From da8c7b8c7b3ab9562637ce32d23490897f09d156 Mon Sep 17 00:00:00 2001 From: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com> Date: Thu, 18 Jun 2026 15:29:36 -0700 Subject: [PATCH 1/2] [nvbugs/6336801][fix] Sync thop kwarg allowlists with call site Signed-off-by: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com> --- tensorrt_llm/_torch/attention_backend/fmha/fallback.py | 1 + tests/unittest/_torch/attention/test_attention_op_sync.py | 8 ++++++++ 2 files changed, 9 insertions(+) diff --git a/tensorrt_llm/_torch/attention_backend/fmha/fallback.py b/tensorrt_llm/_torch/attention_backend/fmha/fallback.py index 73c5778379d8..d3a58cb1660d 100644 --- a/tensorrt_llm/_torch/attention_backend/fmha/fallback.py +++ b/tensorrt_llm/_torch/attention_backend/fmha/fallback.py @@ -39,6 +39,7 @@ "topk_indices", # DSA-only "attention_mask_data", # custom-mask code path "out_scale_sf", # promoted into ``out_scale`` in ``TrtllmAttention.forward`` for NVFP4 path + "multi_item_part_lens", # rejected by TrtllmAttention.forward; FlashInfer-only path } ) diff --git a/tests/unittest/_torch/attention/test_attention_op_sync.py b/tests/unittest/_torch/attention/test_attention_op_sync.py index f85667b5a85e..301fbff79ac6 100644 --- a/tests/unittest/_torch/attention/test_attention_op_sync.py +++ b/tests/unittest/_torch/attention/test_attention_op_sync.py @@ -74,6 +74,14 @@ "metadata", ("max_total_draft_tokens",), ), + "skip_softmax_threshold_scale_factor_decode": ( + "skip_softmax_kernel_params", + ("threshold_scale_factor_decode",), + ), + "skip_softmax_threshold_scale_factor_prefill": ( + "skip_softmax_kernel_params", + ("threshold_scale_factor_prefill",), + ), "workspace_": ("metadata", ("effective_workspace",)), } From 796bc01b54589798b8da4a6988ac271667b5cdc8 Mon Sep 17 00:00:00 2001 From: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com> Date: Fri, 26 Jun 2026 20:56:56 -0700 Subject: [PATCH 2/2] [nvbugs/6336801][fix] Drop stale beam_width alias and unused multi_item_part_lens exclusion Signed-off-by: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com> --- tensorrt_llm/_torch/attention_backend/fmha/fallback.py | 1 - tests/unittest/_torch/attention/test_attention_op_sync.py | 1 - 2 files changed, 2 deletions(-) diff --git a/tensorrt_llm/_torch/attention_backend/fmha/fallback.py b/tensorrt_llm/_torch/attention_backend/fmha/fallback.py index d3a58cb1660d..73c5778379d8 100644 --- a/tensorrt_llm/_torch/attention_backend/fmha/fallback.py +++ b/tensorrt_llm/_torch/attention_backend/fmha/fallback.py @@ -39,7 +39,6 @@ "topk_indices", # DSA-only "attention_mask_data", # custom-mask code path "out_scale_sf", # promoted into ``out_scale`` in ``TrtllmAttention.forward`` for NVFP4 path - "multi_item_part_lens", # rejected by TrtllmAttention.forward; FlashInfer-only path } ) diff --git a/tests/unittest/_torch/attention/test_attention_op_sync.py b/tests/unittest/_torch/attention/test_attention_op_sync.py index 301fbff79ac6..3aeec7fa00a9 100644 --- a/tests/unittest/_torch/attention/test_attention_op_sync.py +++ b/tests/unittest/_torch/attention/test_attention_op_sync.py @@ -63,7 +63,6 @@ } _THOP_KWARG_SOURCE_ALIASES: dict[str, tuple[str, tuple[str, ...]]] = { - "beam_width": ("metadata", ("effective_beam_width",)), "context_lengths": ("metadata", ("prompt_lens_cuda_runtime",)), "head_size": ("attn", ("head_dim",)), "host_context_lengths": ("metadata", ("prompt_lens_cpu_runtime",)),