Skip to content

build cuda 13.0.0 base image with 2.9.0 #3229

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

build cuda 13.0.0 base image with 2.9.0 #3229

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d7f6db4
build cuda 13.0.0 base image with 2.9.0
winglian Oct 23, 2025
6a80917
upgrade causal-conv1d
winglian Oct 24, 2025
20a346a
1.5.4 not in pypi yet
winglian Oct 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 14 additions & 0 deletions .github/workflows/base.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,13 @@ jobs:
pytorch: 2.9.0
torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
dockerfile: "Dockerfile-base"
- cuda: "130"
cuda_version: 13.0.0
cudnn_version: ""
python_version: "3.11"
pytorch: 2.9.0
torch_cuda_arch_list: "9.0+PTX"
dockerfile: "Dockerfile-base"
Comment on lines +63 to +69
Copy link
Contributor

Choose a reason for hiding this comment

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

⚠️ Potential issue | 🔴 Critical

🧩 Analysis chain

Verify torch_cuda_arch_list is intentional for CUDA 13.0.0.

The new CUDA 13.0.0 entry restricts torch_cuda_arch_list to "9.0+PTX", while all existing CUDA 12.x entries support a broader range: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX". This significantly limits hardware compatibility—users with older GPUs (compute capability 7.0–8.9) won't have optimized binaries.

Confirm whether this reduction in supported architectures is:

  • An intentional limitation of CUDA 13.0.0 or PyTorch 2.9.0 with CUDA 13.0.0
  • An oversight that should match the broader architecture support of earlier entries

Also applies to: 153-159

🌐 Web query:

CUDA 13.0.0 supported GPU compute capabilities PyTorch 2.9.0

💡 Result:

Short answer:

  • CUDA 13.0 supports GPU compute capabilities (compute_xx / sm_xx) listed in NVIDIA's nvcc docs: compute_75, 80, 86, 87, 88, 89, 90, 90a, 100, 100f, 100a, 103, 103f, 103a, 110, 110f, 110a, 120, 120f, 120a, 121, 121f, 121a (and corresponding sm_75, sm_80, sm_86, sm_87, sm_88, sm_89, sm_90, sm_90a, sm_100, sm_100f, sm_100a, sm_103, sm_103f, sm_103a, sm_110, sm_110f, sm_110a, sm_120, sm_120f, sm_120a, sm_121, sm_121f, sm_121a). [1][2]

  • CUDA 13.0 supports NVIDIA architectures from Turing (SM 7.5) through Grace/Blackwell; Maxwell/Pascal/Volta (older than Turing) are dropped for several libraries. [3]

  • PyTorch 2.9.0 expands wheel support to include CUDA 13 (i.e., PyTorch 2.9 provides CUDA‑13 build variants). Confirm PyTorch package variant for your platform/installer before installing. [4]

Sources:
[1] NVIDIA nvcc GPU code generation / List of supported GPU codes (CUDA 13.0).
[2] CUDA Toolkit 13.0 release notes (architectures/support details).
[3] NVIDIA developer blog “What’s New in CUDA 13.0”.
[4] PyTorch 2.9 release blog (notes on CUDA 13 wheel support).

Correct torch_cuda_arch_list for CUDA 13.0.0 to match supported GPU architectures.

CUDA 13.0 supports GPU compute capabilities from Turing (SM 7.5) through Grace/Blackwell. The current entry restricts this to "9.0+PTX" only, which excludes Turing (7.5) and Ampere (8.0–8.9) GPUs that CUDA 13.0 officially supports.

Update line 66 (and 156) to:

torch_cuda_arch_list: "7.5 8.0 8.6 8.7 8.8 8.9 9.0+PTX"

Note: Unlike CUDA 12.x entries, omit "7.0" (Pascal) since Maxwell/Pascal/Volta are dropped in CUDA 13.0.

Also applies to: 153-159

🤖 Prompt for AI Agents 
In .github/workflows/base.yml around lines 63-69 (and also apply the same change
at lines 153-159), the torch_cuda_arch_list is incorrectly set to "9.0+PTX"
which excludes supported CUDA 13.0 architectures; update the
torch_cuda_arch_list value to "7.5 8.0 8.6 8.7 8.8 8.9 9.0+PTX" in both places
so Turing and Ampere families are included and Pascal (7.0) is omitted per CUDA
13.0 support.

# - cuda: "128"
# cuda_version: 12.8.1
# cudnn_version: ""
Expand Down Expand Up @@ -143,6 +150,13 @@ jobs:
pytorch: 2.9.0
torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
dockerfile: "Dockerfile-uv-base"
- cuda: "130"
cuda_version: 13.0.0
cudnn_version: ""
python_version: "3.11"
pytorch: 2.9.0
torch_cuda_arch_list: "9.0+PTX"
dockerfile: "Dockerfile-uv-base"
steps:
- name: Checkout
uses: actions/checkout@v4
Expand Down
2 changes: 1 addition & 1 deletion docker/Dockerfile-base
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@ WORKDIR /workspace

RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir causal_conv1d==1.5.2 && \
CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir causal_conv1d==1.5.3.post1 && \
python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
python3 -m pip cache purge

Expand Down
Loading