frost(sdpa): filter arch-specific tests correctly to fix sm120 pipeline failures

[Aneureka]

Before submitting

• I agree to license this contribution under the terms of LICENSE.txt.
• I ran pre-commit run and committed any formatting changes.
• I added GitHub labels: one cat-*, one or more mod-*, and one orig-* (see label list).

Affected area

• CI or test infrastructure

Summary

Filter architecture-specific tests correctly to fix SM120 pipeline failures.

Why

Currently some tests are intended to run in an SM100 environment (either real or mocked), but are currently executed in the SM120 pipeline and fail as a result. This PR filters out those cases accordingly.

Related issues

Related to #381

API and compatibility impact

N/A.

Testing

pytest \
    gemm/frost \
    sdpa/frost \
    linear_attention/frost \
    test_mhas_v2.py \
    -n 4 \
    --tb=short

========= 1507 passed, 8341 skipped, 413 warnings in 222.41s (0:03:42) =========

Summary by CodeRabbit

• Tests
  • Improved GPU architecture coverage by standardizing simulated SM100 and SM103 test environments.
  • Updated matrix multiplication and grouped matrix multiplication tests to use shared architecture fixtures.
  • Restricted specialized compilation and backward-engine tests to environments with the required GPU and runtime support.
  • Consolidated GPU availability checks for dual-GEMM tests.

[coderabbitai]

📝 Walkthrough

Walkthrough

FROST tests now share SM100 and SM103 architecture fixtures. Selected GPU tests also use SM100 and runtime availability markers.

Changes

FROST test environment gating

Layer / File(s)	Summary
Shared architecture fixtures test/python/gemm/frost/conftest.py, test/python/gemm/frost/test_block_scale_matmul.py, test/python/gemm/frost/test_moe_grouped_block_scale_matmul_fwd.py	Added shared SM100 and SM103 fixtures. The SM103 fixture sets the shared-memory budget to 227 KiB. Selected tests now use these fixtures.
GPU and runtime capability gating test/python/gemm/frost/test_multi_gemm.py, test/python/linear_attention/frost/test_gdn_bprop_kernel.py	Applied SM100 GPU and runtime availability markers to the relevant tests and removed a duplicate GPU marker assignment.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• NVIDIA/cudnn-frontend#476: Introduces related FROST architecture-mocking and GPU-gating fixtures and tests.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
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.
Title check	✅ Passed	The title clearly describes the primary change: filtering architecture-specific Frost tests to prevent SM120 pipeline failures.
Description check	✅ Passed	The description includes all required sections, explains the issue and solution, states compatibility impact, and provides testing commands with results.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[coderabbitai]

Actionable comments posted: 2

🤖 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 `@python/cudnn/sdpa/fwd/kernels/prefill_f16_sm120.py`:
- Around line 291-292: Update the initializer containing self.head_tile_qk and
self.head_tile_v to validate both values against SUPPORTED_HEAD_TILES before
deriving any kernel or fragment attributes; reject unsupported widths such as
17, while preserving valid specializations and existing tensor-shape validation.

In `@test/python/sdpa/frost/test_sdpa_graph_analyzer.py`:
- Line 539: Add the `@pytest.mark.L0` decorator to
test_sm120_probe_accepts_mixed_head_dims, preserving the test’s existing
implementation and ensuring the marker is applied directly to this new Python
test.

🪄 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: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 3c55ddaa-6e97-4404-9abf-a96a9c062b52

📥 Commits

Reviewing files that changed from the base of the PR and between bce80fb and 9feae2f.

📒 Files selected for processing (9)

• python/cudnn/sdpa/fwd/api_dsl.py
• python/cudnn/sdpa/fwd/kernels/prefill_f16_sm120.py
• test/python/gemm/frost/conftest.py
• test/python/gemm/frost/test_block_scale_matmul.py
• test/python/gemm/frost/test_moe_grouped_block_scale_matmul_fwd.py
• test/python/gemm/frost/test_multi_gemm.py
• test/python/linear_attention/frost/test_gdn_bprop_kernel.py
• test/python/sdpa/frost/test_sdpa_fwd_dsl_sm120.py
• test/python/sdpa/frost/test_sdpa_graph_analyzer.py

[coderabbitai]

Note

GitHub couldn't provide a complete incremental comparison for this pull request, so CodeRabbit is performing a full review instead. This review may take a little longer.

[Aneureka]

@cudnn-ci-bot run frost

[coderabbitai]

Note

GitHub couldn't provide a complete incremental comparison for this pull request, so CodeRabbit is performing a full review instead. This review may take a little longer.

[Aneureka]

@cudnn-ci-bot run frost

[cudnn-ci-bot]

🚀 Running mirror pipeline

Branch: cudnn-gh/pr-495-3048fd3
Pipeline: 61361135
Targets: frost

[vedaanta]

lets just change the CI internally to only run sm120 for spda tests.
closing