Optimize conflicts between CUDA graph and vocab mask tensors (#1392)

python/sglang/srt/models/chatglm.py

@@ -42,7 +42,6 @@ from sglang.srt.layers.activation import SiluAndMul
 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
 
 LoraConfig = None
@@ -371,7 +370,6 @@ class ChatGLMForCausalLM(nn.Module):
         self.transformer = ChatGLMModel(config, cache_config, quant_config)
         self.lm_head = self.transformer.output_layer
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -381,11 +379,9 @@ class ChatGLMForCausalLM(nn.Module):
         input_metadata: InputMetadata,
     ) -> torch.Tensor:
         hidden_states = self.transformer(input_ids, positions, input_metadata)
-        logits_output = self.logits_processor(
+        return 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]]):
         params_dict = dict(self.named_parameters(remove_duplicate=False))