[Bugfix] Autotune FlashInfer deferred MoE decode kernels before CUDA graph capture (#55377)

Signed-off-by: ming <ming@inferact.ai>
Co-authored-by: Kimi Code <kimi-code@moonshot.cn>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
mingg26
2026-09-08 04:02:23 +00:00
committed by GitHub
co-authored by Kimi Code mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent eb6b619ab2
commit 7fa2c63796
3 changed files with 159 additions and 8 deletions
+4
View File
@@ -247,6 +247,9 @@ flashinfer_convert_sf_to_mma_layout = _lazy_import_wrapper(
flashinfer_b12x_fused_moe = _lazy_import_wrapper(
"flashinfer.fused_moe", "b12x_fused_moe"
)
flashinfer_get_hybrid_num_tokens_buckets = _lazy_import_wrapper(
"flashinfer.fused_moe.utils", "get_hybrid_num_tokens_buckets"
)
trtllm_fp4_block_scale_moe = _lazy_import_wrapper(
"flashinfer", "trtllm_fp4_block_scale_moe"
)
@@ -1286,6 +1289,7 @@ __all__ = [
"nvfp4_block_scale_interleave",
"flashinfer_cute_dsl_fused_moe_nvfp4",
"flashinfer_b12x_fused_moe",
"flashinfer_get_hybrid_num_tokens_buckets",
"flashinfer_convert_sf_to_mma_layout",
"trtllm_fp4_block_scale_moe",
"flashinfer_trtllm_batch_decode_with_kv_cache_mla",