Simplify stream_output (#2398)

This commit is contained in:
Lianmin Zheng
2024-12-08 12:27:13 -08:00
committed by GitHub
parent f62055b528
commit a6ca736c8e
9 changed files with 426 additions and 290 deletions

View File

@@ -400,9 +400,14 @@ class CudaGraphRunner:
forward_mode=ForwardMode.DECODE,
top_logprobs_nums=forward_batch.top_logprobs_nums,
)
logits_output.output_top_logprobs = LogitsProcessor.get_top_logprobs(
(
logits_output.output_top_logprobs_val,
logits_output.output_top_logprobs_idx,
) = LogitsProcessor.get_top_logprobs(
next_token_logprobs, logits_metadata
)[1]
)[
2:4
]
else:
logits_output = LogitsProcessorOutput(
next_token_logits=next_token_logits,