support mtp with deepseek r1 nvfp4 model (#13115)

Co-authored-by: Trevor Morris <tmorris@nvidia.com>
This commit is contained in:
Rain Jiang
2025-12-06 00:45:54 -08:00
committed by GitHub
parent 42fcf5438f
commit ea177372bd
11 changed files with 123 additions and 59 deletions

View File

@@ -764,7 +764,12 @@ class ForwardBatch:
bs = self.batch_size
if self.forward_mode.is_decode():
if (
self.forward_mode.is_decode()
or self.forward_mode.is_target_verify()
or self.forward_mode.is_draft_extend(include_v2=True)
or self.forward_mode.is_idle()
):
if self.is_extend_in_batch and dp_padding_mode.is_max_len():
setattr(self, "_original_forward_mode", self.forward_mode)
self.forward_mode = ForwardMode.EXTEND