[BugFix] Fix use of per-request seed with pipeline parallel (#6698)

This commit is contained in:
Nick Hill
2024-07-30 10:40:08 -07:00
committed by GitHub
parent f058403683
commit 5cf9254a9c
21 changed files with 222 additions and 137 deletions
+9 -5
View File
@@ -1264,11 +1264,15 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
"""
model_input = self._prepare_model_input_tensors(
seq_group_metadata_list, finished_requests_ids)
sampling_metadata = SamplingMetadata.prepare(seq_group_metadata_list,
model_input.seq_lens,
model_input.query_lens,
self.device,
self.pin_memory)
if get_pp_group().is_last_rank:
# Sampling metadata is only required for the final pp group
generators = self.get_generators(finished_requests_ids)
sampling_metadata = SamplingMetadata.prepare(
seq_group_metadata_list, model_input.seq_lens,
model_input.query_lens, self.device, self.pin_memory,
generators)
else:
sampling_metadata = None
is_prompt = (seq_group_metadata_list[0].is_prompt
if seq_group_metadata_list else None)
return dataclasses.replace(model_input,