Move sampler into CUDA graph (#1201)

Co-authored-by: Yineng Zhang <me@zhyncs.com>

python/sglang/srt/models/grok.py

@@ -46,6 +46,7 @@ from sglang.srt.layers.fused_moe import FusedMoE
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -297,6 +298,7 @@ class Grok1ModelForCausalLM(nn.Module):
         self.model = Grok1Model(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
 
         # Monkey patch _prepare_weights to load pre-sharded weights
         setattr(DefaultModelLoader, "_prepare_weights", _prepare_presharded_weights)
@@ -313,9 +315,11 @@ class Grok1ModelForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        return self.logits_processor(
+        logits_output = self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
+        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
+        return sample_output, logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [