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 8, 10:00 UTC
[Bug Fix] Fix `naive_block_assignment` always defaulting to False due to arg misalignment (#33848)
Signed-off-by: Runkai Tao <rt572@physics.rutgers.edu>
Runkai Tao committed
Feb 11, 2026 at 22:30 UTC
e1d97c38f8689da0b11da0fac54cc277c237d5c4
2 files changed
+2
-1
vllm/lora/layers/fused_moe.py
+1
-1
@@ -219,7 +219,7 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
219
self.max_loras,
220
self.adapter_enabled,
221
expert_map,
222
- naive_block_assignment,
222
+ naive_block_assignment=naive_block_assignment,
223
)
224
225
moe_state_dict["sorted_token_ids_lora"] = sorted_token_ids_lora
vllm/lora/punica_wrapper/punica_base.py
+1
@@ -458,6 +458,7 @@ class PunicaWrapperBase(PunicaWrapperABC):
458
adapter_enabled: torch.Tensor,
459
expert_map: torch.Tensor | None = None,
460
pad_sorted_ids: bool = False,
461
+ naive_block_assignment: bool = False,
462
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
463
"""
464
Aligns tokens and experts into block-sized chunks for LoRA-based