Revamp ROCm support

This refines where we extract the LLM libraries to by adding a new OLLAMA_HOME env var, that defaults to `~/.ollama` The logic was already idempotenent, so this should speed up startups after the first time a new release is deployed. It also cleans up after itself. We now build only a single ROCm version (latest major) on both windows and linux. Given the large size of ROCms tensor files, we split the dependency out. It's bundled into the installer on windows, and a separate download on windows. The linux install script is now smart and detects the presence of AMD GPUs and looks to see if rocm v6 is already present, and if not, then downloads our dependency tar file. For Linux discovery, we now use sysfs and check each GPU against what ROCm supports so we can degrade to CPU gracefully instead of having llama.cpp+rocm assert/crash on us. For Windows, we now use go's windows dynamic library loading logic to access the amdhip64.dll APIs to query the GPU information.
wgong · Mar 7, 2024 · 6c5ccb1 · 6c5ccb1
1 parent 2e20110
commit 6c5ccb1
Show file tree

Hide file tree

Showing 27 changed files with 1,091 additions and 588 deletions.
diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml
@@ -7,12 +7,12 @@ jobs:
   generate:
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-latest, macos-latest, windows-2019]
         arch: [amd64, arm64]
         exclude:
           - os: ubuntu-latest
             arch: arm64
-          - os: windows-latest
+          - os: windows-2019
             arch: arm64
     runs-on: ${{ matrix.os }}
     env:
@@ -24,7 +24,18 @@ jobs:
           go-version: '1.22'
           cache: true
       - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$env:PATH"
+          go generate -x ./...
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        name: "Windows Go Generate"
       - run: go generate -x ./...
+        if: ${{ ! startsWith(matrix.os, 'windows-') }}
+        name: "Unix Go Generate"
       - uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
@@ -62,7 +73,6 @@ jobs:
     strategy:
       matrix:
         rocm-version:
-          - '5.7.1'
           - '6.0'
     runs-on: linux
     container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
@@ -91,12 +101,12 @@ jobs:
   lint:
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-latest, macos-latest, windows-2019]
         arch: [amd64, arm64]
         exclude:
           - os: ubuntu-latest
             arch: arm64
-          - os: windows-latest
+          - os: windows-2019
             arch: arm64
           - os: macos-latest
             arch: amd64
@@ -130,12 +140,12 @@ jobs:
     needs: generate
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-latest, macos-latest, windows-2019]
         arch: [amd64]
         exclude:
           - os: ubuntu-latest
             arch: arm64
-          - os: windows-latest
+          - os: windows-2019
             arch: arm64
     runs-on: ${{ matrix.os }}
     env:

diff --git a/Dockerfile b/Dockerfile
@@ -1,6 +1,7 @@
 ARG GOLANG_VERSION=1.22.1
 ARG CMAKE_VERSION=3.22.1
 ARG CUDA_VERSION=11.3.1
+ARG ROCM_VERSION=6.0
 
 # Copy the minimal context we need to run the generate scripts
 FROM scratch AS llm-code
@@ -28,7 +29,7 @@ WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
 ARG CGO_CFLAGS
 RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 
-FROM --platform=linux/amd64 rocm/dev-centos-7:5.7.1-complete AS rocm-5-build-amd64
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
@@ -39,18 +40,14 @@ WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
 ARG CGO_CFLAGS
 ARG AMDGPU_TARGETS
 RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN mkdir /tmp/scratch && \
+    for dep in $(cat /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/x86_64/rocm*/lib/deps.txt) ; do \
+        cp ${dep} /tmp/scratch/ || exit 1 ; \
+    done && \
+    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd /tmp/scratch/ && tar xf - ) && \
+    mkdir -p /go/src/github.com/jmorganca/ollama/dist/deps/ && \
+    (cd /tmp/scratch/ && tar czvf /go/src/github.com/jmorganca/ollama/dist/deps/rocm-amd64-deps.tgz . )
 
-FROM --platform=linux/amd64 rocm/dev-centos-7:6.0-complete AS rocm-6-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV LIBRARY_PATH /opt/amdgpu/lib64
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG AMDGPU_TARGETS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 
 FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
 ARG CMAKE_VERSION
@@ -91,8 +88,8 @@ COPY . .
 COPY --from=cpu_avx-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
 COPY --from=cpu_avx2-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
 COPY --from=cuda-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=rocm-5-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=rocm-6-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/jmorganca/ollama/dist/deps/ ./dist/deps/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN go build .
@@ -117,7 +114,7 @@ RUN apt-get update && apt-get install -y ca-certificates
 COPY --from=build-arm64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
 
 # Radeon images are much larger so we keep it distinct from the CPU/CUDA image
-FROM --platform=linux/amd64 rocm/dev-centos-7:5.7.1-complete as runtime-rocm
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
 RUN update-pciids
 COPY --from=build-amd64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
 EXPOSE 11434

diff --git a/app/ollama.iss b/app/ollama.iss
@@ -91,6 +91,14 @@ Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
 Source: "..\dist\windeps\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
+; Assumes v5.7, may need adjustments for v6
+#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
+
 
 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"

diff --git a/docs/development.md b/docs/development.md
@@ -116,7 +116,7 @@ Note: The windows build for Ollama is still under development.
 
 Install required tools:
 
-- MSVC toolchain - C/C++ and cmake as minimal requirements
+- MSVC toolchain - C/C++ and cmake as minimal requirements - You must build from a "Developer Shell" with the environment variables set
 - go version 1.22 or higher
 - MinGW (pick one variant) with GCC.
   - <https://www.mingw-w64.org/>
@@ -132,6 +132,6 @@ go build .
 
 #### Windows CUDA (NVIDIA)
 
-In addition to the common Windows development tools described above, install:
+In addition to the common Windows development tools described above, install CUDA **AFTER** you install MSVC.
 
 - [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)
diff --git a/docs/linux.md b/docs/linux.md
@@ -10,6 +10,14 @@ Install Ollama running this one-liner:
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
+## AMD Radeon GPU support
+
+While AMD has contributed the `amdgpu` driver upstream to the official linux
+kernel source, the version is older and may not support all ROCm features. We
+recommend you install the latest driver from
+https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+GPU.
+
 ## Manual install
 
 ### Download the `ollama` binary

diff --git a/docs/troubleshooting.md b/docs/troubleshooting.md
@@ -67,6 +67,43 @@ You can see what features your CPU has with the following.
 cat /proc/cpuinfo| grep flags  | head -1
 ```
 
+## AMD Radeon GPU Support
+
+Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
+some cases you can force the system to try to use a close GPU type.  For example
+The Radeon RX 5400 is `gfx1034` (also known as 10.3.4) however, ROCm does not
+support this patch-level, the closest support is `gfx1030`.  You can use the
+environment variable `HSA_OVERRIDE_GFX_VERSION` with `x.y.z` syntax.  So for
+example, to force the system to run on the RX 5400, you would set
+`HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the server.
+
+At this time, the known supported GPU types are the following: (This may change from
+release to release)
+- gfx900
+- gfx906
+- gfx908
+- gfx90a
+- gfx940
+- gfx941
+- gfx942
+- gfx1030
+- gfx1100
+- gfx1101
+- gfx1102
+
+This will not work for all unsupported GPUs.  Reach out on [Discord](https://discord.gg/ollama)
+or file an [issue](https://github.com/ollama/ollama/issues) for additional help.
+
+
+## Installing older versions on Linux
+
+If you run into problems on Linux and want to install an older version you can tell the install script
+which version to install.
+
+```sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.27" sh
+```
+
 ## Known issues
 
 * N/A
diff --git a/docs/windows.md b/docs/windows.md
@@ -4,7 +4,7 @@ Welcome to the Ollama Windows preview.
 
 No more WSL required!
 
-Ollama now runs as a native Windows application, including NVIDIA GPU support.
+Ollama now runs as a native Windows application, including NVIDIA and AMD Radeon GPU support.
 After installing Ollama Windows Preview, Ollama will run in the background and
 the `ollama` command line is available in `cmd`, `powershell` or your favorite
 terminal application. As usual the Ollama [api](./api.md) will be served on
@@ -21,6 +21,7 @@ Logs will often be helpful in dianosing the problem (see
 
 * Windows 10 or newer, Home or Pro
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
+* AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
 
 ## API Access
 

diff --git a/gpu/amd.go b/gpu/amd.go