A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 03:15 UTC
[BugFix] Fix `AssertionError: DCP not support reorder_batch_threshold > 1 now.` (#28751)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Nov 15, 2025 at 17:35 UTC
be263f76451ad8a32baf0b935d3f0432d05300e6
1 file changed
-10
vllm/v1/worker/gpu_model_runner.py
-10
@@ -630,16 +630,6 @@ class GPUModelRunner(
630
return
631
632
if self.reorder_batch_threshold is not None:
633
- # NOTE(lucas): currently no backend supports the custom masking
634
- # required for DCP with q_len > 1, so we assert here. Remove this
635
- # assert once the custom mask is support is added to FA3.
636
- if (
637
- self.dcp_world_size > 1
638
- and envs.VLLM_ATTENTION_BACKEND != "FLASH_ATTN_MLA"
639
- ):
640
- assert self.reorder_batch_threshold == 1, (
641
- "DCP not support reorder_batch_threshold > 1 now."
642
- )
633
reorder_batch_to_split_decodes_and_prefills(
634
self.input_batch,
635
scheduler_output,