[yyihuang]

What

• tests/experimental/test_cake_dense_projection_gemm.py: the GPU tests take every deterministic reference result from the deterministic package's own entry points (projection_det_forward, projection_det_dgrad, projection_det_wgrad, dense_projection_det_gemm, router_det_forward, router_det_dgrad, router_det_wgrad, router_fp32_det_gemm). Before, they called this package's entry points with deterministic left at its default, which forwards to those functions. test_the_deterministic_names_map is removed: it asserted that the deterministic package still exports its former names without det.
• benchmarks/bench_cake_dense_projection_gemm.py: the det arm of --memory calls the deterministic package's convenience entry points directly. The atomic arm still calls this package with deterministic=False.

Why

The naming and mode rules for the experimental training kernels: deterministic entry points carry det in their names, the deterministic package keeps no aliases without det, and the deterministic argument of a non-deterministic package never selects the deterministic package. After this PR, the tests and the benchmark of this package depend on neither the aliases nor the default mode. A follow-up can then remove the aliases from dense_projection_det_gemm. The mode change of this package comes in a separate PR.

Behavior

No library code changes. Each changed call runs the same deterministic function with the same arguments as before. The tests of today's default dispatch stay until the mode change.

Tests

• CPU (aarch64, no GPU), this branch vs main abbe776c: pre-commit on the changed files and on all files, no new failing hook. test_cake_dense_projection_gemm.py, test_cake_dense_projection_det_gemm_names.py and test_cake_dense_projection_det_gemm.py: 0 new failures (test_cake_dense_projection_gemm.py: 393 passed / 161 skipped, main 394 / 161). Collection: one test id removed (test_the_deterministic_names_map), none added.
• GPU, test_cake_dense_projection_gemm.py and test_cake_dense_projection_det_gemm_names.py, this branch vs main on the same node:

GPU	this branch (passed / skipped / failed)	main	new failures	outcome changes
GB200 (sm_100a, 152 SMs)	569 / 23 / 0	570 / 23 / 0	0	0
B200 (sm_100a, 148 SMs)	566 / 26 / 0	567 / 26 / 0	0	0
GB300 (sm_103a, 152 SMs)	pending	pending

The difference of one passed test is the removed test. A coverage run of the changed tests on GB200 executed every changed reference call except the one in test_gpu_router_tmem_head_tail_program_in_its_windows, which skips on sm_100a; GB300 runs it.

• Benchmark on GB200, five rows (projection forward and FP32 weight gradient, MLA forward, router forward and weight gradient) with --memory --accuracy: every row ran; the det arm's convenience memory equals main on every row.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

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

• X
• Mastodon
• Reddit
• LinkedIn

[yyihuang]

@flashinfer-bot run tests/experimental/test_cake_dense_projection_gemm.py tests/experimental/test_cake_dense_projection_det_gemm_names.py

[yyihuang]

Superseded by #8926, which is now the dense GEMM family PR for the #8928 rename (both modes, no deterministic= keyword). Closing this test-only PR.