[Bugfix][Spec Decode] Preserve target parallel config (DCP) for DSpark (#55472)
Signed-off-by: Jing Wang <jingwang96@qq.com>
Jing Wang committed
Sep 9, 2026 at 07:59 UTC
c268198715de0e78a56672c7e9d0b2fd9afdb141
1 file changed
+7
-1
vllm/v1/worker/gpu/spec_decode/dspark/utils.py
+7
-1
@@ -53,7 +53,13 @@ def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mo
53
54
draft_vllm_config = replace(
55
vllm_config,
56
- parallel_config=speculative_config.draft_parallel_config,
56
+ parallel_config=replace(
57
+ vllm_config.parallel_config,
58
+ pipeline_parallel_size=1,
59
+ tensor_parallel_size=(
60
+ speculative_config.draft_parallel_config.tensor_parallel_size
61
+ ),
62
+ ),
63
attention_config=replace(
64
vllm_config.attention_config,
65
use_non_causal=dflash_has_any_non_causal(draft_model_config.hf_config),