From d5f0a6e829faa69d1db289bf62b14dae136c02b2 Mon Sep 17 00:00:00 2001 From: Micah Williamson Date: Fri, 18 Sep 2026 01:33:41 -0500 Subject: [PATCH] [Bugfix][DBO] Fix DeepEP low-latency profiling crash with DP+EP+DBO (#57502) Signed-off-by: Micah Williamson Co-authored-by: Andreas Karatzas --- vllm/v1/worker/gpu_model_runner.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index c943f02212..cedf95e319 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -6164,6 +6164,11 @@ class GPUModelRunner( num_tokens_padded, None, False ) + num_tokens_unpadded = num_tokens_padded if is_profile else 0 + is_padding = self._prepare_padding_mask( + num_tokens_unpadded, num_tokens_padded + ) + if ubatch_slices_padded is not None: # Adjust values to reflect a single ubatch. # TODO(sage,lucas): this is cruft that should be addressed in @@ -6172,8 +6177,6 @@ class GPUModelRunner( if num_tokens_across_dp is not None: num_tokens_across_dp[:] = num_tokens_padded - is_padding = self._prepare_padding_mask(0, num_tokens_padded) - with ( self.maybe_randomize_inputs( input_ids, inputs_embeds, randomize_inputs=randomize_inputs