Fix/mxfp4 a4w4 bm16 streaming accuracy#3
Open
benenzhu wants to merge 2 commits into
Open
Conversation
The FlyDSL gemm2 a4w4 port lost accuracy on the BM16 (M<=128) streaming K-loop for K_TILES_TOTAL==8 (D_INTER=2048, e.g. dsv3_a): the triple-buffered (3-slot) A->LDS pipeline did not reliably drain the async raw.ptr.buffer.load.lds of a streamed slot before that slot's ds_read 3 iterations later, producing non-deterministic garbage (e2e logits_diff ~0.02-0.05 vs the ~0.009 fp4 floor; cosine ~0.90). Other K_TILES (4/6/10/12) happened to schedule cleanly at 3 slots. - Widen the streaming A-slot buffer from 3 to 4 (quad buffer): adds slack between a slot's read and its next overwrite, hiding the load latency and making all K_TILES_TOTAL deterministic. LDS cost is free -- the s_Aq slots stay <= the lds_acc cshuffle union they overlap, so per-WG LDS is unchanged. - Gate the streaming issue_a_load_lds on wave < n_load_waves (matching the prologue): at BM16 only waves 0,1 hold A rows; ungated, waves 2,3 streamed into lds_row 16/24 (past the BM16 slot) from the next m-block's rows. Verified: dsv3_a / dsv4_ep8 / kimi / dsv3_b / minimax_a / dsv4_tp2 all at the ~0.009 fp4 floor and deterministic across runs; large-M (BM32/BM128) unchanged; 33 standalone gemm1/gemm2 port tests pass. Co-Authored-By: Claude Opus 4 (1M context) <noreply@anthropic.com>
gemm_moe_tune.py's run_config benchmarked a4w4 (fp4 act + fp4 weight) with the wrong weight layout and gate_mode, so its accuracy check reported garbage (err_ratio ~1.0, 99.7% elements off) for every a4w4 shape -- the same class of bug already fixed in op_tests/test_moe_2stage.py (commit 6a580ab) but missed in the tuner. - include fp4x2 activations in the a16w4 shuffle branch so a4w4 gets shuffle_weight_a16w4 / shuffle_scale_a16w4 like a16w4/a8w4 (the FlyDSL port the a4w4 MoE runs on expects the gate-up-interleaved a16w4 layout). - pass gate_mode=interleave to fused_moe for mxfp4 weights (a4w4/a8w4/a16w4) and mxfp8, mirroring op_tests _effective_gate_mode (the a4w4 tuned rows are interleave-keyed; the default SEPARATED reaches the wrong dispatch). - add the top-level GateMode import the new code needs. Co-Authored-By: Claude Opus 4 (1M context) <noreply@anthropic.com>
ftyghome
force-pushed
the
dev/randomflow_pr
branch
from
June 24, 2026 01:37
4c1a5c9 to
e333923
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.