Skip to content

Revise PROJECT_STATUS.md with recent updates - #1022

Open
boessu wants to merge 1 commit into
jamiepine:mainfrom
boessu:patch-4
Open

Revise PROJECT_STATUS.md with recent updates#1022
boessu wants to merge 1 commit into
jamiepine:mainfrom
boessu:patch-4

Conversation

@boessu

@boessu boessu commented Aug 9, 2026 •
edited by coderabbitai Bot
Loading

Copy link
Copy Markdown

Updated project status with new issues and fixes related to ROCm Docker builds, audio playback, and other tooling improvements.

Change according (and depending) on pull request #1019

Summary by CodeRabbit

  • Documentation
    • Documented a known limitation affecting ROCm Docker builds and incompatible CUDA wheels.
    • Added guidance on the required PyTorch version pin and related packaging considerations.

@boessu
Revise PROJECT_STATUS.md with recent updates
328ad30 
Updated project status with new issues and fixes related to ROCm Docker builds, audio playback, and other tooling improvements.
@coderabbitai

coderabbitai Bot commented Aug 9, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Review Change Stack

📝 Walkthrough

Walkthrough

The project status documentation now records ROCm Docker and PyTorch wheel compatibility limitations. The Docker tooling entry links to this limitation.

Changes

ROCm documentation

Layer / File(s) Summary
Document ROCm build limitation
docs/PROJECT_STATUS.md		 Documents incompatible CUDA wheels, required PyTorch pins, CUDA dependency cleanup, and the conflicting scripts/package_rocm.py compatibility declaration. The Docker tooling entry references this limitation.

Estimated code review effort: 1 (Trivial) | ~3 minutes

Possibly related PRs

  • jamiepine/voicebox#817: Documents the ROCm torch-wheel conflict and related scripts/package_rocm.py reconciliation.
  • jamiepine/voicebox#976: Addresses ROCm Docker and PyTorch wheel compatibility and the related torch-pin conflict.

Suggested reviewers: jamiepine

🚥 Pre-merge checks | ✅ 5
✅ Passed checks (5 passed)
Check name Status Explanation
Description Check ✅ Passed Check skipped - CodeRabbit’s high-level summary is enabled.
Title check ✅ Passed The title clearly identifies the updated file and accurately summarizes the documentation changes.
Docstring Coverage ✅ Passed No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check ✅ Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check ✅ Passed Check skipped because no linked issues were found for this pull request.
✨ Finishing Touches
🧪 Generate unit tests (beta)
  • Create PR with unit tests

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

coderabbitai[bot]
coderabbitai Bot reviewed Aug 9, 2026
View reviewed changes

@coderabbitai coderabbitai Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/PROJECT_STATUS.md`:
- Line 384: Format the PR entry beginning with “#764” as a Markdown list item by
prefixing the line with “- ”, preserving the existing content and ordering.
- Line 280: Update the ROCm Docker build entry in PROJECT_STATUS.md to replace
the “PR #<your-pr-number>” placeholder with the actual pull request number, or
remove the PR reference if unavailable.
- Line 280: Update the ROCm Docker build guidance in the project status entry to
reflect the existing rocm${ROCM_VERSION} index, /etc/pip.conf configuration,
PYTORCH_VARIANT=rocm path, and versions declared in
backend/requirements-rocm.txt. Remove the unsupported CUDA-default and 2.7.1 pin
claims, describe only the verified failure and required fix, and replace the
unresolved PR placeholder with the actual reference or remove it.
🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

  • Push a commit to this branch (recommended)
  • Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 7712b616-c46e-4cd2-ac77-88eeb4dc9b6c

📥 Commits

Reviewing files that changed from the base of the PR and between 51f49de and 328ad30.

📒 Files selected for processing (1)
  • docs/PROJECT_STATUS.md

Comment thread docs/PROJECT_STATUS.md
- **Blackwell (RTX 50-series) CUDA**: cu128 + sm_120 kernel support shipped (PR #401, #316), but users still report `cudaErrorNoKernelImageForDevice` (#417, #400, #396, #395, #390, #362) — likely a stale CUDA binary on upgraded installs. Needs a follow-up diagnostic / forced re-download path.
- **Long text 50k character limit** (#464, #365, #354): Still hit on GPU despite chunking (PR #266). Chunking reliability needs another pass.
- **ROCm on RDNA 3/4** (#469): `HSA_OVERRIDE_GFX_VERSION` is hardcoded and harms newer cards.
- **ROCm Docker build requires pinned torch (new)**: `docker-compose.rocm.yml`'s build (`torch`/`torchvision`/`torchaudio` unpinned in `Dockerfile`) silently resolves to a CUDA wheel instead of ROCm, because PyTorch 2.9+'s wheel-variant system needs a provider plugin to detect the ROCm environment — detection that can't happen inside a `docker build` context. Symptom: `torch.cuda.is_available()` returns `False`, or (if partially masked by a stale cache) `libcudart.so.13: cannot open shared object file` at model load. Fix requires pinning `torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1` (last pre-variant ROCm release) and explicitly stripping stray CUDA transitive deps pulled in by `qwen-tts` (installed without `--no-deps`). See PR #<your-pr-number>. Also note `scripts/package_rocm.py` declares `torch_compat=">=2.9.0,<2.10.0"` — conflicts with this pin and needs separate reconciliation.

@coderabbitai coderabbitai Bot Aug 9, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Remove the unreplaced PR placeholder.

This entry still contains PR #<your-pr-number>. Replace it with the actual PR number or remove the reference before merge.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/PROJECT_STATUS.md` at line 280, Update the ROCm Docker build entry in
PROJECT_STATUS.md to replace the “PR #<your-pr-number>” placeholder with the
actual pull request number, or remove the PR reference if unavailable.

🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

🧩 Analysis chain

🏁 Script executed:

#!/usr/bin/env bash
set -euo pipefail

for file in Dockerfile docker-compose.rocm.yml backend/requirements-rocm.txt scripts/package_rocm.py; do
  printf '\n### %s\n' "$file"
  rg -n -C 3 'PYTORCH_VARIANT|ROCM_VERSION|index-url|torch(vision)?|torchaudio|torch_compat' "$file" || true
done

printf '\n### Version references\n'
rg -n '2\.7\.1|2\.9\.1|torch_compat|your-pr-number' \
  docs/PROJECT_STATUS.md backend/requirements-rocm.txt scripts/package_rocm.py

Repository: jamiepine/voicebox

Length of output: 5207

🌐 Web query:

PyTorch ROCm docker build install index-url torch wheel variant provider plugin torch.cuda.is_available False

💡 Result:

If your PyTorch ROCm installation inside Docker reports torch.cuda.is_available as False, it almost always indicates a failure to pass the host's GPU hardware and ROCm driver stack into the container correctly [1][2][3]. To resolve this, ensure you are using the correct docker run flags and that your environment supports ROCm. Recommended Docker Execution Flags You must grant the container access to the KFD and DRI devices and the video group [4][5][6]. Use the following command structure [5][6]: docker run -it --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \ --device=/dev/kfd --device=/dev/dri \ --group-add video --group-add render \ --ipc=host --shm-size 8G \ rocm/pytorch: If you are running in rootless mode, this may fail because the group IDs for video/render on the host do not match those inside the container [3]. It is recommended to use standard (rootful) Docker for ROCm workloads to avoid permission and mapping issues [2][3]. Installing PyTorch with Specific Wheels If you need to install PyTorch manually (e.g., in a custom Dockerfile), you must use the ROCm-specific index URL provided by AMD [4][7][6]. The standard PyPI wheel does not include ROCm support [8]. Use the following syntax for your pip install command [4][7]: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocmX.Y Replace rocmX.Y with your required ROCm version (e.g., rocm7.2). For nightly builds, append /nightly: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2 Troubleshooting Checklist 1. Verify Hardware Access: Run rocminfo inside the container. If it fails to list your GPU, the --device or --group-add flags are likely missing or incorrectly configured [2][3][8]. 2. Check Rootless Mode: Ensure you are not using rootless Docker, which often restricts device access required for ROCm [2][3]. 3. Kernel Compatibility: Ensure your host kernel supports the GPU and that the ROCm kernel modules are loaded [2]. 4. Environment Variables: Some setups may require explicit target identification (e.g., HSA_OVERRIDE_GFX_VERSION or PYTORCH_ROCM_ARCH) if using non-standard or unsupported GPUs, though this is not required for officially supported hardware [1].

Citations:

🌐 Web query:

PyTorch backend wheel variant ROCm provider plugin CUDA wheel detection docker build context

💡 Result:

PyTorch has transitioned toward a standardized "wheel variant" system, guided by the WheelNext proposal (PEP 817), which allows installers to automatically select the appropriate prebuilt wheel (CPU, CUDA, or ROCm) for a user's specific environment [1][2][3]. 1. Wheel Variant Selection: Instead of requiring users to manually select and configure index URLs based on their hardware (as traditionally done with e.g., download.pytorch.org/whl/cu121), "variant-aware" installers (like uv) can now automatically detect the system's hardware and software configuration [3][4][5]. This system uses "variant provider" plugins that query the environment (e.g., detecting installed ROCm versions or NVIDIA CUDA capabilities) and match them to compatible prebuilt wheels [2][3]. 2. ROCm Provider Plugin: For AMD environments, tools like rocm-bootstrap provide pure-Python detection of AMD GPUs and ROCm versions without requiring a full ROCm installation [6]. This plugin acts as a variant provider for uv/pip, enabling the automatic selection of device-specific wheels [3][6]. 3. Docker Build Contexts: When building Docker images for PyTorch, the traditional approach remains manual specification of the compute backend through build arguments (e.g., CUDA_PATH, ROCm_VERSION) [7][8][9]. These arguments control which index URL is used during the pip install step within the Dockerfile [7][8]. - For official PyTorch Docker images, build scripts typically use these variables to select the target build stage (e.g., cuda_final or rocm_final) [10][9]. - While the wheel variant system simplifies user-level installation, Docker builds currently still rely on explicit dependency management to ensure the correct image layer is constructed for the target accelerator [10][11]. 4. Key Mechanisms: - Variant Providers: Either "install-time" (dynamic query) or "ahead-of-time" (static metadata), these define the logic for selecting the correct wheel variant [2][4]. - Environment Overrides: Users can force specific variants by setting environment variables (e.g., ROCM_BOOTSTRAP_FORCE_GFX_ARCH or equivalent override mechanisms) to bypass automatic detection when testing or running on non-GPU systems [3][6][12]. This standardization effort aims to reduce the "index hell" associated with managing multiple accelerator-specific PyTorch distributions and ensure consistency across ecosystem libraries like xformers and flash-attn [3][5].

Citations:

🌐 Web query:

PyTorch ROCm 7.2.1 pip index-url rocm7.2 --extra-index-url pypi.org simple Python Package Index default URL documentation

💡 Result:

For ROCm 7.2.1, AMD recommends a specific installation approach depending on your environment rather than using a generic pip index-url [1][2][3]. For native Linux systems, AMD provides direct wheel (.whl) files for ROCm 7.2.1, which should be downloaded and installed via pip [1][2][4]. AMD explicitly advises using these repository-hosted wheels instead of relying on generic PyTorch Foundation wheels, as the official AMD-hosted wheels are tested for ROCm compatibility [2][4]. Example installation command structure for native Linux (Ubuntu): 1. Download the specific wheel files for torch, torchvision, triton, and torchaudio from the AMD repository: wget https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2.1/<wheel_filename>.whl [1][4] 2. Install the downloaded wheels: pip3 install <wheel_filename_1>.whl <wheel_filename_2>.whl... [1][4] Note that for Python 3.12 installations on non-virtual environments, you may need to include the --break-system-packages flag [1][2][4]. For environments requiring PyTorch nightly or upstream builds, the standard PyTorch installation documentation directs users to the PyTorch Get Started matrix [3][5][6]. The specific URL for ROCm 7.2 nightly wheels (if used) is https://download.pytorch.org/whl/nightly/rocm7.2 [3][6]. For Windows environments, AMD also provides specific ROCm 7.2.1 wheels that are installed by downloading and running pip install on the corresponding.whl files from the AMD repository [7]. For the most reliable installation, refer to the official ROCm documentation specific to your hardware (Radeon/Ryzen) or the PyTorch installation matrix for the latest tested configurations [2][3][8].

Citations:

🌐 Web query:

PyTorch docs Get Started Locally ROCm index-url rocm7.2 +rocm7.2 pip

💡 Result:

To install PyTorch with ROCm 7.2 support via pip, you can use the nightlies repository index provided by PyTorch [1][2][3]. The standard installation command is: pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2 Note that AMD documentation recommends this index-url for accessing ROCm 7.2 wheels [1][2][3]. While the general PyTorch "Get Started Locally" page allows you to select ROCm to generate a command [4][1], specific ROCm-compatible wheels are frequently hosted on the PyTorch nightly index when official stable support is transitioning or for specific version requirements [1][5]. For production or highly stable environments, AMD also strongly recommends using their prebuilt, validated Docker images, which include the necessary dependencies and PyTorch pre-installed [1][6][3]. You can find these images on Docker Hub, with tags such as rocm/pytorch:rocm7.2_ubuntu24.04_py3.12_pytorch_release_2.9.1 [6][3][7]. Key recommendations for installing PyTorch on ROCm: 1. Docker-based installation is the recommended approach to avoid dependency and compatibility issues [1][6][3]. 2. For pip installation, ensure you have the appropriate system dependencies installed (e.g., libjpeg-dev, python3-dev, pip) [2][6][8]. 3. Verify compatibility by checking the ROCm-PyTorch support matrices, as ROCm development and PyTorch releases are closely aligned [9][10].

Citations:

🌐 Web query:

"rocm7.3.1" "PyTorch Get Started" "pip install" "torch torchvision torchaudio" "index-url"

💡 Result:

To install PyTorch for ROCm, AMD recommends using the official ROCm AI Ecosystem documentation, which provides the most accurate installation commands for specific ROCm versions and hardware architectures [1]. While general PyTorch instructions can be found at pytorch.org/get-started/locally/ [2], AMD maintains a dedicated repository for ROCm-enabled wheels [1][3]. For a specific version like ROCm 7.3.1, you should consult the official AMD ROCm documentation portal at rocm.docs.amd.com to identify the correct index-url and package strings, as these can vary by hardware architecture (e.g., gfx950, gfx942) and system environment [1][3]. As of the latest available documentation patterns, the standard installation command format follows this structure [1][3]: python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "torch[device-all]==+rocm7.3.1" "torchvision[device-all]==+rocm7.3.1" "torchaudio==+rocm7.3.1" Key steps to ensure a successful installation: 1. Virtual Environment: Always install within a dedicated Python virtual environment to avoid conflicts with system packages [3]. 2. Hardware Specificity: If your specific GPU architecture is known, replace device-all with your target architecture (e.g., device-gfx950) to minimize the installation size [1][3]. 3. Verification: After installation, verify that PyTorch detects the ROCm device by running: python -c "import torch; print(torch.cuda.is_available)" If this returns True, your installation is configured correctly [3]. For the most up-to-date repository URLs and specific package versions for ROCm 7.3.1, please visit the official AMD ROCm AI Ecosystem installation guide [1].

Citations:

🌐 Web query:

PyTorch ROCm 7.3.1 pip index-url https://download.pytorch.org/whl/rocm7.3

💡 Result:

There is no official pip index-url at https://download.pytorch.org/whl/rocm7.3 [1]. PyTorch wheel indexes on the official download site typically only go up to rocm7.2 [1]. For current installations of PyTorch with ROCm support (including on systems running newer ROCm versions like 7.14), the recommended approach is to use the https://repo.amd.com/rocm/whl-multi-arch/ index for architecture-specific wheels, or to fall back to the most recent official PyTorch index, which is https://download.pytorch.org/whl/rocm7.2 [2][3][1]. Key points regarding PyTorch and ROCm versions: 1. Compatibility: PyTorch wheels provided by AMD or the official PyTorch repository include their own ROCm runtime libraries, meaning a rocm7.2 wheel is generally compatible with newer system-level ROCm installations (such as ROCm 7.14) [1]. 2. Installation: For the latest PyTorch versions, AMD recommends installing via the repo.amd.com index [2]. An example command for newer environments is: python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "torch[device-all]==2.12.0+rocm7.14.0" [2] 3. Official PyTorch Website: For standard setups, always refer to the official PyTorch Get Started matrix, which dynamically generates the correct installation command based on your selected platform and ROCm version [3][4].

Citations:

🌐 Web query:

site:pytorch.org/get-started/locally ROCm 7.2.1 pip install torch torchvision torchaudio index-url

💡 Result:

For ROCm 7.2, AMD recommends installing PyTorch, torchvision, and torchaudio using specific wheels hosted on the official AMD repository rather than using a standard pip index-url [1]. AMD advises that the ROCm wheels available on PyTorch.org are not extensively tested by AMD and may change regularly due to nightly builds [2]. Instead, you should download the compatible wheels directly from the AMD repository (repo.radeon.com) for your specific operating system and Python version [2][1]. The recommended installation process involves the following steps [1]: 1. Identify the appropriate wheel URLs for your environment from the official ROCm documentation (e.g., https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/). 2. Use pip to uninstall any existing conflicting packages: pip3 uninstall torch torchvision triton torchaudio 3. Install the downloaded wheels directly: pip3 install <path_or_url_to_torch_whl> <path_or_url_to_torchvision_whl> <path_or_url_to_torchaudio_whl> <path_or_url_to_triton_whl> Note: If you are installing in a non-virtual environment for Python 3.12, you may need to add the --break-system-packages flag [1]. Always check the specific compatibility matrices provided in the ROCm documentation for your version to ensure you are selecting the correct wheel versions for your hardware and Python environment [2][1].

Citations:

Update or clarify the ROCm Docker pin guidance.

The Docker ROCm build already forces the rocm${ROCM_VERSION} PyTorch index during install and writes it to /etc/pip.conf, and docker-compose.rocm.yml enables that path with PYTORCH_VARIANT: rocm. The status entry still claims unpinned installs default to CUDA, pin the 2.7.1 ROCm-era packages, and conflicts with backend/requirements-rocm.txt holding torch==2.9.1+rocm7.2.1, torchaudio==2.9.1+rocm7.2.1, and torchvision==0.24.1+rocm7.2.1. Resolve the version pin against the actual failure path and replace the unresolved PR placeholder.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/PROJECT_STATUS.md` at line 280, Update the ROCm Docker build guidance in
the project status entry to reflect the existing rocm${ROCM_VERSION} index,
/etc/pip.conf configuration, PYTORCH_VARIANT=rocm path, and versions declared in
backend/requirements-rocm.txt. Remove the unsupported CUDA-default and 2.7.1 pin
claims, describe only the verified failure and required fix, and replace the
unresolved PR placeholder with the actual reference or remove it.

Comment thread docs/PROJECT_STATUS.md
### Build / dev tooling / docker (~6)

#764 uv for backend env · #632 docker GPU build + cache + fastmcp (+7.9k) · #630 ROCm docker overlay · #463 ghcr.io auto-publish · #543 / #681 setup-script fixes · #584 docker permission fix
#764 uv for backend env · #632 docker GPU build + cache + fastmcp (+7.9k) · #630 ROCm docker overlay (see Known Limitations — torch pin issue found post-merge)· #463 ghcr.io auto-publish · #543 / #681 setup-script fixes · #584 docker permission fix

@coderabbitai coderabbitai Bot Aug 9, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Format this PR list as a Markdown list item.

Line 384 starts with #764, and markdownlint-cli2 reports MD018. Prefix the line with - or escape the leading hash so the documentation passes lint and renders as intended.

🧰 Tools
🪛 markdownlint-cli2 (0.23.2)

[warning] 384-384: No space after hash on atx style heading

(MD018, no-missing-space-atx)

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/PROJECT_STATUS.md` at line 384, Format the PR entry beginning with
“#764” as a Markdown list item by prefixing the line with “- ”, preserving the
existing content and ordering.

Source: Linters/SAST tools

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@coderabbitai coderabbitai[bot] coderabbitai[bot] left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@boessu