...

Also, add test for DisplayLongest.

Also, plumb fill param to ParseName in MustParseName

.github/ISSUE_TEMPLATE/10_bug_report.yml

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

.github/ISSUE_TEMPLATE/10_model_request.yml

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

.github/ISSUE_TEMPLATE/20_feature_request.md

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

.github/ISSUE_TEMPLATE/20_feature_request.yml

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

.github/ISSUE_TEMPLATE/30_model_request.md

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

.github/ISSUE_TEMPLATE/90_bug_report.yml

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

.github/workflows/release.yaml

@@ -103,7 +103,6 @@ jobs:
           path: |
             llm/build/**/bin/*
             llm/build/**/*.a
-            dist/windows-amd64/**
 
   # ROCm generation step
   generate-windows-rocm:
@@ -174,9 +173,7 @@ jobs:
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-rocm
-          path: |
+          path: llm/build/**/bin/*
-            llm/build/**/bin/*
-            dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
           name: windows-rocm-deps
@@ -256,9 +253,7 @@ jobs:
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cuda
-          path: |
+          path: llm/build/**/bin/*
-            llm/build/**/bin/*
-            dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
           name: windows-cuda-deps
@@ -311,18 +306,23 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cpu
+          path: llm/build
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cuda
+          path: llm/build
       - uses: actions/download-artifact@v4
         with:
           name: windows-cuda-deps
+          path: dist/deps
       - uses: actions/download-artifact@v4
         with:
           name: windows-rocm-deps
+          path: dist/deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
+          path: llm/build
       - run: dir llm/build
       - run: |
           $gopath=(get-command go).source | split-path -parent
@@ -331,13 +331,13 @@ jobs:
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_GENERATE="1"
+          $env:NVIDIA_DIR=$(resolve-path ".\dist\deps")
+          $env:HIP_PATH=$(resolve-path ".\dist\deps")
           & .\scripts\build_windows.ps1
       - uses: actions/upload-artifact@v4
         with:
           name: dist-windows
-          path: |
+          path: dist/*.exe
-            dist/OllamaSetup.exe
-            dist/ollama-windows-*.zip
 
   # Linux x86 assets built using the container based build
   build-linux-amd64:

.github/workflows/test.yaml

@@ -1,15 +1,5 @@
 name: test
 
-concurrency:
-  # For PRs, later CI runs preempt previous ones. e.g. a force push on a PR
-  # cancels running CI jobs and starts all new ones.
-  #
-  # For non-PR pushes, concurrency.group needs to be unique for every distinct
-  # CI run we want to have happen. Use run_id, which in practice means all
-  # non-PR CI runs will be allowed to run without preempting each other.
-  group: ${{ github.workflow }}-$${{ github.pull_request.number || github.run_id }}
-  cancel-in-progress: true
-
 on:
   pull_request:
     paths:
@@ -31,9 +21,7 @@ jobs:
       - id: changes
         run: |
           changed() {
-            git diff-tree -r --no-commit-id --name-only \
+            git diff-tree -r --no-commit-id --name-only ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }} \
-              $(git merge-base ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }}) \
-              ${{ github.event.pull_request.head.sha }} \
               | xargs python3 -c "import sys; print(any([x.startswith('$1') for x in sys.argv[1:]]))"
           }
 
@@ -115,9 +103,7 @@ jobs:
       - uses: actions/upload-artifact@v4
         with:
           name: cuda-${{ matrix.cuda-version }}-libraries
-          path: |
+          path: llm/build/**/bin/*
-            llm/build/**/bin/*
-            dist/windows-amd64/**
   generate-rocm:
     needs: [changes]
     if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
@@ -148,9 +134,7 @@ jobs:
       - uses: actions/upload-artifact@v4
         with:
           name: rocm-${{ matrix.rocm-version }}-libraries
-          path: |
+          path: llm/build/**/bin/*
-            llm/build/**/bin/*
-            dist/windows-amd64/**
 
   # ROCm generation step
   generate-windows-rocm:
@@ -269,9 +253,14 @@ jobs:
           mkdir -p llm/build/darwin/$ARCH/stub/bin
           touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'macos-') }}
+      - run: |
+          mkdir -p llm/build/windows/$ARCH/stub/bin
+          touch llm/build/windows/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        shell: bash
       - uses: golangci/golangci-lint-action@v4
         with:
-          args: --timeout 8m0s -v
+          args: --timeout 8m0s
   test:
     strategy:
       matrix:
@@ -295,6 +284,7 @@ jobs:
         with:
           go-version-file: go.mod
           cache: true
+      - run: go get
       - run: |
           case ${{ matrix.arch }} in
             amd64) echo ARCH=x86_64 ;;
@@ -309,6 +299,10 @@ jobs:
           mkdir -p llm/build/darwin/$ARCH/stub/bin
           touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'macos-') }}
+      - run: |
+          mkdir -p llm/build/windows/$ARCH/stub/bin
+          touch llm/build/windows/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'windows-') }}
         shell: bash
       - run: go generate ./...
       - run: go build

.gitignore

@@ -11,4 +11,4 @@ ggml-metal.metal
 .idea
 test_data
 *.crt
-llm/build
+llm/build

Dockerfile

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

README.md

@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3](https://ollama.com/library/llama3):
+To run and chat with [Llama 2](https://ollama.com/library/llama2):
 
 ```
-ollama run llama3
+ollama run llama2
 ```
 
 ## Model library
@@ -49,14 +49,18 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
+| Llama 2            | 7B         | 3.8GB | `ollama run llama2`            |
-| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
-| Phi-3              | 3,8B       | 2.3GB | `ollama run phi3`              |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
+| Dolphin Phi        | 2.7B       | 1.6GB | `ollama run dolphin-phi`       |
+| Phi-2              | 2.7B       | 1.7GB | `ollama run phi`               |
 | 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` |
+| Llama 2 13B        | 13B        | 7.3GB | `ollama run llama2:13b`        |
+| Llama 2 70B        | 70B        | 39GB  | `ollama run llama2:70b`        |
+| Orca Mini          | 3B         | 1.9GB | `ollama run orca-mini`         |
+| Vicuna             | 7B         | 3.8GB | `ollama run vicuna`            |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 | Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
 | Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
@@ -94,16 +98,16 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama2` model:
 
 ```
-ollama pull llama3
+ollama pull llama2
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3
+FROM llama2
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -138,7 +142,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3
+ollama pull llama2
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -146,13 +150,13 @@ ollama pull llama3
 ### Remove a model
 
 ```
-ollama rm llama3
+ollama rm llama2
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3 my-model
+ollama cp llama2 my-llama2
 ```
 
 ### Multiline input
@@ -176,7 +180,7 @@ The image features a yellow smiley face, which is likely the central focus of th
 ### Pass in prompt as arguments
 
 ```
-$ ollama run llama3 "Summarize this file: $(cat README.md)"
+$ ollama run llama2 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
@@ -223,7 +227,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3
+./ollama run llama2
 ```
 
 ## REST API
@@ -234,7 +238,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama2",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -243,7 +247,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "mistral",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -256,17 +260,16 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
-- [Open WebUI](https://github.com/open-webui/open-webui)
-- [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
 - [LibreChat](https://github.com/danny-avila/LibreChat)
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
+- [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Saddle](https://github.com/jikkuatwork/saddle)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
-- [Chatbot UI v2](https://github.com/mckaywrigley/chatbot-ui)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
+- [Open WebUI](https://github.com/open-webui/open-webui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
 - [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
@@ -288,13 +291,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
-- [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)
 
 ### Terminal
 
@@ -310,13 +309,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Oatmeal](https://github.com/dustinblackman/oatmeal)
 - [cmdh](https://github.com/pgibler/cmdh)
 - [ooo](https://github.com/npahlfer/ooo)
-- [shell-pilot](https://github.com/reid41/shell-pilot)
 - [tenere](https://github.com/pythops/tenere)
 - [llm-ollama](https://github.com/taketwo/llm-ollama) for [Datasette's LLM CLI](https://llm.datasette.io/en/stable/).
 - [typechat-cli](https://github.com/anaisbetts/typechat-cli)
 - [ShellOracle](https://github.com/djcopley/ShellOracle)
 - [tlm](https://github.com/yusufcanb/tlm)
-- [podman-ollama](https://github.com/ericcurtin/podman-ollama)
 
 ### Database
 
@@ -381,7 +378,3 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
-- [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
-
-### Supported backends 
-- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 

api/client.go

@@ -91,13 +91,6 @@ func ClientFromEnvironment() (*Client, error) {
 	}, nil
 }
 
-func NewClient(base *url.URL, http *http.Client) *Client {
-	return &Client{
-		base: base,
-		http: http,
-	}
-}
-
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
 	var reqBody io.Reader
 	var data []byte

api/types.go

@@ -2,7 +2,6 @@ package api
 
 import (
 	"encoding/json"
-	"errors"
 	"fmt"
 	"math"
 	"os"
@@ -308,7 +307,7 @@ func (m *Metrics) Summary() {
 	}
 }
 
-var ErrInvalidOpts = errors.New("invalid options")
+var ErrInvalidOpts = fmt.Errorf("invalid options")
 
 func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
@@ -396,10 +395,8 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 func DefaultOptions() Options {
 	return Options{
 		// options set on request to runner
-		NumPredict: -1,
+		NumPredict:       -1,
-
+		NumKeep:          0,
-		// set a minimal num_keep to avoid issues on context shifts
-		NumKeep:          4,
 		Temperature:      0.8,
 		TopK:             40,
 		TopP:             0.9,

app/.gitignore

@@ -1,2 +1 @@
 ollama.syso
-app

app/AppDelegate.h

@@ -1,7 +0,0 @@
-#import <Cocoa/Cocoa.h>
-
-@interface AppDelegate : NSObject <NSApplicationDelegate>
-
-- (void)applicationDidFinishLaunching:(NSNotification *)aNotification;
-
-@end

app/README.md

@@ -1,6 +1,10 @@
 # Ollama App
 
-## macOS
+## Linux
+
+TODO
+
+## MacOS
 
 TODO
 

app/app_darwin.go

@@ -1,76 +0,0 @@
-package main
-
-// #cgo CFLAGS: -x objective-c
-// #cgo LDFLAGS: -framework Cocoa -framework LocalAuthentication -framework ServiceManagement
-// #include "app_darwin.h"
-import "C"
-import (
-	"context"
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"syscall"
-)
-
-func init() {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		panic(err)
-	}
-
-	ServerLogFile = filepath.Join(home, ".ollama", "logs", "server.log")
-}
-
-func run() {
-	initLogging()
-	slog.Info("ollama macOS app started")
-
-	// Ask to move to applications directory
-	moving := C.askToMoveToApplications()
-	if moving {
-		return
-	}
-
-	C.killOtherInstances()
-
-	code := C.installSymlink()
-	if code != 0 {
-		slog.Error("Failed to install symlink")
-	}
-
-	exe, err := os.Executable()
-	if err != nil {
-		panic(err)
-	}
-
-	var options ServerOptions
-
-	ctx, cancel := context.WithCancel(context.Background())
-	var done chan int
-
-	done, err = SpawnServer(ctx, filepath.Join(filepath.Dir(exe), "..", "Resources", "ollama"), options)
-	if err != nil {
-		slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
-		done = make(chan int, 1)
-		done <- 1
-	}
-
-	// Run the native macOS app
-	// Note: this will block until the app is closed
-	C.run()
-
-	slog.Info("ollama macOS app closed")
-
-	cancel()
-	slog.Info("Waiting for ollama server to shutdown...")
-	if done != nil {
-		<-done
-	}
-	slog.Info("Ollama app exiting")
-}
-
-//export Quit
-func Quit() {
-	syscall.Kill(os.Getpid(), syscall.SIGTERM)
-}

app/app_darwin.h

@@ -1,13 +0,0 @@
-#import <Cocoa/Cocoa.h>
-
-@interface AppDelegate : NSObject <NSApplicationDelegate>
-- (void)applicationDidFinishLaunching:(NSNotification *)aNotification;
-@end
-
-void run();
-void killOtherInstances();
-bool askToMoveToApplications();
-int createSymlinkWithAuthorization();
-int installSymlink();
-extern void Restart();
-extern void Quit();

app/app_darwin.m

@@ -1,282 +0,0 @@
-#import <AppKit/AppKit.h>
-#import <Cocoa/Cocoa.h>
-#import <CoreServices/CoreServices.h>
-#import <Security/Security.h>
-#import <ServiceManagement/ServiceManagement.h>
-#import "app_darwin.h"
-
-@interface AppDelegate ()
-
-@property (strong, nonatomic) NSStatusItem *statusItem;
-
-@end
-
-@implementation AppDelegate
-
-- (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
-    // show status menu
-    NSMenu *menu = [[NSMenu alloc] init];
-
-    NSMenuItem *aboutMenuItem = [[NSMenuItem alloc] initWithTitle:@"About Ollama" action:@selector(aboutOllama) keyEquivalent:@""];
-    [aboutMenuItem setTarget:self];
-    [menu addItem:aboutMenuItem];
-
-    // Settings submenu
-    NSMenu *settingsMenu = [[NSMenu alloc] initWithTitle:@"Settings"];
-
-    // Submenu items
-    NSMenuItem *chooseModelDirectoryItem = [[NSMenuItem alloc] initWithTitle:@"Choose model directory..." action:@selector(chooseModelDirectory) keyEquivalent:@""];
-    [chooseModelDirectoryItem setTarget:self];
-    [chooseModelDirectoryItem setEnabled:YES];
-    [settingsMenu addItem:chooseModelDirectoryItem];
-
-    NSMenuItem *exposeExternallyItem = [[NSMenuItem alloc] initWithTitle:@"Allow external connections" action:@selector(toggleExposeExternally:) keyEquivalent:@""];
-    [exposeExternallyItem setTarget:self];
-    [exposeExternallyItem setState:NSOffState]; // Set initial state to off
-    [exposeExternallyItem setEnabled:YES];
-    [settingsMenu addItem:exposeExternallyItem];
-
-    NSMenuItem *allowCrossOriginItem = [[NSMenuItem alloc] initWithTitle:@"Allow browser requests" action:@selector(toggleCrossOrigin:) keyEquivalent:@""];
-    [allowCrossOriginItem setTarget:self];
-    [allowCrossOriginItem setState:NSOffState]; // Set initial state to off
-    [allowCrossOriginItem setEnabled:YES];
-    [settingsMenu addItem:allowCrossOriginItem];
-
-    NSMenuItem *settingsMenuItem = [[NSMenuItem alloc] initWithTitle:@"Settings" action:nil keyEquivalent:@""];
-    [settingsMenuItem setSubmenu:settingsMenu];
-    [menu addItem:settingsMenuItem];
-
-    [menu addItemWithTitle:@"Quit Ollama" action:@selector(quit) keyEquivalent:@"q"];
-
-    self.statusItem = [[NSStatusBar systemStatusBar] statusItemWithLength:NSVariableStatusItemLength];
-    [self.statusItem addObserver:self forKeyPath:@"button.effectiveAppearance" options:NSKeyValueObservingOptionNew|NSKeyValueObservingOptionInitial context:nil];
-
-    self.statusItem.menu = menu;
-    [self showIcon];
-}
-
-- (void)aboutOllama {
-    [[NSApplication sharedApplication] orderFrontStandardAboutPanel:nil];
-}
-
-- (void)toggleCrossOrigin:(id)sender {
-    NSMenuItem *item = (NSMenuItem *)sender;
-    if ([item state] == NSOffState) {
-        // Do something when cross-origin requests are allowed
-        [item setState:NSOnState];
-    } else {
-        // Do something when cross-origin requests are disallowed
-        [item setState:NSOffState];
-    }
-}
-
-- (void)toggleExposeExternally:(id)sender {
-    NSMenuItem *item = (NSMenuItem *)sender;
-    if ([item state] == NSOffState) {
-        // Do something when Ollama is exposed externally
-        [item setState:NSOnState];
-    } else {
-        // Do something when Ollama is not exposed externally
-        [item setState:NSOffState];
-    }
-}
-
-- (void)chooseModelDirectory {
-    NSOpenPanel *openPanel = [NSOpenPanel openPanel];
-    [openPanel setCanChooseFiles:NO];
-    [openPanel setCanChooseDirectories:YES];
-    [openPanel setAllowsMultipleSelection:NO];
-
-    NSInteger result = [openPanel runModal];
-    if (result == NSModalResponseOK) {
-        NSURL *selectedDirectoryURL = [openPanel URLs].firstObject;
-        // Do something with the selected directory URL
-    }
-}
-
--(void) showIcon {
-    NSAppearance* appearance = self.statusItem.button.effectiveAppearance;
-    NSString* appearanceName = (NSString*)(appearance.name);
-    NSString* iconName = [[appearanceName lowercaseString] containsString:@"dark"] ? @"iconDark" : @"icon";
-    NSImage* statusImage = [NSImage imageNamed:iconName];
-    [statusImage setTemplate:YES];
-    self.statusItem.button.image = statusImage;
-}
-
--(void)observeValueForKeyPath:(NSString *)keyPath ofObject:(id)object change:(NSDictionary<NSKeyValueChangeKey,id> *)change context:(void *)context {
-    [self showIcon];
-}
-
-- (void)quit {
-    [NSApp stop:nil];
-}
-
-@end
-
-void run() {
-    @autoreleasepool {
-        [NSApplication sharedApplication];
-        AppDelegate *appDelegate = [[AppDelegate alloc] init];
-        [NSApp setDelegate:appDelegate];
-        [NSApp run];
-    }
-}
-
-// killOtherInstances kills all other instances of the app currently
-// running. This way we can ensure that only the most recently started
-// instance of Ollama is running
-void killOtherInstances() {
-    pid_t pid = getpid();
-    NSArray *all = [[NSWorkspace sharedWorkspace] runningApplications];
-    NSMutableArray *apps = [NSMutableArray array];
-
-    for (NSRunningApplication *app in all) {
-        if ([app.bundleIdentifier isEqualToString:[[NSBundle mainBundle] bundleIdentifier]] ||
-            [app.bundleIdentifier isEqualToString:@"ai.ollama.ollama"] ||
-            [app.bundleIdentifier isEqualToString:@"com.electron.ollama"]) {
-            if (app.processIdentifier != pid) {
-                [apps addObject:app];
-            }
-        }
-    }
-
-    for (NSRunningApplication *app in apps) {
-        kill(app.processIdentifier, SIGTERM);
-    }
-
-    NSDate *startTime = [NSDate date];
-    for (NSRunningApplication *app in apps) {
-        while (!app.terminated) {
-            if (-[startTime timeIntervalSinceNow] >= 5) {
-                kill(app.processIdentifier, SIGKILL);
-                break;
-            }
-
-            [[NSRunLoop currentRunLoop] runUntilDate:[NSDate dateWithTimeIntervalSinceNow:0.1]];
-        }
-    }
-}
-
-bool askToMoveToApplications() {
-    NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
-    if ([bundlePath hasPrefix:@"/Applications"]) {
-        return false;
-    }
-
-    NSAlert *alert = [[NSAlert alloc] init];
-    [alert setMessageText:@"Move to Applications?"];
-    [alert setInformativeText:@"Ollama works best when run from the Applications directory."];
-    [alert addButtonWithTitle:@"Move to Applications"];
-    [alert addButtonWithTitle:@"Don't move"];
-
-    [NSApp activateIgnoringOtherApps:YES];
-
-    if ([alert runModal] != NSAlertFirstButtonReturn) {
-        return false;
-    }
-
-    // move to applications
-    NSString *applicationsPath = @"/Applications";
-    NSString *newPath = [applicationsPath stringByAppendingPathComponent:@"Ollama.app"];
-    NSFileManager *fileManager = [NSFileManager defaultManager];
-
-    // Check if the newPath already exists
-    if ([fileManager fileExistsAtPath:newPath]) {
-        NSError *removeError = nil;
-        [fileManager removeItemAtPath:newPath error:&removeError];
-        if (removeError) {
-            NSLog(@"Error removing file at %@: %@", newPath, removeError);
-            return false; // or handle the error
-        }
-    }
-
-    NSError *moveError = nil;
-    [fileManager moveItemAtPath:bundlePath toPath:newPath error:&moveError];
-    if (moveError) {
-        NSLog(@"Error moving file from %@ to %@: %@", bundlePath, newPath, moveError);
-        return false;
-    }
-
-    NSLog(@"Opening %@", newPath);
-    NSError *error = nil;
-    NSWorkspace *workspace = [NSWorkspace sharedWorkspace];
-#pragma clang diagnostic ignored "-Wdeprecated-declarations"
-    [workspace launchApplicationAtURL:[NSURL fileURLWithPath:newPath]
-               options:NSWorkspaceLaunchNewInstance | NSWorkspaceLaunchDefault
-               configuration:@{}
-               error:&error];
-
-    return true;
-}
-
-int installSymlink() {
-    NSString *linkPath = @"/usr/local/bin/ollama";
-    NSError *error = nil;
-
-    NSFileManager *fileManager = [NSFileManager defaultManager];
-    NSString *symlinkPath = [fileManager destinationOfSymbolicLinkAtPath:linkPath error:&error];
-    NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
-    NSString *execPath = [[NSBundle mainBundle] executablePath];
-    NSString *resPath = [[NSBundle mainBundle] pathForResource:@"ollama" ofType:nil];
-
-    // if the symlink already exists and points to the right place, don't prompt
-    if ([symlinkPath isEqualToString:resPath]) {
-        NSLog(@"symbolic link already exists and points to the right place");
-        return 0;
-    }
-
-    NSString *authorizationPrompt = @"Ollama is trying to install its command line interface (CLI) tool.";
-
-    AuthorizationRef auth = NULL;
-    OSStatus createStatus = AuthorizationCreate(NULL, kAuthorizationEmptyEnvironment, kAuthorizationFlagDefaults, &auth);
-    if (createStatus != errAuthorizationSuccess) {
-        NSLog(@"Error creating authorization");
-        return -1;
-    }
-
-    NSString * bundleIdentifier = [[NSBundle mainBundle] bundleIdentifier];
-    NSString *rightNameString = [NSString stringWithFormat:@"%@.%@", bundleIdentifier, @"auth3"];
-    const char *rightName = rightNameString.UTF8String;
-
-    OSStatus getRightResult = AuthorizationRightGet(rightName, NULL);
-    if (getRightResult == errAuthorizationDenied) {
-        if (AuthorizationRightSet(auth, rightName, (__bridge CFTypeRef _Nonnull)(@(kAuthorizationRuleAuthenticateAsAdmin)), (__bridge CFStringRef _Nullable)(authorizationPrompt), NULL, NULL) != errAuthorizationSuccess) {
-            NSLog(@"Failed to set right");
-            return -1;
-        }
-    }
-
-    AuthorizationItem right = { .name = rightName, .valueLength = 0, .value = NULL, .flags = 0 };
-    AuthorizationRights rights = { .count = 1, .items = &right };
-    AuthorizationFlags flags = (AuthorizationFlags)(kAuthorizationFlagExtendRights | kAuthorizationFlagInteractionAllowed);
-    AuthorizationItem iconAuthorizationItem = {.name = kAuthorizationEnvironmentIcon, .valueLength = 0, .value = NULL, .flags = 0};
-    AuthorizationEnvironment authorizationEnvironment = {.count = 0, .items = NULL};
-
-    BOOL failedToUseSystemDomain = NO;
-    OSStatus copyStatus = AuthorizationCopyRights(auth, &rights, &authorizationEnvironment, flags, NULL);
-    if (copyStatus != errAuthorizationSuccess) {
-        failedToUseSystemDomain = YES;
-
-        if (copyStatus == errAuthorizationCanceled) {
-            NSLog(@"User cancelled authorization");
-            return -1;
-        } else {
-            NSLog(@"Failed copying system domain rights: %d", copyStatus);
-            return -1;
-        }
-    }
-
-    const char *toolPath = "/bin/ln";
-    const char *args[] = {"-s", "-F", [resPath UTF8String], "/usr/local/bin/ollama", NULL};
-    FILE *pipe = NULL;
-
-#pragma clang diagnostic ignored "-Wdeprecated-declarations"
-    OSStatus status = AuthorizationExecuteWithPrivileges(auth, toolPath, kAuthorizationFlagDefaults, (char *const *)args, &pipe);
-    if (status != errAuthorizationSuccess) {
-        NSLog(@"Failed to create symlink");
-        return -1;
-    }
-
-    AuthorizationFree(auth, kAuthorizationFlagDestroyRights);
-    return 0;
-}

app/app_windows.go

@@ -1,166 +0,0 @@
-package main
-
-import (
-	"context"
-	"errors"
-	"fmt"
-	"log"
-	"log/slog"
-	"os"
-	"os/exec"
-	"os/signal"
-	"path/filepath"
-	"strings"
-	"syscall"
-
-	"github.com/ollama/ollama/app/lifecycle"
-	"github.com/ollama/ollama/app/store"
-	"github.com/ollama/ollama/app/tray"
-	"github.com/ollama/ollama/app/updater"
-)
-
-func init() {
-	AppName += ".exe"
-	CLIName += ".exe"
-	// Logs, configs, downloads go to LOCALAPPDATA
-	localAppData := os.Getenv("LOCALAPPDATA")
-	AppDataDir = filepath.Join(localAppData, "Ollama")
-	AppLogFile = filepath.Join(AppDataDir, "app.log")
-	ServerLogFile = filepath.Join(AppDataDir, "server.log")
-
-	// Executables are stored in APPDATA
-	AppDir = filepath.Join(localAppData, "Programs", "Ollama")
-
-	// Make sure we have PATH set correctly for any spawned children
-	paths := strings.Split(os.Getenv("PATH"), ";")
-	// Start with whatever we find in the PATH/LD_LIBRARY_PATH
-	found := false
-	for _, path := range paths {
-		d, err := filepath.Abs(path)
-		if err != nil {
-			continue
-		}
-		if strings.EqualFold(AppDir, d) {
-			found = true
-		}
-	}
-	if !found {
-		paths = append(paths, AppDir)
-
-		pathVal := strings.Join(paths, ";")
-		slog.Debug("setting PATH=" + pathVal)
-		err := os.Setenv("PATH", pathVal)
-		if err != nil {
-			slog.Error(fmt.Sprintf("failed to update PATH: %s", err))
-		}
-	}
-
-	// Make sure our logging dir exists
-	_, err := os.Stat(AppDataDir)
-	if errors.Is(err, os.ErrNotExist) {
-		if err := os.MkdirAll(AppDataDir, 0o755); err != nil {
-			slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err))
-		}
-	}
-}
-
-func ShowLogs() {
-	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
-	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
-	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
-	err := cmd.Start()
-	if err != nil {
-		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
-	}
-}
-
-func Start() {
-	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
-	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
-	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
-	err := cmd.Start()
-	if err != nil {
-		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
-	}
-}
-
-func run() {
-	initLogging()
-
-	slog.Info("ollama windows app started")
-
-	ctx, cancel := context.WithCancel(context.Background())
-	var done chan int
-
-	t, err := tray.NewTray()
-	if err != nil {
-		log.Fatalf("Failed to start: %s", err)
-	}
-	callbacks := t.GetCallbacks()
-
-	signals := make(chan os.Signal, 1)
-	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
-
-	go func() {
-		slog.Debug("starting callback loop")
-		for {
-			select {
-			case <-callbacks.Quit:
-				slog.Debug("quit called")
-				t.Quit()
-			case <-signals:
-				slog.Debug("shutting down due to signal")
-				t.Quit()
-			case <-callbacks.Update:
-				err := updater.DoUpgrade(cancel, done)
-				if err != nil {
-					slog.Warn(fmt.Sprintf("upgrade attempt failed: %s", err))
-				}
-			case <-callbacks.ShowLogs:
-				ShowLogs()
-			case <-callbacks.DoFirstUse:
-				err := lifecycle.GetStarted()
-				if err != nil {
-					slog.Warn(fmt.Sprintf("Failed to launch getting started shell: %s", err))
-				}
-			}
-		}
-	}()
-
-	if !store.GetFirstTimeRun() {
-		slog.Debug("First time run")
-		err = t.DisplayFirstUseNotification()
-		if err != nil {
-			slog.Debug(fmt.Sprintf("XXX failed to display first use notification %v", err))
-		}
-		store.SetFirstTimeRun(true)
-	} else {
-		slog.Debug("Not first time, skipping first run notification")
-	}
-
-	if isServerRunning(ctx) {
-		slog.Info("Detected another instance of ollama running, exiting")
-		os.Exit(1)
-	}
-
-	done, err = SpawnServer(ctx, CLIName)
-	if err != nil {
-		// TODO - should we retry in a backoff loop?
-		// TODO - should we pop up a warning and maybe add a menu item to view application logs?
-		slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
-		done = make(chan int, 1)
-		done <- 1
-	}
-
-	updater.StartBackgroundUpdaterChecker(ctx, t.UpdateAvailable)
-
-	t.Run()
-	cancel()
-	slog.Info("Waiting for ollama server to shutdown...")
-	if done != nil {
-		<-done
-	}
-	slog.Info("Ollama app exiting")
-}

app/darwin/Ollama.app/Contents/Info.plist

@@ -1,40 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
-<plist version="1.0">
-  <dict>
-    <key>CFBundleDisplayName</key>
-    <string>Ollama</string>
-    <key>CFBundleExecutable</key>
-    <string>Ollama</string>
-    <key>CFBundleIconFile</key>
-    <string>icon.icns</string>
-    <key>CFBundleIdentifier</key>
-    <string>com.ollama.ollama</string>
-    <key>CFBundleInfoDictionaryVersion</key>
-    <string>6.0</string>
-    <key>CFBundleName</key>
-    <string>Ollama</string>
-    <key>CFBundlePackageType</key>
-    <string>APPL</string>
-    <key>CFBundleShortVersionString</key>
-    <string>0.0.0</string>
-    <key>CFBundleVersion</key>
-    <string>0.0.0</string>
-    <key>DTCompiler</key>
-    <string>com.apple.compilers.llvm.clang.1_0</string>
-    <key>DTSDKBuild</key>
-    <string>22E245</string>
-    <key>DTSDKName</key>
-    <string>macosx13.3</string>
-    <key>DTXcode</key>
-    <string>1431</string>
-    <key>DTXcodeBuild</key>
-    <string>14E300c</string>
-    <key>LSApplicationCategoryType</key>
-    <string>public.app-category.developer-tools</string>
-    <key>LSMinimumSystemVersion</key>
-    <string>11.0</string>
-    <key>LSUIElement</key>
-    <true/>
-  </dict>
-</plist>

app/lifecycle/getstarted_nonwindows.go

@@ -1,3 +1,5 @@
+//go:build !windows
+
 package lifecycle
 
 import "fmt"

app/lifecycle/lifecycle.go

@@ -0,0 +1,92 @@
+package lifecycle
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"log/slog"
+	"os"
+	"os/signal"
+	"syscall"
+
+	"github.com/ollama/ollama/app/store"
+	"github.com/ollama/ollama/app/tray"
+)
+
+func Run() {
+	InitLogging()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	var done chan int
+
+	t, err := tray.NewTray()
+	if err != nil {
+		log.Fatalf("Failed to start: %s", err)
+	}
+	callbacks := t.GetCallbacks()
+
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
+
+	go func() {
+		slog.Debug("starting callback loop")
+		for {
+			select {
+			case <-callbacks.Quit:
+				slog.Debug("quit called")
+				t.Quit()
+			case <-signals:
+				slog.Debug("shutting down due to signal")
+				t.Quit()
+			case <-callbacks.Update:
+				err := DoUpgrade(cancel, done)
+				if err != nil {
+					slog.Warn(fmt.Sprintf("upgrade attempt failed: %s", err))
+				}
+			case <-callbacks.ShowLogs:
+				ShowLogs()
+			case <-callbacks.DoFirstUse:
+				err := GetStarted()
+				if err != nil {
+					slog.Warn(fmt.Sprintf("Failed to launch getting started shell: %s", err))
+				}
+			}
+		}
+	}()
+
+	// Are we first use?
+	if !store.GetFirstTimeRun() {
+		slog.Debug("First time run")
+		err = t.DisplayFirstUseNotification()
+		if err != nil {
+			slog.Debug(fmt.Sprintf("XXX failed to display first use notification %v", err))
+		}
+		store.SetFirstTimeRun(true)
+	} else {
+		slog.Debug("Not first time, skipping first run notification")
+	}
+
+	if IsServerRunning(ctx) {
+		slog.Info("Detected another instance of ollama running, exiting")
+		os.Exit(1)
+	} else {
+		done, err = SpawnServer(ctx, CLIName)
+		if err != nil {
+			// TODO - should we retry in a backoff loop?
+			// TODO - should we pop up a warning and maybe add a menu item to view application logs?
+			slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
+			done = make(chan int, 1)
+			done <- 1
+		}
+	}
+
+	StartBackgroundUpdaterChecker(ctx, t.UpdateAvailable)
+
+	t.Run()
+	cancel()
+	slog.Info("Waiting for ollama server to shutdown...")
+	if done != nil {
+		<-done
+	}
+	slog.Info("Ollama app exiting")
+}

app/lifecycle/logging.go

@@ -1,4 +1,4 @@
-package main
+package lifecycle
 
 import (
 	"fmt"
@@ -7,7 +7,7 @@ import (
 	"path/filepath"
 )
 
-func initLogging() {
+func InitLogging() {
 	level := slog.LevelInfo
 
 	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
@@ -41,4 +41,6 @@ func initLogging() {
 	})
 
 	slog.SetDefault(slog.New(handler))
+
+	slog.Info("ollama app started")
 }

app/lifecycle/logging_nonwindows.go

@@ -0,0 +1,9 @@
+//go:build !windows
+
+package lifecycle
+
+import "log/slog"
+
+func ShowLogs() {
+	slog.Warn("ShowLogs not yet implemented")
+}

app/lifecycle/logging_windows.go

@@ -0,0 +1,19 @@
+package lifecycle
+
+import (
+	"fmt"
+	"log/slog"
+	"os/exec"
+	"syscall"
+)
+
+func ShowLogs() {
+	cmd_path := "c:\\Windows\\system32\\cmd.exe"
+	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
+	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
+	err := cmd.Start()
+	if err != nil {
+		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
+	}
+}

app/lifecycle/paths.go

@@ -70,5 +70,10 @@ func init() {
 			}
 		}
 
+	} else if runtime.GOOS == "darwin" {
+		// TODO
+		AppName += ".app"
+		// } else if runtime.GOOS == "linux" {
+		// TODO
 	}
 }

app/lifecycle/server.go

@@ -1,4 +1,4 @@
-package main
+package lifecycle
 
 import (
 	"context"
@@ -14,41 +14,65 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
-type ServerOptions struct {
+func getCLIFullPath(command string) string {
-	Cors       bool
+	cmdPath := ""
-	Expose     bool
+	appExe, err := os.Executable()
-	ModelsPath string
+	if err == nil {
+		cmdPath = filepath.Join(filepath.Dir(appExe), command)
+		_, err := os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
+	}
+	cmdPath, err = exec.LookPath(command)
+	if err == nil {
+		_, err := os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
+	}
+	pwd, err := os.Getwd()
+	if err == nil {
+		cmdPath = filepath.Join(pwd, command)
+		_, err = os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
+	}
+
+	return command
 }
 
-func start(ctx context.Context, command string, options ServerOptions) (*exec.Cmd, error) {
+func SpawnServer(ctx context.Context, command string) (chan int, error) {
-	cmd := getCmd(ctx, command)
+	done := make(chan int)
 
-	// set environment variables
+	logDir := filepath.Dir(ServerLogFile)
-	if options.ModelsPath != "" {
+	_, err := os.Stat(logDir)
-		cmd.Env = append(cmd.Env, fmt.Sprintf("OLLAMA_MODELS=%s", options.ModelsPath))
+	if errors.Is(err, os.ErrNotExist) {
-	}
+		if err := os.MkdirAll(logDir, 0o755); err != nil {
-
+			return done, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
-	if options.Cors {
+		}
-		cmd.Env = append(cmd.Env, "OLLAMA_ORIGINS=*")
-	}
-
-	if options.Expose {
-		cmd.Env = append(cmd.Env, "OLLAMA_HOST=0.0.0.0")
 	}
 
+	cmd := getCmd(ctx, getCLIFullPath(command))
+	// send stdout and stderr to a file
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
-		return nil, fmt.Errorf("failed to spawn server stdout pipe: %w", err)
+		return done, fmt.Errorf("failed to spawn server stdout pipe %s", err)
 	}
 	stderr, err := cmd.StderrPipe()
 	if err != nil {
-		return nil, fmt.Errorf("failed to spawn server stderr pipe: %w", err)
+		return done, fmt.Errorf("failed to spawn server stderr pipe %s", err)
+	}
+	stdin, err := cmd.StdinPipe()
+	if err != nil {
+		return done, fmt.Errorf("failed to spawn server stdin pipe %s", err)
 	}
 
 	// 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)
+		return done, fmt.Errorf("failed to create server log %w", err)
 	}
 	go func() {
 		defer logFile.Close()
@@ -93,38 +117,19 @@ func start(ctx context.Context, command string, options ServerOptions) (*exec.Cm
 
 	// run the command and wait for it to finish
 	if err := cmd.Start(); err != nil {
-		return nil, fmt.Errorf("failed to start server %w", err)
+		return done, fmt.Errorf("failed to start server %w", err)
 	}
 	if cmd.Process != nil {
 		slog.Info(fmt.Sprintf("started ollama server with pid %d", cmd.Process.Pid))
 	}
 	slog.Info(fmt.Sprintf("ollama server logs %s", ServerLogFile))
 
-	return cmd, nil
-}
-
-func SpawnServer(ctx context.Context, command string, options ServerOptions) (chan int, error) {
-	logDir := filepath.Dir(ServerLogFile)
-	_, err := os.Stat(logDir)
-	if errors.Is(err, os.ErrNotExist) {
-		if err := os.MkdirAll(logDir, 0o755); err != nil {
-			return nil, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
-		}
-	}
-
-	done := make(chan int)
-
 	go func() {
 		// Keep the server running unless we're shuttind down the app
 		crashCount := 0
 		for {
-			slog.Info(fmt.Sprintf("starting server..."))
-			cmd, err := start(ctx, command, options)
-			if err != nil {
-				slog.Error(fmt.Sprintf("failed to start server %s", err))
-			}
-
 			cmd.Wait() //nolint:errcheck
+			stdin.Close()
 			var code int
 			if cmd.ProcessState != nil {
 				code = cmd.ProcessState.ExitCode()
@@ -138,16 +143,19 @@ func SpawnServer(ctx context.Context, command string, options ServerOptions) (ch
 			default:
 				crashCount++
 				slog.Warn(fmt.Sprintf("server crash %d - exit code %d - respawning", crashCount, code))
-				time.Sleep(500 * time.Millisecond * time.Duration(crashCount))
+				time.Sleep(500 * time.Millisecond)
-				break
+				if err := cmd.Start(); err != nil {
+					slog.Error(fmt.Sprintf("failed to restart server %s", err))
+					// Keep trying, but back off if we keep failing
+					time.Sleep(time.Duration(crashCount) * time.Second)
+				}
 			}
 		}
 	}()
-
 	return done, nil
 }
 
-func isServerRunning(ctx context.Context) bool {
+func IsServerRunning(ctx context.Context) bool {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		slog.Info("unable to connect to server")

app/lifecycle/server_unix.go

@@ -1,4 +1,6 @@
-package main
+//go:build !windows
+
+package lifecycle
 
 import (
 	"context"

app/lifecycle/server_windows.go

@@ -1,4 +1,4 @@
-package main
+package lifecycle
 
 import (
 	"context"

app/lifecycle/updater.go

@@ -1,4 +1,4 @@
-package updater
+package lifecycle
 
 import (
 	"context"
@@ -22,10 +22,6 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
-var (
-	UpdateStageDir string
-)
-
 var (
 	UpdateCheckURLBase  = "https://ollama.com/api/update"
 	UpdateDownloaded    = false
@@ -127,7 +123,7 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
 		slog.Debug("no etag detected, falling back to filename based dedup")
 		etag = "_"
 	}
-	filename := "OllamaSetup.exe"
+	filename := Installer
 	_, params, err := mime.ParseMediaType(resp.Header.Get("content-disposition"))
 	if err == nil {
 		filename = params["filename"]

app/lifecycle/updater_nonwindows.go

@@ -1,4 +1,6 @@
-package updater
+//go:build !windows
+
+package lifecycle
 
 import (
 	"context"

app/lifecycle/updater_windows.go

@@ -1,4 +1,4 @@
-package updater
+package lifecycle
 
 import (
 	"context"
@@ -9,13 +9,7 @@ import (
 	"path/filepath"
 )
 
-func init() {
-	UpdateStageDir = filepath.Join(os.Getenv("LOCALAPPDATA"), "Ollama", "updates")
-}
-
 func DoUpgrade(cancel context.CancelFunc, done chan int) error {
-	logFile := filepath.Join(os.Getenv("LOCALAPPDATA"), "Ollama", "upgrade.log")
-
 	files, err := filepath.Glob(filepath.Join(UpdateStageDir, "*", "*.exe")) // TODO generalize for multiplatform
 	if err != nil {
 		return fmt.Errorf("failed to lookup downloads: %s", err)
@@ -29,13 +23,13 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 	installerExe := files[0]
 
 	slog.Info("starting upgrade with " + installerExe)
-	slog.Info("upgrade log file " + logFile)
+	slog.Info("upgrade log file " + UpgradeLogFile)
 
 	// When running in debug mode, we'll be "verbose" and let the installer pop up and prompt
 	installArgs := []string{
-		"/CLOSEAPPLICATIONS",             // Quit the tray app if it's still running
+		"/CLOSEAPPLICATIONS",                    // Quit the tray app if it's still running
-		"/LOG=" + filepath.Base(logFile), // Only relative seems reliable, so set pwd
+		"/LOG=" + filepath.Base(UpgradeLogFile), // Only relative seems reliable, so set pwd
-		"/FORCECLOSEAPPLICATIONS",        // Force close the tray app - might be needed
+		"/FORCECLOSEAPPLICATIONS",               // Force close the tray app - might be needed
 	}
 	// When we're not in debug mode, make the upgrade as quiet as possible (no GUI, no prompts)
 	// TODO - temporarily disable since we're pinning in debug mode for the preview
@@ -59,7 +53,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 	}
 
 	slog.Debug(fmt.Sprintf("starting installer: %s %v", installerExe, installArgs))
-	os.Chdir(filepath.Dir(logFile)) //nolint:errcheck
+	os.Chdir(filepath.Dir(UpgradeLogFile)) //nolint:errcheck
 	cmd := exec.Command(installerExe, installArgs...)
 
 	if err := cmd.Start(); err != nil {

app/main.go

@@ -2,15 +2,11 @@ package main
 
 // Compile with the following to get rid of the cmd pop up on windows
 // go build -ldflags="-H windowsgui" .
-var (
+
-	AppName       string
+import (
-	CLIName       string
+	"github.com/ollama/ollama/app/lifecycle"
-	AppDir        string
-	AppDataDir    string
-	AppLogFile    string
-	ServerLogFile string
 )
 
 func main() {
-	run()
+	lifecycle.Run()
 }

app/ollama.iss

@@ -88,12 +88,15 @@ DialogFontSize=12
 [Files]
 Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
 Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-amd64\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
+Source: "..\dist\windeps\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-amd64\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\rocm")
+; Assumes v5.7, may need adjustments for v6
-  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
+#if GetEnv("HIP_PATH") != ""
+  Source: "{#GetEnv('HIP_PATH')}\bin\hipblas.dll"; DestDir: "{app}\rocm\"; Flags: ignoreversion
+  Source: "{#GetEnv('HIP_PATH')}\bin\rocblas.dll"; DestDir: "{app}\rocm\"; Flags: ignoreversion
+  ; amdhip64.dll dependency comes from the driver and must be installed already
+  Source: "{#GetEnv('HIP_PATH')}\bin\rocblas\library\*"; DestDir: "{app}\rocm\rocblas\library\"; Flags: ignoreversion
 #endif
 
 
@@ -129,7 +132,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
 
 
 ;FinishedHeadingLabel=Run your first model
-;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama2
 ;ClickFinish=%n
 
 [Registry]

app/tray/tray_nonwindows.go

@@ -1,3 +1,5 @@
+//go:build !windows
+
 package tray
 
 import (

cmd/cmd.go

@@ -17,7 +17,6 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
-	"regexp"
 	"runtime"
 	"strings"
 	"syscall"
@@ -54,6 +53,8 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
+	bars := make(map[string]*progress.Bar)
+
 	modelfile, err := os.ReadFile(filename)
 	if err != nil {
 		return err
@@ -94,16 +95,95 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				return err
 			}
 
+			// TODO make this work w/ adapters
 			if fi.IsDir() {
-				// this is likely a safetensors or pytorch directory
+				tf, err := os.CreateTemp("", "ollama-tf")
-				// TODO make this work w/ adapters
-				tempfile, err := tempZipFiles(path)
 				if err != nil {
 					return err
 				}
-				defer os.RemoveAll(tempfile)
+				defer os.RemoveAll(tf.Name())
 
-				path = tempfile
+				zf := zip.NewWriter(tf)
+
+				files := []string{}
+
+				tfiles, err := filepath.Glob(filepath.Join(path, "pytorch_model-*.bin"))
+				if err != nil {
+					return err
+				} else if len(tfiles) == 0 {
+					tfiles, err = filepath.Glob(filepath.Join(path, "model-*.safetensors"))
+					if err != nil {
+						return err
+					}
+				}
+
+				files = append(files, tfiles...)
+
+				if len(files) == 0 {
+					return fmt.Errorf("no models were found in '%s'", path)
+				}
+
+				// add the safetensor/torch config file + tokenizer
+				files = append(files, filepath.Join(path, "config.json"))
+				files = append(files, filepath.Join(path, "params.json"))
+				files = append(files, filepath.Join(path, "added_tokens.json"))
+				files = append(files, filepath.Join(path, "tokenizer.model"))
+
+				for _, fn := range files {
+					f, err := os.Open(fn)
+
+					// just skip whatever files aren't there
+					if os.IsNotExist(err) {
+						if strings.HasSuffix(fn, "tokenizer.model") {
+							// try the parent dir before giving up
+							parentDir := filepath.Dir(path)
+							newFn := filepath.Join(parentDir, "tokenizer.model")
+							f, err = os.Open(newFn)
+							if os.IsNotExist(err) {
+								continue
+							} else if err != nil {
+								return err
+							}
+						} else {
+							continue
+						}
+					} else if err != nil {
+						return err
+					}
+
+					fi, err := f.Stat()
+					if err != nil {
+						return err
+					}
+
+					h, err := zip.FileInfoHeader(fi)
+					if err != nil {
+						return err
+					}
+
+					h.Name = filepath.Base(fn)
+					h.Method = zip.Store
+
+					w, err := zf.CreateHeader(h)
+					if err != nil {
+						return err
+					}
+
+					_, err = io.Copy(w, f)
+					if err != nil {
+						return err
+					}
+
+				}
+
+				if err := zf.Close(); err != nil {
+					return err
+				}
+
+				if err := tf.Close(); err != nil {
+					return err
+				}
+				path = tf.Name()
 			}
 
 			digest, err := createBlob(cmd, client, path)
@@ -111,17 +191,10 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				return err
 			}
 
-			name := c.Name
+			modelfile = bytes.ReplaceAll(modelfile, []byte(c.Args), []byte("@"+digest))
-			if c.Name == "model" {
-				name = "from"
-			}
-
-			re := regexp.MustCompile(fmt.Sprintf(`(?im)^(%s)\s+%s\s*$`, name, c.Args))
-			modelfile = re.ReplaceAll(modelfile, []byte("$1 @"+digest))
 		}
 	}
 
-	bars := make(map[string]*progress.Bar)
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
 			spinner.Stop()
@@ -155,114 +228,6 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
-func tempZipFiles(path string) (string, error) {
-	tempfile, err := os.CreateTemp("", "ollama-tf")
-	if err != nil {
-		return "", err
-	}
-	defer tempfile.Close()
-
-	zipfile := zip.NewWriter(tempfile)
-	defer zipfile.Close()
-
-	detectContentType := func(path string) (string, error) {
-		f, err := os.Open(path)
-		if err != nil {
-			return "", err
-		}
-		defer f.Close()
-
-		var b bytes.Buffer
-		b.Grow(512)
-
-		if _, err := io.CopyN(&b, f, 512); err != nil && !errors.Is(err, io.EOF) {
-			return "", err
-		}
-
-		contentType, _, _ := strings.Cut(http.DetectContentType(b.Bytes()), ";")
-		return contentType, nil
-	}
-
-	glob := func(pattern, contentType string) ([]string, error) {
-		matches, err := filepath.Glob(pattern)
-		if err != nil {
-			return nil, err
-		}
-
-		for _, safetensor := range matches {
-			if ct, err := detectContentType(safetensor); err != nil {
-				return nil, err
-			} else if ct != contentType {
-				return nil, fmt.Errorf("invalid content type: expected %s for %s", ct, safetensor)
-			}
-		}
-
-		return matches, nil
-	}
-
-	var files []string
-	if st, _ := glob(filepath.Join(path, "model*.safetensors"), "application/octet-stream"); len(st) > 0 {
-		// safetensors files might be unresolved git lfs references; skip if they are
-		// covers model-x-of-y.safetensors, model.fp32-x-of-y.safetensors, model.safetensors
-		files = append(files, st...)
-	} else if pt, _ := glob(filepath.Join(path, "pytorch_model*.bin"), "application/zip"); len(pt) > 0 {
-		// pytorch files might also be unresolved git lfs references; skip if they are
-		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
-		files = append(files, pt...)
-	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/octet-stream"); len(pt) > 0 {
-		// pytorch files might also be unresolved git lfs references; skip if they are
-		// covers consolidated.x.pth, consolidated.pth
-		files = append(files, pt...)
-	} else {
-		return "", errors.New("no safetensors or torch files found")
-	}
-
-	// add configuration files, json files are detected as text/plain
-	js, err := glob(filepath.Join(path, "*.json"), "text/plain")
-	if err != nil {
-		return "", err
-	}
-	files = append(files, js...)
-
-	if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
-		// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
-		// tokenizer.model might be a unresolved git lfs reference; error if it is
-		files = append(files, tks...)
-	} else if tks, _ := glob(filepath.Join(path, "**/tokenizer.model"), "text/plain"); len(tks) > 0 {
-		// some times tokenizer.model is in a subdirectory (e.g. meta-llama/Meta-Llama-3-8B)
-		files = append(files, tks...)
-	}
-
-	for _, file := range files {
-		f, err := os.Open(file)
-		if err != nil {
-			return "", err
-		}
-		defer f.Close()
-
-		fi, err := f.Stat()
-		if err != nil {
-			return "", err
-		}
-
-		zfi, err := zip.FileInfoHeader(fi)
-		if err != nil {
-			return "", err
-		}
-
-		zf, err := zipfile.CreateHeader(zfi)
-		if err != nil {
-			return "", err
-		}
-
-		if _, err := io.Copy(zf, f); err != nil {
-			return "", err
-		}
-	}
-
-	return tempfile.Name(), nil
-}
-
 func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, error) {
 	bin, err := os.Open(path)
 	if err != nil {

convert/convert.go

@@ -18,23 +18,19 @@ import (
 )
 
 type Params struct {
-	Architectures     []string `json:"architectures"`
+	Architectures    []string `json:"architectures"`
-	VocabSize         int      `json:"vocab_size"`
+	VocabSize        int      `json:"vocab_size"`
-	HiddenSize        int      `json:"hidden_size"`       // n_embd
+	HiddenSize       int      `json:"hidden_size"`       // n_embd
-	HiddenLayers      int      `json:"num_hidden_layers"` // n_layer
+	HiddenLayers     int      `json:"num_hidden_layers"` // n_layer
-	ContextSize       int      `json:"max_position_embeddings"`
+	ContextSize      int      `json:"max_position_embeddings"`
-	IntermediateSize  int      `json:"intermediate_size"`
+	IntermediateSize int      `json:"intermediate_size"`
-	AttentionHeads    int      `json:"num_attention_heads"` // n_head
+	AttentionHeads   int      `json:"num_attention_heads"` // n_head
-	KeyValHeads       int      `json:"num_key_value_heads"`
+	KeyValHeads      int      `json:"num_key_value_heads"`
-	NormEPS           float64  `json:"rms_norm_eps"`
+	NormEPS          float64  `json:"rms_norm_eps"`
-	BoSTokenID        int      `json:"bos_token_id"`
+	BoSTokenID       int      `json:"bos_token_id"`
-	EoSTokenID        int      `json:"eos_token_id"`
+	EoSTokenID       int      `json:"eos_token_id"`
-	HeadDimension     int      `json:"head_dim"`
+	HeadDimension    int      `json:"head_dim"`
-	PaddingTokenID    int      `json:"pad_token_id"`
+	PaddingTokenID   int      `json:"pad_token_id"`
-	RopeFrequencyBase float64  `json:"rope_theta"`
-
-	Experts     int `json:"num_local_experts"`
-	ExpertsUsed int `json:"num_experts_per_tok"`
 
 	ByteOrder
 }

convert/mixtral.go

@@ -1,96 +0,0 @@
-package convert
-
-import (
-	"os"
-	"regexp"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type MixtralModel struct {
-	ModelData
-}
-
-func (m *MixtralModel) GetTensors() error {
-	t, err := m.Format.GetTensors(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-
-	m.Tensors = []llm.Tensor{}
-
-	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
-	re, err := regexp.Compile(pattern)
-	if err != nil {
-		return err
-	}
-
-	for _, l := range t {
-		matches := re.FindAllStringSubmatch(l.Name, -1)
-		if len(matches) > 0 {
-			wt := l.WriterTo.(safetensorWriterTo)
-			wt.handler = mistralLayerHandler
-			l.WriterTo = wt
-		}
-		m.Tensors = append(m.Tensors, l)
-	}
-
-	return nil
-}
-
-func (m *MixtralModel) LoadVocab() error {
-	v, err := LoadSentencePieceTokens(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-	m.Vocab = v
-	return nil
-}
-
-func (m *MixtralModel) WriteGGUF() (string, error) {
-	kv := llm.KV{
-		"general.architecture":          "llama",
-		"general.name":                  m.Name,
-		"llama.block_count":             uint32(m.Params.HiddenLayers),
-		"llama.context_length":          uint32(m.Params.ContextSize),
-		"llama.embedding_length":        uint32(m.Params.HiddenSize),
-		"llama.feed_forward_length":     uint32(m.Params.IntermediateSize),
-		"llama.attention.head_count":    uint32(m.Params.AttentionHeads),
-		"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
-
-		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
-		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
-
-		"llama.expert_count":      uint32(m.Params.Experts),
-		"llama.expert_used_count": uint32(m.Params.ExpertsUsed),
-
-		"llama.vocab_size":           uint32(len(m.Vocab.Tokens)),
-		"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
-
-		"general.file_type":    uint32(1),
-		"tokenizer.ggml.model": "llama",
-
-		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.scores":     m.Vocab.Scores,
-		"tokenizer.ggml.token_type": m.Vocab.Types,
-
-		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
-		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
-		"tokenizer.ggml.unknown_token_id": uint32(0),
-		"tokenizer.ggml.add_bos_token":    true,
-		"tokenizer.ggml.add_eos_token":    false,
-	}
-
-	f, err := os.CreateTemp("", "ollama-gguf")
-	if err != nil {
-		return "", err
-	}
-	defer f.Close()
-
-	mod := llm.NewGGUFV3(m.Params.ByteOrder)
-	if err := mod.Encode(f, kv, m.Tensors); err != nil {
-		return "", err
-	}
-
-	return f.Name(), nil
-}

convert/safetensors.go

@@ -93,6 +93,7 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
 	}
 
 	slices.Sort(keys)
+
 	slog.Info("converting layers")
 
 	var tensors []llm.Tensor
@@ -104,6 +105,7 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
 			return nil, 0, err
 		}
 
+		slog.Debug(fmt.Sprintf("metadata = %#v", data))
 		var size uint64
 		var kind uint32
 		switch len(data.Shape) {
@@ -148,13 +150,11 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
 			padding:  8 + jsonSize,
 		}
 
-		offset += size
 		tensors = append(tensors, t)
+		offset += size
 	}
-
 	slog.Debug(fmt.Sprintf("total tensors for file = %d", len(tensors)))
 	slog.Debug(fmt.Sprintf("offset = %d", offset))
-
 	return tensors, offset, nil
 }
 
@@ -185,19 +185,15 @@ func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
 	}
 
 	tMap := map[string]string{
-		"model.layers.(\\d+).input_layernorm.weight":                    "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
-		"model.layers.(\\d+).mlp.down_proj.weight":                      "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
-		"model.layers.(\\d+).mlp.gate_proj.weight":                      "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
-		"model.layers.(\\d+).mlp.up_proj.weight":                        "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
-		"model.layers.(\\d+).post_attention_layernorm.weight":           "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
-		"model.layers.(\\d+).self_attn.k_proj.weight":                   "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
-		"model.layers.(\\d+).self_attn.o_proj.weight":                   "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
-		"model.layers.(\\d+).self_attn.q_proj.weight":                   "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
-		"model.layers.(\\d+).self_attn.v_proj.weight":                   "blk.$1.attn_v.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
-		"model.layers.(\\d+).block_sparse_moe.gate.weight":              "blk.$1.ffn_gate_inp.weight",
-		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w1.weight": "blk.$1.ffn_gate.$2.weight",
-		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w2.weight": "blk.$1.ffn_down.$2.weight",
-		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w3.weight": "blk.$1.ffn_up.$2.weight",
 	}
 
 	v, ok := directMap[n]
@@ -290,15 +286,6 @@ func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (M
 					Format: m,
 				},
 			}, nil
-		case "MixtralForCausalLM":
-			return &MixtralModel{
-				ModelData{
-					Name:   name,
-					Path:   dirPath,
-					Params: params,
-					Format: m,
-				},
-			}, nil
 		case "GemmaForCausalLM":
 			return &GemmaModel{
 				ModelData{

docs/api.md

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

docs/faq.md

@@ -228,7 +228,3 @@ To unload the model and free up memory use:
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "llama2", "keep_alive": 0}'
 ```
-
-Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
-
-If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` API parameter with the `/api/generate` or `/api/chat` API endpoints.

docs/tutorials/nvidia-jetson.md

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

docs/windows.md

@@ -14,7 +14,7 @@ As this is a preview release, you should expect a few bugs here and there.  If
 you run into a problem you can reach out on
 [Discord](https://discord.gg/ollama), or file an 
 [issue](https://github.com/ollama/ollama/issues).
-Logs will often be helpful in diagnosing the problem (see
+Logs will often be helpful in dianosing the problem (see
 [Troubleshooting](#troubleshooting) below)
 
 ## System Requirements

format/bytes.go

@@ -15,7 +15,6 @@ const (
 
 	KibiByte = Byte * 1024
 	MebiByte = KibiByte * 1024
-	GibiByte = MebiByte * 1024
 )
 
 func HumanBytes(b int64) string {

gpu/amd_common.go

@@ -7,7 +7,7 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
-	"runtime"
+	"strconv"
 	"strings"
 )
 
@@ -35,64 +35,22 @@ func GetSupportedGFX(libDir string) ([]string, error) {
 	return ret, nil
 }
 
-func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
+func amdSetVisibleDevices(ids []int, skip map[int]interface{}) {
-	ids := []string{}
+	// Set the visible devices if not already set
-	for _, info := range gpuInfo {
+	// TODO - does sort order matter?
-		if info.Library != "rocm" {
+	devices := []string{}
-			// TODO shouldn't happen if things are wired correctly...
+	for i := range ids {
-			slog.Debug("rocmGetVisibleDevicesEnv skipping over non-rocm device", "library", info.Library)
+		if _, skipped := skip[i]; skipped {
 			continue
 		}
-		ids = append(ids, info.ID)
+		devices = append(devices, strconv.Itoa(i))
+	}
+
+	val := strings.Join(devices, ",")
+	err := os.Setenv("HIP_VISIBLE_DEVICES", val)
+	if err != nil {
+		slog.Warn(fmt.Sprintf("failed to set env: %s", err))
+	} else {
+		slog.Info("Setting HIP_VISIBLE_DEVICES=" + val)
 	}
-	return "HIP_VISIBLE_DEVICES", strings.Join(ids, ",")
-}
-
-func commonAMDValidateLibDir() (string, error) {
-	// We try to favor system paths first, so that we can wire up the subprocess to use
-	// the system version.  Only use our bundled version if the system version doesn't work
-	// This gives users a more recovery options if versions have subtle problems at runtime
-
-	// Prefer explicit HIP env var
-	hipPath := os.Getenv("HIP_PATH")
-	if hipPath != "" {
-		hipLibDir := filepath.Join(hipPath, "bin")
-		if rocmLibUsable(hipLibDir) {
-			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
-			return hipLibDir, nil
-		}
-	}
-
-	// Scan the LD_LIBRARY_PATH or PATH
-	pathEnv := "LD_LIBRARY_PATH"
-	if runtime.GOOS == "windows" {
-		pathEnv = "PATH"
-	}
-
-	paths := os.Getenv(pathEnv)
-	for _, path := range filepath.SplitList(paths) {
-		d, err := filepath.Abs(path)
-		if err != nil {
-			continue
-		}
-		if rocmLibUsable(d) {
-			return d, nil
-		}
-	}
-
-	// Well known location(s)
-	if rocmLibUsable(RocmStandardLocation) {
-		return RocmStandardLocation, nil
-	}
-
-	// Installer payload location if we're running the installed binary
-	exe, err := os.Executable()
-	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
-		if rocmLibUsable(rocmTargetDir) {
-			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-			return rocmTargetDir, nil
-		}
-	}
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
 }

gpu/amd_hip_windows.go

@@ -69,7 +69,7 @@ func NewHipLib() (*HipLib, error) {
 func (hl *HipLib) Release() {
 	err := windows.FreeLibrary(hl.dll)
 	if err != nil {
-		slog.Warn("failed to unload amdhip64.dll", "error", err)
+		slog.Warn(fmt.Sprintf("failed to unload amdhip64.dll: %s", err))
 	}
 	hl.dll = 0
 }
@@ -98,7 +98,7 @@ func (hl *HipLib) HipGetDeviceCount() int {
 		return 0
 	}
 	if status != hipSuccess {
-		slog.Warn("failed call to hipGetDeviceCount", "status", status, "error", err)
+		slog.Warn(fmt.Sprintf("failed call to hipGetDeviceCount: %d %s", status, err))
 	}
 	return count
 }

gpu/amd_linux.go

@@ -11,8 +11,6 @@ import (
 	"slices"
 	"strconv"
 	"strings"
-
-	"github.com/ollama/ollama/format"
 )
 
 // Discovery logic for AMD/ROCm GPUs
@@ -26,6 +24,9 @@ const (
 	GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
 	GPUUsedMemoryFileGlob  = "mem_banks/*/used_memory"
 	RocmStandardLocation   = "/opt/rocm/lib"
+
+	// TODO find a better way to detect iGPU instead of minimum memory
+	IGPUMemLimit = 1024 * 1024 * 1024 // 512G is what they typically report, so anything less than 1G must be iGPU
 )
 
 var (
@@ -34,11 +35,14 @@ var (
 )
 
 // Gather GPU information from the amdgpu driver if any supported GPUs are detected
-func AMDGetGPUInfo() []GpuInfo {
+// HIP_VISIBLE_DEVICES will be set if we detect a mix of unsupported and supported devices
-	resp := []GpuInfo{}
+// and the user hasn't already set this variable
+func AMDGetGPUInfo(resp *GpuInfo) {
+	// TODO - DRY this out with windows
 	if !AMDDetected() {
-		return resp
+		return
 	}
+	skip := map[int]interface{}{}
 
 	// Opportunistic logging of driver version to aid in troubleshooting
 	ver, err := AMDDriverVersion()
@@ -46,117 +50,160 @@ func AMDGetGPUInfo() []GpuInfo {
 		slog.Info("AMD Driver: " + ver)
 	} else {
 		// TODO - if we see users crash and burn with the upstreamed kernel this can be adjusted to hard-fail rocm support and fallback to CPU
-		slog.Warn("ollama recommends running the https://www.amd.com/en/support/linux-drivers", "error", err)
+		slog.Warn(fmt.Sprintf("ollama recommends running the https://www.amd.com/en/support/linux-drivers: %s", err))
 	}
 
-	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
+	// If the user has specified exactly which GPUs to use, look up their memory
-	var visibleDevices []string
+	visibleDevices := os.Getenv("HIP_VISIBLE_DEVICES")
-	hipVD := os.Getenv("HIP_VISIBLE_DEVICES")   // zero based index only
+	if visibleDevices != "" {
-	rocrVD := os.Getenv("ROCR_VISIBLE_DEVICES") // zero based index or UUID, but consumer cards seem to not support UUID
+		ids := []int{}
-	gpuDO := os.Getenv("GPU_DEVICE_ORDINAL")    // zero based index
+		for _, idStr := range strings.Split(visibleDevices, ",") {
-	switch {
+			id, err := strconv.Atoi(idStr)
-	// TODO is this priorty order right?
+			if err != nil {
-	case hipVD != "":
+				slog.Warn(fmt.Sprintf("malformed HIP_VISIBLE_DEVICES=%s %s", visibleDevices, err))
-		visibleDevices = strings.Split(hipVD, ",")
+			} else {
-	case rocrVD != "":
+				ids = append(ids, id)
-		visibleDevices = strings.Split(rocrVD, ",")
+			}
-		// TODO - since we don't yet support UUIDs, consider detecting and reporting here
+		}
-		// all our test systems show GPU-XX indicating UUID is not supported
+		amdProcMemLookup(resp, nil, ids)
-	case gpuDO != "":
+		return
-		visibleDevices = strings.Split(gpuDO, ",")
 	}
 
+	// Gather GFX version information from all detected cards
+	gfx := AMDGFXVersions()
+	verStrings := []string{}
+	for i, v := range gfx {
+		verStrings = append(verStrings, v.ToGFXString())
+		if v.Major == 0 {
+			// Silently skip CPUs
+			skip[i] = struct{}{}
+			continue
+		}
+		if v.Major < 9 {
+			// TODO consider this a build-time setting if we can support 8xx family GPUs
+			slog.Warn(fmt.Sprintf("amdgpu [%d] too old %s", i, v.ToGFXString()))
+			skip[i] = struct{}{}
+		}
+	}
+	slog.Info(fmt.Sprintf("detected amdgpu versions %v", verStrings))
+
+	// Abort if all GPUs are skipped
+	if len(skip) >= len(gfx) {
+		slog.Info("all detected amdgpus are skipped, falling back to CPU")
+		return
+	}
+
+	// If we got this far, then we have at least 1 GPU that's a ROCm candidate, so make sure we have a lib
+	libDir, err := AMDValidateLibDir()
+	if err != nil {
+		slog.Warn(fmt.Sprintf("unable to verify rocm library, will use cpu: %s", err))
+		return
+	}
+
+	updateLibPath(libDir)
+
 	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
-	var supported []string
+	if gfxOverride == "" {
-	libDir := ""
+		supported, err := GetSupportedGFX(libDir)
-
-	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
-	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
-	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
-	cpuCount := 0
-	for _, match := range matches {
-		slog.Debug("evaluating amdgpu node " + match)
-		fp, err := os.Open(match)
 		if err != nil {
-			slog.Debug("failed to open sysfs node", "file", match, "error", err)
+			slog.Warn(fmt.Sprintf("failed to lookup supported GFX types, falling back to CPU mode: %s", err))
-			continue
+			return
-		}
-		defer fp.Close()
-		nodeID, err := strconv.Atoi(filepath.Base(filepath.Dir(match)))
-		if err != nil {
-			slog.Debug("failed to parse node ID", "error", err)
-			continue
 		}
+		slog.Debug(fmt.Sprintf("rocm supported GPU types %v", supported))
 
-		scanner := bufio.NewScanner(fp)
+		for i, v := range gfx {
-		isCPU := false
+			if !slices.Contains[[]string, string](supported, v.ToGFXString()) {
-		var major, minor, patch uint64
+				slog.Warn(fmt.Sprintf("amdgpu [%d] %s is not supported by %s %v", i, v.ToGFXString(), libDir, supported))
-		for scanner.Scan() {
+				// TODO - consider discrete markdown just for ROCM troubleshooting?
-			line := strings.TrimSpace(scanner.Text())
+				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
-			// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
+				skip[i] = struct{}{}
-			if strings.HasPrefix(line, "gfx_target_version") {
+			} else {
-				ver := strings.Fields(line)
+				slog.Info(fmt.Sprintf("amdgpu [%d] %s is supported", i, v.ToGFXString()))
-
-				// Detect CPUs
-				if len(ver) == 2 && ver[1] == "0" {
-					slog.Debug("detected CPU " + match)
-					isCPU = true
-					break
-				}
-
-				if len(ver) != 2 || len(ver[1]) < 5 {
-					slog.Warn("malformed "+match, "gfx_target_version", line)
-					// If this winds up being a CPU, our offsets may be wrong
-					continue
-				}
-				l := len(ver[1])
-				var err1, err2, err3 error
-				patch, err1 = strconv.ParseUint(ver[1][l-2:l], 10, 32)
-				minor, err2 = strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
-				major, err3 = strconv.ParseUint(ver[1][:l-4], 10, 32)
-				if err1 != nil || err2 != nil || err3 != nil {
-					slog.Debug("malformed int " + line)
-					continue
-				}
 			}
-
-			// TODO - any other properties we want to extract and record?
-			// vendor_id + device_id -> pci lookup for "Name"
-			// Other metrics that may help us understand relative performance between multiple GPUs
 		}
+	} else {
+		slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
+	}
 
-		if isCPU {
+	if len(skip) >= len(gfx) {
-			cpuCount++
+		slog.Info("all detected amdgpus are skipped, falling back to CPU")
+		return
+	}
+
+	ids := make([]int, len(gfx))
+	i := 0
+	for k := range gfx {
+		ids[i] = k
+		i++
+	}
+	amdProcMemLookup(resp, skip, ids)
+	if resp.memInfo.DeviceCount == 0 {
+		return
+	}
+	if len(skip) > 0 {
+		amdSetVisibleDevices(ids, skip)
+	}
+}
+
+func updateLibPath(libDir string) {
+	ldPaths := []string{}
+	if val, ok := os.LookupEnv("LD_LIBRARY_PATH"); ok {
+		ldPaths = strings.Split(val, ":")
+	}
+	for _, d := range ldPaths {
+		if d == libDir {
+			return
+		}
+	}
+	val := strings.Join(append(ldPaths, libDir), ":")
+	slog.Debug("updated lib path", "LD_LIBRARY_PATH", val)
+	os.Setenv("LD_LIBRARY_PATH", val)
+}
+
+// Walk the sysfs nodes for the available GPUs and gather information from them
+// skipping over any devices in the skip map
+func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
+	resp.memInfo.DeviceCount = 0
+	resp.memInfo.TotalMemory = 0
+	resp.memInfo.FreeMemory = 0
+	slog.Debug("discovering VRAM for amdgpu devices")
+	if len(ids) == 0 {
+		entries, err := os.ReadDir(AMDNodesSysfsDir)
+		if err != nil {
+			slog.Warn(fmt.Sprintf("failed to read amdgpu sysfs %s - %s", AMDNodesSysfsDir, err))
+			return
+		}
+		for _, node := range entries {
+			if !node.IsDir() {
+				continue
+			}
+			id, err := strconv.Atoi(node.Name())
+			if err != nil {
+				slog.Warn("malformed amdgpu sysfs node id " + node.Name())
+				continue
+			}
+			ids = append(ids, id)
+		}
+	}
+	slog.Debug(fmt.Sprintf("amdgpu devices %v", ids))
+
+	for _, id := range ids {
+		if _, skipped := skip[id]; skipped {
 			continue
 		}
-
-		// CPUs are always first in the list
-		gpuID := nodeID - cpuCount
-
-		// Shouldn't happen, but just in case...
-		if gpuID < 0 {
-			slog.Error("unexpected amdgpu sysfs data resulted in negative GPU ID, please set OLLAMA_DEBUG=1 and report an issue")
-			return []GpuInfo{}
-		}
-
-		if int(major) < RocmComputeMin {
-			slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%d%x", major, minor, patch), "gpu", gpuID)
-			continue
-		}
-
-		// Look up the memory for the current node
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
-		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUTotalMemoryFileGlob)
+		// Adjust for sysfs vs HIP ids
+		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id+1), GPUTotalMemoryFileGlob)
 		propFiles, err := filepath.Glob(propGlob)
 		if err != nil {
-			slog.Warn("error looking up total GPU memory", "glob", propGlob, "error", err)
+			slog.Warn(fmt.Sprintf("error looking up total GPU memory: %s %s", propGlob, err))
 		}
 		// 1 or more memory banks - sum the values of all of them
 		for _, propFile := range propFiles {
 			fp, err := os.Open(propFile)
 			if err != nil {
-				slog.Warn("failed to open sysfs node", "file", propFile, "erroir", err)
+				slog.Warn(fmt.Sprintf("failed to open sysfs node file %s: %s", propFile, err))
 				continue
 			}
 			defer fp.Close()
@@ -179,113 +226,49 @@ func AMDGetGPUInfo() []GpuInfo {
 			}
 		}
 		if totalMemory == 0 {
-			slog.Warn("amdgpu reports zero total memory", "gpu", gpuID)
+			slog.Warn(fmt.Sprintf("amdgpu [%d] reports zero total memory, skipping", id))
+			skip[id] = struct{}{}
 			continue
 		}
-		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUUsedMemoryFileGlob)
+		if totalMemory < IGPUMemLimit {
+			slog.Info(fmt.Sprintf("amdgpu [%d] appears to be an iGPU with %dM reported total memory, skipping", id, totalMemory/1024/1024))
+			skip[id] = struct{}{}
+			continue
+		}
+		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id), GPUUsedMemoryFileGlob)
 		usedFiles, err := filepath.Glob(usedGlob)
 		if err != nil {
-			slog.Warn("error looking up used GPU memory", "glob", usedGlob, "error", err)
+			slog.Warn(fmt.Sprintf("error looking up used GPU memory: %s %s", usedGlob, err))
 			continue
 		}
 		for _, usedFile := range usedFiles {
 			fp, err := os.Open(usedFile)
 			if err != nil {
-				slog.Warn("failed to open sysfs node", "file", usedFile, "error", err)
+				slog.Warn(fmt.Sprintf("failed to open sysfs node file %s: %s", usedFile, err))
 				continue
 			}
 			defer fp.Close()
 			data, err := io.ReadAll(fp)
 			if err != nil {
-				slog.Warn("failed to read sysfs node", "file", usedFile, "error", err)
+				slog.Warn(fmt.Sprintf("failed to read sysfs node file %s: %s", usedFile, err))
 				continue
 			}
 			used, err := strconv.ParseUint(strings.TrimSpace(string(data)), 10, 64)
 			if err != nil {
-				slog.Warn("malformed used memory", "data", string(data), "error", err)
+				slog.Warn(fmt.Sprintf("malformed used memory %s: %s", string(data), err))
 				continue
 			}
 			usedMemory += used
 		}
-
+		slog.Info(fmt.Sprintf("[%d] amdgpu totalMemory %dM", id, totalMemory/1024/1024))
-		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
+		slog.Info(fmt.Sprintf("[%d] amdgpu freeMemory  %dM", id, (totalMemory-usedMemory)/1024/1024))
-		if totalMemory < IGPUMemLimit {
+		resp.memInfo.DeviceCount++
-			slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
+		resp.memInfo.TotalMemory += totalMemory
-			continue
+		resp.memInfo.FreeMemory += (totalMemory - usedMemory)
-		}
-
-		slog.Info("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
-		slog.Info("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
-		gpuInfo := GpuInfo{
-			Library: "rocm",
-			memInfo: memInfo{
-				TotalMemory: totalMemory,
-				FreeMemory:  (totalMemory - usedMemory),
-			},
-			ID: fmt.Sprintf("%d", gpuID),
-			// Name: not exposed in sysfs directly, would require pci device id lookup
-			Major:         int(major),
-			Minor:         int(minor),
-			Patch:         int(patch),
-			MinimumMemory: rocmMinimumMemory,
-		}
-
-		// If the user wants to filter to a subset of devices, filter out if we aren't a match
-		if len(visibleDevices) > 0 {
-			include := false
-			for _, visible := range visibleDevices {
-				if visible == gpuInfo.ID {
-					include = true
-					break
-				}
-			}
-			if !include {
-				slog.Info("filtering out device per user request", "id", gpuInfo.ID, "visible_devices", visibleDevices)
-				continue
-			}
-		}
-
-		// Final validation is gfx compatibility - load the library if we haven't already loaded it
-		// even if the user overrides, we still need to validate the library
-		if libDir == "" {
-			libDir, err = AMDValidateLibDir()
-			if err != nil {
-				slog.Warn("unable to verify rocm library, will use cpu", "error", err)
-				return []GpuInfo{}
-			}
-		}
-		gpuInfo.DependencyPath = libDir
-
-		if gfxOverride == "" {
-			// Only load supported list once
-			if len(supported) == 0 {
-				supported, err = GetSupportedGFX(libDir)
-				if err != nil {
-					slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
-					return []GpuInfo{}
-				}
-				slog.Debug("rocm supported GPUs", "types", supported)
-			}
-			gfx := fmt.Sprintf("gfx%d%d%x", gpuInfo.Major, gpuInfo.Minor, gpuInfo.Patch)
-			if !slices.Contains[[]string, string](supported, gfx) {
-				slog.Warn("amdgpu is not supported", "gpu", gpuInfo.ID, "gpu_type", gfx, "library", libDir, "supported_types", supported)
-				// TODO - consider discrete markdown just for ROCM troubleshooting?
-				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
-				continue
-			} else {
-				slog.Info("amdgpu is supported", "gpu", gpuInfo.ID, "gpu_type", gfx)
-			}
-		} else {
-			slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
-		}
-
-		// The GPU has passed all the verification steps and is supported
-		resp = append(resp, gpuInfo)
 	}
-	if len(resp) == 0 {
+	if resp.memInfo.DeviceCount > 0 {
-		slog.Info("no compatible amdgpu devices detected")
+		resp.Library = "rocm"
 	}
-	return resp
 }
 
 // Quick check for AMD driver so we can skip amdgpu discovery if not present
@@ -297,24 +280,87 @@ func AMDDetected() bool {
 		slog.Debug("amdgpu driver not detected " + sysfsDir)
 		return false
 	} else if err != nil {
-		slog.Debug("error looking up amd driver", "path", sysfsDir, "error", err)
+		slog.Debug(fmt.Sprintf("error looking up amd driver %s %s", sysfsDir, err))
 		return false
 	}
 	return true
 }
 
+func setupLink(source, target string) error {
+	if err := os.RemoveAll(target); err != nil {
+		return fmt.Errorf("failed to remove old rocm directory %s %w", target, err)
+	}
+	if err := os.Symlink(source, target); err != nil {
+		return fmt.Errorf("failed to create link %s => %s %w", source, target, err)
+	}
+	slog.Debug(fmt.Sprintf("host rocm linked %s => %s", source, target))
+	return nil
+}
+
+// Ensure the AMD rocm lib dir is wired up
 // Prefer to use host installed ROCm, as long as it meets our minimum requirements
 // failing that, tell the user how to download it on their own
 func AMDValidateLibDir() (string, error) {
-	libDir, err := commonAMDValidateLibDir()
+	// We rely on the rpath compiled into our library to find rocm
+	// so we establish a symlink to wherever we find it on the system
+	// to <payloads>/rocm
+	payloadsDir, err := PayloadsDir()
+	if err != nil {
+		return "", err
+	}
+
+	// If we already have a rocm dependency wired, nothing more to do
+	rocmTargetDir := filepath.Clean(filepath.Join(payloadsDir, "..", "rocm"))
+	if rocmLibUsable(rocmTargetDir) {
+		return rocmTargetDir, nil
+	}
+
+	// next to the running binary
+	exe, err := os.Executable()
 	if err == nil {
-		return libDir, nil
+		peerDir := filepath.Dir(exe)
+		if rocmLibUsable(peerDir) {
+			slog.Debug("detected ROCM next to ollama executable " + peerDir)
+			return rocmTargetDir, setupLink(peerDir, rocmTargetDir)
+		}
+		peerDir = filepath.Join(filepath.Dir(exe), "rocm")
+		if rocmLibUsable(peerDir) {
+			slog.Debug("detected ROCM next to ollama executable " + peerDir)
+			return rocmTargetDir, setupLink(peerDir, rocmTargetDir)
+		}
 	}
 
 	// Well known ollama installer path
 	installedRocmDir := "/usr/share/ollama/lib/rocm"
 	if rocmLibUsable(installedRocmDir) {
-		return installedRocmDir, nil
+		return rocmTargetDir, setupLink(installedRocmDir, rocmTargetDir)
+	}
+
+	// Prefer explicit HIP env var
+	hipPath := os.Getenv("HIP_PATH")
+	if hipPath != "" {
+		hipLibDir := filepath.Join(hipPath, "lib")
+		if rocmLibUsable(hipLibDir) {
+			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
+			return rocmTargetDir, setupLink(hipLibDir, rocmTargetDir)
+		}
+	}
+
+	// Scan the library path for potential matches
+	ldPaths := strings.Split(os.Getenv("LD_LIBRARY_PATH"), ":")
+	for _, ldPath := range ldPaths {
+		d, err := filepath.Abs(ldPath)
+		if err != nil {
+			continue
+		}
+		if rocmLibUsable(d) {
+			return rocmTargetDir, setupLink(d, rocmTargetDir)
+		}
+	}
+
+	// Well known location(s)
+	if rocmLibUsable("/opt/rocm/lib") {
+		return rocmTargetDir, setupLink("/opt/rocm/lib", rocmTargetDir)
 	}
 
 	// If we still haven't found a usable rocm, the user will have to install it on their own
@@ -338,3 +384,68 @@ func AMDDriverVersion() (string, error) {
 	}
 	return strings.TrimSpace(string(verString)), nil
 }
+
+func AMDGFXVersions() map[int]Version {
+	// The amdgpu driver always exposes the host CPU as node 0, but we have to skip that and subtract one
+	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
+	res := map[int]Version{}
+	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
+	for _, match := range matches {
+		fp, err := os.Open(match)
+		if err != nil {
+			slog.Debug(fmt.Sprintf("failed to open sysfs node file %s: %s", match, err))
+			continue
+		}
+		defer fp.Close()
+		i, err := strconv.Atoi(filepath.Base(filepath.Dir(match)))
+		if err != nil {
+			slog.Debug(fmt.Sprintf("failed to parse node ID %s", err))
+			continue
+		}
+
+		if i == 0 {
+			// Skipping the CPU
+			continue
+		}
+		// Align with HIP IDs (zero is first GPU, not CPU)
+		i -= 1
+
+		scanner := bufio.NewScanner(fp)
+		for scanner.Scan() {
+			line := strings.TrimSpace(scanner.Text())
+			if strings.HasPrefix(line, "gfx_target_version") {
+				ver := strings.Fields(line)
+				if len(ver) != 2 || len(ver[1]) < 5 {
+					if ver[1] != "0" {
+						slog.Debug("malformed " + line)
+					}
+					res[i] = Version{
+						Major: 0,
+						Minor: 0,
+						Patch: 0,
+					}
+					continue
+				}
+				l := len(ver[1])
+				patch, err1 := strconv.ParseUint(ver[1][l-2:l], 10, 32)
+				minor, err2 := strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
+				major, err3 := strconv.ParseUint(ver[1][:l-4], 10, 32)
+				if err1 != nil || err2 != nil || err3 != nil {
+					slog.Debug("malformed int " + line)
+					continue
+				}
+
+				res[i] = Version{
+					Major: uint(major),
+					Minor: uint(minor),
+					Patch: uint(patch),
+				}
+			}
+		}
+	}
+	return res
+}
+
+func (v Version) ToGFXString() string {
+	return fmt.Sprintf("gfx%d%d%d", v.Major, v.Minor, v.Patch)
+}

gpu/amd_windows.go

@@ -7,10 +7,7 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
-	"strconv"
 	"strings"
-
-	"github.com/ollama/ollama/format"
 )
 
 const (
@@ -25,32 +22,36 @@ var (
 	ROCmLibGlobs = []string{"hipblas.dll", "rocblas"} // TODO - probably include more coverage of files here...
 )
 
-func AMDGetGPUInfo() []GpuInfo {
+func AMDGetGPUInfo(resp *GpuInfo) {
-	resp := []GpuInfo{}
 	hl, err := NewHipLib()
 	if err != nil {
 		slog.Debug(err.Error())
-		return nil
+		return
 	}
 	defer hl.Release()
+	skip := map[int]interface{}{}
+	ids := []int{}
+	resp.memInfo.DeviceCount = 0
+	resp.memInfo.TotalMemory = 0
+	resp.memInfo.FreeMemory = 0
 
 	ver, err := hl.AMDDriverVersion()
 	if err == nil {
 		slog.Info("AMD Driver: " + ver)
 	} else {
 		// For now this is benign, but we may eventually need to fail compatibility checks
-		slog.Debug("error looking up amd driver version", "error", err)
+		slog.Debug(fmt.Sprintf("error looking up amd driver version: %s", err))
 	}
 
-	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
+	// Note: the HIP library automatically handles HIP_VISIBLE_DEVICES
 	count := hl.HipGetDeviceCount()
 	if count == 0 {
-		return nil
+		return
 	}
 	libDir, err := AMDValidateLibDir()
 	if err != nil {
-		slog.Warn("unable to verify rocm library, will use cpu", "error", err)
+		slog.Warn(fmt.Sprintf("unable to verify rocm library, will use cpu: %s", err))
-		return nil
+		return
 	}
 
 	var supported []string
@@ -58,120 +59,95 @@ func AMDGetGPUInfo() []GpuInfo {
 	if gfxOverride == "" {
 		supported, err = GetSupportedGFX(libDir)
 		if err != nil {
-			slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
+			slog.Warn(fmt.Sprintf("failed to lookup supported GFX types, falling back to CPU mode: %s", err))
-			return nil
+			return
 		}
 	} else {
 		slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
 	}
 
-	slog.Info("detected hip devices", "count", count)
+	slog.Info(fmt.Sprintf("detected %d hip devices", count))
-	// TODO how to determine the underlying device ID when visible devices is causing this to subset?
 	for i := 0; i < count; i++ {
+		ids = append(ids, i)
 		err = hl.HipSetDevice(i)
 		if err != nil {
-			slog.Warn("set device", "id", i, "error", err)
+			slog.Warn(fmt.Sprintf("[%d] %s", i, err))
+			skip[i] = struct{}{}
 			continue
 		}
 
 		props, err := hl.HipGetDeviceProperties(i)
 		if err != nil {
-			slog.Warn("get properties", "id", i, "error", err)
+			slog.Warn(fmt.Sprintf("[%d] %s", i, err))
+			skip[i] = struct{}{}
 			continue
 		}
 		n := bytes.IndexByte(props.Name[:], 0)
 		name := string(props.Name[:n])
-		// TODO is UUID actually populated on windows?
+		slog.Info(fmt.Sprintf("[%d] Name: %s", i, name))
-		// Can luid be used on windows for setting visible devices (and is it actually set?)
 		n = bytes.IndexByte(props.GcnArchName[:], 0)
 		gfx := string(props.GcnArchName[:n])
-		slog.Info("hip device", "id", i, "name", name, "gfx", gfx)
+		slog.Info(fmt.Sprintf("[%d] GcnArchName: %s", i, gfx))
-		var major, minor, patch string
-		switch len(gfx) {
-		case 6:
-			major, minor, patch = gfx[3:4], gfx[4:5], gfx[5:]
-		case 7:
-			major, minor, patch = gfx[3:5], gfx[5:6], gfx[6:]
-		}
 		//slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
 		if strings.EqualFold(name, iGPUName) {
-			slog.Info("iGPU detected skipping", "id", i)
+			slog.Info(fmt.Sprintf("iGPU detected [%d] skipping", i))
+			skip[i] = struct{}{}
 			continue
 		}
 		if gfxOverride == "" {
 			if !slices.Contains[[]string, string](supported, gfx) {
-				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
+				slog.Warn(fmt.Sprintf("amdgpu [%d] %s is not supported by %s %v", i, gfx, libDir, supported))
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
+				skip[i] = struct{}{}
 				continue
 			} else {
-				slog.Info("amdgpu is supported", "gpu", i, "gpu_type", gfx)
+				slog.Info(fmt.Sprintf("amdgpu [%d] %s is supported", i, gfx))
 			}
 		}
 
-		freeMemory, totalMemory, err := hl.HipMemGetInfo()
+		totalMemory, freeMemory, err := hl.HipMemGetInfo()
 		if err != nil {
-			slog.Warn("get mem info", "id", i, "error", err)
+			slog.Warn(fmt.Sprintf("[%d] %s", i, err))
 			continue
 		}
 
-		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
+		// TODO according to docs, freeMem may lie on windows!
-		if totalMemory < IGPUMemLimit {
+		slog.Info(fmt.Sprintf("[%d] Total Mem: %d", i, totalMemory))
-			slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", i, "total", format.HumanBytes2(totalMemory))
+		slog.Info(fmt.Sprintf("[%d] Free Mem:  %d", i, freeMemory))
-			continue
+		resp.memInfo.DeviceCount++
-		}
+		resp.memInfo.TotalMemory += totalMemory
-
+		resp.memInfo.FreeMemory += freeMemory
-		// TODO revisit this once ROCm v6 is available on windows.
-		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
-		slog.Info("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
-		slog.Info("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
-		gpuInfo := GpuInfo{
-			Library: "rocm",
-			memInfo: memInfo{
-				TotalMemory: totalMemory,
-				FreeMemory:  freeMemory,
-			},
-			ID:             fmt.Sprintf("%d", i), // TODO this is probably wrong if we specify visible devices
-			DependencyPath: libDir,
-			MinimumMemory:  rocmMinimumMemory,
-		}
-		if major != "" {
-			gpuInfo.Major, err = strconv.Atoi(major)
-			if err != nil {
-				slog.Info("failed to parse version", "version", gfx, "error", err)
-			}
-		}
-		if minor != "" {
-			gpuInfo.Minor, err = strconv.Atoi(minor)
-			if err != nil {
-				slog.Info("failed to parse version", "version", gfx, "error", err)
-			}
-		}
-		if patch != "" {
-			// Patch rev is hex; e.g. gfx90a
-			p, err := strconv.ParseInt(patch, 16, 0)
-			if err != nil {
-				slog.Info("failed to parse version", "version", gfx, "error", err)
-			} else {
-				gpuInfo.Patch = int(p)
-			}
-		}
-		if gpuInfo.Major < RocmComputeMin {
-			slog.Warn(fmt.Sprintf("amdgpu [%s] too old gfx%d%d%x", gpuInfo.ID, gpuInfo.Major, gpuInfo.Minor, gpuInfo.Patch))
-			continue
-		}
-
-		resp = append(resp, gpuInfo)
 	}
-
+	if resp.memInfo.DeviceCount > 0 {
-	return resp
+		resp.Library = "rocm"
+	}
+	// Abort if all GPUs are skipped
+	if len(skip) >= count {
+		slog.Info("all detected amdgpus are skipped, falling back to CPU")
+		return
+	}
+	if len(skip) > 0 {
+		amdSetVisibleDevices(ids, skip)
+	}
+	UpdatePath(libDir)
 }
 
 func AMDValidateLibDir() (string, error) {
-	libDir, err := commonAMDValidateLibDir()
+	// On windows non-admins typically can't create links
+	// so instead of trying to rely on rpath and a link in
+	// $LibDir/rocm, we instead rely on setting PATH to point
+	// to the location of the ROCm library
+
+	// Installer payload location if we're running the installed binary
+	exe, err := os.Executable()
 	if err == nil {
-		return libDir, nil
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
 	}
 
 	// Installer payload (if we're running from some other location)
@@ -183,6 +159,21 @@ func AMDValidateLibDir() (string, error) {
 		return rocmTargetDir, nil
 	}
 
+	// Prefer explicit HIP env var
+	hipPath := os.Getenv("HIP_PATH")
+	if hipPath != "" {
+		hipLibDir := filepath.Join(hipPath, "bin")
+		if rocmLibUsable(hipLibDir) {
+			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
+			return hipLibDir, nil
+		}
+	}
+
+	// Well known location(s)
+	if rocmLibUsable(RocmStandardLocation) {
+		return RocmStandardLocation, nil
+	}
+
 	// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
 	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")

gpu/assets.go

@@ -24,51 +24,6 @@ func PayloadsDir() (string, error) {
 	defer lock.Unlock()
 	var err error
 	if payloadsDir == "" {
-		runnersDir := os.Getenv("OLLAMA_RUNNERS_DIR")
-		// On Windows we do not carry the payloads inside the main executable
-		if runtime.GOOS == "windows" && runnersDir == "" {
-			appExe, err := os.Executable()
-			if err != nil {
-				slog.Error("failed to lookup executable path", "error", err)
-				return "", err
-			}
-
-			cwd, err := os.Getwd()
-			if err != nil {
-				slog.Error("failed to lookup working directory", "error", err)
-				return "", err
-			}
-
-			var paths []string
-			for _, root := range []string{appExe, cwd} {
-				paths = append(paths,
-					filepath.Join(root),
-					filepath.Join(root, "windows-"+runtime.GOARCH),
-					filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
-				)
-			}
-
-			// Try a few variations to improve developer experience when building from source in the local tree
-			for _, p := range paths {
-				candidate := filepath.Join(p, "ollama_runners")
-				_, err := os.Stat(candidate)
-				if err == nil {
-					runnersDir = candidate
-					break
-				}
-			}
-			if runnersDir == "" {
-				err = fmt.Errorf("unable to locate llm runner directory.  Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
-				slog.Error("incomplete distribution", "error", err)
-				return "", err
-			}
-		}
-		if runnersDir != "" {
-			payloadsDir = runnersDir
-			return payloadsDir, nil
-		}
-
-		// The remainder only applies on non-windows where we still carry payloads in the main executable
 		cleanupTmpDirs()
 		tmpDir := os.Getenv("OLLAMA_TMPDIR")
 		if tmpDir == "" {
@@ -125,7 +80,7 @@ func cleanupTmpDirs() {
 		}
 		err = os.RemoveAll(d)
 		if err != nil {
-			slog.Debug("unable to cleanup stale tmpdir", "path", d, "error", err)
+			slog.Debug(fmt.Sprintf("unable to cleanup stale tmpdir %s: %s", d, err))
 		}
 	}
 }
@@ -133,8 +88,7 @@ func cleanupTmpDirs() {
 func Cleanup() {
 	lock.Lock()
 	defer lock.Unlock()
-	runnersDir := os.Getenv("OLLAMA_RUNNERS_DIR")
+	if payloadsDir != "" {
-	if payloadsDir != "" && runnersDir == "" && runtime.GOOS != "windows" {
 		// We want to fully clean up the tmpdir parent of the payloads dir
 		tmpDir := filepath.Clean(filepath.Join(payloadsDir, ".."))
 		slog.Debug("cleaning up", "dir", tmpDir)
@@ -166,7 +120,7 @@ func UpdatePath(dir string) {
 			}
 		}
 		newPath := strings.Join(append([]string{dir}, pathComponents...), ";")
-		slog.Info("updating", "PATH", newPath)
+		slog.Info(fmt.Sprintf("Updating PATH to %s", newPath))
 		os.Setenv("PATH", newPath)
 	}
 	// linux and darwin rely on rpath

gpu/cuda_common.go

@@ -1,22 +0,0 @@
-//go:build linux || windows
-
-package gpu
-
-import (
-	"log/slog"
-	"strings"
-)
-
-func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
-	ids := []string{}
-	for _, info := range gpuInfo {
-		if info.Library != "cuda" {
-			// TODO shouldn't happen if things are wired correctly...
-			slog.Debug("cudaGetVisibleDevicesEnv skipping over non-cuda device", "library", info.Library)
-			continue
-		}
-		ids = append(ids, info.ID)
-	}
-	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
-
-}

gpu/gpu.go

@@ -16,6 +16,7 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"sync"
 	"unsafe"
@@ -24,8 +25,8 @@ import (
 )
 
 type handles struct {
-	deviceCount int
+	nvml   *C.nvml_handle_t
-	cudart      *C.cudart_handle_t
+	cudart *C.cudart_handle_t
 }
 
 const (
@@ -38,10 +39,26 @@ var gpuMutex sync.Mutex
 // With our current CUDA compile flags, older than 5.0 will not work properly
 var CudaComputeMin = [2]C.int{5, 0}
 
-var RocmComputeMin = 9
+// Possible locations for the nvidia-ml library
+var NvmlLinuxGlobs = []string{
+	"/usr/local/cuda/lib64/libnvidia-ml.so*",
+	"/usr/lib/x86_64-linux-gnu/nvidia/current/libnvidia-ml.so*",
+	"/usr/lib/x86_64-linux-gnu/libnvidia-ml.so*",
+	"/usr/lib/wsl/lib/libnvidia-ml.so*",
+	"/usr/lib/wsl/drivers/*/libnvidia-ml.so*",
+	"/opt/cuda/lib64/libnvidia-ml.so*",
+	"/usr/lib*/libnvidia-ml.so*",
+	"/usr/lib/aarch64-linux-gnu/nvidia/current/libnvidia-ml.so*",
+	"/usr/lib/aarch64-linux-gnu/libnvidia-ml.so*",
+	"/usr/local/lib*/libnvidia-ml.so*",
 
-// TODO find a better way to detect iGPU instead of minimum memory
+	// TODO: are these stubs ever valid?
-const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
+	"/opt/cuda/targets/x86_64-linux/lib/stubs/libnvidia-ml.so*",
+}
+
+var NvmlWindowsGlobs = []string{
+	"c:\\Windows\\System32\\nvml.dll",
+}
 
 var CudartLinuxGlobs = []string{
 	"/usr/local/cuda/lib64/libcudart.so*",
@@ -71,18 +88,26 @@ func initGPUHandles() *handles {
 
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 
-	gpuHandles := &handles{}
+	gpuHandles := &handles{nil, nil}
+	var nvmlMgmtName string
+	var nvmlMgmtPatterns []string
 	var cudartMgmtName string
 	var cudartMgmtPatterns []string
 
 	tmpDir, _ := PayloadsDir()
 	switch runtime.GOOS {
 	case "windows":
+		nvmlMgmtName = "nvml.dll"
+		nvmlMgmtPatterns = make([]string, len(NvmlWindowsGlobs))
+		copy(nvmlMgmtPatterns, NvmlWindowsGlobs)
 		cudartMgmtName = "cudart64_*.dll"
 		localAppData := os.Getenv("LOCALAPPDATA")
 		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", cudartMgmtName)}
 		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartWindowsGlobs...)
 	case "linux":
+		nvmlMgmtName = "libnvidia-ml.so"
+		nvmlMgmtPatterns = make([]string, len(NvmlLinuxGlobs))
+		copy(nvmlMgmtPatterns, NvmlLinuxGlobs)
 		cudartMgmtName = "libcudart.so*"
 		if tmpDir != "" {
 			// TODO - add "payloads" for subprocess
@@ -93,21 +118,31 @@ func initGPUHandles() *handles {
 		return gpuHandles
 	}
 
-	slog.Info("Detecting GPUs")
+	slog.Info("Detecting GPU type")
 	cudartLibPaths := FindGPULibs(cudartMgmtName, cudartMgmtPatterns)
 	if len(cudartLibPaths) > 0 {
-		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
+		cudart := LoadCUDARTMgmt(cudartLibPaths)
 		if cudart != nil {
-			slog.Info("detected GPUs", "library", libPath, "count", deviceCount)
+			slog.Info("Nvidia GPU detected via cudart")
 			gpuHandles.cudart = cudart
-			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
+		}
+	}
+
+	// TODO once we build confidence, remove this and the gpu_info_nvml.[ch] files
+	nvmlLibPaths := FindGPULibs(nvmlMgmtName, nvmlMgmtPatterns)
+	if len(nvmlLibPaths) > 0 {
+		nvml := LoadNVMLMgmt(nvmlLibPaths)
+		if nvml != nil {
+			slog.Info("Nvidia GPU detected via nvidia-ml")
+			gpuHandles.nvml = nvml
 			return gpuHandles
 		}
 	}
 	return gpuHandles
 }
 
-func GetGPUInfo() GpuInfoList {
+func GetGPUInfo() GpuInfo {
 	// TODO - consider exploring lspci (and equivalent on windows) to check for
 	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
 	gpuMutex.Lock()
@@ -115,6 +150,9 @@ func GetGPUInfo() GpuInfoList {
 
 	gpuHandles := initGPUHandles()
 	defer func() {
+		if gpuHandles.nvml != nil {
+			C.nvml_release(*gpuHandles.nvml)
+		}
 		if gpuHandles.cudart != nil {
 			C.cudart_release(*gpuHandles.cudart)
 		}
@@ -127,63 +165,72 @@ func GetGPUInfo() GpuInfoList {
 	}
 
 	var memInfo C.mem_info_t
-	resp := []GpuInfo{}
+	resp := GpuInfo{}
-
+	if gpuHandles.nvml != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
-	// NVIDIA first
+		C.nvml_check_vram(*gpuHandles.nvml, &memInfo)
-	for i := 0; i < gpuHandles.deviceCount; i++ {
-		// TODO once we support CPU compilation variants of GPU libraries refine this...
-		if cpuVariant == "" && runtime.GOARCH == "amd64" {
-			continue
-		}
-		gpuInfo := GpuInfo{
-			Library: "cuda",
-		}
-		C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
 		if memInfo.err != nil {
-			slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+			slog.Info(fmt.Sprintf("[nvidia-ml] error looking up NVML GPU memory: %s", C.GoString(memInfo.err)))
 			C.free(unsafe.Pointer(memInfo.err))
-			continue
+		} else if memInfo.count > 0 {
+			// Verify minimum compute capability
+			var cc C.nvml_compute_capability_t
+			C.nvml_compute_capability(*gpuHandles.nvml, &cc)
+			if cc.err != nil {
+				slog.Info(fmt.Sprintf("[nvidia-ml] error looking up NVML GPU compute capability: %s", C.GoString(cc.err)))
+				C.free(unsafe.Pointer(cc.err))
+			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
+				slog.Info(fmt.Sprintf("[nvidia-ml] NVML CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
+				resp.Library = "cuda"
+				resp.MinimumMemory = cudaMinimumMemory
+			} else {
+				slog.Info(fmt.Sprintf("[nvidia-ml] CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
+			}
 		}
-		if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+	} else if gpuHandles.cudart != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
-			slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
+		C.cudart_check_vram(*gpuHandles.cudart, &memInfo)
-			continue
-		}
-		gpuInfo.TotalMemory = uint64(memInfo.total)
-		gpuInfo.FreeMemory = uint64(memInfo.free)
-		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-		gpuInfo.Major = int(memInfo.major)
-		gpuInfo.Minor = int(memInfo.minor)
-		gpuInfo.MinimumMemory = cudaMinimumMemory
-
-		// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
-		resp = append(resp, gpuInfo)
-	}
-
-	// Then AMD
-	resp = append(resp, AMDGetGPUInfo()...)
-
-	if len(resp) == 0 {
-		C.cpu_check_ram(&memInfo)
 		if memInfo.err != nil {
-			slog.Info("error looking up CPU memory", "error", C.GoString(memInfo.err))
+			slog.Info(fmt.Sprintf("[cudart] error looking up CUDART GPU memory: %s", C.GoString(memInfo.err)))
 			C.free(unsafe.Pointer(memInfo.err))
+		} else if memInfo.count > 0 {
+			// Verify minimum compute capability
+			var cc C.cudart_compute_capability_t
+			C.cudart_compute_capability(*gpuHandles.cudart, &cc)
+			if cc.err != nil {
+				slog.Info(fmt.Sprintf("[cudart] error looking up CUDA compute capability: %s", C.GoString(cc.err)))
+				C.free(unsafe.Pointer(cc.err))
+			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
+				slog.Info(fmt.Sprintf("[cudart] CUDART CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
+				resp.Library = "cuda"
+				resp.MinimumMemory = cudaMinimumMemory
+			} else {
+				slog.Info(fmt.Sprintf("[cudart] CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
+			}
+		}
+	} else {
+		AMDGetGPUInfo(&resp)
+		if resp.Library != "" {
+			resp.MinimumMemory = rocmMinimumMemory
 			return resp
 		}
-		gpuInfo := GpuInfo{
+	}
-			Library: "cpu",
+	if resp.Library == "" {
-			Variant: cpuVariant,
+		C.cpu_check_ram(&memInfo)
-		}
+		resp.Library = "cpu"
-		gpuInfo.TotalMemory = uint64(memInfo.total)
+		resp.Variant = cpuVariant
-		gpuInfo.FreeMemory = uint64(memInfo.free)
+	}
-		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+	if memInfo.err != nil {
-
+		slog.Info(fmt.Sprintf("error looking up CPU memory: %s", C.GoString(memInfo.err)))
-		resp = append(resp, gpuInfo)
+		C.free(unsafe.Pointer(memInfo.err))
+		return resp
 	}
 
+	resp.DeviceCount = uint32(memInfo.count)
+	resp.FreeMemory = uint64(memInfo.free)
+	resp.TotalMemory = uint64(memInfo.total)
 	return resp
 }
 
-func GetCPUMem() (memInfo, error) {
+func getCPUMem() (memInfo, error) {
 	var ret memInfo
 	var info C.mem_info_t
 	C.cpu_check_ram(&info)
@@ -196,11 +243,29 @@ func GetCPUMem() (memInfo, error) {
 	return ret, nil
 }
 
+func CheckVRAM() (uint64, error) {
+	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
+	if userLimit != "" {
+		avail, err := strconv.ParseInt(userLimit, 10, 64)
+		if err != nil {
+			return 0, fmt.Errorf("Invalid OLLAMA_MAX_VRAM setting %s: %s", userLimit, err)
+		}
+		slog.Info(fmt.Sprintf("user override OLLAMA_MAX_VRAM=%d", avail))
+		return uint64(avail), nil
+	}
+	gpuInfo := GetGPUInfo()
+	if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") {
+		return gpuInfo.FreeMemory, nil
+	}
+
+	return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation
+}
+
 func FindGPULibs(baseLibName string, patterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
 	gpuLibPaths := []string{}
-	slog.Debug("Searching for GPU library", "name", baseLibName)
+	slog.Info(fmt.Sprintf("Searching for GPU management library %s", baseLibName))
 
 	switch runtime.GOOS {
 	case "windows":
@@ -218,7 +283,7 @@ func FindGPULibs(baseLibName string, patterns []string) []string {
 		}
 		patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
 	}
-	slog.Debug("gpu library search", "globs", patterns)
+	slog.Debug(fmt.Sprintf("gpu management search paths: %v", patterns))
 	for _, pattern := range patterns {
 		// Ignore glob discovery errors
 		matches, _ := filepath.Glob(pattern)
@@ -246,11 +311,28 @@ func FindGPULibs(baseLibName string, patterns []string) []string {
 			}
 		}
 	}
-	slog.Debug("discovered GPU libraries", "paths", gpuLibPaths)
+	slog.Info(fmt.Sprintf("Discovered GPU libraries: %v", gpuLibPaths))
 	return gpuLibPaths
 }
 
-func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
+func LoadNVMLMgmt(nvmlLibPaths []string) *C.nvml_handle_t {
+	var resp C.nvml_init_resp_t
+	resp.ch.verbose = getVerboseState()
+	for _, libPath := range nvmlLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.nvml_init(lib, &resp)
+		if resp.err != nil {
+			slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return &resp.ch
+		}
+	}
+	return nil
+}
+
+func LoadCUDARTMgmt(cudartLibPaths []string) *C.cudart_handle_t {
 	var resp C.cudart_init_resp_t
 	resp.ch.verbose = getVerboseState()
 	for _, libPath := range cudartLibPaths {
@@ -258,13 +340,13 @@ func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
 		defer C.free(unsafe.Pointer(lib))
 		C.cudart_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load cudart", "library", libPath, "error", C.GoString(resp.err))
+			slog.Info(fmt.Sprintf("Unable to load cudart CUDA management library %s: %s", libPath, C.GoString(resp.err)))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			return int(resp.num_devices), &resp.ch, libPath
+			return &resp.ch
 		}
 	}
-	return 0, nil, ""
+	return nil
 }
 
 func getVerboseState() C.uint16_t {
@@ -273,22 +355,3 @@ func getVerboseState() C.uint16_t {
 	}
 	return C.uint16_t(0)
 }
-
-// Given the list of GPUs this instantiation is targeted for,
-// figure out the visible devices environment variable
-//
-// If different libraries are detected, the first one is what we use
-func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
-	if len(l) == 0 {
-		return "", ""
-	}
-	switch l[0].Library {
-	case "cuda":
-		return cudaGetVisibleDevicesEnv(l)
-	case "rocm":
-		return rocmGetVisibleDevicesEnv(l)
-	default:
-		slog.Debug("no filter required for library " + l[0].Library)
-		return "", ""
-	}
-}

gpu/gpu_darwin.go

@@ -1,3 +1,5 @@
+//go:build darwin
+
 package gpu
 
 /*
@@ -7,41 +9,52 @@ package gpu
 */
 import "C"
 import (
+	"fmt"
+	"log/slog"
+	"os"
 	"runtime"
+	"strconv"
 )
 
-func GetGPUInfo() GpuInfoList {
+// CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
-	mem, _ := GetCPUMem()
+func CheckVRAM() (uint64, error) {
-	if runtime.GOARCH == "amd64" {
+	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
-		return []GpuInfo{
+	if userLimit != "" {
-			{
+		avail, err := strconv.ParseInt(userLimit, 10, 64)
-				Library: "cpu",
+		if err != nil {
-				Variant: GetCPUVariant(),
+			return 0, fmt.Errorf("Invalid OLLAMA_MAX_VRAM setting %s: %s", userLimit, err)
-				memInfo: mem,
-			},
 		}
+		slog.Info(fmt.Sprintf("user override OLLAMA_MAX_VRAM=%d", avail))
+		return uint64(avail), nil
 	}
-	info := GpuInfo{
+
-		Library: "metal",
+	if runtime.GOARCH == "amd64" {
-		ID:      "0",
+		// gpu not supported, this may not be metal
+		return 0, nil
 	}
-	info.TotalMemory = uint64(C.getRecommendedMaxVRAM())
 
-	// TODO is there a way to gather actual allocated video memory? (currentAllocatedSize doesn't work)
+	return uint64(C.getRecommendedMaxVRAM()), nil
-	info.FreeMemory = info.TotalMemory
-
-	info.MinimumMemory = 0
-	return []GpuInfo{info}
 }
 
-func GetCPUMem() (memInfo, error) {
+func GetGPUInfo() GpuInfo {
+	mem, _ := getCPUMem()
+	if runtime.GOARCH == "amd64" {
+		return GpuInfo{
+			Library: "cpu",
+			Variant: GetCPUVariant(),
+			memInfo: mem,
+		}
+	}
+	return GpuInfo{
+		Library: "metal",
+		memInfo: mem,
+	}
+}
+
+func getCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
 		FreeMemory:  0,
+		DeviceCount: 0,
 	}, nil
 }
-
-func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
-	// No-op on darwin
-	return "", ""
-}

gpu/gpu_info.h

@@ -38,17 +38,12 @@
 extern "C" {
 #endif
 
-#define GPU_ID_LEN 64
-
 typedef struct mem_info {
-  char *err;  // If non-nill, caller responsible for freeing
-  char gpu_id[GPU_ID_LEN];
   uint64_t total;
   uint64_t free;
-
+  unsigned int count;
-  // Compute Capability
+  int igpu_index; // If >= 0, we detected an integrated GPU to ignore
-  int major; 
+  char *err;  // If non-nill, caller responsible for freeing
-  int minor;
 } mem_info_t;
 
 void cpu_check_ram(mem_info_t *resp);
@@ -57,6 +52,7 @@ void cpu_check_ram(mem_info_t *resp);
 }
 #endif
 
+#include "gpu_info_nvml.h"
 #include "gpu_info_cudart.h"
 
 #endif  // __GPU_INFO_H__

gpu/gpu_info_cpu.c

@@ -8,11 +8,9 @@ void cpu_check_ram(mem_info_t *resp) {
   MEMORYSTATUSEX info;
   info.dwLength = sizeof(info);
   if (GlobalMemoryStatusEx(&info) != 0) {
+    resp->count = 1;
     resp->total = info.ullTotalPhys;
     resp->free = info.ullAvailPhys;
-    resp->major = 0;
-    resp->minor = 0;
-    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
   } else {
     resp->err = LOAD_ERR();
   }
@@ -29,11 +27,9 @@ void cpu_check_ram(mem_info_t *resp) {
   if (sysinfo(&info) != 0) {
     resp->err = strdup(strerror(errno));
   } else {
+    resp->count = 1;
     resp->total = info.totalram * info.mem_unit;
     resp->free = info.freeram * info.mem_unit;
-    resp->major = 0;
-    resp->minor = 0;
-    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
   }
   return;
 }

gpu/gpu_info_cudart.c

@@ -6,7 +6,6 @@
 void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
   cudartReturn_t ret;
   resp->err = NULL;
-  resp->num_devices = 0;
   const int buflen = 256;
   char buf[buflen + 1];
   int i;
@@ -22,7 +21,6 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
       {"cudaGetDeviceCount", (void *)&resp->ch.cudaGetDeviceCount},
       {"cudaDeviceGetAttribute", (void *)&resp->ch.cudaDeviceGetAttribute},
       {"cudaDriverGetVersion", (void *)&resp->ch.cudaDriverGetVersion},
-      {"cudaGetDeviceProperties", (void *)&resp->ch.cudaGetDeviceProperties},
       {NULL, NULL},
   };
 
@@ -38,7 +36,13 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
     return;
   }
 
+  // TODO once we've squashed the remaining corner cases remove this log
+  LOG(resp->ch.verbose, "wiring cudart library functions in %s\n", cudart_lib_path);
+  
   for (i = 0; l[i].s != NULL; i++) {
+    // TODO once we've squashed the remaining corner cases remove this log
+    LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
+
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
     if (!l[i].p) {
       char *msg = LOAD_ERR();
@@ -59,7 +63,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
     if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
-      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
+      resp->err = strdup("your nvidia driver is too old or missing, please upgrade to run ollama");
       return;
     }
     snprintf(buf, buflen, "cudart init failure: %d", ret);
@@ -81,95 +85,110 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
     driverVersion.minor = (version - (driverVersion.major * 1000)) / 10;
     LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", driverVersion.major, driverVersion.minor);
   }
-
-  ret = (*resp->ch.cudaGetDeviceCount)(&resp->num_devices);
-  if (ret != CUDART_SUCCESS) {
-    LOG(resp->ch.verbose, "cudaGetDeviceCount err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "unable to get device count: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
 }
 
 
-void cudart_check_vram(cudart_handle_t h, int i, mem_info_t *resp) {
+void cudart_check_vram(cudart_handle_t h, mem_info_t *resp) {
   resp->err = NULL;
   cudartMemory_t memInfo = {0,0,0};
   cudartReturn_t ret;
   const int buflen = 256;
   char buf[buflen + 1];
+  int i;
 
   if (h.handle == NULL) {
     resp->err = strdup("cudart handle isn't initialized");
     return;
   }
 
-  ret = (*h.cudaSetDevice)(i);
+  // cudaGetDeviceCount takes int type, resp-> count is uint
+  int deviceCount;
+  ret = (*h.cudaGetDeviceCount)(&deviceCount);
   if (ret != CUDART_SUCCESS) {
-    snprintf(buf, buflen, "cudart device failed to initialize");
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
     return;
-  }
-
-  cudaDeviceProp_t props;
-  ret = (*h.cudaGetDeviceProperties)(&props, i);
-  if (ret != CUDART_SUCCESS) {
-    LOG(h.verbose, "[%d] device properties lookup failure: %d\n", i, ret);
-    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
-    resp->major = 0;
-    resp->minor = 0;
   } else {
-    int allNull = 1;
+    resp->count = (unsigned int)deviceCount;
-    for (int j = 0; j < 16; j++) {
-      if (props.uuid.bytes[j] != 0) {
-        allNull = 0;
-        break;
-      }
-    }
-    if (allNull != 0) {
-      snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
-    } else {
-      // GPU-d110a105-ac29-1d54-7b49-9c90440f215b
-      snprintf(&resp->gpu_id[0], GPU_ID_LEN,
-          "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
-          props.uuid.bytes[0],
-          props.uuid.bytes[1],
-          props.uuid.bytes[2],
-          props.uuid.bytes[3],
-          props.uuid.bytes[4],
-          props.uuid.bytes[5],
-          props.uuid.bytes[6],
-          props.uuid.bytes[7],
-          props.uuid.bytes[8],
-          props.uuid.bytes[9],
-          props.uuid.bytes[10],
-          props.uuid.bytes[11],
-          props.uuid.bytes[12],
-          props.uuid.bytes[13],
-          props.uuid.bytes[14],
-          props.uuid.bytes[15]
-        );
-    }
-    resp->major = props.major;
-    resp->minor = props.minor;
-
-    // TODO add other useful properties from props
   }
-  ret = (*h.cudaMemGetInfo)(&memInfo.free, &memInfo.total);
+
+  resp->total = 0;
+  resp->free = 0;
+  for (i = 0; i < resp-> count; i++) {  
+    ret = (*h.cudaSetDevice)(i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "cudart device failed to initialize");
+      resp->err = strdup(buf);
+      return;
+    }
+    ret = (*h.cudaMemGetInfo)(&memInfo.free, &memInfo.total);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "cudart device memory info lookup failure %d", ret);
+      resp->err = strdup(buf);
+      return;
+    }
+
+    LOG(h.verbose, "[%d] CUDA totalMem %lu\n", i, memInfo.total);
+    LOG(h.verbose, "[%d] CUDA freeMem %lu\n", i, memInfo.free);
+
+    resp->total += memInfo.total;
+    resp->free += memInfo.free;
+  }
+}
+
+void cudart_compute_capability(cudart_handle_t h, cudart_compute_capability_t *resp) {
+  resp->err = NULL;
+  resp->major = 0;
+  resp->minor = 0;
+  int major = 0;
+  int minor = 0;
+  cudartReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("cudart handle not initialized");
+    return;
+  }
+
+  int devices;
+  ret = (*h.cudaGetDeviceCount)(&devices);
   if (ret != CUDART_SUCCESS) {
-    snprintf(buf, buflen, "cudart device memory info lookup failure %d", ret);
+    snprintf(buf, buflen, "unable to get cudart device count: %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
-  resp->total = memInfo.total;
+  for (i = 0; i < devices; i++) {
-  resp->free = memInfo.free;
+    ret = (*h.cudaSetDevice)(i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "cudart device failed to initialize");
+      resp->err = strdup(buf);
+      return;
+    }
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
+    ret = (*h.cudaDeviceGetAttribute)(&major, cudartDevAttrComputeCapabilityMajor, i);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
+    if (ret != CUDART_SUCCESS) {
-  LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
+      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    ret = (*h.cudaDeviceGetAttribute)(&minor, cudartDevAttrComputeCapabilityMinor, i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+      
+    // Report the lowest major.minor we detect as that limits our compatibility
+    if (resp->major == 0 || resp->major > major ) {
+      resp->major = major;
+      resp->minor = minor;
+    } else if ( resp->major == major && resp->minor > minor ) {
+      resp->minor = minor;
+    }
+  }
 }
 
 void cudart_release(cudart_handle_t h) {

gpu/gpu_info_cudart.h

@@ -6,8 +6,7 @@
 // Just enough typedef's to dlopen/dlsym for memory information
 typedef enum cudartReturn_enum {
   CUDART_SUCCESS = 0,
-  CUDA_ERROR_INVALID_VALUE = 1,
+  CUDART_UNSUPPORTED = 1,
-  CUDA_ERROR_MEMORY_ALLOCATION = 2,
   CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
   // Other values omitted for now...
 } cudartReturn_t;
@@ -15,11 +14,6 @@ typedef enum cudartReturn_enum {
 typedef enum cudartDeviceAttr_enum {
   cudartDevAttrComputeCapabilityMajor = 75,
   cudartDevAttrComputeCapabilityMinor = 76,
-
-  // TODO - not yet wired up but may be useful for Jetson or other
-  // integrated GPU scenarios with shared memory
-  cudaDevAttrIntegrated = 18
-
 } cudartDeviceAttr_t;
 
 typedef void *cudartDevice_t;  // Opaque is sufficient
@@ -34,92 +28,6 @@ typedef struct cudartDriverVersion {
   int minor;
 } cudartDriverVersion_t;
 
-typedef struct cudaUUID {
-    unsigned char bytes[16];
-} cudaUUID_t;
-typedef struct cudaDeviceProp {
-    char         name[256];                  /**< ASCII string identifying device */
-    cudaUUID_t   uuid;                       /**< 16-byte unique identifier */
-    char         luid[8];                    /**< 8-byte locally unique identifier. Value is undefined on TCC and non-Windows platforms */
-    unsigned int luidDeviceNodeMask;         /**< LUID device node mask. Value is undefined on TCC and non-Windows platforms */
-    size_t       totalGlobalMem;             /**< Global memory available on device in bytes */
-    size_t       sharedMemPerBlock;          /**< Shared memory available per block in bytes */
-    int          regsPerBlock;               /**< 32-bit registers available per block */
-    int          warpSize;                   /**< Warp size in threads */
-    size_t       memPitch;                   /**< Maximum pitch in bytes allowed by memory copies */
-    int          maxThreadsPerBlock;         /**< Maximum number of threads per block */
-    int          maxThreadsDim[3];           /**< Maximum size of each dimension of a block */
-    int          maxGridSize[3];             /**< Maximum size of each dimension of a grid */
-    int          clockRate;                  /**< Clock frequency in kilohertz */
-    size_t       totalConstMem;              /**< Constant memory available on device in bytes */
-    int          major;                      /**< Major compute capability */
-    int          minor;                      /**< Minor compute capability */
-    size_t       textureAlignment;           /**< Alignment requirement for textures */
-    size_t       texturePitchAlignment;      /**< Pitch alignment requirement for texture references bound to pitched memory */
-    int          deviceOverlap;              /**< Device can concurrently copy memory and execute a kernel. Deprecated. Use instead asyncEngineCount. */
-    int          multiProcessorCount;        /**< Number of multiprocessors on device */
-    int          kernelExecTimeoutEnabled;   /**< Specified whether there is a run time limit on kernels */
-    int          integrated;                 /**< Device is integrated as opposed to discrete */
-    int          canMapHostMemory;           /**< Device can map host memory with cudaHostAlloc/cudaHostGetDevicePointer */
-    int          computeMode;                /**< Compute mode (See ::cudaComputeMode) */
-    int          maxTexture1D;               /**< Maximum 1D texture size */
-    int          maxTexture1DMipmap;         /**< Maximum 1D mipmapped texture size */
-    int          maxTexture1DLinear;         /**< Deprecated, do not use. Use cudaDeviceGetTexture1DLinearMaxWidth() or cuDeviceGetTexture1DLinearMaxWidth() instead. */
-    int          maxTexture2D[2];            /**< Maximum 2D texture dimensions */
-    int          maxTexture2DMipmap[2];      /**< Maximum 2D mipmapped texture dimensions */
-    int          maxTexture2DLinear[3];      /**< Maximum dimensions (width, height, pitch) for 2D textures bound to pitched memory */
-    int          maxTexture2DGather[2];      /**< Maximum 2D texture dimensions if texture gather operations have to be performed */
-    int          maxTexture3D[3];            /**< Maximum 3D texture dimensions */
-    int          maxTexture3DAlt[3];         /**< Maximum alternate 3D texture dimensions */
-    int          maxTextureCubemap;          /**< Maximum Cubemap texture dimensions */
-    int          maxTexture1DLayered[2];     /**< Maximum 1D layered texture dimensions */
-    int          maxTexture2DLayered[3];     /**< Maximum 2D layered texture dimensions */
-    int          maxTextureCubemapLayered[2];/**< Maximum Cubemap layered texture dimensions */
-    int          maxSurface1D;               /**< Maximum 1D surface size */
-    int          maxSurface2D[2];            /**< Maximum 2D surface dimensions */
-    int          maxSurface3D[3];            /**< Maximum 3D surface dimensions */
-    int          maxSurface1DLayered[2];     /**< Maximum 1D layered surface dimensions */
-    int          maxSurface2DLayered[3];     /**< Maximum 2D layered surface dimensions */
-    int          maxSurfaceCubemap;          /**< Maximum Cubemap surface dimensions */
-    int          maxSurfaceCubemapLayered[2];/**< Maximum Cubemap layered surface dimensions */
-    size_t       surfaceAlignment;           /**< Alignment requirements for surfaces */
-    int          concurrentKernels;          /**< Device can possibly execute multiple kernels concurrently */
-    int          ECCEnabled;                 /**< Device has ECC support enabled */
-    int          pciBusID;                   /**< PCI bus ID of the device */
-    int          pciDeviceID;                /**< PCI device ID of the device */
-    int          pciDomainID;                /**< PCI domain ID of the device */
-    int          tccDriver;                  /**< 1 if device is a Tesla device using TCC driver, 0 otherwise */
-    int          asyncEngineCount;           /**< Number of asynchronous engines */
-    int          unifiedAddressing;          /**< Device shares a unified address space with the host */
-    int          memoryClockRate;            /**< Peak memory clock frequency in kilohertz */
-    int          memoryBusWidth;             /**< Global memory bus width in bits */
-    int          l2CacheSize;                /**< Size of L2 cache in bytes */
-    int          persistingL2CacheMaxSize;   /**< Device's maximum l2 persisting lines capacity setting in bytes */
-    int          maxThreadsPerMultiProcessor;/**< Maximum resident threads per multiprocessor */
-    int          streamPrioritiesSupported;  /**< Device supports stream priorities */
-    int          globalL1CacheSupported;     /**< Device supports caching globals in L1 */
-    int          localL1CacheSupported;      /**< Device supports caching locals in L1 */
-    size_t       sharedMemPerMultiprocessor; /**< Shared memory available per multiprocessor in bytes */
-    int          regsPerMultiprocessor;      /**< 32-bit registers available per multiprocessor */
-    int          managedMemory;              /**< Device supports allocating managed memory on this system */
-    int          isMultiGpuBoard;            /**< Device is on a multi-GPU board */
-    int          multiGpuBoardGroupID;       /**< Unique identifier for a group of devices on the same multi-GPU board */
-    int          hostNativeAtomicSupported;  /**< Link between the device and the host supports native atomic operations */
-    int          singleToDoublePrecisionPerfRatio; /**< Ratio of single precision performance (in floating-point operations per second) to double precision performance */
-    int          pageableMemoryAccess;       /**< Device supports coherently accessing pageable memory without calling cudaHostRegister on it */
-    int          concurrentManagedAccess;    /**< Device can coherently access managed memory concurrently with the CPU */
-    int          computePreemptionSupported; /**< Device supports Compute Preemption */
-    int          canUseHostPointerForRegisteredMem; /**< Device can access host registered memory at the same virtual address as the CPU */
-    int          cooperativeLaunch;          /**< Device supports launching cooperative kernels via ::cudaLaunchCooperativeKernel */
-    int          cooperativeMultiDeviceLaunch; /**< Deprecated, cudaLaunchCooperativeKernelMultiDevice is deprecated. */
-    size_t       sharedMemPerBlockOptin;     /**< Per device maximum shared memory per block usable by special opt in */
-    int          pageableMemoryAccessUsesHostPageTables; /**< Device accesses pageable memory via the host's page tables */
-    int          directManagedMemAccessFromHost; /**< Host can directly access managed memory on the device without migration. */
-    int          maxBlocksPerMultiProcessor; /**< Maximum number of resident blocks per multiprocessor */
-    int          accessPolicyMaxWindowSize;  /**< The maximum value of ::cudaAccessPolicyWindow::num_bytes. */
-    size_t       reservedSharedMemPerBlock;  /**< Shared memory reserved by CUDA driver per block in bytes */
-  } cudaDeviceProp_t;
-
 typedef struct cudart_handle {
   void *handle;
   uint16_t verbose;
@@ -130,17 +38,23 @@ typedef struct cudart_handle {
   cudartReturn_t (*cudaGetDeviceCount)(int *);
   cudartReturn_t (*cudaDeviceGetAttribute)(int* value, cudartDeviceAttr_t attr, int device);
   cudartReturn_t (*cudaDriverGetVersion) (int *driverVersion);
-  cudartReturn_t (*cudaGetDeviceProperties) (cudaDeviceProp_t* prop, int device);
 } cudart_handle_t;
 
 typedef struct cudart_init_resp {
   char *err;  // If err is non-null handle is invalid
   cudart_handle_t ch;
-  int num_devices;
 } cudart_init_resp_t;
 
+typedef struct cudart_compute_capability {
+  char *err;
+  int major;
+  int minor;
+} cudart_compute_capability_t;
+
+
 void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
-void cudart_check_vram(cudart_handle_t ch, int device_id, mem_info_t *resp);
+void cudart_check_vram(cudart_handle_t ch, mem_info_t *resp);
+void cudart_compute_capability(cudart_handle_t th, cudart_compute_capability_t *cc);
 void cudart_release(cudart_handle_t ch);
 
 #endif  // __GPU_INFO_CUDART_H__

gpu/gpu_info_nvml.c

@@ -0,0 +1,221 @@
+#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
+
+#include <string.h>
+
+#include "gpu_info_nvml.h"
+
+void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
+  nvmlReturn_t ret;
+  resp->err = NULL;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  struct lookup {
+    char *s;
+    void **p;
+  } l[] = {
+      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
+      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
+      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.nvmlDeviceGetHandleByIndex},
+      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
+      {"nvmlDeviceGetCount_v2", (void *)&resp->ch.nvmlDeviceGetCount_v2},
+      {"nvmlDeviceGetCudaComputeCapability", (void *)&resp->ch.nvmlDeviceGetCudaComputeCapability},
+      {"nvmlSystemGetDriverVersion", (void *)&resp->ch.nvmlSystemGetDriverVersion},
+      {"nvmlDeviceGetName", (void *)&resp->ch.nvmlDeviceGetName},
+      {"nvmlDeviceGetSerial", (void *)&resp->ch.nvmlDeviceGetSerial},
+      {"nvmlDeviceGetVbiosVersion", (void *)&resp->ch.nvmlDeviceGetVbiosVersion},
+      {"nvmlDeviceGetBoardPartNumber", (void *)&resp->ch.nvmlDeviceGetBoardPartNumber},
+      {"nvmlDeviceGetBrand", (void *)&resp->ch.nvmlDeviceGetBrand},
+      {NULL, NULL},
+  };
+
+  resp->ch.handle = LOAD_LIBRARY(nvml_lib_path, RTLD_LAZY);
+  if (!resp->ch.handle) {
+    char *msg = LOAD_ERR();
+    LOG(resp->ch.verbose, "library %s load err: %s\n", nvml_lib_path, msg);
+    snprintf(buf, buflen,
+             "Unable to load %s library to query for Nvidia GPUs: %s",
+             nvml_lib_path, msg);
+    free(msg);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  // TODO once we've squashed the remaining corner cases remove this log
+  LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", nvml_lib_path);
+  
+  for (i = 0; l[i].s != NULL; i++) {
+    // TODO once we've squashed the remaining corner cases remove this log
+    LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
+
+    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
+    if (!l[i].p) {
+      resp->ch.handle = NULL;
+      char *msg = LOAD_ERR();
+      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->ch.handle);
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+               msg);
+      free(msg);
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+
+  ret = (*resp->ch.nvmlInit_v2)();
+  if (ret != NVML_SUCCESS) {
+    LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  // Report driver version if we're in verbose mode, ignore errors
+  ret = (*resp->ch.nvmlSystemGetDriverVersion)(buf, buflen);
+  if (ret != NVML_SUCCESS) {
+    LOG(resp->ch.verbose, "nvmlSystemGetDriverVersion failed: %d\n", ret);
+  } else {
+    LOG(resp->ch.verbose, "CUDA driver version: %s\n", buf);
+  }
+}
+
+void nvml_check_vram(nvml_handle_t h, mem_info_t *resp) {
+  resp->err = NULL;
+  nvmlDevice_t device;
+  nvmlMemory_t memInfo = {0};
+  nvmlReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("nvml handle isn't initialized");
+    return;
+  }
+
+  ret = (*h.nvmlDeviceGetCount_v2)(&resp->count);
+  if (ret != NVML_SUCCESS) {
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  resp->total = 0;
+  resp->free = 0;
+  for (i = 0; i < resp->count; i++) {
+    ret = (*h.nvmlDeviceGetHandleByIndex)(i, &device);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "unable to get device handle %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+
+    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "device memory info lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    if (h.verbose) {
+      nvmlBrandType_t brand = 0;
+      // When in verbose mode, report more information about
+      // the card we discover, but don't fail on error
+      ret = (*h.nvmlDeviceGetName)(device, buf, buflen);
+      if (ret != NVML_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetName failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA device name: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetBoardPartNumber)(device, buf, buflen);
+      if (ret != NVML_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetBoardPartNumber failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA part number: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetSerial)(device, buf, buflen);
+      if (ret != NVML_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetSerial failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA S/N: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetVbiosVersion)(device, buf, buflen);
+      if (ret != NVML_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetVbiosVersion failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA vbios version: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetBrand)(device, &brand);
+      if (ret != NVML_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetBrand failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA brand: %d\n", i, brand);
+      }
+    }
+
+    LOG(h.verbose, "[%d] CUDA totalMem %ld\n", i, memInfo.total);
+    LOG(h.verbose, "[%d] CUDA freeMem %ld\n", i, memInfo.free);
+
+    resp->total += memInfo.total;
+    resp->free += memInfo.free;
+  }
+}
+
+void nvml_compute_capability(nvml_handle_t h, nvml_compute_capability_t *resp) {
+  resp->err = NULL;
+  resp->major = 0;
+  resp->minor = 0;
+  nvmlDevice_t device;
+  int major = 0;
+  int minor = 0;
+  nvmlReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("nvml handle not initialized");
+    return;
+  }
+
+  unsigned int devices;
+  ret = (*h.nvmlDeviceGetCount_v2)(&devices);
+  if (ret != NVML_SUCCESS) {
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  for (i = 0; i < devices; i++) {
+    ret = (*h.nvmlDeviceGetHandleByIndex)(i, &device);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "unable to get device handle %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+
+    ret = (*h.nvmlDeviceGetCudaComputeCapability)(device, &major, &minor);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    // Report the lowest major.minor we detect as that limits our compatibility
+    if (resp->major == 0 || resp->major > major ) {
+      resp->major = major;
+      resp->minor = minor;
+    } else if ( resp->major == major && resp->minor > minor ) {
+      resp->minor = minor;
+    }
+  }
+}
+
+void nvml_release(nvml_handle_t h) {
+  LOG(h.verbose, "releasing nvml library\n");
+  UNLOAD_LIBRARY(h.handle);
+  h.handle = NULL;
+}
+
+#endif  // __APPLE__

gpu/gpu_info_nvml.h

@@ -0,0 +1,57 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_NVML_H__
+#define __GPU_INFO_NVML_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum nvmlReturn_enum {
+  NVML_SUCCESS = 0,
+  // Other values omitted for now...
+} nvmlReturn_t;
+typedef void *nvmlDevice_t;  // Opaque is sufficient
+typedef struct nvmlMemory_st {
+  unsigned long long total;
+  unsigned long long free;
+  unsigned long long used;
+} nvmlMemory_t;
+
+typedef enum nvmlBrandType_enum
+{
+    NVML_BRAND_UNKNOWN          = 0,
+} nvmlBrandType_t;
+
+typedef struct nvml_handle {
+  void *handle;
+  uint16_t verbose;
+  nvmlReturn_t (*nvmlInit_v2)(void);
+  nvmlReturn_t (*nvmlShutdown)(void);
+  nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(unsigned int, nvmlDevice_t *);
+  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
+  nvmlReturn_t (*nvmlDeviceGetCount_v2)(unsigned int *);
+  nvmlReturn_t (*nvmlDeviceGetCudaComputeCapability)(nvmlDevice_t, int* major, int* minor);
+  nvmlReturn_t (*nvmlSystemGetDriverVersion) (char* version, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetName) (nvmlDevice_t device, char* name, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetSerial) (nvmlDevice_t device, char* serial, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetVbiosVersion) (nvmlDevice_t device, char* version, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetBoardPartNumber) (nvmlDevice_t device, char* partNumber, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetBrand) (nvmlDevice_t device, nvmlBrandType_t* type);
+} nvml_handle_t;
+
+typedef struct nvml_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  nvml_handle_t ch;
+} nvml_init_resp_t;
+
+typedef struct nvml_compute_capability {
+  char *err;
+  int major;
+  int minor;
+} nvml_compute_capability_t;
+
+void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp);
+void nvml_check_vram(nvml_handle_t ch, mem_info_t *resp);
+void nvml_compute_capability(nvml_handle_t ch, nvml_compute_capability_t *cc);
+void nvml_release(nvml_handle_t ch);
+
+#endif  // __GPU_INFO_NVML_H__
+#endif  // __APPLE__

gpu/gpu_test.go

@@ -9,16 +9,23 @@ import (
 
 func TestBasicGetGPUInfo(t *testing.T) {
 	info := GetGPUInfo()
-	assert.Greater(t, len(info), 0)
+	assert.Contains(t, "cuda rocm cpu metal", info.Library)
-	assert.Contains(t, "cuda rocm cpu metal", info[0].Library)
+
-	if info[0].Library != "cpu" {
+	switch runtime.GOOS {
-		assert.Greater(t, info[0].TotalMemory, uint64(0))
+	case "darwin":
-		assert.Greater(t, info[0].FreeMemory, uint64(0))
+		// TODO - remove this once MacOS returns some size for CPU
+		return
+	case "linux", "windows":
+		assert.Greater(t, info.TotalMemory, uint64(0))
+		assert.Greater(t, info.FreeMemory, uint64(0))
+		assert.Greater(t, info.DeviceCount, uint32(0))
+	default:
+		return
 	}
 }
 
 func TestCPUMemInfo(t *testing.T) {
-	info, err := GetCPUMem()
+	info, err := getCPUMem()
 	assert.NoError(t, err)
 	switch runtime.GOOS {
 	case "darwin":

gpu/types.go

@@ -3,6 +3,7 @@ package gpu
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
+	DeviceCount uint32 `json:"device_count,omitempty"`
 }
 
 // Beginning of an `ollama info` command
@@ -16,49 +17,11 @@ type GpuInfo struct {
 	// MinimumMemory represents the minimum memory required to use the GPU
 	MinimumMemory uint64 `json:"-"`
 
-	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
+	// TODO add other useful attributes about the card here for discovery information
-	DependencyPath string `json:"lib_path,omitempty"`
-
-	// GPU information
-	ID    string `json:"gpu_id"`          // string to use for selection of this specific GPU
-	Name  string `json:"name"`            // user friendly name if available
-	Major int    `json:"major,omitempty"` // Major compatibility version (CC or gfx)
-	Minor int    `json:"minor,omitempty"` // Minor compatibility version (CC or gfx)
-	Patch int    `json:"patch,omitempty"` // Patch compatibility only matters on AMD
-
-	// TODO other performance capability info to help in scheduling decisions
 }
 
-type GpuInfoList []GpuInfo
+type Version struct {
-
+	Major uint
-// Split up the set of gpu info's by Library and variant
+	Minor uint
-func (l GpuInfoList) ByLibrary() []GpuInfoList {
+	Patch uint
-	resp := []GpuInfoList{}
-	libs := []string{}
-	for _, info := range l {
-		found := false
-		requested := info.Library
-		if info.Variant != "" {
-			requested += "_" + info.Variant
-		}
-		for i, lib := range libs {
-			if lib == requested {
-				resp[i] = append(resp[i], info)
-				found = true
-				break
-			}
-		}
-		if !found {
-			libs = append(libs, info.Library)
-			resp = append(resp, []GpuInfo{info})
-		}
-	}
-	return resp
 }
-
-// Sort by Free Space
-type ByFreeMemory []GpuInfo
-
-func (a ByFreeMemory) Len() int           { return len(a) }
-func (a ByFreeMemory) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
-func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }

integration/basic_test.go

@@ -4,14 +4,11 @@ package integration
 
 import (
 	"context"
-	"log/slog"
+	"net/http"
-	"os"
-	"runtime"
 	"testing"
 	"time"
 
 	"github.com/ollama/ollama/api"
-	"github.com/stretchr/testify/require"
 )
 
 func TestOrcaMiniBlueSky(t *testing.T) {
@@ -27,44 +24,5 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+	GenerateTestHelper(ctx, t, &http.Client{}, req, []string{"rayleigh", "scattering"})
-}
-
-func TestUnicodeModelDir(t *testing.T) {
-	// This is only useful for Windows with utf-16 characters, so skip this test for other platforms
-	if runtime.GOOS != "windows" {
-		t.Skip("Unicode test only applicable to windows")
-	}
-	// Only works for local testing
-	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
-		t.Skip("TestUnicodeModelDir only works for local testing, skipping")
-	}
-
-	modelDir, err := os.MkdirTemp("", "ollama_埃")
-	require.NoError(t, err)
-	defer os.RemoveAll(modelDir)
-	slog.Info("unicode", "OLLAMA_MODELS", modelDir)
-
-	oldModelsDir := os.Getenv("OLLAMA_MODELS")
-	if oldModelsDir == "" {
-		defer os.Unsetenv("OLLAMA_MODELS")
-	} else {
-		defer os.Setenv("OLLAMA_MODELS", oldModelsDir)
-	}
-	err = os.Setenv("OLLAMA_MODELS", modelDir)
-	require.NoError(t, err)
-
-	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
-	defer cancel()
-
-	req := api.GenerateRequest{
-		Model:  "orca-mini",
-		Prompt: "why is the sky blue?",
-		Stream: &stream,
-		Options: map[string]interface{}{
-			"temperature": 0,
-			"seed":        123,
-		},
-	}
-	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
 }

integration/concurrency_test.go

@@ -1,225 +0,0 @@
-//go:build integration
-
-package integration
-
-import (
-	"context"
-	"log/slog"
-	"os"
-	"strconv"
-	"sync"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-	"github.com/stretchr/testify/require"
-)
-
-func TestMultiModelConcurrency(t *testing.T) {
-	var (
-		req = [2]api.GenerateRequest{
-			{
-				Model:  "orca-mini",
-				Prompt: "why is the ocean blue?",
-				Stream: &stream,
-				Options: map[string]interface{}{
-					"seed":        42,
-					"temperature": 0.0,
-				},
-			}, {
-				Model:  "tinydolphin",
-				Prompt: "what is the origin of the us thanksgiving holiday?",
-				Stream: &stream,
-				Options: map[string]interface{}{
-					"seed":        42,
-					"temperature": 0.0,
-				},
-			},
-		}
-		resp = [2][]string{
-			[]string{"sunlight"},
-			[]string{"england", "english", "massachusetts", "pilgrims"},
-		}
-	)
-	var wg sync.WaitGroup
-	wg.Add(len(req))
-	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
-	defer cancel()
-	for i := 0; i < len(req); i++ {
-		go func(i int) {
-			defer wg.Done()
-			GenerateTestHelper(ctx, t, req[i], resp[i])
-		}(i)
-	}
-	wg.Wait()
-}
-
-func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute) // GTX 750 2G card takes ~9 minutes
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	req, resp := GenerateRequests()
-	// Get the server running (if applicable) warm the model up with a single initial request
-	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 5*time.Second)
-
-	var wg sync.WaitGroup
-	wg.Add(len(req))
-	for i := 0; i < len(req); i++ {
-		go func(i int) {
-			defer wg.Done()
-			for j := 0; j < 5; j++ {
-				slog.Info("Starting", "req", i, "iter", j)
-				// On slower GPUs it can take a while to process the 4 concurrent requests
-				// so we allow a much longer initial timeout
-				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
-			}
-		}(i)
-	}
-	wg.Wait()
-}
-
-// Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
-func TestMultiModelStress(t *testing.T) {
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
-	if vram == "" {
-		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
-	}
-	max, err := strconv.ParseUint(vram, 10, 64)
-	require.NoError(t, err)
-	const MB = uint64(1024 * 1024)
-	type model struct {
-		name string
-		size uint64 // Approximate amount of VRAM they typically use when fully loaded in VRAM
-	}
-
-	smallModels := []model{
-		{
-			name: "orca-mini",
-			size: 2992 * MB,
-		},
-		{
-			name: "phi",
-			size: 2616 * MB,
-		},
-		{
-			name: "gemma:2b",
-			size: 2364 * MB,
-		},
-		{
-			name: "stable-code:3b",
-			size: 2608 * MB,
-		},
-		{
-			name: "starcoder2:3b",
-			size: 2166 * MB,
-		},
-	}
-	mediumModels := []model{
-		{
-			name: "llama2",
-			size: 5118 * MB,
-		},
-		{
-			name: "mistral",
-			size: 4620 * MB,
-		},
-		{
-			name: "orca-mini:7b",
-			size: 5118 * MB,
-		},
-		{
-			name: "dolphin-mistral",
-			size: 4620 * MB,
-		},
-		{
-			name: "gemma:7b",
-			size: 5000 * MB,
-		},
-		// TODO - uncomment this once #3565 is merged and this is rebased on it
-		// {
-		// 	name: "codellama:7b",
-		// 	size: 5118 * MB,
-		// },
-	}
-
-	// These seem to be too slow to be useful...
-	// largeModels := []model{
-	// 	{
-	// 		name: "llama2:13b",
-	// 		size: 7400 * MB,
-	// 	},
-	// 	{
-	// 		name: "codellama:13b",
-	// 		size: 7400 * MB,
-	// 	},
-	// 	{
-	// 		name: "orca-mini:13b",
-	// 		size: 7400 * MB,
-	// 	},
-	// 	{
-	// 		name: "gemma:7b",
-	// 		size: 5000 * MB,
-	// 	},
-	// 	{
-	// 		name: "starcoder2:15b",
-	// 		size: 9100 * MB,
-	// 	},
-	// }
-
-	var chosenModels []model
-	switch {
-	case max < 10000*MB:
-		slog.Info("selecting small models")
-		chosenModels = smallModels
-	// case max < 30000*MB:
-	default:
-		slog.Info("selecting medium models")
-		chosenModels = mediumModels
-		// default:
-		// 	slog.Info("selecting large models")
-		// 	chosenModels = largModels
-	}
-
-	req, resp := GenerateRequests()
-
-	for i := range req {
-		if i > len(chosenModels) {
-			break
-		}
-		req[i].Model = chosenModels[i].name
-	}
-
-	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute) // TODO baseline -- 10m too short
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// Make sure all the models are pulled before we get started
-	for _, r := range req {
-		require.NoError(t, PullIfMissing(ctx, client, r.Model))
-	}
-
-	var wg sync.WaitGroup
-	consumed := uint64(256 * MB) // Assume some baseline usage
-	for i := 0; i < len(req); i++ {
-		// Always get at least 2 models, but dont' overshoot VRAM too much or we'll take too long
-		if i > 1 && consumed > max {
-			slog.Info("achieved target vram exhaustion", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
-			break
-		}
-		consumed += chosenModels[i].size
-		slog.Info("target vram", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
-
-		wg.Add(1)
-		go func(i int) {
-			defer wg.Done()
-			for j := 0; j < 3; j++ {
-				slog.Info("Starting", "req", i, "iter", j, "model", req[i].Model)
-				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
-			}
-		}(i)
-	}
-	wg.Wait()
-}

integration/context_test.go

@@ -4,6 +4,7 @@ package integration
 
 import (
 	"context"
+	"net/http"
 	"testing"
 	"time"
 
@@ -24,5 +25,5 @@ func TestContextExhaustion(t *testing.T) {
 			"num_ctx":     128,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"once", "upon", "lived"})
+	GenerateTestHelper(ctx, t, &http.Client{}, req, []string{"once", "upon", "lived"})
 }

integration/llm_image_test.go

@@ -5,6 +5,7 @@ package integration
 import (
 	"context"
 	"encoding/base64"
+	"net/http"
 	"testing"
 	"time"
 
@@ -28,11 +29,10 @@ func TestIntegrationMultimodal(t *testing.T) {
 		},
 	}
 
-	// Note: sometimes it returns "the ollamas" sometimes "the ollams"
+	resp := "the ollamas"
-	resp := "the ollam"
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
-	GenerateTestHelper(ctx, t, req, []string{resp})
+	GenerateTestHelper(ctx, t, &http.Client{}, req, []string{resp})
 }
 
 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb

integration/llm_test.go

@@ -4,6 +4,8 @@ package integration
 
 import (
 	"context"
+	"net/http"
+	"sync"
 	"testing"
 	"time"
 
@@ -43,5 +45,25 @@ var (
 func TestIntegrationSimpleOrcaMini(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
 	defer cancel()
-	GenerateTestHelper(ctx, t, req[0], resp[0])
+	GenerateTestHelper(ctx, t, &http.Client{}, req[0], resp[0])
 }
+
+// TODO
+// The server always loads a new runner and closes the old one, which forces serial execution
+// At present this test case fails with concurrency problems.  Eventually we should try to
+// get true concurrency working with n_parallel support in the backend
+func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
+	var wg sync.WaitGroup
+	wg.Add(len(req))
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	defer cancel()
+	for i := 0; i < len(req); i++ {
+		go func(i int) {
+			defer wg.Done()
+			GenerateTestHelper(ctx, t, &http.Client{}, req[i], resp[i])
+		}(i)
+	}
+	wg.Wait()
+}
+
+// TODO - create a parallel test with 2 different models once we support concurrency

integration/utils_test.go

@@ -5,14 +5,13 @@ package integration
 import (
 	"bytes"
 	"context"
-	"errors"
+	"encoding/json"
 	"fmt"
 	"io"
 	"log/slog"
 	"math/rand"
 	"net"
 	"net/http"
-	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -24,13 +23,9 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/app/lifecycle"
-	"github.com/stretchr/testify/require"
+	"github.com/stretchr/testify/assert"
 )
 
-func Init() {
-	lifecycle.InitLogging()
-}
-
 func FindPort() string {
 	port := 0
 	if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
@@ -46,7 +41,7 @@ func FindPort() string {
 	return strconv.Itoa(port)
 }
 
-func GetTestEndpoint() (*api.Client, string) {
+func GetTestEndpoint() (string, string) {
 	defaultPort := "11434"
 	ollamaHost := os.Getenv("OLLAMA_HOST")
 
@@ -72,20 +67,16 @@ func GetTestEndpoint() (*api.Client, string) {
 		port = FindPort()
 	}
 
-	slog.Info("server connection", "host", host, "port", port)
+	url := fmt.Sprintf("%s:%s", host, port)
-
+	slog.Info("server connection", "url", url)
-	return api.NewClient(
+	return scheme, url
-		&url.URL{
-			Scheme: scheme,
-			Host:   net.JoinHostPort(host, port),
-		},
-		http.DefaultClient), fmt.Sprintf("%s:%s", host, port)
 }
 
+// TODO make fanicier, grab logs, etc.
 var serverMutex sync.Mutex
 var serverReady bool
 
-func startServer(ctx context.Context, ollamaHost string) error {
+func StartServer(ctx context.Context, ollamaHost string) error {
 	// Make sure the server has been built
 	CLIName, err := filepath.Abs("../ollama")
 	if err != nil {
@@ -134,76 +125,67 @@ func startServer(ctx context.Context, ollamaHost string) error {
 	return nil
 }
 
-func PullIfMissing(ctx context.Context, client *api.Client, modelName string) error {
+func PullIfMissing(ctx context.Context, client *http.Client, scheme, testEndpoint, modelName string) error {
 	slog.Info("checking status of model", "model", modelName)
 	showReq := &api.ShowRequest{Name: modelName}
-
+	requestJSON, err := json.Marshal(showReq)
-	showCtx, cancel := context.WithDeadlineCause(
+	if err != nil {
-		ctx,
-		time.Now().Add(5*time.Second),
-		fmt.Errorf("show for existing model %s took too long", modelName),
-	)
-	defer cancel()
-	_, err := client.Show(showCtx, showReq)
-	var statusError api.StatusError
-	switch {
-	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
-		break
-	case err != nil:
 		return err
-	default:
+	}
+
+	req, err := http.NewRequest("POST", scheme+"://"+testEndpoint+"/api/show", bytes.NewReader(requestJSON))
+	if err != nil {
+		return err
+	}
+
+	// Make the request with the HTTP client
+	response, err := client.Do(req.WithContext(ctx))
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode == 200 {
 		slog.Info("model already present", "model", modelName)
 		return nil
 	}
-	slog.Info("model missing", "model", modelName)
+	slog.Info("model missing", "status", response.StatusCode)
 
-	stallDuration := 30 * time.Second // This includes checksum verification, which can take a while on larger models
-	stallTimer := time.NewTimer(stallDuration)
-	fn := func(resp api.ProgressResponse) error {
-		// fmt.Print(".")
-		if !stallTimer.Reset(stallDuration) {
-			return fmt.Errorf("stall was detected, aborting status reporting")
-		}
-		return nil
-	}
-
-	stream := true
 	pullReq := &api.PullRequest{Name: modelName, Stream: &stream}
-
+	requestJSON, err = json.Marshal(pullReq)
-	var pullError error
+	if err != nil {
-
+		return err
-	done := make(chan int)
-	go func() {
-		pullError = client.Pull(ctx, pullReq, fn)
-		done <- 0
-	}()
-
-	select {
-	case <-stallTimer.C:
-		return fmt.Errorf("download stalled")
-	case <-done:
-		return pullError
 	}
+
+	req, err = http.NewRequest("POST", scheme+"://"+testEndpoint+"/api/pull", bytes.NewReader(requestJSON))
+	if err != nil {
+		return err
+	}
+	slog.Info("pulling", "model", modelName)
+
+	response, err = client.Do(req.WithContext(ctx))
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode != 200 {
+		return fmt.Errorf("failed to pull model") // TODO more details perhaps
+	}
+	slog.Info("model pulled", "model", modelName)
+	return nil
 }
 
 var serverProcMutex sync.Mutex
 
-// Returns an Client, the testEndpoint, and a cleanup function, fails the test on errors
+func GenerateTestHelper(ctx context.Context, t *testing.T, client *http.Client, genReq api.GenerateRequest, anyResp []string) {
-// Starts the server if needed
-func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, string, func()) {
-	client, testEndpoint := GetTestEndpoint()
-	if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
-		serverProcMutex.Lock()
-		fp, err := os.CreateTemp("", "ollama-server-*.log")
-		if err != nil {
-			t.Fatalf("failed to generate log file: %s", err)
-		}
-		lifecycle.ServerLogFile = fp.Name()
-		fp.Close()
-		require.NoError(t, startServer(ctx, testEndpoint))
-	}
 
-	return client, testEndpoint, func() {
+	// TODO maybe stuff in an init routine?
+	lifecycle.InitLogging()
+
+	requestJSON, err := json.Marshal(genReq)
+	if err != nil {
+		t.Fatalf("Error serializing request: %v", err)
+	}
+	defer func() {
 		if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
 			defer serverProcMutex.Unlock()
 			if t.Failed() {
@@ -221,118 +203,63 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 				os.Stderr.Write(data)
 				slog.Warn("END OF SERVER")
 			}
-			err := os.Remove(lifecycle.ServerLogFile)
+			err = os.Remove(lifecycle.ServerLogFile)
 			if err != nil && !os.IsNotExist(err) {
 				slog.Warn("failed to cleanup", "logfile", lifecycle.ServerLogFile, "error", err)
 			}
 		}
-	}
-}
-
-func GenerateTestHelper(ctx context.Context, t *testing.T, genReq api.GenerateRequest, anyResp []string) {
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, genReq.Model))
-	DoGenerate(ctx, t, client, genReq, anyResp, 30*time.Second, 10*time.Second)
-}
-
-func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq api.GenerateRequest, anyResp []string, initialTimeout, streamTimeout time.Duration) {
-	stallTimer := time.NewTimer(initialTimeout)
-	var buf bytes.Buffer
-	fn := func(response api.GenerateResponse) error {
-		// fmt.Print(".")
-		buf.Write([]byte(response.Response))
-		if !stallTimer.Reset(streamTimeout) {
-			return fmt.Errorf("stall was detected while streaming response, aborting")
-		}
-		return nil
-	}
-
-	stream := true
-	genReq.Stream = &stream
-	done := make(chan int)
-	var genErr error
-	go func() {
-		genErr = client.Generate(ctx, &genReq, fn)
-		done <- 0
 	}()
+	scheme, testEndpoint := GetTestEndpoint()
 
-	select {
+	if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
-	case <-stallTimer.C:
+		serverProcMutex.Lock()
-		if buf.Len() == 0 {
+		fp, err := os.CreateTemp("", "ollama-server-*.log")
-			t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
+		if err != nil {
-		} else {
+			t.Fatalf("failed to generate log file: %s", err)
-			t.Errorf("generate stalled.  Response so far:%s", buf.String())
 		}
-	case <-done:
+		lifecycle.ServerLogFile = fp.Name()
-		require.NoError(t, genErr, "failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
+		fp.Close()
-		// Verify the response contains the expected data
+		assert.NoError(t, StartServer(ctx, testEndpoint))
-		response := buf.String()
-		atLeastOne := false
-		for _, resp := range anyResp {
-			if strings.Contains(strings.ToLower(response), resp) {
-				atLeastOne = true
-				break
-			}
-		}
-		require.True(t, atLeastOne, "none of %v found in %s", anyResp, response)
-		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
-	case <-ctx.Done():
-		t.Error("outer test context done while waiting for generate")
 	}
-}
 
-// Generate a set of requests
+	err = PullIfMissing(ctx, client, scheme, testEndpoint, genReq.Model)
-// By default each request uses orca-mini as the model
+	if err != nil {
-func GenerateRequests() ([]api.GenerateRequest, [][]string) {
+		t.Fatalf("Error pulling model: %v", err)
-	return []api.GenerateRequest{
+	}
-			{
+
-				Model:  "orca-mini",
+	// Make the request and get the response
-				Prompt: "why is the ocean blue?",
+	req, err := http.NewRequest("POST", scheme+"://"+testEndpoint+"/api/generate", bytes.NewReader(requestJSON))
-				Stream: &stream,
+	if err != nil {
-				Options: map[string]interface{}{
+		t.Fatalf("Error creating request: %v", err)
-					"seed":        42,
+	}
-					"temperature": 0.0,
+
-				},
+	// Set the content type for the request
-			}, {
+	req.Header.Set("Content-Type", "application/json")
-				Model:  "orca-mini",
+
-				Prompt: "why is the color of dirt brown?",
+	// Make the request with the HTTP client
-				Stream: &stream,
+	response, err := client.Do(req.WithContext(ctx))
-				Options: map[string]interface{}{
+	if err != nil {
-					"seed":        42,
+		t.Fatalf("Error making request: %v", err)
-					"temperature": 0.0,
+	}
-				},
+	defer response.Body.Close()
-			}, {
+	body, err := io.ReadAll(response.Body)
-				Model:  "orca-mini",
+	assert.NoError(t, err)
-				Prompt: "what is the origin of the us thanksgiving holiday?",
+	assert.Equal(t, response.StatusCode, 200, string(body))
-				Stream: &stream,
+
-				Options: map[string]interface{}{
+	// Verify the response is valid JSON
-					"seed":        42,
+	var payload api.GenerateResponse
-					"temperature": 0.0,
+	err = json.Unmarshal(body, &payload)
-				},
+	if err != nil {
-			}, {
+		assert.NoError(t, err, body)
-				Model:  "orca-mini",
+	}
-				Prompt: "what is the origin of independence day?",
+
-				Stream: &stream,
+	// Verify the response contains the expected data
-				Options: map[string]interface{}{
+	atLeastOne := false
-					"seed":        42,
+	for _, resp := range anyResp {
-					"temperature": 0.0,
+		if strings.Contains(strings.ToLower(payload.Response), resp) {
-				},
+			atLeastOne = true
-			}, {
+			break
-				Model:  "orca-mini",
-				Prompt: "what is the composition of air?",
-				Stream: &stream,
-				Options: map[string]interface{}{
-					"seed":        42,
-					"temperature": 0.0,
-				},
-			},
-		},
-		[][]string{
-			[]string{"sunlight"},
-			[]string{"soil", "organic", "earth", "black", "tan"},
-			[]string{"england", "english", "massachusetts", "pilgrims"},
-			[]string{"fourth", "july", "declaration", "independence"},
-			[]string{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
+	}
+	assert.True(t, atLeastOne, "none of %v found in %s", anyResp, payload.Response)
 }

llm/ext_server/server.cpp

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

llm/generate/gen_common.sh

@@ -21,7 +21,7 @@ init_vars() {
         # TODO - add additional optimization flags...
         CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off ${CMAKE_DEFS}"
     fi
-    case $(uname -s) in
+    case $(uname -s) in 
     "Darwin")
         LIB_EXT="dylib"
         WHOLE_ARCHIVE="-Wl,-force_load"

llm/generate/gen_linux.sh

@@ -57,21 +57,21 @@ init_vars
 git_module_setup
 apply_patches
 
-init_vars
-if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
-    # Builds by default, allows skipping, forces build if OLLAMA_CPU_TARGET="static"
-    # Enables optimized Dockerfile builds using a blanket skip and targeted overrides
-    # Static build for linking into the Go binary
-    init_vars
-    CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-    BUILD_DIR="../build/linux/${ARCH}_static"
-    echo "Building static library"
-    build
-fi
 
 init_vars
 if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
+
+    if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
+        # Static build for linking into the Go binary
+        init_vars
+        CMAKE_TARGETS="--target llama --target ggml"
+        CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        BUILD_DIR="../build/linux/${ARCH}_static"
+        echo "Building static library"
+        build
+    fi
+
+
     # Users building from source can tune the exact flags we pass to cmake for configuring
     # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
@@ -165,22 +165,14 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
     fi
     if [ "${ARCH}" == "arm64" ]; then
         echo "ARM CPU detected - disabling unsupported AVX instructions"
-
+        
         # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
         #
-        # CUDA compute < 6.0 lacks proper FP16 support on ARM.
+        # CUDA compute < 6.0 lacks proper FP16 support on ARM. 
-        # Disabling has minimal performance effect while maintaining compatibility.
+        # Disabling has minimal performance effect while maintaining compatibility. 
         ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
     fi
-    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
+    CMAKE_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS}"
-    if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
-        echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
-        echo "Building custom CUDA GPU"
-    else
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
-    fi
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
     EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
     build
@@ -225,12 +217,6 @@ if [ -d "${ROCM_PATH}" ]; then
     fi
     init_vars
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
-    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
-    if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
-        echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
-        CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
-        echo "Building custom ROCM GPU"
-    fi
     BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
     EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build

llm/generate/gen_windows.ps1

@@ -26,25 +26,15 @@ function amdGPUs {
     $GPU_LIST -join ';'
 }
 
-
 function init_vars {
-    if (!$script:SRC_DIR) {
+    $script:SRC_DIR = $(resolve-path "..\..\")
-        $script:SRC_DIR = $(resolve-path "..\..\")
+    $script:llamacppDir = "../llama.cpp"
-    }
-    if (!$script:llamacppDir) {
-        $script:llamacppDir = "../llama.cpp"
-    }
-    if (!$script:cmakeTargets) {
-        $script:cmakeTargets = @("ollama_llama_server")
-    }
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
         "-DLLAMA_NATIVE=off"
         )
-    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
+    $script:cmakeTargets = @("ollama_llama_server")
     $script:ARCH = "amd64" # arm not yet supported.
-    $script:DIST_BASE = "${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"
@@ -65,6 +55,7 @@ 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"
@@ -143,18 +134,21 @@ function sign {
     }
 }
 
-function install {
+function compress {
-    write-host "Installing binaries to dist dir ${script:distDir}"
+    if ($script:GZIP -eq $null) {
-    mkdir ${script:distDir} -ErrorAction SilentlyContinue
+        write-host "gzip not installed, not compressing files"
+        return
+    }
+    write-host "Compressing binaries..."
     $binaries = dir "${script:buildDir}/bin/*.exe"
     foreach ($file in $binaries) {
-        copy-item -Path $file -Destination ${script:distDir} -Force
+        & "$script:GZIP" --best -f $file
     }
 
-    write-host "Installing dlls to dist dir ${script:distDir}"
+    write-host "Compressing dlls..."
     $dlls = dir "${script:buildDir}/bin/*.dll"
     foreach ($file in $dlls) {
-        copy-item -Path $file -Destination ${script:distDir} -Force
+        & "$script:GZIP" --best -f $file
     }
 }
 
@@ -175,191 +169,123 @@ 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")
 
-function build_static() {
+if ($null -eq ${env:OLLAMA_SKIP_CPU_GENERATE}) {
-    if ((-not "${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 = $oldTargets
-    } else {
-        write-host "Skipping CPU generation step as requested"
-    }
-}
-
-function build_cpu() {
-    if ((-not "${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", "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"
-        $script:distDir="$script:DIST_BASE\cpu"
-        write-host "Building LCD CPU"
-        build
-        sign
-        install
-    } else {
-        write-host "Skipping CPU generation step as requested"
-    }
-}
-
-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_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-${script:ARCH}\"
-    } else {
-        write-host "Skipping CUDA generation step"
-    }
-}
-
-function build_rocm() {
-    if ((-not "${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
-        }
-
-        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"
-    }
-}
 
+# GCC build for direct linking into the Go binary
 init_vars
-if ($($args.count) -eq 0) {
+# cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
-    git_module_setup
+# as we need this to be compiled by gcc for golang to be able to link with itx
-    apply_patches
+write-host "Checking for MinGW..."
-    build_static
+# error action ensures we exit on failure
-    build_cpu
+get-command gcc
-    build_cpu_avx
+get-command mingw32-make
-    build_cpu_avx2
+$script:cmakeTargets = @("llama", "ggml")
-    build_cuda
+$script:cmakeDefs = @(
-    build_rocm
+    "-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
 
-    cleanup
+# remaining llama.cpp builds use MSVC 
-    write-host "`ngo generate completed.  LLM runners: $(get-childitem -path $script:DIST_BASE)"
+    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 {
-    for ( $i = 0; $i -lt $args.count; $i++ ) {
+    write-host "Skipping CPU generation step as requested"
-        write-host "performing $($args[$i])"
+}
-        & $($args[$i])
+
-    } 
+if ($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
+    }
+    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=${script:CMAKE_CUDA_ARCHITECTURES}")
+    build
+    sign
+    compress
+}
+
+if ($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
+    }
+
+    init_vars
+    $script:buildDir="../build/windows/${script:ARCH}/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=""
+
+    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
+    compress
+}
+
+
+cleanup
+write-host "`ngo generate completed.  LLM runners: $(get-childitem -path ${script:SRC_DIR}\llm\build\windows\${script:ARCH})"

llm/ggml.go

@@ -164,8 +164,7 @@ func (ts Tensors) Layers() map[string]Layer {
 	for _, t := range ts {
 		parts := strings.Split(t.Name, ".")
 		if parts[0] == "blk" {
-			// join first and second part, e.g. blk.%d
+			parts = parts[1:]
-			parts = append([]string{fmt.Sprintf("%s.%s", parts[0], parts[1])}, parts[2:]...)
 		}
 
 		if _, ok := layers[parts[0]]; !ok {
@@ -343,15 +342,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
 
-		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
+		if ffnGateWeight, ok := layers["0"]["ffn_gate.0.weight"]; ok {
-			// mixtral 8x22b
-			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
-			partialOffload = max(
-				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(
@@ -389,12 +380,6 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 		)
 
 		partialOffload = 4*batch*(2*embedding+vocab) + embedding*vocab*105/128
-	case "stablelm":
-		fullOffload = 4 * batch * (context*(1+heads) + 3*embedding + 2)
-		partialOffload = max(
-			4*batch*(vocab+2*embedding),
-			fullOffload,
-		)
 	}
 
 	return

llm/gguf.go

@@ -190,6 +190,8 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		llm.kv[k] = v
 	}
 
+	slog.Debug(fmt.Sprintf("general.architecture = %s", llm.kv["general.architecture"]))
+
 	// decode tensors
 	for i := 0; uint64(i) < llm.numTensor(); i++ {
 		name, err := readGGUFString(llm, rs)
@@ -246,17 +248,13 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	padding := llm.padding(offset, int64(alignment))
-	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
+	if _, err := rs.Seek(padding-offset, io.SeekCurrent); err != nil {
 		return err
 	}
 
 	for _, tensor := range llm.tensors {
-		if _, err := rs.Seek(int64(tensor.size()), io.SeekCurrent); err != nil {
+		padded := (int64(tensor.size()) + int64(alignment) - 1) & ^(int64(alignment) - 1)
-			return err
+		if _, err := rs.Seek(padded, io.SeekCurrent); err != nil {
-		}
-
-		padding := llm.padding(int64(tensor.size()), int64(alignment))
-		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
 			return err
 		}
 	}
@@ -463,13 +461,11 @@ var ggufKVOrder = map[string][]string{
 		"llama.embedding_length",
 		"llama.block_count",
 		"llama.feed_forward_length",
+		"llama.rope.dimension_count",
 		"llama.attention.head_count",
 		"llama.attention.head_count_kv",
 		"llama.attention.layer_norm_rms_epsilon",
 		"llama.rope.freq_base",
-		"llama.rope.dimension_count",
-		"llama.expert_count",
-		"llama.expert_used_count",
 		"gemma.context_length",
 		"gemma.embedding_length",
 		"gemma.block_count",
@@ -577,8 +573,6 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 					return err
 				}
 			}
-		default:
-			return fmt.Errorf("improper type for '%s'", k)
 		}
 		if err != nil {
 			return err
@@ -600,11 +594,9 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 			return err
 		}
 
-		dims := 0
+		dims := 1
-		for cnt := 0; cnt < len(tensor.Shape); cnt++ {
+		if tensor.Shape[1] > 0 {
-			if tensor.Shape[cnt] > 0 {
+			dims = 2
-				dims++
-			}
 		}
 
 		if err := binary.Write(ws, llm.ByteOrder, uint32(dims)); err != nil {
@@ -631,9 +623,8 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 		return err
 	}
 
-	var alignment int64 = 32
+	padding := llm.padding(offset, 32)
-	padding := llm.padding(offset, alignment)
+	if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding-offset))); err != nil {
-	if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
 		return err
 	}
 
@@ -647,8 +638,8 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 			return err
 		}
 
-		padding := llm.padding(offset, alignment)
+		padding := llm.padding(offset, 32)
-		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
+		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding-offset))); err != nil {
 			return err
 		}
 	}
@@ -657,5 +648,5 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 }
 
 func (gguf) padding(offset, align int64) int64 {
-	return (align - offset%align) % align
+	return (offset + align - 1) / align * align
 }

llm/llm_windows.go

@@ -2,5 +2,5 @@ package llm
 
 import "embed"
 
-// unused on windows
+//go:embed build/windows/*/*/bin/*
 var libEmbed embed.FS

llm/memory.go

@@ -1,188 +0,0 @@
-package llm
-
-import (
-	"fmt"
-	"log/slog"
-	"os"
-	"strconv"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/format"
-	"github.com/ollama/ollama/gpu"
-)
-
-// This algorithm looks for a complete fit to determine if we need to unload other models
-func PredictServerFit(allGpus gpu.GpuInfoList, ggml *GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
-	var estimatedVRAM uint64
-	if opts.NumCtx > int(ggml.KV().ContextLength()) {
-		slog.Warn("requested context length is greater than model max context length", "requested", opts.NumCtx, "model", ggml.KV().ContextLength())
-		opts.NumCtx = int(ggml.KV().ContextLength())
-	}
-
-	if opts.NumCtx < 4 {
-		opts.NumCtx = 4
-	}
-
-	// Split up the GPUs by type and try them
-	for _, gpus := range allGpus.ByLibrary() {
-		var layerCount int
-		layerCount, estimatedVRAM = EstimateGPULayers(gpus, ggml, projectors, opts)
-		if opts.NumGPU < 0 {
-			if layerCount > 0 && layerCount >= int(ggml.KV().BlockCount()+1) {
-				return true, estimatedVRAM
-			}
-		} else {
-			if layerCount > 0 && layerCount >= opts.NumGPU {
-				return true, estimatedVRAM
-			}
-		}
-	}
-	return false, estimatedVRAM
-}
-
-// Given a model and one or more GPU targets, predict how many layers and bytes we can load
-// The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts api.Options) (int, uint64) {
-	if gpus[0].Library == "cpu" {
-		return 0, 0
-	}
-	var memoryAvailable uint64
-	for _, info := range gpus {
-		memoryAvailable += info.FreeMemory
-	}
-	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			slog.Info("user override memory limit", "OLLAMA_MAX_VRAM", avail, "actual", memoryAvailable)
-			memoryAvailable = avail
-		}
-	}
-
-	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", format.HumanBytes2(memoryAvailable))
-
-	// TODO - this is probably wrong, first GPU vs secondaries will have different overheads
-	memoryMinimum := gpus[0].MinimumMemory
-
-	for _, projector := range projectors {
-		memoryMinimum += projectorMemoryRequirements(projector)
-
-		// multimodal models require at least 2048 context
-		opts.NumCtx = max(opts.NumCtx, 2048)
-	}
-
-	// 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()
-
-	graphPartialOffload, graphFullOffload := ggml.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)))
-	if graphPartialOffload == 0 {
-		graphPartialOffload = ggml.KV().GQA() * kv / 6
-	}
-
-	if graphFullOffload == 0 {
-		graphFullOffload = graphPartialOffload
-	}
-
-	graphFullOffload *= uint64(len(gpus))
-	graphPartialOffload *= uint64(len(gpus))
-
-	// memoryRequiredTotal represents the memory required for full GPU offloading (all layers)
-	memoryRequiredTotal := memoryMinimum + graphFullOffload
-
-	// memoryRequiredPartial represents the memory required for partial GPU offloading (n > 0, n < layers)
-	memoryRequiredPartial := memoryMinimum + graphPartialOffload
-
-	if memoryRequiredPartial > memoryAvailable {
-		slog.Debug("insufficient VRAM to load any model layers")
-		return 0, 0
-	}
-
-	layers := ggml.Tensors().Layers()
-
-	var memoryLayerOutput uint64
-	if layer, ok := layers["output_norm"]; ok {
-		memoryLayerOutput += layer.size()
-	}
-
-	if layer, ok := layers["output"]; ok {
-		memoryLayerOutput += layer.size()
-	} else if layer, ok := layers["token_embd"]; ok {
-		memoryLayerOutput += layer.size()
-	}
-
-	if gpus[0].Library == "metal" && opts.UseMMap {
-		// memory is preallocated for output tensors
-		memoryRequiredTotal += memoryLayerOutput
-		memoryRequiredPartial += memoryLayerOutput
-	}
-
-	var layerCount int
-	for i := 0; i < int(ggml.KV().BlockCount()); i++ {
-		memoryLayer := layers[fmt.Sprintf("blk.%d", i)].size()
-
-		// KV is proportional to the number of layers
-		memoryLayer += kv / ggml.KV().BlockCount()
-
-		memoryRequiredTotal += memoryLayer
-		if memoryAvailable > memoryRequiredPartial+memoryLayer {
-			memoryRequiredPartial += memoryLayer
-			layerCount++
-		}
-	}
-
-	if gpus[0].Library != "metal" || !opts.UseMMap {
-		// memory was not preallocated for output tensors
-		memoryRequiredTotal += memoryLayerOutput
-	}
-
-	if memoryAvailable > memoryRequiredTotal {
-		layerCount = int(ggml.KV().BlockCount()) + 1
-		memoryRequiredPartial = memoryRequiredTotal
-	}
-
-	memoryWeights := memoryRequiredTotal - memoryMinimum - graphFullOffload - kv
-
-	slog.Info(
-		"offload to gpu",
-		slog.Group(
-			"layers",
-			// actual number of layers offloaded
-			"real", opts.NumGPU,
-			// estimated number of layers that can be offloaded
-			"estimate", layerCount,
-		),
-		slog.Group(
-			"memory",
-			// memory available for offloading
-			"available", format.HumanBytes2(memoryAvailable),
-			slog.Group(
-				"required",
-				// memory required for full offloading
-				"full", format.HumanBytes2(memoryRequiredTotal),
-				// memory required to offload layers.estimate layers
-				"partial", format.HumanBytes2(memoryRequiredPartial),
-				// memory of KV cache
-				"kv", format.HumanBytes2(kv),
-			),
-			slog.Group(
-				"weights",
-				// memory of the weights
-				"total", format.HumanBytes2(memoryWeights),
-				// memory of repeating layers
-				"repeating", format.HumanBytes2(memoryWeights-memoryLayerOutput),
-				// memory of non-repeating layers
-				"nonrepeating", format.HumanBytes2(memoryLayerOutput),
-			),
-			slog.Group(
-				"graph",
-				// memory of graph when fully offloaded
-				"full", format.HumanBytes2(graphFullOffload),
-				// memory of graph when not fully offloaded
-				"partial", format.HumanBytes2(graphPartialOffload),
-			),
-		),
-	)
-	return layerCount, uint64(memoryRequiredPartial)
-}

llm/patches/02-clip-log.diff

@@ -1,12 +0,0 @@
-diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index e431c7f7..f077e688 100644
---- a/examples/llava/clip.cpp
-+++ b/examples/llava/clip.cpp
-@@ -3,6 +3,7 @@
- // I'll gradually clean and extend it
- // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
- #include "clip.h"
-+#include "common.h"
- #include "log.h"
- #include "ggml.h"
- #include "ggml-alloc.h"

llm/patches/04-metal.diff

@@ -1,45 +0,0 @@
-diff --git a/ggml-metal.m b/ggml-metal.m
-index 0207b787..b5e9884b 100644
---- a/ggml-metal.m
-+++ b/ggml-metal.m
-@@ -1396,27 +1396,23 @@ static enum ggml_status ggml_metal_graph_compute(
-                         // to the matrix-vector kernel
-                         int ne11_mm_min = 1;
- 
--#if 0
-                         // the numbers below are measured on M2 Ultra for 7B and 13B models
-                         // these numbers do not translate to other devices or model sizes
-                         // TODO: need to find a better approach
--                        if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
--                            switch (src0t) {
--                                case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
--                                case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
--                                case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
--                                case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
--                                case GGML_TYPE_Q4_0:
--                                case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
--                                case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
--                                case GGML_TYPE_Q5_0:                          // not tested yet
--                                case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
--                                case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
--                                case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
--                                default:             ne11_mm_min = 1;  break;
--                            }
-+                        switch (src0t) {
-+                            case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
-+                            case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
-+                            case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
-+                            case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
-+                            case GGML_TYPE_Q4_0:
-+                            case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
-+                            case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
-+                            case GGML_TYPE_Q5_0:                          // not tested yet
-+                            case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
-+                            case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
-+                            case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
-+                            default:             ne11_mm_min = 1;  break;
-                         }
--#endif
- 
-                         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel

llm/payload.go

@@ -9,7 +9,6 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
-	"runtime"
 	"strings"
 
 	"golang.org/x/exp/slices"
@@ -18,7 +17,7 @@ import (
 	"github.com/ollama/ollama/gpu"
 )
 
-var errPayloadMissing = errors.New("expected payloads not included in this build of ollama")
+var errPayloadMissing = fmt.Errorf("expected payloads not included in this build of ollama")
 
 func Init() error {
 	payloadsDir, err := gpu.PayloadsDir()
@@ -26,15 +25,13 @@ func Init() error {
 		return err
 	}
 
-	if runtime.GOOS != "windows" {
+	slog.Info("extracting embedded files", "dir", payloadsDir)
-		slog.Info("extracting embedded files", "dir", payloadsDir)
+	binGlob := "build/*/*/*/bin/*"
-		binGlob := "build/*/*/*/bin/*"
 
-		// extract server libraries
+	// extract server libraries
-		err = extractFiles(payloadsDir, binGlob)
+	err = extractFiles(payloadsDir, binGlob)
-		if err != nil {
+	if err != nil {
-			return fmt.Errorf("extract binaries: %v", err)
+		return fmt.Errorf("extract binaries: %v", err)
-		}
 	}
 
 	var variants []string
@@ -141,23 +138,6 @@ func serversForGpu(info gpu.GpuInfo) []string {
 	return servers
 }
 
-// Return the optimal server for this CPU architecture
-func serverForCpu() string {
-	if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
-		return "metal"
-	}
-	variant := gpu.GetCPUVariant()
-	availableServers := availableServers()
-	if variant != "" {
-		for cmp := range availableServers {
-			if cmp == "cpu_"+variant {
-				return cmp
-			}
-		}
-	}
-	return "cpu"
-}
-
 // extract extracts the embedded files to the target directory
 func extractFiles(targetDir string, glob string) error {
 	files, err := fs.Glob(libEmbed, glob)

llm/server.go

@@ -21,43 +21,21 @@ import (
 	"strings"
 	"time"
 
-	"golang.org/x/sync/semaphore"
-
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
 )
 
-type LlamaServer interface {
+// LlamaServer is an instance of the llama.cpp server
-	Ping(ctx context.Context) error
+type LlamaServer struct {
-	WaitUntilRunning(ctx context.Context) error
-	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embedding(ctx context.Context, prompt string) ([]float64, error)
-	Tokenize(ctx context.Context, content string) ([]int, error)
-	Detokenize(ctx context.Context, tokens []int) (string, error)
-	Close() error
-	EstimatedVRAM() uint64
-}
-
-// llmServer is an instance of the llama.cpp server
-type llmServer struct {
 	port    int
 	cmd     *exec.Cmd
 	done    chan error // Channel to signal when the process exits
 	status  *StatusWriter
 	options api.Options
-
-	// TODO - this should be broken down by GPU
-	estimatedVRAM uint64 // Estimated usage of VRAM by the loaded model
-
-	sem *semaphore.Weighted
 }
 
-func LoadModel(model string) (*GGML, error) {
+func NewLlamaServer(model string, adapters, projectors []string, opts api.Options) (*LlamaServer, error) {
-	if _, err := os.Stat(model); err != nil {
-		return nil, err
-	}
-
 	f, err := os.Open(model)
 	if err != nil {
 		return nil, err
@@ -65,13 +43,10 @@ func LoadModel(model string) (*GGML, error) {
 	defer f.Close()
 
 	ggml, _, err := DecodeGGML(f)
-	return ggml, err
+	if err != nil {
-}
+		return nil, err
+	}
 
-// NewLlamaServer will run a server for the given GPUs
-// The gpu list must be a single family.
-func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options) (LlamaServer, error) {
-	var err error
 	if opts.NumCtx > int(ggml.KV().ContextLength()) {
 		slog.Warn("requested context length is greater than model max context length", "requested", opts.NumCtx, "model", ggml.KV().ContextLength())
 		opts.NumCtx = int(ggml.KV().ContextLength())
@@ -81,51 +56,91 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		opts.NumCtx = 4
 	}
 
-	cpuRunner := ""
+	memoryAvailable, _ := gpu.CheckVRAM()
-	var estimatedVRAM uint64
+	info := gpu.GetGPUInfo()
-	var systemMemory uint64
-	if (len(gpus) == 1 && gpus[0].Library == "cpu") || opts.NumGPU == 0 {
 
-		// TODO evaluate system memory to see if we should block the load, or force an unload of another CPU runner
+	memoryMinimum := info.MinimumMemory
+	for _, projector := range projectors {
+		memoryMinimum += projectorMemoryRequirements(projector)
 
-		cpuRunner = serverForCpu()
+		// multimodal models require at least 2048 context
-	} else {
+		opts.NumCtx = max(opts.NumCtx, 2048)
-		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))
-			}
-		}
-		var layers int
-		layers, estimatedVRAM = EstimateGPULayers(gpus, ggml, projectors, opts)
 
-		if gpus[0].Library == "metal" && estimatedVRAM > systemMemory {
+	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
-			// disable partial offloading when model is greater than total system memory as this
+	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()
-			// can lead to locking up the system
+
-			opts.NumGPU = 0
+	graphPartialOffload, graphFullOffload := ggml.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)))
-		} else if opts.NumGPU < 0 && layers > 0 && gpus[0].Library != "cpu" {
+	if graphPartialOffload == 0 {
-			opts.NumGPU = layers
+		graphPartialOffload = ggml.KV().GQA() * kv / 6
+	}
+
+	if graphFullOffload == 0 {
+		graphFullOffload = graphPartialOffload
+	}
+
+	// memoryRequiredTotal represents the memory required for full GPU offloading (all layers)
+	memoryRequiredTotal := memoryMinimum + graphFullOffload
+
+	// memoryRequiredPartial represents the memory required for partial GPU offloading (n > 0, n < layers)
+	memoryRequiredPartial := memoryMinimum + graphPartialOffload
+
+	if info.Library != "metal" {
+		if memoryRequiredPartial > memoryAvailable {
+			info.Library = "cpu"
 		}
 	}
 
-	// Loop through potential servers
+	var layerCount int
-	finalErr := fmt.Errorf("no suitable llama servers found")
+	layers := ggml.Tensors().Layers()
+	for i := 0; i < int(ggml.KV().BlockCount()); i++ {
+		memoryLayer := layers[fmt.Sprintf("%d", i)].size()
+
+		// KV is proportional to the number of layers
+		memoryLayer += kv / ggml.KV().BlockCount()
+
+		memoryRequiredTotal += memoryLayer
+		if memoryAvailable > memoryRequiredPartial+memoryLayer {
+			memoryRequiredPartial += memoryLayer
+			layerCount++
+		}
+	}
+
+	memoryLayerOutput := layers["output"].size()
+	memoryRequiredTotal += memoryLayerOutput
+
+	if info.Library == "metal" && memoryRequiredTotal > info.TotalMemory {
+		// disable partial offloading when model is greater than total system memory
+		opts.NumGPU = 0
+	} else if memoryAvailable > memoryRequiredTotal {
+		layerCount = int(ggml.KV().BlockCount()) + 1
+		memoryRequiredPartial = memoryRequiredTotal
+	}
+
+	if opts.NumGPU < 0 {
+		opts.NumGPU = layerCount
+	}
+
+	slog.Info(
+		"offload to gpu",
+		"reallayers", opts.NumGPU,
+		"layers", layerCount,
+		"required", format.HumanBytes2(memoryRequiredTotal),
+		"used", format.HumanBytes2(memoryRequiredPartial),
+		"available", format.HumanBytes2(memoryAvailable),
+		"kv", format.HumanBytes2(kv),
+		"fulloffload", format.HumanBytes2(graphFullOffload),
+		"partialoffload", format.HumanBytes2(graphPartialOffload),
+	)
 
 	if len(adapters) > 1 {
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}
 
 	availableServers := availableServers()
-	var servers []string
+	servers := serversForGpu(info)
-	if cpuRunner != "" {
+
-		servers = []string{cpuRunner}
+	demandLib := os.Getenv("OLLAMA_LLM_LIBRARY")
-	} else {
-		servers = serversForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
-	}
-	demandLib := strings.Trim(os.Getenv("OLLAMA_LLM_LIBRARY"), "\"' ")
 	if demandLib != "" {
 		serverPath := availableServers[demandLib]
 		if serverPath == "" {
@@ -137,7 +152,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	}
 
 	if len(servers) == 0 {
-		return nil, fmt.Errorf("no servers found for %v", gpus)
+		return nil, fmt.Errorf("no servers found for %v", info)
 	}
 
 	params := []string{
@@ -194,26 +209,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--numa")
 	}
 
-	// "--cont-batching", // TODO - doesn't seem to have any noticeable perf change for multiple requests
+	// Loop through potential servers
-	numParallel := 1
+	var finalErr error
-	if onp := os.Getenv("OLLAMA_NUM_PARALLEL"); onp != "" {
-		numParallel, err = strconv.Atoi(onp)
-		if err != nil || numParallel <= 0 {
-			err = fmt.Errorf("invalid OLLAMA_NUM_PARALLEL=%s must be greater than zero - %w", onp, err)
-			slog.Error("misconfiguration", "error", err)
-			return nil, err
-		}
-	}
-	params = append(params, "--parallel", fmt.Sprintf("%d", numParallel))
-
 	for i := 0; i < len(servers); i++ {
 		dir := availableServers[servers[i]]
-		if dir == "" {
-			// Shouldn't happen
-			finalErr = fmt.Errorf("[%d] server %s not listed in available servers %v", i, servers[i], availableServers)
-			slog.Error("sever list inconsistent", "error", finalErr)
-			continue
-		}
 
 		// Find an availableServers  port, retry on each iterration in case the failure was a port conflict race
 		port := 0
@@ -236,60 +235,30 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 		// append 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
 			// This will favor system libraries over our bundled library dependencies
 			libraryPaths = append(filepath.SplitList(libraryPath), libraryPaths...)
 		}
 
-		// Note: we always put the dependency path first
-		// since this was the exact version we verified for AMD GPUs
-		// and we favor what the user had in their path
-		if gpus[0].DependencyPath != "" {
-			// TODO refine for multi-gpu support
-			libraryPaths = append([]string{gpus[0].DependencyPath}, libraryPaths...)
-		}
-
 		server := filepath.Join(dir, "ollama_llama_server")
 		if runtime.GOOS == "windows" {
 			server = server + ".exe"
 		}
 
-		// Detect tmp cleaners wiping out the file
+		s := &LlamaServer{
-		_, err := os.Stat(server)
+			port:    port,
-		if errors.Is(err, os.ErrNotExist) {
+			cmd:     exec.Command(server, finalParams...),
-			slog.Warn("llama server disappeared, reinitializing payloads", "path", server, "error", err)
+			status:  NewStatusWriter(os.Stderr),
-			err = Init()
+			options: opts,
-			if err != nil {
-				slog.Warn("failed to reinitialize payloads", "error", err)
-				return nil, err
-			}
 		}
-
-		s := &llmServer{
-			port:          port,
-			cmd:           exec.Command(server, finalParams...),
-			status:        NewStatusWriter(os.Stderr),
-			options:       opts,
-			estimatedVRAM: estimatedVRAM,
-			sem:           semaphore.NewWeighted(int64(numParallel)),
-		}
-
 		libEnv := fmt.Sprintf("%s=%s", pathEnv, strings.Join(libraryPaths, string(filepath.ListSeparator)))
+		slog.Debug(libEnv)
 		s.cmd.Env = append(os.Environ(), libEnv)
 		s.cmd.Stdout = os.Stdout
 		s.cmd.Stderr = s.status
 
-		// TODO - multiple GPU selection logic...
-		key, val := gpu.GpuInfoList(gpus).GetVisibleDevicesEnv()
-		if key != "" {
-			s.cmd.Env = append(s.cmd.Env, key+"="+val)
-		}
-
 		slog.Info("starting llama server", "cmd", s.cmd.String())
-		// Log at debug as the environment is inherited and might contain sensitive information
-		slog.Debug("subprocess", "environment", s.cmd.Env)
 
 		if err = s.cmd.Start(); err != nil {
 			msg := ""
@@ -307,13 +276,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			_ = s.cmd.Wait()
 		}()
 
-		// TODO - make sure this is all wired up correctly
-		// if err = s.WaitUntilRunning(); err != nil {
-		// 	slog.Error("error starting llama server", "server", servers[i], "error", err)
-		// 	s.Close()
-		// 	finalErr = err
-		// 	continue
-		// }
 		return s, nil
 	}
 
@@ -351,21 +313,6 @@ const ( // iota is reset to 0
 	ServerStatusError
 )
 
-func (s ServerStatus) ToString() string {
-	switch s {
-	case ServerStatusReady:
-		return "llm server ready"
-	case ServerStatusNoSlotsAvaialble:
-		return "llm busy - no slots available"
-	case ServerStatusLoadingModel:
-		return "llm server loading model"
-	case ServerStatusNotResponding:
-		return "llm server not responding"
-	default:
-		return "llm server error"
-	}
-}
-
 type ServerStatusResp struct {
 	Status          string `json:"status"`
 	SlotsIdle       int    `json:"slots_idle"`
@@ -373,7 +320,7 @@ type ServerStatusResp struct {
 	Error           string `json:"error"`
 }
 
-func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
+func (s *LlamaServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 	// Fail fast if its exited
 	if s.cmd.ProcessState != nil {
 		msg := ""
@@ -420,7 +367,7 @@ func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 	}
 }
 
-func (s *llmServer) Ping(ctx context.Context) error {
+func (s *LlamaServer) Ping(ctx context.Context) error {
 	_, err := s.getServerStatus(ctx)
 	if err != nil {
 		slog.Debug("server unhealthy", "error", err)
@@ -429,7 +376,7 @@ func (s *llmServer) Ping(ctx context.Context) error {
 	return nil
 }
 
-func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
+func (s *LlamaServer) WaitUntilRunning() error {
 	start := time.Now()
 	// TODO we need to wire up a better way to detect hangs during model load and startup of the server
 	expiresAt := time.Now().Add(10 * time.Minute) // be generous with timeout, large models can take a while to load
@@ -440,9 +387,6 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 	var lastStatus ServerStatus = -1
 	for {
 		select {
-		case <-ctx.Done():
-			slog.Info("context expired before server started")
-			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
 		case err := <-s.done:
 			msg := ""
 			if s.status != nil && s.status.LastErrMsg != "" {
@@ -466,9 +410,9 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 				return fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
 			}
 
-			c, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
+			ctx, cancel := context.WithTimeout(context.Background(), 200*time.Millisecond)
 			defer cancel()
-			status, err := s.getServerStatus(c)
+			status, err := s.getServerStatus(ctx)
 			if err != nil && lastStatus != status {
 				slog.Debug("server not yet available", "error", err)
 				lastStatus = status
@@ -554,19 +498,7 @@ type CompletionResponse struct {
 	EvalDuration       time.Duration
 }
 
-func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
+func (s *LlamaServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
-	if err := s.sem.Acquire(ctx, 1); err != nil {
-		slog.Error("Failed to acquire semaphore", "error", err)
-		return err
-	}
-	defer s.sem.Release(1)
-
-	// only allow maximum 10 "context shifts" to avoid infinite generation
-	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
-		req.Options.NumPredict = 10 * s.options.NumCtx
-		slog.Debug("setting token limit to 10x num_ctx", "num_ctx", s.options.NumCtx, "num_predict", req.Options.NumPredict)
-	}
-
 	request := map[string]any{
 		"prompt":            req.Prompt,
 		"stream":            true,
@@ -597,7 +529,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	if err != nil {
 		return err
 	} else if status != ServerStatusReady {
-		return fmt.Errorf("unexpected server status: %s", status.ToString())
+		return fmt.Errorf("unexpected server status: %d", status)
 	}
 
 	if req.Format == "json" {
@@ -744,18 +676,13 @@ type EmbeddingResponse struct {
 	Embedding []float64 `json:"embedding"`
 }
 
-func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, error) {
+func (s *LlamaServer) Embedding(ctx context.Context, prompt string) ([]float64, error) {
-	if err := s.sem.Acquire(ctx, 1); err != nil {
-		slog.Error("Failed to acquire semaphore", "error", err)
-		return nil, err
-	}
-	defer s.sem.Release(1)
 	// Make sure the server is ready
 	status, err := s.getServerStatus(ctx)
 	if err != nil {
 		return nil, err
 	} else if status != ServerStatusReady {
-		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
+		return nil, fmt.Errorf("unexpected server status: %d", status)
 	}
 
 	data, err := json.Marshal(TokenizeRequest{Content: prompt})
@@ -801,13 +728,13 @@ type TokenizeResponse struct {
 	Tokens []int `json:"tokens"`
 }
 
-func (s *llmServer) Tokenize(ctx context.Context, content string) ([]int, error) {
+func (s *LlamaServer) Tokenize(ctx context.Context, content string) ([]int, error) {
 	// Make sure the server is ready
 	status, err := s.getServerStatus(ctx)
 	if err != nil {
 		return nil, err
-	} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvaialble {
+	} else if status != ServerStatusReady {
-		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
+		return nil, fmt.Errorf("unexpected server status: %d", status)
 	}
 
 	data, err := json.Marshal(TokenizeRequest{Content: content})
@@ -853,13 +780,13 @@ type DetokenizeResponse struct {
 	Content string `json:"content"`
 }
 
-func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error) {
+func (s *LlamaServer) Detokenize(ctx context.Context, tokens []int) (string, error) {
 	// Make sure the server is ready
 	status, err := s.getServerStatus(ctx)
 	if err != nil {
 		return "", err
-	} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvaialble {
+	} else if status != ServerStatusReady {
-		return "", fmt.Errorf("unexpected server status: %s", status.ToString())
+		return "", fmt.Errorf("unexpected server status: %d", status)
 	}
 
 	data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
@@ -897,7 +824,7 @@ func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error
 	return decoded.Content, nil
 }
 
-func (s *llmServer) Close() error {
+func (s *LlamaServer) Close() error {
 	if s.cmd != nil {
 		slog.Debug("stopping llama server")
 		return s.cmd.Process.Kill()
@@ -906,10 +833,6 @@ func (s *llmServer) Close() error {
 	return nil
 }
 
-func (s *llmServer) EstimatedVRAM() uint64 {
-	return s.estimatedVRAM
-}
-
 func parseDurationMs(ms float64) time.Duration {
 	dur, err := time.ParseDuration(fmt.Sprintf("%fms", ms))
 	if err != nil {

macapp/.eslintrc.json

@@ -0,0 +1,16 @@
+{
+  "env": {
+    "browser": true,
+    "es6": true,
+    "node": true
+  },
+  "extends": [
+    "eslint:recommended",
+    "plugin:@typescript-eslint/eslint-recommended",
+    "plugin:@typescript-eslint/recommended",
+    "plugin:import/recommended",
+    "plugin:import/electron",
+    "plugin:import/typescript"
+  ],
+  "parser": "@typescript-eslint/parser"
+}

macapp/.gitignore

@@ -0,0 +1,92 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+lerna-debug.log*
+
+# Diagnostic reports (https://nodejs.org/api/report.html)
+report.[0-9]*.[0-9]*.[0-9]*.[0-9]*.json
+
+# Runtime data
+pids
+*.pid
+*.seed
+*.pid.lock
+.DS_Store
+
+# Directory for instrumented libs generated by jscoverage/JSCover
+lib-cov
+
+# Coverage directory used by tools like istanbul
+coverage
+*.lcov
+
+# nyc test coverage
+.nyc_output
+
+# node-waf configuration
+.lock-wscript
+
+# Compiled binary addons (https://nodejs.org/api/addons.html)
+build/Release
+
+# Dependency directories
+node_modules/
+jspm_packages/
+
+# TypeScript v1 declaration files
+typings/
+
+# TypeScript cache
+*.tsbuildinfo
+
+# Optional npm cache directory
+.npm
+
+# Optional eslint cache
+.eslintcache
+
+# Optional REPL history
+.node_repl_history
+
+# Output of 'npm pack'
+*.tgz
+
+# Yarn Integrity file
+.yarn-integrity
+
+# dotenv environment variables file
+.env
+.env.test
+
+# parcel-bundler cache (https://parceljs.org/)
+.cache
+
+# next.js build output
+.next
+
+# nuxt.js build output
+.nuxt
+
+# vuepress build output
+.vuepress/dist
+
+# Serverless directories
+.serverless/
+
+# FuseBox cache
+.fusebox/
+
+# DynamoDB Local files
+.dynamodb/
+
+# Webpack
+.webpack/
+
+# Vite
+.vite/
+
+# Electron-Forge
+out/

macapp/README.md

@@ -0,0 +1,21 @@
+# Desktop
+
+This app builds upon Ollama to provide a desktop experience for running models.
+
+## Developing
+
+First, build the `ollama` binary:
+
+```
+cd ..
+go build .
+```
+
+Then run the desktop app with `npm start`:
+
+```
+cd macapp
+npm install
+npm start
+```
+

macapp/forge.config.ts

@@ -0,0 +1,78 @@
+import type { ForgeConfig } from '@electron-forge/shared-types'
+import { MakerSquirrel } from '@electron-forge/maker-squirrel'
+import { MakerZIP } from '@electron-forge/maker-zip'
+import { PublisherGithub } from '@electron-forge/publisher-github'
+import { AutoUnpackNativesPlugin } from '@electron-forge/plugin-auto-unpack-natives'
+import { WebpackPlugin } from '@electron-forge/plugin-webpack'
+import * as path from 'path'
+import * as fs from 'fs'
+
+import { mainConfig } from './webpack.main.config'
+import { rendererConfig } from './webpack.renderer.config'
+
+const packageJson = JSON.parse(fs.readFileSync(path.resolve(__dirname, './package.json'), 'utf8'))
+
+const config: ForgeConfig = {
+  packagerConfig: {
+    appVersion: process.env.VERSION || packageJson.version,
+    asar: true,
+    icon: './assets/icon.icns',
+    extraResource: [
+      '../dist/ollama',
+      path.join(__dirname, './assets/iconTemplate.png'),
+      path.join(__dirname, './assets/iconTemplate@2x.png'),
+      path.join(__dirname, './assets/iconUpdateTemplate.png'),
+      path.join(__dirname, './assets/iconUpdateTemplate@2x.png'),
+      path.join(__dirname, './assets/iconDarkTemplate.png'),
+      path.join(__dirname, './assets/iconDarkTemplate@2x.png'),
+      path.join(__dirname, './assets/iconDarkUpdateTemplate.png'),
+      path.join(__dirname, './assets/iconDarkUpdateTemplate@2x.png'),
+    ],
+    ...(process.env.SIGN
+      ? {
+          osxSign: {
+            identity: process.env.APPLE_IDENTITY,
+          },
+          osxNotarize: {
+            tool: 'notarytool',
+            appleId: process.env.APPLE_ID || '',
+            appleIdPassword: process.env.APPLE_PASSWORD || '',
+            teamId: process.env.APPLE_TEAM_ID || '',
+          },
+        }
+      : {}),
+    osxUniversal: {
+      x64ArchFiles: '**/ollama',
+    },
+  },
+  rebuildConfig: {},
+  makers: [new MakerSquirrel({}), new MakerZIP({}, ['darwin'])],
+  hooks: {
+    readPackageJson: async (_, packageJson) => {
+      return { ...packageJson, version: process.env.VERSION || packageJson.version }
+    },
+  },
+  plugins: [
+    new AutoUnpackNativesPlugin({}),
+    new WebpackPlugin({
+      mainConfig,
+      devContentSecurityPolicy: `default-src * 'unsafe-eval' 'unsafe-inline'; img-src data: 'self'`,
+      renderer: {
+        config: rendererConfig,
+        nodeIntegration: true,
+        entryPoints: [
+          {
+            html: './src/index.html',
+            js: './src/renderer.tsx',
+            name: 'main_window',
+            preload: {
+              js: './src/preload.ts',
+            },
+          },
+        ],
+      },
+    }),
+  ],
+}
+
+export default config