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, 18:15 UTC
[DCP] check return_lse for all layers in dcp (#27929)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Chen Zhang committed
Nov 5, 2025 at 06:27 UTC
5d16d0fa629eb8ff3d92b0f1eed06548c5277e81
1 file changed
+1
-4
vllm/v1/worker/gpu_model_runner.py
+1
-4
@@ -4680,10 +4680,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
4680
kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
4681
4682
if self.dcp_world_size > 1:
4683
- layer_names = self.attn_groups[0][0].layer_names
4684
- layers = get_layers_from_vllm_config(
4685
- self.vllm_config, AttentionLayerBase, layer_names
4686
- )
4683
+ layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
4684
for layer in layers.values():
4685
assert layer.impl.need_to_return_lse_for_decode, (
4686
"DCP requires attention impls to return"