EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[OAI] Support non-normalized logprobs in OpenAI server (#5961)

Browse Source
This commit is contained in:
Chang Su
2025-05-24 21:35:55 -07:00
committed by GitHub
parent 681fdc264b
commit 681e7af32b
1 changed files with 4 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/sglang/srt/layers/sampler.py
View File

@@ -86,11 +86,9 @@ class Sampler(nn.Module):
# NOTE: the top_p_renorm_prob from flashinfer has numerical problems,
# https://github.com/flashinfer-ai/flashinfer/issues/708
# so we use the torch implementation.
# clamp to avoid -inf
logprobs = torch.log(
top_p_normalize_probs_torch(probs, sampling_info.top_ps)
).clamp(min=torch.finfo(probs.dtype).min)
# NOTE: OpenAI's logprobs is independent of top-p, we use the
# same rule.
logprobs = torch.log(probs).clamp(min=torch.finfo(probs.dtype).min)
max_top_k_round, batch_size = 32, probs.shape[0]
if sampling_info.need_min_p_sampling:
@@ -121,10 +119,7 @@ class Sampler(nn.Module):
)
if return_logprob:
# clamp to avoid -inf
logprobs = torch.log(
top_p_normalize_probs_torch(probs, sampling_info.top_ps)
).clamp(min=torch.finfo(probs.dtype).min)
logprobs = torch.log(probs).clamp(min=torch.finfo(probs.dtype).min)
else:
raise ValueError(
f"Invalid sampling backend: {global_server_args_dict['sampling_backend']}"