Add customized sampler registration (#15423)

python/sglang/srt/sampling/sampling_batch_info.py

@@ -89,8 +89,15 @@ class SamplingBatchInfo:
         )
         sampling_seed = (
             torch.tensor(
-                [r.sampling_params.sampling_seed for r in reqs],
-                dtype=torch.int32,
+                [
+                    (
+                        r.sampling_params.sampling_seed
+                        if r.sampling_params.sampling_seed is not None
+                        else 42
+                    )
+                    for r in reqs
+                ],
+                dtype=torch.int64,
                 device=device,
             )
             if enable_deterministic
@@ -173,8 +180,13 @@ class SamplingBatchInfo:
             device=device,
             logit_bias=logit_bias,
         )
+        ret.adjusted_from_schedule_batch(batch, vocab_size)
         return ret
 
+    # placeholder for override
+    def adjusted_from_schedule_batch(self, batch: ScheduleBatch, vocab_size: int):
+        pass
+
     def __len__(self):
         return len(self.temperatures)
 

python/sglang/srt/sampling/sampling_params.py

@@ -58,7 +58,7 @@ class SamplingParams:
         custom_params: Optional[Dict[str, Any]] = None,
         stream_interval: Optional[int] = None,
         logit_bias: Optional[Dict[str, float]] = None,
-        sampling_seed: int = 42,
+        sampling_seed: Optional[int] = None,
     ) -> None:
         self.max_new_tokens = max_new_tokens
         self.stop_strs = stop
@@ -146,8 +146,6 @@ class SamplingParams:
                         f"logit_bias must has keys in [0, {vocab_size - 1}], got "
                         f"{token_id}."
                     )
-        if self.sampling_seed is None:
-            raise ValueError("sampling_seed should not be None")
 
         grammars = [
             self.json_schema,