-platform builds,
-	// e.g., in CI to target a different platform than the build matrix
-	// default. These allows us to run generate without a separate build
-	// step for building the script binary for the host ARCH and then
-	// runing the generate script for the target ARCH. Instead, we can
-	// just run `go run build.go -target=$GOARCH` to generate the
-	// deps.
-	flagGOARCH = flag.String("target", "", "sets GOARCH to use when generating dependencies and building")
-)
-
-func buildEnv() []string {
-	return append(os.Environ(), "GOARCH="+cmp.Or(
-		*flagGOARCH,
-		os.Getenv("OLLAMA_BUILD_TARGET_ARCH"),
-		runtime.GOARCH,
-	))
-}
-
-func main() {
-	log.SetFlags(0)
-	flag.Usage = func() {
-		log.Printf("Usage: go run build.go [flags]")
-		log.Println()
-		log.Println("Flags:")
-		flag.PrintDefaults()
-		log.Println()
-		log.Println("This script builds the Ollama server binary and generates the llama.cpp")
-		log.Println("bindings for the current platform. It assumes that the current working")
-		log.Println("directory is the root directory of the Ollama project.")
-		log.Println()
-		log.Println("If the -d flag is provided, the script will force regeneration of the")
-		log.Println("dependencies; removing the 'llm/build' directory before starting.")
-		log.Println()
-		log.Println("If the -g flag is provided, the script will regenerate the dependencies")
-		log.Println("without removing the 'llm/build' directory.")
-		log.Println()
-		log.Println("If the -s flag is provided, the script will skip building the Ollama binary")
-		log.Println()
-		log.Println("If the -target flag is provided, the script will set GOARCH to the value")
-		log.Println("of the flag. This is useful for cross-platform builds.")
-		log.Println()
-		log.Println("The script will check for the required dependencies (cmake, gcc) and")
-		log.Println("print their version.")
-		log.Println()
-		log.Println("The script will also check if it is being run from the root directory of")
-		log.Println("the Ollama project.")
-		log.Println()
-		os.Exit(1)
-	}
-	flag.Parse()
-
-	log.Printf("=== Building Ollama ===")
-	defer func() {
-		log.Printf("=== Done building Ollama ===")
-		if !*flagSkipBuild {
-			log.Println()
-			log.Println("To run the Ollama server, use:")
-			log.Println()
-			log.Println("    ./ollama serve")
-			log.Println()
-		}
-	}()
-
-	if flag.NArg() > 0 {
-		flag.Usage()
-	}
-
-	if !inRootDir() {
-		log.Fatalf("Please run this script from the root directory of the Ollama project.")
-	}
-
-	if err := checkDependencies(); err != nil {
-		log.Fatalf("Failed dependency check: %v", err)
-	}
-	if err := buildLlammaCPP(); err != nil {
-		log.Fatalf("Failed to build llama.cpp: %v", err)
-	}
-	if err := goBuildOllama(); err != nil {
-		log.Fatalf("Failed to build ollama Go binary: %v", err)
-	}
-}
-
-// checkDependencies does a quick check to see if the required dependencies are
-// installed on the system and functioning enough to print their version.
-//
-// TODO(bmizerany): Check the actual version of the dependencies? Seems a
-// little daunting given diff versions might print diff things. This should
-// be good enough for now.
-func checkDependencies() error {
-	var err error
-	check := func(name string, args ...string) {
-		log.Printf("=== Checking for %s ===", name)
-		defer log.Printf("=== Done checking for %s ===\n\n", name)
-		cmd := exec.Command(name, args...)
-		cmd.Stdout = os.Stdout
-		cmd.Stderr = os.Stderr
-		err = errors.Join(err, cmd.Run())
-	}
-
-	check("cmake", "--version")
-	check("gcc", "--version")
-	return err
-}
-
-func goBuildOllama() error {
-	log.Println("=== Building Ollama binary ===")
-	defer log.Printf("=== Done building Ollama binary ===\n\n")
-	if *flagSkipBuild {
-		log.Println("Skipping 'go build -o ollama .'")
-		return nil
-	}
-	cmd := exec.Command("go", "build", "-o", "ollama", ".")
-	cmd.Stdout = os.Stdout
-	cmd.Stderr = os.Stderr
-	cmd.Env = buildEnv()
-	return cmd.Run()
-}
-
-// buildLlammaCPP generates the llama.cpp bindings for the current platform.
-//
-// It assumes that the current working directory is the root directory of the
-// Ollama project.
-func buildLlammaCPP() error {
-	log.Println("=== Generating dependencies ===")
-	defer log.Printf("=== Done generating dependencies ===\n\n")
-	if *flagRegenerateDestroy {
-		if err := os.RemoveAll(filepath.Join("llm", "build")); err != nil {
-			return err
-		}
-	}
-	if isDirectory(filepath.Join("llm", "build")) && !*flagRegenerateGently {
-		log.Println("llm/build already exists; skipping.  Use -d or -g to re-generate.")
-		return nil
-	}
-
-	scriptDir, err := filepath.Abs(filepath.Join("llm", "generate"))
-	if err != nil {
-		return err
-	}
-
-	var cmd *exec.Cmd
-	switch runtime.GOOS {
-	case "windows":
-		script := filepath.Join(scriptDir, "gen_windows.ps1")
-		cmd = exec.Command("powershell", "-ExecutionPolicy", "Bypass", "-File", script)
-	case "linux":
-		script := filepath.Join(scriptDir, "gen_linux.sh")
-		cmd = exec.Command("bash", script)
-	case "darwin":
-		script := filepath.Join(scriptDir, "gen_darwin.sh")
-		cmd = exec.Command("bash", script)
-	default:
-		log.Fatalf("Unsupported OS: %s", runtime.GOOS)
-	}
-	cmd.Dir = filepath.Join("llm", "generate")
-	cmd.Stdout = os.Stdout
-	cmd.Stderr = os.Stderr
-	cmd.Env = buildEnv()
-
-	log.Printf("Running GOOS=%s GOARCH=%s %s", runtime.GOOS, runtime.GOARCH, cmd.Args)
-
-	return cmd.Run()
-}
-
-func isDirectory(path string) bool {
-	info, err := os.Stat(path)
-	if err != nil {
-		return false
-	}
-	return info.IsDir()
-}
-
-// inRootDir returns true if the current working directory is the root
-// directory of the Ollama project. It looks for a file named "go.mod".
-func inRootDir() bool {
-	_, err := os.Stat("go.mod")
-	return err == nil
-}
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -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 {
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@@ -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.")
--- a/docs/api.md
+++ b/docs/api.md
@@ -26,7 +26,7 @@ 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

@@ -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`
  }
 }
 ```
--- a/docs/development.md
+++ b/docs/development.md
@@ -25,7 +25,13 @@ export OLLAMA_DEBUG=1
 Get the required libraries and build the native LLM code:

 ```bash
-go run build.go
+go generate ./...
+```
+
+Then build ollama:
+
+```bash
+go build .
 ```

 Now you can run `ollama`:
@@ -34,16 +40,6 @@ Now you can run `ollama`:
 ./ollama
 ```

-### Rebuilding the native code
-
-If at any point you need to rebuild the native code, you can run the
-build.go script again using the `-f` flag to force a rebuild, and,
-optionally, the `-d` flag to skip building the Go binary:
-
-```bash
-go run build.go -d -s
-```
-
 ### Linux

 #### Linux CUDA (NVIDIA)
@@ -59,10 +55,16 @@ specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
 libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
 a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")

+Then generate dependencies:
+
+```
+go generate ./...
+```
+
 Then build the binary:

 ```
-go run build.go
+go build .
 ```

 #### Linux ROCm (AMD)
@@ -78,17 +80,21 @@ install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
 CLBlast install (typically `/usr/lib/cmake/CLBlast`). You can also customize
 the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)

+```
+go generate ./...
+```
+
 Then build the binary:

 ```
-go run build.go
+go build .
 ```

 ROCm requires elevated privileges to access the GPU at runtime. On most distros you can add your user account to the `render` group, or run as root.

 #### Advanced CPU Settings

-By default, running `go run build.go` will compile a few different variations
+By default, running `go generate ./...` will compile a few different variations
 of the LLM library based on common CPU families and vector math capabilities,
 including a lowest-common-denominator which should run on almost any 64 bit CPU
 somewhat slowly. At runtime, Ollama will auto-detect the optimal variation to
@@ -98,7 +104,8 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:

 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go run build.go
+OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
+go build .
 ```

 #### Containerized Linux Build
@@ -122,7 +129,8 @@ Then, build the `ollama` binary:

 ```powershell
 $env:CGO_ENABLED="1"
-go run build.go
+go generate ./...
+go build .
 ```

 #### Windows CUDA (NVIDIA)
--- a/docs/gpu.md
+++ b/docs/gpu.md
@@ -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`                                                                                 |
--- a/docs/import.md
+++ b/docs/import.md
@@ -47,13 +47,19 @@ success

 ### Supported Quantizations

+<details>
+<summary>Legacy Quantization</summary>
+
 - `Q4_0`
 - `Q4_1`
 - `Q5_0`
 - `Q5_1`
 - `Q8_0`

-#### K-means Quantizations
+</details>
+
+<details>
+<summary>K-means Quantization</summary>`

 - `Q3_K_S`
 - `Q3_K_M`
@@ -64,6 +70,11 @@ success
 - `Q5_K_M`
 - `Q6_K`

+</details>
+
+> [!NOTE]
+> Activation-aware Weight Quantization (i.e. IQ) are not currently supported for automatic quantization however you can still import the quantized model into Ollama, see [Import GGUF](#import-gguf).
+
 ## Template Detection

 > [!NOTE]
--- a/docs/openai.md
+++ b/docs/openai.md
@@ -104,6 +104,7 @@ curl http://localhost:11434/v1/chat/completions \

 #### Notes

+- `finish_reason` will always be `stop`
 - `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached

 ## Models
--- a/docs/troubleshooting.md
+++ b/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
--- a/docs/windows.md
+++ b/docs/windows.md
@@ -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
--- a/gpu/assets.go
+++ b/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(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)
 		}
 	}
 }
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -231,7 +231,7 @@ func GetGPUInfo() GpuInfoList {
 		// 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")
+			depPath = filepath.Dir(envconfig.RunnersDir)
 		}

 		// Load ALL libraries
@@ -282,12 +282,6 @@ func GetGPUInfo() GpuInfoList {
 		// 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
@@ -312,7 +306,7 @@ func GetGPUInfo() GpuInfoList {
 					gpuInfo.FreeMemory = uint64(memInfo.free)
 					gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 					gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-					gpuInfo.DependencyPath = depPath
+					// TODO dependency path?
 					oneapiGPUs = append(oneapiGPUs, gpuInfo)
 				}
 			}
--- a/gpu/gpu_info_cudart.c
+++ b/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);
--- a/gpu/gpu_info_nvcuda.c
+++ b/gpu/gpu_info_nvcuda.c
@@ -43,7 +43,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_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);
--- a/gpu/gpu_info_nvml.c
+++ b/gpu/gpu_info_nvml.c
@@ -42,7 +42,7 @@ void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
    // 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)) {
+    if (!l[i].p) {
      resp->ch.handle = NULL;
      char *msg = LOAD_ERR();
      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
--- a/gpu/gpu_info_oneapi.c
+++ b/gpu/gpu_info_oneapi.c
@@ -50,7 +50,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
    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);
@@ -98,7 +98,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
  }

  for (d = 0; d < resp->oh.num_drivers; d++) {
-    LOG(resp->oh.verbose, "calling zesDeviceGet count %d: %p\n", d, resp->oh.drivers[d]);
+    LOG(resp->oh.verbose, "calling zesDeviceGet %d\n", resp->oh.drivers[d]);
    ret = (*resp->oh.zesDeviceGet)(resp->oh.drivers[d],
                                   &resp->oh.num_devices[d], NULL);
    if (ret != ZE_RESULT_SUCCESS) {
--- a/llm/ext_server/server.cpp
+++ b/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;
@@ -426,6 +427,16 @@ 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;
@@ -1650,41 +1661,26 @@ struct llama_server_context
                    }
                    slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);

-                    char buf[256];
-                    llama_model_meta_val_str(model, "general.architecture", buf, 256);
-                    bool gemma2 = strcmp(buf, "gemma2") == 0;
-
-                    int32_t truncate_at = slot.n_ctx;
-
-                    // truncate at 2/3 of the context length for gemma2 models
-                    // as they do not support context shifts (from the sliding window implementation).
-                    // this way, prompts that almost fit the context length can still generate a full
-                    // response without a sudden stop from hitting the context limit
-                    if (gemma2) {
-                        truncate_at = 2 * slot.n_ctx / 3;
-                    }
-
                    // if input prompt is too big, truncate it, if group attention self-extend is disabled
-                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= truncate_at)
+                    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;
@@ -1693,19 +1689,6 @@ struct llama_server_context
                        GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                    }

-                    // Models with sliding window attention do not work with context shifts, so
-                    // limit their prediction to the context length
-                    if (gemma2) {
-                        int32_t limit = slot.n_ctx - slot.n_prompt_tokens;
-                        slot.n_predict = limit;
-                        slot.params.n_predict = limit;
-                        LOG_INFO("model does not support sliding window, limiting generation", {
-                            {"n_ctx", slot.n_ctx},
-                            {"n_prompt_tokens", slot.n_prompt_tokens},
-                            {"n_predict", slot.n_predict}
-                        });
-                    }
-
                    if (!slot.params.cache_prompt)
                    {
                        llama_sampling_reset(slot.ctx_sampling);
@@ -2552,6 +2535,7 @@ 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")
        {
@@ -3024,6 +3008,11 @@ 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
--- a/llm/generate/gen_darwin.sh
+++ b/llm/generate/gen_darwin.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
-# This script is intended to run inside the `go run build.go` script, which
-# sets the working directory to the correct location: ./llm/generate/.
+# This script is intended to run inside the go generate
+# working directory must be ./llm/generate/

 # TODO - add hardening to detect missing tools (cmake, etc.)

@@ -18,7 +18,7 @@ sign() {
    fi
 }

-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 -DLLAMA_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 "${GOARCH}" in
 "amd64")
@@ -27,7 +27,7 @@ case "${GOARCH}" in
    # Static build for linking into the Go binary
    init_vars
    CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_BLAS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_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
@@ -37,7 +37,7 @@ case "${GOARCH}" in
        # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
        #
        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${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
@@ -49,7 +49,7 @@ case "${GOARCH}" in
        # Approximately 400% faster than LCD on same CPU
        #
        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${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
@@ -61,7 +61,7 @@ case "${GOARCH}" in
        # Approximately 10% faster than AVX on same CPU
        #
        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${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="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
@@ -75,7 +75,7 @@ case "${GOARCH}" in
    # Static build for linking into the Go binary
    init_vars
    CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_BLAS=off -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 ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${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
@@ -92,10 +92,10 @@ case "${GOARCH}" in
    ;;
 *)
    echo "GOARCH must be set"
-    echo "this script is meant to be run from within 'go run build.go'"
+    echo "this script is meant to be run from within go generate"
    exit 1
    ;;
 esac

 cleanup
-echo "code generation completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
+echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
-# This script is intended to run with the `go run build.go` script, which
-# sets the working directory to the correct location: ./llm/generate/.
+# This script is intended to run inside the go generate
+# working directory must be llm/generate/

 # First we build one or more CPU based LLM libraries
 #
@@ -51,7 +51,7 @@ if [ -z "${CUDACXX}" ]; then
        export CUDACXX=$(command -v nvcc)
    fi
 fi
-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 -DLLAMA_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 -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_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
@@ -93,7 +93,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake

-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_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)
@@ -178,7 +178,7 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
        echo "Building custom CUDA GPU"
    else
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        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${CUDA_VARIANT}"
@@ -281,4 +281,4 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
 fi

 cleanup
-echo "code generation completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
+echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@@ -1,7 +1,5 @@
 #!powershell

-$ErrorActionPreference = "Stop"
-
 function amdGPUs {
    if ($env:AMDGPU_TARGETS) {
        return $env:AMDGPU_TARGETS
@@ -26,15 +24,25 @@ function amdGPUs {
    $GPU_LIST -join ';'
 }

+
 function init_vars {
-    $script:SRC_DIR = $(resolve-path "..\..\")
-    $script:llamacppDir = "../llama.cpp"
+    if (!$script:SRC_DIR) {
+        $script:SRC_DIR = $(resolve-path "..\..\")
+    }
+    if (!$script:llamacppDir) {
+        $script:llamacppDir = "../llama.cpp"
+    }
+    if (!$script:cmakeTargets) {
+        $script:cmakeTargets = @("ollama_llama_server")
+    }
    $script:cmakeDefs = @(
        "-DBUILD_SHARED_LIBS=on",
        "-DLLAMA_NATIVE=off"
        )
-    $script:cmakeTargets = @("ollama_llama_server")
-    $script:ARCH = "amd64" # arm not yet supported.
+    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
+    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
+    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
+    md "$script:DIST_BASE" -ea 0 > $null
    if ($env:CGO_CFLAGS -contains "-g") {
        $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
        $script:config = "RelWithDebInfo"
@@ -55,7 +63,6 @@ function init_vars {
    } else {
        $script:CUDA_LIB_DIR=$env:CUDA_LIB_DIR
    }
-    $script:GZIP=(get-command -ea 'silentlycontinue' gzip).path
    $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
    if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
        $script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
@@ -76,9 +83,9 @@ function init_vars {
 function git_module_setup {
    # TODO add flags to skip the init/patch logic to make it easier to mod llama.cpp code in-repo
    & git submodule init
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
    & git submodule update --force "${script:llamacppDir}"
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
 }

 function apply_patches {
@@ -112,10 +119,15 @@ function build {
    write-host "generating config with: cmake -S ${script:llamacppDir} -B $script:buildDir $script:cmakeDefs"
    & cmake --version
    & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
-    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    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($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
+    if ($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", $_ }) $extra
+    if ($LASTEXITCODE -ne 0) { write-host "cmake build exit status $LASTEXITCODE"; throw($LASTEXITCODE)}
    # Rearrange output to be consistent between different generators
    if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
        mv -force "${script:buildDir}/bin/${script:config}/*" "${script:buildDir}/bin/"
@@ -129,26 +141,23 @@ function sign {
        foreach ($file in @(get-childitem "${script:buildDir}/bin/*.exe") + @(get-childitem "${script:buildDir}/bin/*.dll")){
            & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                /csp "Google Cloud KMS Provider" /kc "${env:KEY_CONTAINER}" $file
-            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
        }
    }
 }

-function compress {
-    if ($script:GZIP -eq $null) {
-        write-host "gzip not installed, not compressing files"
-        return
-    }
-    write-host "Compressing binaries..."
+function install {
+    write-host "Installing binaries to dist dir ${script:distDir}"
+    mkdir ${script:distDir} -ErrorAction SilentlyContinue
    $binaries = dir "${script:buildDir}/bin/*.exe"
    foreach ($file in $binaries) {
-        & "$script:GZIP" --best -f $file
+        copy-item -Path $file -Destination ${script:distDir} -Force
    }

-    write-host "Compressing dlls..."
+    write-host "Installing dlls to dist dir ${script:distDir}"
    $dlls = dir "${script:buildDir}/bin/*.dll"
    foreach ($file in $dlls) {
-        & "$script:GZIP" --best -f $file
+        copy-item -Path $file -Destination ${script:distDir} -Force
    }
 }

@@ -169,132 +178,266 @@ function cleanup {
    }
 }

-init_vars
-git_module_setup
-apply_patches

 # -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

-$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")

-if ($null -eq ${env:OLLAMA_SKIP_CPU_GENERATE}) {
-
-# GCC build for direct linking into the Go binary
-init_vars
-# cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
-# as we need this to be compiled by gcc for golang to be able to link with itx
-write-host "Checking for MinGW..."
-# error action ensures we exit on failure
-get-command gcc
-get-command mingw32-make
-$script:cmakeTargets = @("llama", "ggml")
-$script:cmakeDefs = @(
-    "-G", "MinGW Makefiles"
-    "-DCMAKE_C_COMPILER=gcc.exe",
-    "-DCMAKE_CXX_COMPILER=g++.exe",
-    "-DBUILD_SHARED_LIBS=off",
-    "-DLLAMA_NATIVE=off",
-    "-DLLAMA_AVX=off",
-    "-DLLAMA_AVX2=off",
-    "-DLLAMA_AVX512=off",
-    "-DLLAMA_F16C=off",
-    "-DLLAMA_FMA=off")
-$script:buildDir="../build/windows/${script:ARCH}_static"
-write-host "Building static library"
-build
-
-# remaining llama.cpp builds use MSVC 
-    init_vars
-    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
-    $script:buildDir="../build/windows/${script:ARCH}/cpu"
-    write-host "Building LCD CPU"
-    build
-    sign
-    compress
-
-    init_vars
-    $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"
-    write-host "Building AVX CPU"
-    build
-    sign
-    compress
-
-    init_vars
-    $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"
-    write-host "Building AVX2 CPU"
-    build
-    sign
-    compress
-} else {
-    write-host "Skipping CPU generation step as requested"
+function build_static() {
+    if ((-not "${env:OLLAMA_SKIP_STATIC_GENERATE}") -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "static"))) {
+        # GCC build for direct linking into the Go binary
+        init_vars
+        # cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
+        # as we need this to be compiled by gcc for golang to be able to link with itx
+        write-host "Checking for MinGW..."
+        # error action ensures we exit on failure
+        get-command gcc
+        get-command mingw32-make
+        $oldTargets = $script:cmakeTargets
+        $script:cmakeTargets = @("llama", "ggml")
+        $script:cmakeDefs = @(
+            "-G", "MinGW Makefiles"
+            "-DCMAKE_C_COMPILER=gcc.exe",
+            "-DCMAKE_CXX_COMPILER=g++.exe",
+            "-DBUILD_SHARED_LIBS=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
+        $script:cmakeTargets = $oldTargets
+    } else {
+        write-host "Skipping CPU generation step as requested"
+    }
 }

-if ($null -ne $script:CUDA_LIB_DIR) {
-    # Then build cuda as a dynamically loaded library
-    $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
-    $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
-    if ($null -ne $script:CUDA_VERSION) {
-        $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
+function build_cpu() {
+    if ($script:ARCH -eq "arm64") {
+        $gen_arch = "ARM64"
+    } else { # amd64
+        $gen_arch = "x64"
    }
-    init_vars
-    $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
-    $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUDA=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$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}")
-        write-host "building custom CUDA GPU"
+
+    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, "-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"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU generation step as requested"
    }
-    build
-    sign
-    compress
 }

-if ($null -ne $env:HIP_PATH) {
-    $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
-    if ($null -ne $script:ROCM_VERSION) {
-        $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
+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", "-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"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU AVX generation step as requested"
    }
+}

+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", "-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"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU AVX2 generation step as requested"
+    }
+}
+
+function build_cuda() {
+    if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${script:CUDA_LIB_DIR}")) {
+        # Then build cuda as a dynamically loaded library
+        $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
+        $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
+        if ($null -ne $script:CUDA_VERSION) {
+            $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
+        }
+        init_vars
+        $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
+        $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
+        $script:cmakeDefs += @(
+            "-A", "x64",
+            "-DLLAMA_CUDA=ON",
+            "-DLLAMA_AVX=on",
+            "-DLLAMA_AVX2=off",
+            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
+            "-DCMAKE_CUDA_FLAGS=-t8"
+            "-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}")
+            write-host "building custom CUDA GPU"
+        }
+        build
+        sign
+        install
+
+        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}"))  {
+    # 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
+    if ($null -ne $script:ONEAPI_VERSION) {
+      $script:ONEAPI_VARIANT = "_v" + $script:ONEAPI_VERSION
+    }
    init_vars
-    $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
+    $script:buildDir = "../build/windows/${script:ARCH}/oneapi$script:ONEAPI_VARIANT"
+    $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
    $script:cmakeDefs += @(
-        "-G", "Ninja", 
-        "-DCMAKE_C_COMPILER=clang.exe",
-        "-DCMAKE_CXX_COMPILER=clang++.exe",
-        "-DLLAMA_HIPBLAS=on",
-        "-DHIP_PLATFORM=amd",
-        "-DLLAMA_AVX=on",
-        "-DLLAMA_AVX2=off",
-        "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
-        "-DAMDGPU_TARGETS=$(amdGPUs)",
-        "-DGPU_TARGETS=$(amdGPUs)"
-        )
+      "-G", "MinGW Makefiles",
+      "-DLLAMA_SYCL=ON",
+      "-DCMAKE_C_COMPILER=icx",
+      "-DCMAKE_CXX_COMPILER=icx",
+      "-DCMAKE_BUILD_TYPE=Release"
+    )

-    # Make sure the ROCm binary dir is first in the path
-    $env:PATH="$env:HIP_PATH\bin;$env:PATH"
-
-    # We have to clobber the LIB var from the developer shell for clang to work properly
-    $env:LIB=""
-    if ($null -ne $env:OLLAMA_CUSTOM_ROCM_DEFS) {
-        write-host "OLLAMA_CUSTOM_ROCM_DEFS=`"${env:OLLAMA_CUSTOM_ROCM_DEFS}`""
-        $script:cmakeDefs += @("${env:OLLAMA_CUSTOM_ROCM_DEFS}")
-        write-host "building custom ROCM GPU"
-    }
-    write-host "Building ROCm"
+    Write-Host "Building oneAPI"
    build
    # Ninja doesn't prefix with config name
-    ${script:config}=""
    if ($null -ne $script:DUMPBIN) {
-        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
+      & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | Select-String ".dll"
    }
    sign
-    compress
+    install
+
+    cp "${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"
+  }
 }

+function build_rocm() {
+    if ((-not "${env:OLLAMA_SKIP_ROCM_GENERATE}") -and ("${env:HIP_PATH}")) {
+        $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
+        if ($null -ne $script:ROCM_VERSION) {
+            $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
+        }

-cleanup
-write-host "`code generation completed.  LLM runners: $(get-childitem -path ${script:SRC_DIR}\llm\build\windows\${script:ARCH})"
+        init_vars
+        $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
+        $script:distDir="$script:DIST_BASE\rocm$script:ROCM_VARIANT"
+        $script:cmakeDefs += @(
+            "-G", "Ninja", 
+            "-DCMAKE_C_COMPILER=clang.exe",
+            "-DCMAKE_CXX_COMPILER=clang++.exe",
+            "-DLLAMA_HIPBLAS=on",
+            "-DHIP_PLATFORM=amd",
+            "-DLLAMA_AVX=on",
+            "-DLLAMA_AVX2=off",
+            "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
+            "-DAMDGPU_TARGETS=$(amdGPUs)",
+            "-DGPU_TARGETS=$(amdGPUs)"
+            )
+
+        # Make sure the ROCm binary dir is first in the path
+        $env:PATH="$env:HIP_PATH\bin;$env:PATH"
+
+        # We have to clobber the LIB var from the developer shell for clang to work properly
+        $env:LIB=""
+        if ($null -ne $env:OLLAMA_CUSTOM_ROCM_DEFS) {
+            write-host "OLLAMA_CUSTOM_ROCM_DEFS=`"${env:OLLAMA_CUSTOM_ROCM_DEFS}`""
+            $script:cmakeDefs += @("${env:OLLAMA_CUSTOM_ROCM_DEFS}")
+            write-host "building custom ROCM GPU"
+        }
+        write-host "Building ROCm"
+        build
+        # Ninja doesn't prefix with config name
+        ${script:config}=""
+        if ($null -ne $script:DUMPBIN) {
+            & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
+        }
+        sign
+        install
+
+        # Assumes v5.7, may need adjustments for v6
+        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
+        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
+        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
+    } else {
+        write-host "Skipping ROCm generation step"
+    }
+}
+
+init_vars
+if ($($args.count) -eq 0) {
+    git_module_setup
+    apply_patches
+
+    $tasks = @("build_static", "build_cpu")
+    $jobs = @()
+    if ($script:ARCH -ne "arm64") {
+        $tasks += $("build_cpu_avx", "build_cpu_avx2", "build_cuda", "build_oneapi", "build_rocm")
+    }
+    foreach ($t in $tasks) {
+        $jobs += @(Start-ThreadJob -ThrottleLimit 12 -FilePath .\gen_windows.ps1 -ArgumentList $t -Name $t)
+    }
+    get-job
+    foreach ($job in $jobs) {
+        write-host "----" $job.Name output follows
+        receive-job -wait -job $job
+        write-host "----" $job.Name $job.State
+        write-host ""
+        if ($job.State -contains 'Failed') {
+            cleanup
+            write-host "Terminating remaining jobs (this takes a while, you can ^C)"
+            # TODO find some way to kill the spawned cmake processes faster
+            remove-job -force -job $jobs
+            exit(-1)
+        }
+        get-job
+    }
+
+    cleanup
+    write-host "`ngo generate completed.  LLM runners: $(get-childitem -path $script:DIST_BASE)"
+} else {
+    for ( $i = 0; $i -lt $args.count; $i++ ) {
+        write-host "performing $($args[$i])"
+        & $($args[$i])
+    } 
+}
--- a/llm/generate/generate_darwin.go
+++ b/llm/generate/generate_darwin.go
@@ -0,0 +1,3 @@
+package generate
+
+//go:generate bash ./gen_darwin.sh
--- a/llm/generate/generate_linux.go
+++ b/llm/generate/generate_linux.go
@@ -0,0 +1,3 @@
+package generate
+
+//go:generate bash ./gen_linux.sh
--- a/llm/generate/generate_windows.go
+++ b/llm/generate/generate_windows.go
@@ -0,0 +1,3 @@
+package generate
+
+//go:generate powershell -ExecutionPolicy Bypass -File ./gen_windows.ps1
--- a/llm/ggla.go
+++ b/llm/ggla.go
@@ -53,7 +53,7 @@ func (llm *ggla) Tensors() Tensors {
 	return llm.tensors
 }

-func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
+func (llm *ggla) decode(rs io.ReadSeeker) error {
 	var r uint32
 	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
 		return err
@@ -69,18 +69,9 @@ func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
-			if errors.Is(err, io.EOF) {
-				return nil
-			}
 			return err
 		}

-		defer func() {
-			if errors.Is(retErr, io.EOF) {
-				retErr = io.ErrUnexpectedEOF
-			}
-		}()
-
 		var namesize uint32
 		if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
 			return err
@@ -117,7 +108,7 @@ func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
 			return err
 		}

-		if _, err := rs.Seek((offset+31)&-32-offset, io.SeekCurrent); err != nil {
+		if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
 			return err
 		}

--- a/llm/ggml.go
+++ b/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()
 }
@@ -280,18 +254,7 @@ 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
@@ -304,15 +267,17 @@ 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
 	}

@@ -332,10 +297,7 @@ 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()

@@ -346,7 +308,7 @@ 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,
 		)

@@ -354,30 +316,21 @@ 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),
@@ -414,16 +367,6 @@ 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
--- a/llm/ggml_test.go
+++ b/llm/ggml_test.go
@@ -1 +0,0 @@
-package llm
--- a/llm/gguf.go
+++ b/llm/gguf.go
@@ -3,10 +3,11 @@ package llm
 import (
 	"bytes"
 	"encoding/binary"
-	"encoding/json"
 	"fmt"
 	"io"
 	"strings"
+
+	"log/slog"
 )

 type containerGGUF struct {
@@ -28,12 +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 {
@@ -59,6 +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
 	}
@@ -89,8 +85,6 @@ type gguf struct {
 	tensors []*Tensor

 	parameters uint64
-
-	scratch [16 << 10]byte
 }

 func newGGUF(container *containerGGUF) *gguf {
@@ -187,34 +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 := [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{
@@ -236,19 +230,24 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		alignment = 32
 	}

+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	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 := llm.padding(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
 		}
 	}

@@ -286,48 +285,22 @@ 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(llm *gguf, w io.Writer, s string) error {
@@ -343,16 +316,7 @@ func writeGGUFString(llm *gguf, 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
@@ -363,12 +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:
@@ -402,15 +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)
 	}
@@ -425,12 +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:
@@ -456,11 +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)
 		}
@@ -468,12 +416,10 @@ 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](llm *gguf, w io.Writer, t uint32, s S) error {
--- a/llm/memory.go
+++ b/llm/memory.go
@@ -1,7 +1,6 @@
 package llm

 import (
-	"fmt"
 	"log/slog"
 	"strconv"
 	"strings"
@@ -50,18 +49,6 @@ type MemoryEstimate struct {

 	// For multi-GPU scenarios, this is the size in bytes per GPU
 	GPUSizes []uint64
-
-	// internal fields for logging purposes
-	inferenceLibrary    string
-	layersRequested     int
-	layersModel         int
-	availableList       []string
-	kv                  uint64
-	allocationsList     []string
-	memoryWeights       uint64
-	memoryLayerOutput   uint64
-	graphFullOffload    uint64
-	graphPartialOffload uint64
 }

 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
@@ -115,8 +102,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		slog.Warn("model missing blk.0 layer size")
 	}

-	// fp16 k,v = sizeof(float16) * n_ctx * n_layer * (n_embd_head_k + n_embd_head_v) * n_head_kv
-	var kv uint64 = 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * (ggml.KV().EmbeddingHeadCountK() + ggml.KV().EmbeddingHeadCountV()) * ggml.KV().HeadCountKV()
+	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
+	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()

 	// KV is proportional to the number of layers
 	layerSize += kv / ggml.KV().BlockCount()
@@ -180,11 +167,6 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts

 	// For all the layers, find where they can fit on the GPU(s)
 	for i := range int(ggml.KV().BlockCount()) {
-		// Some models have inconsistent layer sizes
-		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
-			layerSize = blk.size()
-			layerSize += kv / ggml.KV().BlockCount()
-		}
 		memoryWeights += layerSize

 		if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
@@ -270,86 +252,78 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		allocationsList = append(allocationsList, format.HumanBytes2(a))
 	}

-	estimate := MemoryEstimate{
-		TotalSize: memoryRequiredTotal,
-		Layers:    0,
-		Graph:     0,
-		VRAMSize:  0,
-		GPUSizes:  []uint64{},
-
-		inferenceLibrary:    gpus[0].Library,
-		layersRequested:     opts.NumGPU,
-		layersModel:         int(ggml.KV().BlockCount()) + 1,
-		availableList:       availableList,
-		kv:                  kv,
-		allocationsList:     allocationsList,
-		memoryWeights:       memoryWeights,
-		memoryLayerOutput:   memoryLayerOutput,
-		graphFullOffload:    graphFullOffload,
-		graphPartialOffload: graphPartialOffload,
-	}
-
-	if gpus[0].Library == "cpu" {
-		return estimate
-	}
-	if layerCount == 0 {
-		slog.Debug("insufficient VRAM to load any model layers")
-		return estimate
-	}
-	estimate.Layers = layerCount
-	estimate.Graph = graphOffload
-	estimate.VRAMSize = memoryRequiredPartial
-	estimate.TotalSize = memoryRequiredTotal
-	estimate.TensorSplit = tensorSplit
-	estimate.GPUSizes = gpuAllocations
-	return estimate
-}
-
-func (m MemoryEstimate) log() {
 	slog.Info(
-		"offload to "+m.inferenceLibrary,
+		"offload to gpu",
 		slog.Group(
 			"layers",
 			// requested number of layers to offload
-			"requested", m.layersRequested,
+			"requested", opts.NumGPU,
 			// The number of layers the model has (including output)
-			"model", m.layersModel,
+			"model", int(ggml.KV().BlockCount())+1,
 			// estimated number of layers that can be offloaded
-			"offload", m.Layers,
-			// multi-gpu split for tensors
-			"split", m.TensorSplit,
+			"offload", layerCount,
+			// multi-gpu split for tesnors
+			"split", tensorSplit,
 		),
 		slog.Group(
 			"memory",
 			// memory available by GPU for offloading
-			"available", m.availableList,
+			"available", availableList,
 			slog.Group(
 				"required",
 				// memory required for full offloading
-				"full", format.HumanBytes2(m.TotalSize),
+				"full", format.HumanBytes2(memoryRequiredTotal),
 				// memory required to offload layers.estimate layers
-				"partial", format.HumanBytes2(m.VRAMSize),
+				"partial", format.HumanBytes2(memoryRequiredPartial),
 				// memory of KV cache
-				"kv", format.HumanBytes2(m.kv),
+				"kv", format.HumanBytes2(kv),
 				// Allocations across the GPUs
-				"allocations", m.allocationsList,
+				"allocations", allocationsList,
 			),
 			slog.Group(
 				"weights",
 				// memory of the weights
-				"total", format.HumanBytes2(m.memoryWeights),
+				"total", format.HumanBytes2(memoryWeights),
 				// memory of repeating layers
-				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
+				"repeating", format.HumanBytes2(memoryWeights-memoryLayerOutput),
 				// memory of non-repeating layers
-				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
+				"nonrepeating", format.HumanBytes2(memoryLayerOutput),
 			),
 			slog.Group(
 				"graph",
 				// memory of graph when fully offloaded
-				"full", format.HumanBytes2(m.graphFullOffload),
+				"full", format.HumanBytes2(graphFullOffload),
 				// memory of graph when not fully offloaded
-				"partial", format.HumanBytes2(m.graphPartialOffload),
+				"partial", format.HumanBytes2(graphPartialOffload),
 			),
 		),
 	)
+	if gpus[0].Library == "cpu" {
+		return MemoryEstimate{
+			Layers:    0,
+			Graph:     0,
+			VRAMSize:  0,
+			TotalSize: memoryRequiredTotal,
+			GPUSizes:  []uint64{},
+		}
+	}
+	if layerCount == 0 {
+		slog.Debug("insufficient VRAM to load any model layers")
+		return MemoryEstimate{
+			Layers:    0,
+			Graph:     0,
+			VRAMSize:  0,
+			TotalSize: memoryRequiredTotal,
+			GPUSizes:  []uint64{},
+		}
+	}
+
+	return MemoryEstimate{
+		Layers:      layerCount,
+		Graph:       graphOffload,
+		VRAMSize:    memoryRequiredPartial,
+		TotalSize:   memoryRequiredTotal,
+		TensorSplit: tensorSplit,
+		GPUSizes:    gpuAllocations,
+	}
 }
--- a/llm/memory_test.go
+++ b/llm/memory_test.go
@@ -22,14 +22,13 @@ func TestEstimateGPULayers(t *testing.T) {
 	defer f.Close()
 	gguf := NewGGUFV3(binary.LittleEndian)
 	inputLayerCount := 5
-
 	tensors := []Tensor{
-		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-		{Name: "blk.1.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-		{Name: "blk.2.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-		{Name: "blk.3.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-		{Name: "blk.4.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.1.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.2.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.3.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.4.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
 	}
 	assert.Len(t, tensors, inputLayerCount+1)
 	err = gguf.Encode(f, KV{
@@ -46,10 +45,8 @@ func TestEstimateGPULayers(t *testing.T) {
 	}, tensors)
 	require.NoError(t, err)

-	ggml, err := LoadModel(f.Name(), 0)
-	if err != nil {
-		t.Fatal(err)
-	}
+	ggml, err := LoadModel(f.Name())
+	require.NoError(t, err)

 	// Simple CPU scenario
 	gpus := []gpu.GpuInfo{
--- a/llm/patches/01-load-progress.diff
+++ b/llm/patches/01-load-progress.diff
@@ -1,8 +1,8 @@
 diff --git a/common/common.cpp b/common/common.cpp
-index 73ff0e85..6adb1a92 100644
+index ba1ecf0e..cead57cc 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -2447,6 +2447,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
+@@ -1836,6 +1836,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
@@ -12,20 +12,20 @@ index 73ff0e85..6adb1a92 100644
         mparams.kv_overrides = NULL;
     } else {
 diff --git a/common/common.h b/common/common.h
-index 58ed72f4..0bb2605e 100644
+index d80344f2..71e84834 100644
 --- a/common/common.h
 +++ b/common/common.h
-@@ -180,6 +180,13 @@ struct gpt_params {
+@@ -174,6 +174,13 @@ struct gpt_params {
+     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector
     std::vector<std::string> image; // path to image file(s)
- 
+
 +    // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
 +    // If the provided progress_callback returns true, model loading continues.
 +    // If it returns false, model loading is immediately aborted.
 +    llama_progress_callback progress_callback = NULL;
 +    // context pointer passed to the progress callback
 +    void * progress_callback_user_data;
-+
-     // server params
-     int32_t port           = 8080;         // server listens on this network port
-     int32_t timeout_read   = 600;          // http read timeout in seconds
+ };
+ 
+ void gpt_params_handle_model_default(gpt_params & params);
--- a/llm/patches/05-default-pretokenizer.diff
+++ b/llm/patches/05-default-pretokenizer.diff
@@ -1,8 +1,8 @@
 diff --git a/llama.cpp b/llama.cpp
-index 61948751..4b72a293 100644
+index 40d2ec2c..74f3ee9c 100644
 --- a/llama.cpp
 +++ b/llama.cpp
-@@ -4824,16 +4824,7 @@ static void llm_load_vocab(
+@@ -4642,16 +4642,7 @@ static void llm_load_vocab(
 
         // for now, only BPE models have pre-tokenizers
         if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
@@ -15,14 +15,14 @@ index 61948751..4b72a293 100644
 -                LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
 -                LLAMA_LOG_WARN("%s:                                             \n", __func__);
 -                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
-            } else if (tokenizer_pre == "default") {
-+            if (tokenizer_pre == "default") {
+-            } else if (
+            if (
+                     tokenizer_pre == "default") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             } else if (
-                     tokenizer_pre == "llama3"   ||
-@@ -4888,7 +4879,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "poro-chat") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
+@@ -4703,7 +4694,8 @@ static void llm_load_vocab(
+                 tokenizer_pre == "smaug-bpe") {
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG;
             } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
--- a/llm/patches/07-gemma.diff
+++ b/llm/patches/07-gemma.diff
@@ -1,305 +0,0 @@
-From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001
-From: Ollama maintainers <hello@ollama.com>
-Date: Wed, 26 Jun 2024 16:18:09 -0700
-Subject: [PATCH] Architecture support
-
---
- llama.cpp | 194 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
- 1 file changed, 193 insertions(+), 1 deletion(-)
-
-diff --git a/llama.cpp b/llama.cpp
-index 61948751..3b4196f5 100644
--- a/llama.cpp
-+++ b/llama.cpp
-@@ -217,6 +217,7 @@ enum llm_arch {
-     LLM_ARCH_INTERNLM2,
-     LLM_ARCH_MINICPM,
-     LLM_ARCH_GEMMA,
-+    LLM_ARCH_GEMMA2,
-     LLM_ARCH_STARCODER2,
-     LLM_ARCH_MAMBA,
-     LLM_ARCH_XVERSE,
-@@ -255,6 +256,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
-     { LLM_ARCH_INTERNLM2,       "internlm2"    },
-     { LLM_ARCH_MINICPM,         "minicpm"      },
-     { LLM_ARCH_GEMMA,           "gemma"        },
-+    { LLM_ARCH_GEMMA2,          "gemma2"       },
-     { LLM_ARCH_STARCODER2,      "starcoder2"   },
-     { LLM_ARCH_MAMBA,           "mamba"        },
-     { LLM_ARCH_XVERSE,          "xverse"       },
-@@ -464,10 +466,12 @@ enum llm_tensor {
-     LLM_TENSOR_ATTN_NORM,
-     LLM_TENSOR_ATTN_NORM_2,
-     LLM_TENSOR_ATTN_OUT_NORM,
-+    LLM_TENSOR_ATTN_POST_NORM,
-     LLM_TENSOR_ATTN_ROT_EMBD,
-     LLM_TENSOR_FFN_GATE_INP,
-     LLM_TENSOR_FFN_GATE_INP_SHEXP,
-     LLM_TENSOR_FFN_NORM,
-+    LLM_TENSOR_FFN_POST_NORM,
-     LLM_TENSOR_FFN_GATE,
-     LLM_TENSOR_FFN_DOWN,
-     LLM_TENSOR_FFN_UP,
-@@ -960,6 +964,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
-             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-         },
-     },
-+    {
-+        LLM_ARCH_GEMMA2,
-+        {
-+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
-+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
-+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
-+        },
-+    },
-     {
-         LLM_ARCH_STARCODER2,
-         {
-@@ -1941,6 +1963,8 @@ enum e_model {
-     MODEL_8x22B,
-     MODEL_16x12B,
-     MODEL_10B_128x3_66B,
-+    MODEL_9B,
-+    MODEL_27B,
- };
- 
- static const size_t kiB = 1024;
-@@ -2114,6 +2138,7 @@ struct llama_layer {
-     struct ggml_tensor * attn_out_norm_b;
-     struct ggml_tensor * attn_q_a_norm;
-     struct ggml_tensor * attn_kv_a_norm;
-+    struct ggml_tensor * attn_post_norm;
- 
-     // attention
-     struct ggml_tensor * wq;
-@@ -2136,6 +2161,7 @@ struct llama_layer {
-     // normalization
-     struct ggml_tensor * ffn_norm;
-     struct ggml_tensor * ffn_norm_b;
-+    struct ggml_tensor * ffn_post_norm;
-     struct ggml_tensor * layer_out_norm;
-     struct ggml_tensor * layer_out_norm_b;
-     struct ggml_tensor * ffn_norm_exps;
-@@ -4529,6 +4555,16 @@ static void llm_load_hparams(
-                 }
-             } break;
-         case LLM_ARCH_GEMMA:
-+            {
-+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-+
-+                switch (hparams.n_layer) {
-+                    case 18: model.type = e_model::MODEL_9B; break;
-+                    case 28: model.type = e_model::MODEL_27B; break;
-+                    default: model.type = e_model::MODEL_UNKNOWN;
-+               }
-+            } break;
-+        case LLM_ARCH_GEMMA2:
-             {
-                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
- 
-@@ -6305,6 +6341,40 @@ static bool llm_load_tensors(
-                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-                     }
-                 } break;
-+            case LLM_ARCH_GEMMA2:
-+                {
-+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-+
-+                    // output
-+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
-+
-+                    const int64_t n_ff          = hparams.n_ff;
-+                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
-+                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
-+                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
-+
-+                    for (uint32_t i = 0; i < n_layer; ++i) {
-+                        ggml_context * ctx_layer = ctx_for_layer(i);
-+                        ggml_context * ctx_split = ctx_for_layer_split(i);
-+
-+                        auto & layer = model.layers[i];
-+
-+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-+
-+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
-+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
-+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
-+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
-+                        layer.attn_post_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd});
-+
-+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-+                        layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd});
-+                    }
-+                } break;
-             case LLM_ARCH_STARCODER2:
-                 {
-                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-@@ -10614,6 +10684,123 @@ struct llm_build_context {
-         return gf;
-     }
- 
-+    struct ggml_cgraph * build_gemma2() {
-+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
-+
-+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-+
-+        struct ggml_tensor * cur;
-+        struct ggml_tensor * inpL;
-+
-+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
-+
-+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-+        cb(inpL, "inp_scaled", -1);
-+
-+        // inp_pos - contains the positions
-+        struct ggml_tensor * inp_pos = build_inp_pos();
-+
-+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-+
-+        for (int il = 0; il < n_layer; ++il) {
-+            // norm
-+            cur = llm_build_norm(ctx0, inpL, hparams,
-+                    model.layers[il].attn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_norm", il);
-+
-+            // self-attention
-+            {
-+                // compute Q and K and RoPE them
-+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-+                cb(Qcur, "Qcur", il);
-+
-+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-+                cb(Kcur, "Kcur", il);
-+
-+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-+                cb(Vcur, "Vcur", il);
-+
-+                Qcur = ggml_rope_ext(
-+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
-+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                        ext_factor, attn_factor, beta_fast, beta_slow);
-+                cb(Qcur, "Qcur", il);
-+
-+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
-+                cb(Qcur, "Qcur_scaled", il);
-+
-+                Kcur = ggml_rope_ext(
-+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
-+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                        ext_factor, attn_factor, beta_fast, beta_slow);
-+                cb(Kcur, "Kcur", il);
-+
-+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
-+                        model.layers[il].wo, NULL,
-+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-+            }
-+
-+            if (il == n_layer - 1) {
-+                // skip computing output for unused tokens
-+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-+            }
-+
-+            cur = llm_build_norm(ctx0, cur, hparams,
-+                    model.layers[il].attn_post_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_post_norm", il);
-+
-+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-+            cb(sa_out, "sa_out", il);
-+
-+            cur = llm_build_norm(ctx0, sa_out, hparams,
-+                    model.layers[il].ffn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "ffn_norm", il);
-+
-+            // feed-forward network
-+            {
-+                cur = llm_build_ffn(ctx0, cur,
-+                        model.layers[il].ffn_up, NULL,
-+                        model.layers[il].ffn_gate, NULL,
-+                        model.layers[il].ffn_down, NULL,
-+                        NULL,
-+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-+                cb(cur, "ffn_out", il);
-+            }
-+
-+            cur = llm_build_norm(ctx0, cur, hparams,
-+                model.layers[il].ffn_post_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+            cb(cur, "ffn_post_norm", -1);
-+
-+            cur = ggml_add(ctx0, cur, sa_out);
-+            cb(cur, "l_out", il);
-+
-+            // input for next layer
-+            inpL = cur;
-+        }
-+
-+        cur = inpL;
-+
-+        cur = llm_build_norm(ctx0, cur, hparams,
-+                model.output_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+        cb(cur, "result_norm", -1);
-+
-+        // lm_head
-+        cur = ggml_mul_mat(ctx0, model.output, cur);
-+        cb(cur, "result_output", -1);
-+
-+        ggml_build_forward_expand(gf, cur);
-+
-+        return gf;
-+    }
-+
-     struct ggml_cgraph * build_starcoder2() {
-         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
- 
-@@ -11847,6 +12034,10 @@ static struct ggml_cgraph * llama_build_graph(
-             {
-                 result = llm.build_gemma();
-             } break;
-+        case LLM_ARCH_GEMMA2:
-+            {
-+                result = llm.build_gemma2();
-+            } break;
-         case LLM_ARCH_STARCODER2:
-             {
-                 result = llm.build_starcoder2();
-@@ -16671,6 +16862,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
-         case LLM_ARCH_PHI2:
-         case LLM_ARCH_PHI3:
-         case LLM_ARCH_GEMMA:
-+        case LLM_ARCH_GEMMA2:
-         case LLM_ARCH_STARCODER2:
-         case LLM_ARCH_GPTNEOX:
-             return LLAMA_ROPE_TYPE_NEOX;
-@@ -18551,7 +18743,7 @@ static int32_t llama_chat_apply_template_internal(
-         if (add_ass) {
-             ss << "<s>assistant\n";
-         }
-    } else if (tmpl == "gemma" || tmpl.find("<start_of_turn>") != std::string::npos) {
-+    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("<start_of_turn>") != std::string::npos) {
-         // google/gemma-7b-it
-         std::string system_prompt = "";
-         for (auto message : chat) {
-- 
-2.45.2
-
--- a/llm/payload.go
+++ b/llm/payload.go
@@ -58,7 +58,7 @@ func availableServers() map[string]string {
 	}

 	// glob payloadsDir for files that start with ollama_
-	pattern := filepath.Join(payloadsDir, "*", "ollama_*")
+	pattern := filepath.Join(payloadsDir, "*")

 	files, err := filepath.Glob(pattern)
 	if err != nil {
@@ -69,7 +69,7 @@ func availableServers() map[string]string {
 	servers := make(map[string]string)
 	for _, file := range files {
 		slog.Debug("availableServers : found", "file", file)
-		servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
+		servers[filepath.Base(file)] = file
 	}

 	return servers
--- a/llm/server.go
+++ b/llm/server.go
@@ -60,12 +60,7 @@ type llmServer struct {
 	sem *semaphore.Weighted
 }

-// LoadModel will load a model from disk. The model must be in the GGML format.
-//
-// 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 LoadModel(model string, maxArraySize int) (*GGML, error) {
+func LoadModel(model string) (*GGML, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -76,7 +71,7 @@ func LoadModel(model string, maxArraySize int) (*GGML, error) {
 	}
 	defer f.Close()

-	ggml, _, err := DecodeGGML(f, maxArraySize)
+	ggml, _, err := DecodeGGML(f)
 	return ggml, err
 }

@@ -86,17 +81,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	var err error
 	var cpuRunner string
 	var estimate MemoryEstimate
-	var systemTotalMemory uint64
-	var systemFreeMemory uint64
-
-	systemMemInfo, err := gpu.GetCPUMem()
-	if err != nil {
-		slog.Error("failed to lookup system memory", "error", err)
-	} else {
-		systemTotalMemory = systemMemInfo.TotalMemory
-		systemFreeMemory = systemMemInfo.FreeMemory
-		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", systemFreeMemory)
-	}
+	var systemMemory uint64

 	// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
 	if opts.NumGPU == 0 {
@@ -106,10 +91,19 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		cpuRunner = serverForCpu()
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
+		if gpus[0].Library == "metal" {
+			memInfo, err := gpu.GetCPUMem()
+			if err != nil {
+				slog.Error("failed to lookup system memory", "error", err)
+			} else {
+				systemMemory = memInfo.TotalMemory
+				slog.Debug("system memory", "total", format.HumanBytes2(systemMemory))
+			}
+		}
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)

 		switch {
-		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:
+		case gpus[0].Library == "metal" && estimate.VRAMSize > systemMemory:
 			// disable partial offloading when model is greater than total system memory as this
 			// can lead to locking up the system
 			opts.NumGPU = 0
@@ -122,8 +116,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 	}

-	estimate.log()
-
 	// Loop through potential servers
 	finalErr := errors.New("no suitable llama servers found")

@@ -208,7 +200,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if g.Library == "metal" &&
 			uint64(opts.NumGPU) > 0 &&
 			uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
-			opts.UseMMap = api.TriStateFalse
+			opts.UseMMap = false
 		}
 	}

@@ -216,11 +208,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--flash-attn")
 	}

-	// Windows CUDA should not use mmap for best performance
-	// Linux  with a model larger than free space, mmap leads to thrashing
-	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == api.TriStateUndefined) ||
-		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == api.TriStateUndefined) ||
-		opts.UseMMap == api.TriStateFalse {
+	if !opts.UseMMap {
 		params = append(params, "--no-mmap")
 	}

@@ -283,8 +271,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if runtime.GOOS == "windows" {
 			pathEnv = "PATH"
 		}
-		// prepend the server directory to LD_LIBRARY_PATH/PATH and the parent dir for common dependencies
-		libraryPaths := []string{dir, filepath.Dir(dir)}
+		// prepend the server directory to LD_LIBRARY_PATH/PATH
+		libraryPaths := []string{dir}

 		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
 			// Append our runner directory to the path
@@ -417,7 +405,7 @@ func projectorMemoryRequirements(filename string) uint64 {
 	}
 	defer file.Close()

-	ggml, _, err := DecodeGGML(file, 0)
+	ggml, _, err := DecodeGGML(file)
 	if err != nil {
 		return 0
 	}
--- a/main.go
+++ b/main.go
@@ -1,7 +1,5 @@
 package main

-//go:generate go run build.go -g -s
-
 import (
 	"context"

--- a/scripts/build_windows.ps1
+++ b/scripts/build_windows.ps1
@@ -103,19 +103,19 @@ function buildApp() {
 function gatherDependencies() {
    write-host "Gathering runtime dependencies"
    cd "${script:SRC_DIR}"
-    md "${script:DEPS_DIR}\ollama_runners" -ea 0 > $null
+    md "${script:DEPS_DIR}" -ea 0 > $null

    # TODO - this varies based on host build system and MSVC version - drive from dumpbin output
    # currently works for Win11 + MSVC 2019 + Cuda V11
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140.dll" "${script:DEPS_DIR}\ollama_runners\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\ollama_runners\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\ollama_runners\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140.dll" "${script:DEPS_DIR}\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\"


    cp "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"
    if ("${env:KEY_CONTAINER}") {
        write-host "about to sign"
-        foreach ($file in (get-childitem "${script:DEPS_DIR}\cuda\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
+        foreach ($file in (get-childitem "${script:DEPS_DIR}/cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
            write-host "signing $file"
            & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} $file
--- a/scripts/install.sh
+++ b/scripts/install.sh
@@ -279,7 +279,7 @@ if ! check_gpu nvidia-smi || [ -z "$(nvidia-smi | grep -o "CUDA Version: [0-9]*\
    case $OS_NAME in
        centos|rhel) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -d '.' -f 1) ;;
        rocky) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -c1) ;;
-        fedora) [ $OS_VERSION -lt '39' ] && install_cuda_driver_yum $OS_NAME $OS_VERSION || install_cuda_driver_yum $OS_NAME '39';;
+        fedora) [ $OS_VERSION -lt '37' ] && install_cuda_driver_yum $OS_NAME $OS_VERSION || install_cuda_driver_yum $OS_NAME '37';;
        amzn) install_cuda_driver_yum 'fedora' '37' ;;
        debian) install_cuda_driver_apt $OS_NAME $OS_VERSION ;;
        ubuntu) install_cuda_driver_apt $OS_NAME $(echo $OS_VERSION | sed 's/\.//') ;;
--- a/server/images.go
+++ b/server/images.go
@@ -414,22 +414,17 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 							return err
 						}

-						layer, err := NewLayer(temp, baseLayer.MediaType)
+						layers, err := parseFromFile(ctx, temp, "", fn)
 						if err != nil {
 							return err
 						}

-						if _, err := temp.Seek(0, io.SeekStart); err != nil {
-							return err
+						if len(layers) != 1 {
+							return errors.New("quantization failed")
 						}

-						ggml, _, err := llm.DecodeGGML(temp, 0)
-						if err != nil {
-							return err
-						}
-
-						baseLayer.Layer = layer
-						baseLayer.GGML = ggml
+						baseLayer.Layer = layers[0].Layer
+						baseLayer.GGML = layers[0].GGML
 					}
 				}

--- a/server/model.go
+++ b/server/model.go
@@ -11,7 +11,6 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
-	"strings"

 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
@@ -64,7 +63,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 			}
 			defer blob.Close()

-			ggml, _, err := llm.DecodeGGML(blob, 0)
+			ggml, _, err := llm.DecodeGGML(blob)
 			if err != nil {
 				return nil, err
 			}
@@ -78,80 +77,62 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	return layers, nil
 }

-func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse)) error {
+func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
 	stat, err := file.Stat()
 	if err != nil {
-		return err
+		return nil, err
 	}

 	r, err := zip.NewReader(file, stat.Size())
 	if err != nil {
-		return err
+		return nil, err
 	}

+	tempdir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
+	if err != nil {
+		return nil, err
+	}
+	defer os.RemoveAll(tempdir)
+
 	fn(api.ProgressResponse{Status: "unpacking model metadata"})
 	for _, f := range r.File {
-		n := filepath.Join(p, f.Name)
-		if !strings.HasPrefix(n, p) {
-			slog.Warn("skipped extracting file outside of context", "name", f.Name)
-			continue
-		}
-
-		if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil {
-			return err
-		}
-
 		// TODO(mxyng): this should not write out all files to disk
-		outfile, err := os.Create(n)
+		outfile, err := os.Create(filepath.Join(tempdir, f.Name))
 		if err != nil {
-			return err
+			return nil, err
 		}
 		defer outfile.Close()

 		infile, err := f.Open()
 		if err != nil {
-			return err
+			return nil, err
 		}
 		defer infile.Close()

 		if _, err = io.Copy(outfile, infile); err != nil {
-			return err
+			return nil, err
 		}

 		if err := outfile.Close(); err != nil {
-			return err
+			return nil, err
 		}

 		if err := infile.Close(); err != nil {
-			return err
+			return nil, err
 		}
 	}

-	return nil
-}
-
-func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
-	tempDir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
-	if err != nil {
-		return nil, err
-	}
-	defer os.RemoveAll(tempDir)
-
-	if err := extractFromZipFile(tempDir, file, fn); err != nil {
-		return nil, err
-	}
-
-	mf, err := convert.GetModelFormat(tempDir)
+	mf, err := convert.GetModelFormat(tempdir)
 	if err != nil {
 		return nil, err
 	}

-	params, err := mf.GetParams(tempDir)
+	params, err := mf.GetParams(tempdir)
 	if err != nil {
 		return nil, err
 	}

-	mArch, err := mf.GetModelArch("", tempDir, params)
+	mArch, err := mf.GetModelArch("", tempdir, params)
 	if err != nil {
 		return nil, err
 	}
@@ -169,7 +150,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a

 	// TODO(mxyng): this should write directly into a layer
 	// e.g. NewLayer(arch.Reader(), "application/vnd.ollama.image.model")
-	temp, err := os.CreateTemp(tempDir, "fp16")
+	temp, err := os.CreateTemp(tempdir, "fp16")
 	if err != nil {
 		return nil, err
 	}
@@ -195,7 +176,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a
 	}
 	defer bin.Close()

-	ggml, _, err := llm.DecodeGGML(bin, 0)
+	ggml, _, err := llm.DecodeGGML(bin)
 	if err != nil {
 		return nil, err
 	}
@@ -229,7 +210,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap

 	var offset int64
 	for offset < stat.Size() {
-		ggml, n, err := llm.DecodeGGML(file, 0)
+		ggml, n, err := llm.DecodeGGML(file)
 		if errors.Is(err, io.EOF) {
 			break
 		} else if err != nil {
--- a/server/model_test.go
+++ b/server/model_test.go
@@ -1,92 +0,0 @@
-package server
-
-import (
-	"archive/zip"
-	"bytes"
-	"io"
-	"os"
-	"path/filepath"
-	"slices"
-	"testing"
-
-	"github.com/ollama/ollama/api"
-)
-
-func createZipFile(t *testing.T, name string) *os.File {
-	t.Helper()
-
-	f, err := os.CreateTemp(t.TempDir(), "")
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	zf := zip.NewWriter(f)
-	defer zf.Close()
-
-	zh, err := zf.CreateHeader(&zip.FileHeader{Name: name})
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	if _, err := io.Copy(zh, bytes.NewReader([]byte(""))); err != nil {
-		t.Fatal(err)
-	}
-
-	return f
-}
-
-func TestExtractFromZipFile(t *testing.T) {
-	cases := []struct {
-		name   string
-		expect []string
-	}{
-		{
-			name:   "good",
-			expect: []string{"good"},
-		},
-		{
-			name: filepath.Join("..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"),
-		},
-	}
-
-	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			f := createZipFile(t, tt.name)
-			defer f.Close()
-
-			tempDir := t.TempDir()
-			if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); err != nil {
-				t.Fatal(err)
-			}
-
-			var matches []string
-			if err := filepath.Walk(tempDir, func(p string, fi os.FileInfo, err error) error {
-				if err != nil {
-					return err
-				}
-
-				if !fi.IsDir() {
-					matches = append(matches, p)
-				}
-
-				return nil
-			}); err != nil {
-				t.Fatal(err)
-			}
-
-			var actual []string
-			for _, match := range matches {
-				rel, err := filepath.Rel(tempDir, match)
-				if err != nil {
-					t.Error(err)
-				}
-
-				actual = append(actual, rel)
-			}
-
-			if !slices.Equal(actual, tt.expect) {
-				t.Fatalf("expected %d files, got %d", len(tt.expect), len(matches))
-			}
-		})
-	}
-}
--- a/server/routes.go
+++ b/server/routes.go
@@ -734,48 +734,9 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	fmt.Fprint(&sb, m.String())
 	resp.Modelfile = sb.String()

-	kvData, err := getKVData(m.ModelPath, req.Verbose)
-	if err != nil {
-		return nil, err
-	}
-	delete(kvData, "general.name")
-	delete(kvData, "tokenizer.chat_template")
-	resp.ModelInfo = kvData
-
-	if len(m.ProjectorPaths) > 0 {
-		projectorData, err := getKVData(m.ProjectorPaths[0], req.Verbose)
-		if err != nil {
-			return nil, err
-		}
-		resp.ProjectorInfo = projectorData
-	}
-
 	return resp, nil
 }

-func getKVData(digest string, verbose bool) (llm.KV, error) {
-	maxArraySize := 0
-	if verbose {
-		maxArraySize = -1
-	}
-	kvData, err := llm.LoadModel(digest, maxArraySize)
-	if err != nil {
-		return nil, err
-	}
-
-	kv := kvData.KV()
-
-	if !verbose {
-		for k := range kv {
-			if t, ok := kv[k].([]any); len(t) > 5 && ok {
-				kv[k] = []any{}
-			}
-		}
-	}
-
-	return kv, nil
-}
-
 func (s *Server) ListModelsHandler(c *gin.Context) {
 	ms, err := Manifests()
 	if err != nil {
@@ -1105,20 +1066,11 @@ func Serve(ln net.Listener) error {
 	schedCtx, schedDone := context.WithCancel(ctx)
 	sched := InitScheduler(schedCtx)
 	s := &Server{addr: ln.Addr(), sched: sched}
-
-	http.Handle("/", s.GenerateRoutes())
+	r := s.GenerateRoutes()

 	slog.Info(fmt.Sprintf("Listening on %s (version %s)", ln.Addr(), version.Version))
 	srvr := &http.Server{
-		// Use http.DefaultServeMux so we get net/http/pprof for
-		// free.
-		//
-		// TODO(bmizerany): Decide if we want to make this
-		// configurable so it is not exposed by default, or allow
-		// users to bind it to a different port. This was a quick
-		// and easy way to get pprof, but it may not be the best
-		// way.
-		Handler: nil,
+		Handler: r,
 	}

 	// listen for a ctrl+c and stop any loaded llm
--- a/server/routes_test.go
+++ b/server/routes_test.go
@@ -19,7 +19,6 @@ import (

 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -213,7 +212,6 @@ func Test_Routes(t *testing.T) {
 					"top_p 0.9",
 				}
 				assert.Equal(t, expectedParams, params)
-				assert.InDelta(t, 0, showResp.ModelInfo["general.parameter_count"], 1e-9, "Parameter count should be 0")
 			},
 		},
 	}
@@ -327,40 +325,3 @@ func TestCase(t *testing.T) {
 		})
 	}
 }
-
-func TestShow(t *testing.T) {
-	t.Setenv("OLLAMA_MODELS", t.TempDir())
-	envconfig.LoadConfig()
-
-	var s Server
-
-	createRequest(t, s.CreateModelHandler, api.CreateRequest{
-		Name: "show-model",
-		Modelfile: fmt.Sprintf(
-			"FROM %s\nFROM %s",
-			createBinFile(t, llm.KV{"general.architecture": "test"}, nil),
-			createBinFile(t, llm.KV{"general.architecture": "clip"}, nil),
-		),
-	})
-
-	w := createRequest(t, s.ShowModelHandler, api.ShowRequest{
-		Name: "show-model",
-	})
-
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status code 200, actual %d", w.Code)
-	}
-
-	var resp api.ShowResponse
-	if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
-		t.Fatal(err)
-	}
-
-	if resp.ModelInfo["general.architecture"] != "test" {
-		t.Fatal("Expected model architecture to be 'test', but got", resp.ModelInfo["general.architecture"])
-	}
-
-	if resp.ProjectorInfo["general.architecture"] != "clip" {
-		t.Fatal("Expected projector architecture to be 'clip', but got", resp.ProjectorInfo["general.architecture"])
-	}
-}
--- a/server/sched.go
+++ b/server/sched.go
@@ -144,7 +144,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					}

 					// Load model for fitting
-					ggml, err := llm.LoadModel(pending.model.ModelPath, 0)
+					ggml, err := llm.LoadModel(pending.model.ModelPath)
 					if err != nil {
 						pending.errCh <- err
 						break
--- a/server/sched_test.go
+++ b/server/sched_test.go
@@ -128,14 +128,14 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 		"tokenizer.ggml.scores":         []float32{0},
 		"tokenizer.ggml.token_type":     []int32{0},
 	}, []llm.Tensor{
-		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
 	})
 	require.NoError(t, err)

 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	scenario.ggml, err = llm.LoadModel(model.ModelPath, 0)
+	scenario.ggml, err = llm.LoadModel(model.ModelPath)
 	require.NoError(t, err)

 	scenario.req = &LlmRequest{
--- a/types/model/name.go
+++ b/types/model/name.go
@@ -4,6 +4,7 @@ package model

 import (
 	"cmp"
+	"encoding/hex"
 	"errors"
 	"fmt"
 	"log/slog"
@@ -370,3 +371,57 @@ func cutPromised(s, sep string) (before, after string, ok bool) {
 	}
 	return cmp.Or(before, MissingPart), cmp.Or(after, MissingPart), true
 }
+
+type DigestType byte
+
+const (
+	DigestTypeInvalid DigestType = iota
+	DigestTypeSHA256
+)
+
+func (t DigestType) String() string {
+	switch t {
+	case DigestTypeSHA256:
+		return "sha256"
+	default:
+		return "invalid"
+	}
+}
+
+type Digest struct {
+	Type DigestType
+	Sum  [32]byte
+}
+
+func ParseDigest(s string) (Digest, error) {
+	i := strings.IndexAny(s, "-:")
+	if i < 0 {
+		return Digest{}, fmt.Errorf("invalid digest %q", s)
+	}
+	typ, encSum := s[:i], s[i+1:]
+	if typ != "sha256" {
+		return Digest{}, fmt.Errorf("unsupported digest type %q", typ)
+	}
+	d := Digest{
+		Type: DigestTypeSHA256,
+	}
+	n, err := hex.Decode(d.Sum[:], []byte(encSum))
+	if err != nil {
+		return Digest{}, err
+	}
+	if n != 32 {
+		return Digest{}, fmt.Errorf("digest %q decoded to %d bytes; want 32", encSum, n)
+	}
+	return d, nil
+}
+
+func (d Digest) String() string {
+	if d.Type == DigestTypeInvalid {
+		return ""
+	}
+	return fmt.Sprintf("sha256-%x", d.Sum)
+}
+
+func (d Digest) IsValid() bool {
+	return d.Type != DigestTypeInvalid
+}
--- a/types/model/name_test.go
+++ b/types/model/name_test.go
@@ -284,6 +284,40 @@ func TestFilepathAllocs(t *testing.T) {
 	}
 }

+const (
+	validSha256    = "sha256-1000000000000000000000000000000000000000000000000000000000000000"
+	validSha256Old = "sha256:1000000000000000000000000000000000000000000000000000000000000000"
+)
+
+func TestParseDigest(t *testing.T) {
+	cases := []struct {
+		in   string
+		want string
+	}{
+		{"", ""},           // empty
+		{"sha123-12", ""},  // invalid type
+		{"sha256-", ""},    // invalid sum
+		{"sha256-123", ""}, // invalid odd length sum
+
+		{validSha256, validSha256},
+		{validSha256Old, validSha256},
+	}
+	for _, tt := range cases {
+		t.Run(tt.in, func(t *testing.T) {
+			got, err := ParseDigest(tt.in)
+			if err != nil {
+				if tt.want != "" {
+					t.Errorf("parseDigest(%q) = %v; want %v", tt.in, err, tt.want)
+				}
+				return
+			}
+			if got.String() != tt.want {
+				t.Errorf("parseDigest(%q).String() = %q; want %q", tt.in, got, tt.want)
+			}
+		})
+	}
+}
+
 func TestParseNameFromFilepath(t *testing.T) {
 	cases := map[string]Name{
 		filepath.Join("host", "namespace", "model", "tag"):      {Host: "host", Namespace: "namespace", Model: "model", Tag: "tag"},
--- a/util/bufioutil/buffer_seeker.go
+++ b/util/bufioutil/buffer_seeker.go
@@ -1,34 +0,0 @@
-package bufioutil
-
-import (
-	"bufio"
-	"io"
-)
-
-type BufferedSeeker struct {
-	rs io.ReadSeeker
-	br *bufio.Reader
-}
-
-func NewBufferedSeeker(rs io.ReadSeeker, size int) *BufferedSeeker {
-	return &BufferedSeeker{
-		rs: rs,
-		br: bufio.NewReaderSize(rs, size),
-	}
-}
-
-func (b *BufferedSeeker) Read(p []byte) (int, error) {
-	return b.br.Read(p)
-}
-
-func (b *BufferedSeeker) Seek(offset int64, whence int) (int64, error) {
-	if whence == io.SeekCurrent {
-		offset -= int64(b.br.Buffered())
-	}
-	n, err := b.rs.Seek(offset, whence)
-	if err != nil {
-		return 0, err
-	}
-	b.br.Reset(b.rs)
-	return n, nil
-}
--- a/util/bufioutil/buffer_seeker_test.go
+++ b/util/bufioutil/buffer_seeker_test.go
@@ -1,64 +0,0 @@
-package bufioutil
-
-import (
-	"bytes"
-	"io"
-	"strings"
-	"testing"
-)
-
-func TestBufferedSeeker(t *testing.T) {
-	const alphabet = "abcdefghijklmnopqrstuvwxyz"
-
-	bs := NewBufferedSeeker(strings.NewReader(alphabet), 0) // minReadBufferSize = 16
-
-	checkRead := func(buf []byte, expected string) {
-		t.Helper()
-		_, err := bs.Read(buf)
-		if err != nil {
-			t.Fatal(err)
-		}
-		if !bytes.Equal(buf, []byte(expected)) {
-			t.Fatalf("expected %s, got %s", expected, buf)
-		}
-	}
-
-	// Read the first 5 bytes
-	buf := make([]byte, 5)
-
-	checkRead(buf, "abcde")
-
-	// Seek back to the beginning
-	_, err := bs.Seek(0, io.SeekStart)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	// read 'a'
-	checkRead(buf[:1], "a")
-
-	if bs.br.Buffered() == 0 {
-		t.Fatalf("totally unexpected sanity check failed")
-	}
-
-	// Seek past 'b'
-	_, err = bs.Seek(1, io.SeekCurrent)
-	if err != nil {
-		t.Fatal(err)
-	}
-	checkRead(buf, "cdefg")
-
-	// Seek back to the beginning
-	_, err = bs.Seek(0, io.SeekStart)
-	if err != nil {
-		t.Fatal(err)
-	}
-	checkRead(buf, "abcde")
-
-	// Seek to the end
-	_, err = bs.Seek(-5, io.SeekEnd)
-	if err != nil {
-		t.Fatal(err)
-	}
-	checkRead(buf, "vwxyz")
-}