[Feat] Return hidden states (experimental) (#3364)

Co-authored-by: Chayenne <zhaochen20@outlook.com>
This commit is contained in:
Jackmin801
2025-02-10 15:54:37 -08:00
committed by GitHub
parent 2f47d710ae
commit 5f0e7de339
12 changed files with 204 additions and 5 deletions

View File

@@ -156,6 +156,10 @@ class TpModelWorkerClient:
logits_output.input_token_logprobs = (
logits_output.input_token_logprobs.to("cpu", non_blocking=True)
)
if logits_output.hidden_states is not None:
logits_output.hidden_states = logits_output.hidden_states.to(
"cpu", non_blocking=True
)
next_token_ids = next_token_ids.to("cpu", non_blocking=True)
copy_done.record()