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

Author: CatherineSue

python/sglang/srt/layers/sampler.py

@@ -86,11 +86,9 @@ def forward(
                     # 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 @@ def forward(
                 )
 
                 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']}"