Add pyproject and CI wheel build configuration

[MagellaX]

Summary

• add pyproject.toml with pinned CUDA and Triton build requirements
• expand .env.example with GPU arch flags and cache paths
• set up CI workflow to build wheels for SM80 and SM90

Testing

• pytest (fails: TabError, ModuleNotFoundError: yaml)

---

https://chatgpt.com/codex/tasks/task_e_68ab6a20aff08322808d5c4d81ddc7fd

---

Summary by cubic

Adds pyproject.toml and a GitHub Actions workflow to build CUDA wheels for SM80 and SM90. Updates .env.example with GPU arch and cache settings to standardize local and CI builds.

• New Features
  • Added pyproject.toml with setuptools build backend; pins torch==2.1.0 and triton==2.1.0 (CUDA 12.1 noted in tool config).
  • Introduced CI workflow to build wheels on push/PR for SM 80 and 90 (Python 3.10), using TORCH_CUDA_ARCH_LIST from a matrix.
  • Expanded .env.example with TORCH_CUDA_ARCH_LIST, CUDA_HOME, and CUDA/TRITON cache paths.

[cubic-dev-ai]

5 issues found across 3 files

_{React with 👍 or 👎 to teach cubic. You can also tag @cubic-dev-ai to give feedback, ask questions, or re-run the review.}

[cubic-dev-ai]

PyYAML is imported but not declared as a dependency; add "pyyaml" to [project].dependencies to fix ModuleNotFoundError.

(Based on your team's feedback about pytest failing with ModuleNotFoundError: yaml.)

Prompt for AI agents

Address the following comment on pyproject.toml at line 18:

<comment>PyYAML is imported but not declared as a dependency; add &quot;pyyaml&quot; to [project].dependencies to fix ModuleNotFoundError.

(Based on your team&#39;s feedback about pytest failing with ModuleNotFoundError: yaml.)</comment>

<file context>
@@ -0,0 +1,25 @@
+[build-system]
+requires = [
+    &quot;setuptools&gt;=61&quot;,
+    &quot;wheel&quot;,
+    &quot;torch==2.1.0&quot;,
+    &quot;triton==2.1.0&quot;,
+]
+build-backend = &quot;setuptools.build_meta&quot;
+
</file context>

[cubic-dev-ai]

Triton is listed in build requirements; move it to project dependencies to avoid heavyweight installs during build isolation.

Prompt for AI agents

Address the following comment on pyproject.toml at line 6:

<comment>Triton is listed in build requirements; move it to project dependencies to avoid heavyweight installs during build isolation.</comment>

<file context>
@@ -0,0 +1,25 @@
+[build-system]
+requires = [
+    &quot;setuptools&gt;=61&quot;,
+    &quot;wheel&quot;,
+    &quot;torch==2.1.0&quot;,
+    &quot;triton==2.1.0&quot;,
+]
+build-backend = &quot;setuptools.build_meta&quot;
</file context>

[cubic-dev-ai]

PyTorch is listed as a build requirement, which will trigger heavyweight installs in the isolated build env; keep build-system requires minimal (build tools only).

Prompt for AI agents

Address the following comment on pyproject.toml at line 5:

<comment>PyTorch is listed as a build requirement, which will trigger heavyweight installs in the isolated build env; keep build-system requires minimal (build tools only).</comment>

<file context>
@@ -0,0 +1,25 @@
+[build-system]
+requires = [
+    &quot;setuptools&gt;=61&quot;,
+    &quot;wheel&quot;,
+    &quot;torch==2.1.0&quot;,
+    &quot;triton==2.1.0&quot;,
+]
</file context>

[MagellaX]

@cursoragent fix those issues

[cursor]

I refactored dependency management by moving torch and triton to project dependencies in pyproject.toml and added pyyaml. I also enhanced the .env.example with TORCH_CUDA_ARCH_LIST for build configuration and cleaned up a duplicate function call in the accuracy test.

• Moved torch==2.1.0 and triton==2.1.0 from build-system.requires to project.dependencies in pyproject.toml. This clarifies them as runtime dependencies rather than just build-time requirements.
• Added pyyaml>=6.0 as a new dependency to project.dependencies in pyproject.toml.
• Introduced TORCH_CUDA_ARCH_LIST=80;90 to .env.example to explicitly define target CUDA architectures for compilation, improving build configuration.
• Removed a duplicate main() function call in stream_attention/benchmarks/accuracy_test.py to fix a minor bug.

---

4 tasks completed

• Update .env.example to use ${HOME} for cache paths
• Fix pyproject: remove torch/triton from build-system; add pyyaml dependency
• Remove stray tab-indented main() in accuracy_test.py to resolve TabError
• Run Python compileall to verify no syntax/indentation errors

_{Learn more about Cursor Agents}