[Bugfix][DBO] Fix DeepEP low-latency profiling crash with DP+EP+DBO (#57502)
Signed-off-by: Micah Williamson <micah.williamson@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Micah Williamson committed
Sep 18, 2026 at 01:33 UTC
d5f0a6e829faa69d1db289bf62b14dae136c02b2
1 file changed
+5
-2
vllm/v1/worker/gpu_model_runner.py
+5
-2
@@ -6164,6 +6164,11 @@ class GPUModelRunner(
6164
num_tokens_padded, None, False
6165
)
6166
6167
+ num_tokens_unpadded = num_tokens_padded if is_profile else 0
6168
+ is_padding = self._prepare_padding_mask(
6169
+ num_tokens_unpadded, num_tokens_padded
6170
+ )
6171
+
6172
if ubatch_slices_padded is not None:
6173
# Adjust values to reflect a single ubatch.
6174
# TODO(sage,lucas): this is cruft that should be addressed in
@@ -6172,8 +6177,6 @@ class GPUModelRunner(
6177
if num_tokens_across_dp is not None:
6178
num_tokens_across_dp[:] = num_tokens_padded
6179
6175
- is_padding = self._prepare_padding_mask(0, num_tokens_padded)
6176
-
6180
with (
6181
self.maybe_randomize_inputs(
6182
input_ids, inputs_embeds, randomize_inputs=randomize_inputs