use static signature for sfd_col_d_srelu_tensor

[jiemingz]

Summary by CodeRabbit

• Bug Fixes
  • Improved caching efficiency for grouped matrix multiplication operations with dynamic activation functions, resulting in more precise instance selection and better performance optimization.

[Anerudhan]

@cudnn-ci-bot run

[coderabbitai]

🧹 Nitpick comments (1)

python/cudnn/grouped_gemm/grouped_gemm_dsrelu/api.py (1)

1483-1493: ⚡ Quick win

Add documentation explaining the dimension selection logic.

The helper omits dimension 4 from the static shape and marks stride dimensions 2 and 5 as dynamic, but the rationale for this specific selection is not documented. Consider adding a docstring or inline comment explaining:

• Why dimension 4 (rest_m) is excluded from the static shape
• Why stride dimensions 2 and 5 are treated as dynamic
• How this relates to the SFD column tensor structure

This would improve maintainability and help future developers understand the cache key granularity design. As per coding guidelines, documentation is a key focus area for this module.

🤖 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 `@python/cudnn/grouped_gemm/grouped_gemm_dsrelu/api.py` around lines 1483 -
1493, Add a short docstring or inline comment to
dynamic_sfd_col_tensor_signature explaining the dimension-selection rationale:
state that static_shape intentionally omits dimension 4 (rest_m) because rest_m
varies per-column and should not be part of the cache key, and that
dynamic_stride_dims=(2, 5) marks the stride-related dimensions (the inner M
chunk and the leading stride for packed layout) as dynamic because their strides
can vary even when logical sizes match; also mention how this choice maps to the
SFD column tensor layout and why it yields the desired cache key granularity
when delegating to dynamic_m_tensor_signature.

Source: Coding guidelines

🤖 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.

Nitpick comments:
In `@python/cudnn/grouped_gemm/grouped_gemm_dsrelu/api.py`:
- Around line 1483-1493: Add a short docstring or inline comment to
dynamic_sfd_col_tensor_signature explaining the dimension-selection rationale:
state that static_shape intentionally omits dimension 4 (rest_m) because rest_m
varies per-column and should not be part of the cache key, and that
dynamic_stride_dims=(2, 5) marks the stride-related dimensions (the inner M
chunk and the leading stride for packed layout) as dynamic because their strides
can vary even when logical sizes match; also mention how this choice maps to the
SFD column tensor layout and why it yields the desired cache key granularity
when delegating to dynamic_m_tensor_signature.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 941018c2-14e8-4ce1-a369-5b4cb4767865

📥 Commits

Reviewing files that changed from the base of the PR and between 035b520 and bf8b797.

📒 Files selected for processing (1)

• python/cudnn/grouped_gemm/grouped_gemm_dsrelu/api.py