[Model][NVIDIA] Route DSA models to the CUDA non-compiled path (#52861)
This commit is contained in:
@@ -1018,7 +1018,11 @@ class SpecDecodeBaseProposer:
|
||||
# feedback into the next draft step.
|
||||
architectures = self.draft_model_config.hf_config.architectures or []
|
||||
return bool(
|
||||
{"DeepSeekMTPModel", "KimiK3MTPModel"}.intersection(architectures)
|
||||
{
|
||||
"DeepSeekMTPModel",
|
||||
"DeepseekV32MTPModel",
|
||||
"KimiK3MTPModel",
|
||||
}.intersection(architectures)
|
||||
)
|
||||
return self.method not in ("mtp", "draft_model", "dflash")
|
||||
|
||||
|
||||
Reference in New Issue
Block a user