Tyler Michael SmithandClaude Opus 4.6 0d92fc63fd [Bugfix] Fix FlashInfer NVFP4 cross-row scale corruption in MoE quant ...

FlashInfer's silu_and_mul_scaled_nvfp4_experts_quantize and
scaled_fp4_grouped_quantize kernels corrupt real token scales when
padding rows (beyond masked_m) contain NaN or garbage data.

This affects the FlashInferCuteDSLBatchedExperts MoE path used with
NVFP4 weights (e.g. nvidia/DeepSeek-R1-0528-NVFP4-v2 with DeepEP LL).
The corruption produces wrong finite values (silent accuracy
degradation) that compound across layers.

Fix: zero-fill padding rows in flashinfer_cutedsl_moe_masked before
calling the FlashInfer quantization kernels.

Tests:
- test_silu_quant_cross_row_corruption: direct kernel test (xfail,
  proving the underlying FlashInfer kernel bug exists)
- test_grouped_quant_cross_row_corruption: direct kernel test (xfail)
- test_cutedsl_wrapper_nan_padding: wrapper test (PASSES with fix)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

2026-04-13 17:58:14 -04:00

..

attention

Add platform manual_seed_all API (#38468)

2026-04-10 13:43:50 +08:00

core

[Kernel] Porting the TRTLLM minimax_allreduce_rms kernels (#37045)

2026-04-11 00:20:20 +08:00

helion

[release 2.11] Update to torch 2.11 (#34644)

2026-04-07 18:55:48 -07:00

ir

[vLLM IR] 1/N Implement IR skeleton and rms_norm op (#33825)

2026-03-31 22:15:05 -04:00

mamba

[Bugfix] Use null block (0) for padded block table entries (#35431)

2026-03-30 14:02:51 -07:00

moe

[Bugfix] Fix FlashInfer NVFP4 cross-row scale corruption in MoE quant

2026-04-13 17:58:14 -04:00

quantization

[Perf] Fuse Zero Initializer for FP8 DeepGemm Block Quant Kernel (#39547)

2026-04-11 07:16:51 -07:00

__init__.py

[CI/Build] Move test_utils.py to tests/utils.py (#4425)

2024-05-13 23:50:09 +09:00

allclose_default.py

Convert formatting to use ruff instead of yapf + isort (#26247)

2025-10-05 07:06:22 -07:00

quant_utils.py

[Perf] Create TMA-aligned input scale tensor for DeepGemm on Hopper (#32619)

2026-01-22 15:47:04 -05:00

test_apply_repetition_penalties.py

[Platform] Deprecate seed_everything (#31659)

2026-01-04 18:34:04 -08:00

test_awq_int4_to_int8.py

[CPU] Support int8 compute mode in CPU AWQ (#35697)

2026-03-31 15:27:37 +08:00

test_cache_kernels.py

[Hardware] Replace torch.cuda.device_count/current_device/set_device API (#36145)

2026-03-12 07:57:47 -07:00

test_concat_mla_q.py

[Attention][Perf][Kernel] Replace torch.cat with vectorized CUDA kernel MLA query concat - DeepSeek-V3.2 (#34917)

2026-03-09 09:50:36 -07:00

test_cp_gather_fp8.py

[Attention][Perf] Optimize cp_gather_and_upconvert_fp8_kv_cache - DeepSeek-v3.2 (#35290)

2026-03-09 09:46:57 -07:00

test_fla_layernorm_guard.py

[Bugfix] Fix dtype mismatch in RMSNormGated.forward_native() during torch.compile (#35256)

2026-03-01 15:14:46 -05:00

test_flex_attention.py

full cudagraph for flex-attn (#36298)

2026-04-02 21:15:01 -07:00

test_fused_gdn_post_conv.py

[Perf] fuse kernels in gdn (#37813)

2026-04-02 11:52:18 +00:00

test_fused_quant_activation.py

Add platform manual_seed_all API (#38468)

2026-04-10 13:43:50 +08:00

test_fused_recurrent_packed_decode.py

[Bugfix] fix main branch pre-commit error (1 line change) (#36897)

2026-03-12 09:08:37 -07:00

test_fused_sigmoid_gating_delta_rule.py

[Kernel] Add fused_sigmoid_gating_delta_rule_update kernel for Qwen3 Next (#35777)

2026-03-08 23:41:01 -07:00

test_onednn.py

[CPU] Refactor CPU attention backend (#27954)

2025-11-12 09:43:06 +08:00

test_shuffle_rows.py

Convert formatting to use ruff instead of yapf + isort (#26247)

2025-10-05 07:06:22 -07:00

test_top_k_per_row.py

[Perf][Kernel] Persistent TopK scheduler: unified CUDAGraph-safe kernel with dynamic per-row dispatch - DeepSeek-V3.2 DSA decode (#37421)

2026-04-08 13:35:57 -04:00

utils.py

[ROCm] Fix MoE kernel test failures on gfx950 (#37833)

2026-03-25 13:46:40 -05:00