feat: add AMD ROCm GPU support (#30)

* feat: initial amd gpu support with rocm 7.1

* formatting

* fix: resolve bugs and clean up ROCm support from PR #28

Cherry-picked pyqlsa's ROCm implementation (commits 7e6f796, e2343fe)
and applied fixes on top:

- Fix ROCm app incorrectly named "cuda" in nix/apps.nix (would shadow CUDA)
- Fix mkPython passing boolean `false` instead of string `"none"` for gpuSupport
- Fix flake description still referencing v0.12.2 instead of v0.14.2
- Fix README Podman ROCm example referencing latest-cuda instead of latest-rocm
- Fix ROCm torchaudio missing FFmpeg/sox ignore deps (matching CUDA pattern)
- Resolve all TODO/XXX review comments left by pyqlsa
- Update CHANGELOG, CLAUDE.md, and README with ROCm documentation
- Fix CHANGELOG footer links (missing v0.14.2 link, Unreleased pointing to v0.12.2)

Closes #27
Co-authored-by: pyqlsa <26353308+pyqlsa@users.noreply.github.com>

---------

Co-authored-by: pyqlsa <26353308+pyqlsa@users.noreply.github.com>

Author: jamesbrink

.github/workflows/build.yml

@@ -105,6 +105,13 @@ jobs:
             platform: linux/amd64
             tag_suffix: '-cuda'
             runner: ubuntu-latest
+          - variant: rocm
+            arch: x86_64-linux
+            nix_target: dockerImageRocm
+            package_target: rocm 
+            platform: linux/amd64
+            tag_suffix: '-rocm'
+            runner: ubuntu-latest
 
     steps:
       - name: Checkout repository
@@ -160,8 +167,8 @@ jobs:
           sudo rm -rf /var/lib/apt/lists/*
           sudo apt-get clean
           docker system prune -af || true
-          # Extra cleanup for CUDA builds (needs maximum disk space)
-          if [[ "${{ matrix.variant }}" == "cuda" ]]; then
+          # Extra cleanup for CUDA/ROCm builds (needs maximum disk space)
+          if [[ "${{ matrix.variant }}" =~ ^cuda$|^rocm$ ]]; then
             sudo rm -rf /usr/local/share/powershell
             sudo rm -rf /usr/local/share/chromium
             sudo rm -rf /usr/local/share/vcpkg
@@ -217,8 +224,8 @@ jobs:
       - name: Build Docker image with Nix
         run: |
           echo "Building ${{ matrix.variant }} variant for ${{ matrix.arch }}..."
-          # For CUDA builds, limit parallelism to reduce memory pressure
-          if [[ "${{ matrix.variant }}" == "cuda" ]]; then
+          # For CUDA/ROCm builds, limit parallelism to reduce memory pressure
+          if [[ "${{ matrix.variant }}" =~ ^cuda$|^rocm$ ]]; then
             nix build .#packages.${{ matrix.arch }}.${{ matrix.nix_target }} --print-build-logs --max-jobs 1 --cores 2
           else
             nix build .#packages.${{ matrix.arch }}.${{ matrix.nix_target }} --print-build-logs
@@ -264,8 +271,8 @@ jobs:
           # Build the actual package (not Docker image) for caching
           nix build .#packages.${{ matrix.arch }}.${{ matrix.package_target }} -o result-pkg --print-build-logs
 
-          # For CUDA builds, include build-time dependencies (magma, triton, etc.)
-          if [[ "${{ matrix.variant }}" == "cuda" ]]; then
+          # For CUDA/ROCm builds, include build-time dependencies (magma, triton, etc.)
+          if [[ "${{ matrix.variant }}" =~ ^cuda$|^rocm$ ]]; then
             ./scripts/push-to-cachix.sh --build-deps ./result-pkg
           else
             ./scripts/push-to-cachix.sh ./result-pkg
@@ -367,6 +374,31 @@ jobs:
           docker buildx imagetools create --tag "$CUDA_TAG" "$CUDA_AMD64_TAG"
           docker buildx imagetools create --tag "$VERSION_CUDA_TAG" "$CUDA_AMD64_TAG"
 
+      - name: Create and push ROCm manifest (x86_64 only)
+        run: |
+          if [[ "${{ github.ref }}" == "refs/heads/main" ]]; then
+            ROCM_TAG="${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest-rocm"
+            VERSION_ROCM_TAG="${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ steps.meta.outputs.version }}-rocm"
+            ROCM_AMD64_TAG="${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest-rocm-amd64"
+          elif [[ "${{ github.ref }}" == refs/tags/v* ]]; then
+            TAG_VERSION=${GITHUB_REF#refs/tags/v}
+            ROCM_TAG="${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${TAG_VERSION}-rocm"
+            VERSION_ROCM_TAG="${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest-rocm"
+            ROCM_AMD64_TAG="${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${TAG_VERSION}-rocm-amd64"
+          fi
+
+          echo "Verifying ROCm image exists..."
+          if ! docker manifest inspect "$ROCM_AMD64_TAG" &>/dev/null; then
+            echo "ERROR: Required ROCm image not found: $ROCM_AMD64_TAG"
+            exit 1
+          fi
+          echo "Found: $ROCM_AMD64_TAG"
+
+          # Use buildx imagetools to create tags without pulling the full image
+          echo "Creating ROCm tags via registry (no local pull required)..."
+          docker buildx imagetools create --tag "$ROCM_TAG" "$ROCM_AMD64_TAG"
+          docker buildx imagetools create --tag "$VERSION_ROCM_TAG" "$ROCM_AMD64_TAG"
+
       - name: Generate manifest summary
         run: |
           echo "## Multi-Architecture Docker Images Published" >> $GITHUB_STEP_SUMMARY
@@ -382,6 +414,10 @@ jobs:
           echo '```bash' >> $GITHUB_STEP_SUMMARY
           echo "docker pull ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest-cuda" >> $GITHUB_STEP_SUMMARY
           echo '```' >> $GITHUB_STEP_SUMMARY
+          echo "### ROCM (amd64 only)" >> $GITHUB_STEP_SUMMARY
+          echo '```bash' >> $GITHUB_STEP_SUMMARY
+          echo "docker pull ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest-rocm" >> $GITHUB_STEP_SUMMARY
+          echo '```' >> $GITHUB_STEP_SUMMARY
 
   update-description:
     needs: create-manifests

CHANGELOG.md

@@ -9,8 +9,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [0.14.2] - 2026-02-19
 
+### Added
+- AMD ROCm GPU support via pre-built PyTorch wheels (ROCm 7.1, tested on gfx1100/7900 XTX) (#27)
+- `nix run .#rocm` app, `dockerImageRocm`, and NixOS module `gpuSupport = "rocm"` option
+- ROCm Docker images and CI pipeline (`ghcr.io/utensils/comfyui-nix:latest-rocm`)
+- ROCm dev shell (`nix develop .#rocm`)
+
 ### Changed
 - Upgraded ComfyUI from v0.12.2 to v0.14.2 (5 upstream releases)
+- Replaced `cudaSupport` boolean with `gpuSupport` enum (`"cuda"`, `"rocm"`, `"none"`) across flake, packages, and NixOS module
 - Updated `comfyui-frontend-package` 1.37.11 → 1.38.14
 - Updated `comfyui-workflow-templates` 0.8.31 → 0.8.43
 - Updated `comfyui-workflow-templates-core` 0.3.124 → 0.3.147
@@ -36,6 +43,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - More efficient rope implementation for LLaMA, torch RMSNorm for Flux models
 - New API nodes: Magnific Upscalers, Bria RMBG, Recraft V4, Vidu Q3 Turbo, Kling V3/O3, Tencent 3D
 
+### Fixed
+- Pin template input URLs to commit SHA instead of mutable `refs/heads/main` branch ref, preventing `hash mismatch in fixed-output derivation` errors when upstream changes files (#25)
+
 ## [0.12.2] - 2026-02-07
 
 ### Changed
@@ -176,7 +186,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Apple Silicon (M-series) support
 - Basic persistence for user data
 
-[Unreleased]: https://github.com/utensils/comfyui-nix/compare/v0.12.2...HEAD
+[Unreleased]: https://github.com/utensils/comfyui-nix/compare/v0.14.2...HEAD
+[0.14.2]: https://github.com/utensils/comfyui-nix/compare/v0.12.2...v0.14.2
 [0.12.2]: https://github.com/utensils/comfyui-nix/compare/v0.7.0-2...v0.12.2
 [0.7.0-2]: https://github.com/utensils/comfyui-nix/compare/v0.7.0-1...v0.7.0-2
 [0.7.0-1]: https://github.com/utensils/comfyui-nix/compare/v0.7.0...v0.7.0-1

CLAUDE.md

@@ -12,6 +12,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 - **Run application**: `nix run` (default)
 - **Run with browser**: `nix run -- --open` (automatically opens browser)
 - **Run with CUDA**: `nix run .#cuda` (Linux/NVIDIA only, uses pre-built PyTorch CUDA wheels)
+- **Run with ROCm**: `nix run .#rocm` (Linux/AMD only, uses pre-built PyTorch ROCm 7.1 wheels)
 - **Run with custom port**: `nix run -- --port=8080`
 - **Run with network access**: `nix run -- --listen 0.0.0.0`
 - **Run with debug logging**: `nix run -- --debug` or `nix run -- --verbose`
@@ -23,10 +24,12 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 - **Build Docker image**: `nix run .#buildDocker` (creates `comfy-ui:latest`)
 - **Build CUDA Docker**: `nix run .#buildDockerCuda` (creates `comfy-ui:cuda`)
-- **Cross-build Linux images from macOS**: `nix run .#buildDockerLinux`, `nix run .#buildDockerLinuxCuda`, `nix run .#buildDockerLinuxArm64`
-- **Pull pre-built**: `docker pull ghcr.io/utensils/comfyui-nix:latest` (or `:latest-cuda`)
+- **Build ROCm Docker**: `nix run .#buildDockerRocm` (creates `comfy-ui:rocm`)
+- **Cross-build Linux images from macOS**: `nix run .#buildDockerLinux`, `nix run .#buildDockerLinuxCuda`, `nix run .#buildDockerLinuxRocm`, `nix run .#buildDockerLinuxArm64`
+- **Pull pre-built**: `docker pull ghcr.io/utensils/comfyui-nix:latest` (or `:latest-cuda`, `:latest-rocm`)
 - **Run container**: `docker run -p 8188:8188 -v $PWD/data:/data comfy-ui:latest`
 - **Run CUDA container**: `docker run --gpus all -p 8188:8188 -v $PWD/data:/data comfy-ui:cuda`
+- **Run ROCm container**: `docker run --device /dev/kfd --device /dev/dri -p 8188:8188 -v $PWD/data:/data comfy-ui:rocm`
 
 ## Linting and Code Quality
 
@@ -46,13 +49,13 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## Version Management
 
-- Current ComfyUI version: v0.12.2 (pinned in `nix/versions.nix`)
+- Current ComfyUI version: v0.14.2 (pinned in `nix/versions.nix`)
 - To update ComfyUI: modify `version`, `rev`, and `hash` in `nix/versions.nix`
 - Vendored wheels (spandrel, frontend, docs, etc.) also pinned in `nix/versions.nix`
 - Template input files: auto-generated in `nix/template-inputs.nix`
   - Update with: `./scripts/update-template-inputs.sh && git add nix/template-inputs.nix`
 - Python version: 3.12
-- PyTorch: macOS uses pre-built wheels (2.5.1, pinned to work around MPS bugs on macOS 26); CUDA uses pre-built wheels from pytorch.org (cu124); Linux CPU uses nixpkgs
+- PyTorch: macOS uses pre-built wheels (2.5.1, pinned to work around MPS bugs on macOS 26); CUDA uses pre-built wheels from pytorch.org (cu124); ROCm uses pre-built wheels from pytorch.org (rocm7.1); Linux CPU uses nixpkgs
 
 ## Project Architecture
 
@@ -127,6 +130,7 @@ fonts/         - Bundled fonts for nodes requiring system fonts
 
 - macOS: PyTorch pinned to 2.5.1 to work around MPS bugs on macOS 26 (Tahoe); browser opens via `/usr/bin/open`
 - CUDA: Pre-built wheels from pytorch.org with CUDA 12.4 runtime bundled (no separate toolkit needed); supports Pascal through Hopper
+- ROCm: Pre-built wheels from pytorch.org with ROCm 7.1 runtime bundled; tested on gfx1100 (7900 XTX); `/run/opengl-driver/lib` provides AMD drivers on NixOS
 - Linux CPU: Uses nixpkgs PyTorch; browser opens via `xdg-open`
 - Cross-platform Docker builds work from any system via `nix run .#buildDockerLinux` etc.
 
@@ -139,6 +143,7 @@ fonts/         - Bundled fonts for nodes requiring system fonts
 - Triggers: push to main, version tags (v*), pull requests
 - CPU images: multi-arch (amd64 + arm64 via QEMU)
 - CUDA images: x86_64 only
+- ROCm images: x86_64 only
 - Published to `ghcr.io/utensils/comfyui-nix` (`:latest`, `:latest-cuda`, `:X.Y.Z`)
 
 **Claude Code Integration** (`.github/workflows/claude.yml`, `.github/workflows/claude-code-review.yml`):

README.md

@@ -26,6 +26,14 @@ nix run github:utensils/comfyui-nix#cuda
 
 CUDA builds use pre-built PyTorch wheels from pytorch.org, so builds are fast (~2GB download) and support all GPU architectures from Pascal (GTX 1080) through Hopper (H100) in a single package.
 
+For ROCm (Linux/AMD):
+
+```bash
+nix run github:utensils/comfyui-nix#rocm
+```
+
+ROCm builds use pre-built PyTorch wheels from pytorch.org, so builds are fast (~2GB download); so far, GPU support has only been validated on `gfx1100`.
+
 ## Options
 
 All [ComfyUI CLI options] are supported. Common examples:
@@ -62,6 +70,21 @@ nix run github:utensils/comfyui-nix#cuda
 
 This single package works on any NVIDIA GPU from the past ~8 years.
 
+## ROCm GPU Support
+
+ROCm builds are available for Linux with AMD GPUs. The `#rocm` package uses pre-built PyTorch wheels from pytorch.org which:
+
+- **Fast builds**: Downloads ~2GB of pre-built wheels instead of compiling for hours
+- **Low memory**: No 30-60GB RAM requirement for compilation
+- **Supported architectures**: To date, only `gfx1100` (7900XTX) has been tested
+- **Bundled runtime**: ROCm 7.1 libraries included in wheels (no separate toolkit needed)
+
+```bash
+nix run github:utensils/comfyui-nix#rocm
+```
+
+> ROCm support contributed by [@pyqlsa](https://github.com/pyqlsa) — thank you!
+
 ## Why a Nix Flake?
 
 ComfyUI's standard installation relies on pip and manual dependency management, which doesn't integrate well with NixOS's declarative approach. This flake provides:
@@ -241,6 +264,7 @@ Add ComfyUI as a package in your system configuration:
         nixpkgs.overlays = [ comfyui-nix.overlays.default ];
         environment.systemPackages = [ pkgs.comfy-ui ];
         # Or for CUDA: pkgs.comfy-ui-cuda
+        # Or for ROCm: pkgs.comfy-ui-rocm
       }];
     };
 
@@ -263,6 +287,7 @@ Add ComfyUI as a package in your system configuration:
   environment.systemPackages = [
     inputs.comfyui-nix.packages.${pkgs.system}.default  # CPU
     # inputs.comfyui-nix.packages.${pkgs.system}.cuda   # CUDA (Linux)
+    # inputs.comfyui-nix.packages.${pkgs.system}.rocm   # ROCm (Linux)
   ];
 }
 ```
@@ -275,6 +300,7 @@ The overlay provides these packages:
 | -------------------- | ------------------------------------------------ |
 | `pkgs.comfy-ui`      | CPU + Apple Silicon (Metal) - use this for macOS |
 | `pkgs.comfy-ui-cuda` | NVIDIA GPUs (Linux only, all architectures)      |
+| `pkgs.comfy-ui-rocm` | AMD GPUs (Linux only, `gfx1100`)      |
 
 > **Note:** On macOS with Apple Silicon, the base `comfy-ui` package automatically uses Metal for GPU acceleration. No separate CUDA package is needed.
 
@@ -288,6 +314,9 @@ nix profile add github:utensils/comfyui-nix
 
 # CUDA (Linux/NVIDIA only)
 nix profile add github:utensils/comfyui-nix#cuda
+
+# ROCm (Linux/AMD only)
+nix profile add github:utensils/comfyui-nix#rocm
 ```
 
 > **Note:** Profile installation is convenient for trying ComfyUI but isn't declarative. For reproducible setups, add the package to your NixOS/nix-darwin configuration instead.
@@ -301,7 +330,8 @@ nix profile add github:utensils/comfyui-nix#cuda
 
   services.comfyui = {
     enable = true;
-    cuda = true;  # Enable NVIDIA GPU acceleration (recommended for most users)
+    gpuSupport = "cuda";  # Enable NVIDIA GPU acceleration (recommended for most users)
+    # gpuSupport = "rocm";  # Enable AMD GPU acceleration
     # cudaCapabilities = [ "8.9" ];  # Optional: optimize system CUDA packages for RTX 40xx
     #   Note: Pre-built PyTorch wheels already support all GPU architectures
     enableManager = true;  # Enable the built-in ComfyUI Manager
@@ -320,7 +350,7 @@ nix profile add github:utensils/comfyui-nix#cuda
 | Option          | Default              | Description                                      |
 | --------------- | -------------------- | ------------------------------------------------ |
 | `enable`        | `false`              | Enable the ComfyUI service                       |
-| `cuda`          | `false`              | Enable NVIDIA GPU acceleration                   |
+| `gpuSupport`    | `"none"`             | Enable NVIDIA or AMD GPU acceleration            |
 | `cudaCapabilities` | `null`           | Optional CUDA compute capability list            |
 | `enableManager` | `false`              | Enable the built-in ComfyUI Manager              |
 | `port`          | `8188`               | Port for the web interface                       |
@@ -347,7 +377,7 @@ To run ComfyUI with data in a user's home directory:
 ```nix
 services.comfyui = {
   enable = true;
-  cuda = true;
+  gpuSupport = "cuda";
   dataDir = "/home/myuser/comfyui-data";
   user = "myuser";
   group = "users";
@@ -393,6 +423,10 @@ docker run -p 8188:8188 -v "$PWD/data:/data" ghcr.io/utensils/comfyui-nix:latest
 # CUDA (x86_64 only, requires nvidia-container-toolkit)
 # Supports ALL GPU architectures: Pascal, Volta, Turing, Ampere, Ada, Hopper
 docker run --gpus all -p 8188:8188 -v "$PWD/data:/data" ghcr.io/utensils/comfyui-nix:latest-cuda
+
+# ROCm (x86_64 only)
+# Supports `gfx1100` (and possibly others)
+docker run --gpus all -p 8188:8188 -v "$PWD/data:/data" ghcr.io/utensils/comfyui-nix:latest-rocm
 ```
 
 **Podman:**
@@ -403,6 +437,10 @@ podman run -p 8188:8188 -v "$PWD/data:/data:Z" ghcr.io/utensils/comfyui-nix:late
 
 # CUDA (requires nvidia-container-toolkit and CDI configured)
 podman run --device nvidia.com/gpu=all -p 8188:8188 -v "$PWD/data:/data:Z" ghcr.io/utensils/comfyui-nix:latest-cuda
+
+# ROCm
+podman run --device /dev/kfd --device /dev/dri -p 8188:8188 -v "$PWD/data:/data:rw" -v "/etc/passwd:/etc/passwd:ro" ghcr.io/utensils/comfyui-nix:latest-rocm
+
 ```
 
 **Passing additional arguments:**
@@ -417,19 +455,42 @@ docker run --gpus all -p 8188:8188 -v "$PWD/data:/data" \
 # Podman with manager enabled
 podman run --device nvidia.com/gpu=all -p 8188:8188 -v "$PWD/data:/data:Z" \
   ghcr.io/utensils/comfyui-nix:latest-cuda --listen 0.0.0.0 --enable-manager
+
+# Podman and ROCm with manager enabled and some recommended settings
+podman run \
+  --device /dev/kfd \
+  --device /dev/dri \
+  -p 8188:8188 \
+  -v "$PWD/data:/data:rw" \
+  -v "/etc/passwd:/etc/passwd:ro" \
+  ghcr.io/utensils/comfyui-nix:latest-rocm \
+  --listen 0.0.0.0 \
+  --enable-manager \
+  --disable-xformers \
+  --use-pytorch-cross-attention
 ```
 
 **Build locally:**
 
 ```bash
-nix run .#buildDocker      # CPU
-nix run .#buildDockerCuda  # CUDA
+nix build .#dockerImage      # CPU
+nix build .#dockerImageCuda  # CUDA
+nix build .#dockerImageRocm  # ROCm
 
 # Load into Docker/Podman
 docker load < result
 podman load < result
 ```
 
+**Build locally and load in a single step:**
+
+```bash
+# builds image and loads into docker
+nix run .#buildDocker      # CPU
+nix run .#buildDockerCuda  # CUDA
+nix run .#buildDockerRocm  # ROCm
+```
+
 **Note:** Docker/Podman on macOS runs CPU-only. For GPU acceleration on Apple Silicon, use `nix run` directly.
 
 ## Development