EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

correct the tp_plan logic (#8850)

Browse Source
This commit is contained in:
Stefan He
2025-08-07 16:53:34 -07:00
committed by GitHub
parent 3e7ff1ab1f
commit d3be97104b
1 changed files with 2 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/models/transformers.py
View File

@@ -211,16 +211,13 @@ class TransformersForCausalLM(nn.Module):
Apply the model's tensor parallelization plan. Apply the model's tensor parallelization plan.
Currently only supports linear layers. Currently only supports linear layers.
""" """
if not self.model.supports_tp_plan: tp_plan = getattr(self.model.config, "base_model_tp_plan", None) or {}
if tp_size <= 1:
return
if not tp_plan and self.tp_size > 1:
raise ValueError( raise ValueError(
f"{type(self.model)} does not support tensor parallel yet!" f"{type(self.model)} does not support tensor parallel yet!"
) )
tp_plan = getattr(self.model.config, "base_model_tp_plan", None) or {}
def _tensor_parallel(module: nn.Module, prefix: str = ""): def _tensor_parallel(module: nn.Module, prefix: str = ""):
for child_name, child_module in module.named_children(): for child_name, child_module in module.named_children():
qual_name = maybe_prefix(prefix, child_name) qual_name = maybe_prefix(prefix, child_name)