[BugFix] Fix mtp empty run issue in overlap schedule and EP model (#7300)

This commit is contained in:
sunxin
2026-04-10 18:29:45 +08:00
committed by GitHub
parent 627f0d9cc8
commit 00005c92e0
+6
View File
@@ -2116,6 +2116,12 @@ class GPUModelRunner(ModelRunnerBase):
self._cached_sampler_output = sampler_output
self._cached_post_process_event = post_process_event
else:
if (
self.fd_config.speculative_config.method == SpecMethod.MTP
and hasattr(self.proposer.model, "empty_input_forward")
and self.parallel_config.use_ep
):
self._execute_empty_mtp_input(self.forward_meta)
self._cached_model_output_data = None
self._cached_sampler_output = None
self._cached_post_process_event = None