Tiny skip_sample adjust (#11225)

python/sglang/srt/managers/schedule_batch.py

@@ -663,7 +663,11 @@ class Req:
     @property
     def is_prefill_only(self) -> bool:
         """Check if this request is prefill-only (no token generation needed)."""
-        return self.sampling_params.max_new_tokens == 0
+        # NOTE: when spec is enabled, prefill_only optimizations are disabled
+        return (
+            self.sampling_params.max_new_tokens == 0
+            and global_server_args_dict["speculative_algorithm"] is None
+        )
 
     def add_latency(self, stage: RequestStage):
         if self.metrics_collector is None:

python/sglang/srt/managers/tp_worker.py

@@ -237,7 +237,7 @@ class TpModelWorker:
         self,
         model_worker_batch: ModelWorkerBatch,
         launch_done: Optional[threading.Event] = None,
-        skip_sample: bool = False,
+        is_verify: bool = False,
     ) -> ForwardBatchOutput:
         # update the consumer index of hicache to the running batch
         self.set_hicache_consumer(model_worker_batch.hicache_consumer_index)
@@ -259,19 +259,16 @@ class TpModelWorker:
             if launch_done is not None:
                 launch_done.set()
 
-            if skip_sample:
+            skip_sample = is_verify or model_worker_batch.is_prefill_only
-                next_token_ids = None
+            next_token_ids = None
-                # For prefill-only requests, we still need to compute logprobs even when sampling is skipped
+
-                if (
+            if not skip_sample:
-                    model_worker_batch.is_prefill_only
-                    and model_worker_batch.return_logprob
-                ):
-                    # Compute logprobs without full sampling
-                    self.model_runner.compute_logprobs_only(
-                        logits_output, model_worker_batch
-                    )
-            else:
                 next_token_ids = self.model_runner.sample(logits_output, forward_batch)
+            elif model_worker_batch.return_logprob and not is_verify:
+                # NOTE: Compute logprobs without full sampling
+                self.model_runner.compute_logprobs_only(
+                    logits_output, model_worker_batch
+                )
 
             return ForwardBatchOutput(
                 logits_output=logits_output,

python/sglang/srt/managers/tp_worker_overlap_thread.py

@@ -164,8 +164,6 @@ class TpModelWorkerClient:
             forward_batch_output = self.worker.forward_batch_generation(
                 model_worker_batch,
                 model_worker_batch.launch_done,
-                #  Skip sampling for prefill-only requests
-                skip_sample=model_worker_batch.is_prefill_only,
             )
 
             logits_output, next_token_ids, can_run_cuda_graph = (

python/sglang/srt/speculative/eagle_worker.py

@@ -823,7 +823,7 @@ class EAGLEWorker(TpModelWorker):
 
         # Forward
         forward_batch_output = self.target_worker.forward_batch_generation(
-            model_worker_batch, skip_sample=True
+            model_worker_batch, is_verify=True
         )
         logits_output, can_run_cuda_graph = (
             forward_batch_output.logits_output,

python/sglang/srt/speculative/ngram_worker.py

@@ -214,7 +214,7 @@ class NGRAMWorker:
 
         if model_worker_batch.forward_mode.is_target_verify():
             forward_batch_output = self.target_worker.forward_batch_generation(
-                model_worker_batch, skip_sample=True
+                model_worker_batch, is_verify=True
             )
             logits_output, can_run_cuda_graph = (
                 forward_batch_output.logits_output,