From 75d011342e32448bc2611a3c7b5c9b7e1f2b8102 Mon Sep 17 00:00:00 2001 From: Jin Li <59594262+liji-nv@users.noreply.github.com> Date: Mon, 1 Dec 2025 23:45:14 -0800 Subject: [PATCH 1/3] [https://nvbugs/5467531][fix] Unwaive fused_moe all to all test with DeepEPLowLatency Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> --- tests/unittest/_torch/modules/test_fused_moe.py | 4 ---- 1 file changed, 4 deletions(-) diff --git a/tests/unittest/_torch/modules/test_fused_moe.py b/tests/unittest/_torch/modules/test_fused_moe.py index 14acccf452f6..cea8ff57e299 100644 --- a/tests/unittest/_torch/modules/test_fused_moe.py +++ b/tests/unittest/_torch/modules/test_fused_moe.py @@ -335,10 +335,6 @@ def per_rank_test_fused_moe_alltoall(job_id): ], ids=lambda s: s.name) def test_fused_moe_alltoall_fp4(alltoall_method_type): - - if alltoall_method_type == AlltoallMethodType.DeepEPLowLatency: - pytest.skip("Skipped due to https://nvbugs/5467531") - world_size = 4 dtype = torch.bfloat16 HIDDEN_SIZE = 4096 From 610d74f18dd87959a8de2f3a6826de7c0ca59967 Mon Sep 17 00:00:00 2001 From: Jin Li <59594262+liji-nv@users.noreply.github.com> Date: Tue, 2 Dec 2025 01:44:31 -0800 Subject: [PATCH 2/3] Fix MI merge conflict Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> --- tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py index 62fee47852e9..fcc5ffbf6233 100755 --- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py +++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py @@ -976,7 +976,7 @@ def forward_fake( all_rank_num_tokens=all_rank_num_tokens, use_dp_padding=use_dp_padding, **kwargs) - if self.alltoall_method_type == AlltoallMethodType.MNNVL: + if self.alltoall_method_type == AlltoallMethodType.NVLinkTwoSided: shape = moe_output.shape top_k = self.routing_method.experts_per_token new_shape = [shape[0], top_k, shape[1]] From 2544475f64d3646f1f98fe2de8d4ba6e5e16c4e2 Mon Sep 17 00:00:00 2001 From: Jin Li <59594262+liji-nv@users.noreply.github.com> Date: Wed, 3 Dec 2025 02:03:19 -0800 Subject: [PATCH 3/3] Fix another conflict on shape Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> --- tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py index fcc5ffbf6233..8fab75d77e26 100755 --- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py +++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py @@ -966,6 +966,7 @@ def forward_fake( output_dtype: Optional[torch.dtype] = None, all_rank_num_tokens: Optional[List[int]] = None, use_dp_padding: Optional[bool] = None, + alltoall_result_do_sum: bool = True, **kwargs, ) -> Union[torch.Tensor, List[torch.Tensor]]: moe_output = super().forward_fake( @@ -976,7 +977,7 @@ def forward_fake( all_rank_num_tokens=all_rank_num_tokens, use_dp_padding=use_dp_padding, **kwargs) - if self.alltoall_method_type == AlltoallMethodType.NVLinkTwoSided: + if self.alltoall_method_type == AlltoallMethodType.NVLinkTwoSided and not alltoall_result_do_sum: shape = moe_output.shape top_k = self.routing_method.experts_per_token new_shape = [shape[0], top_k, shape[1]]