compatible with L40

vllm_hacked/v1/sample/sampler.py

@@ -103,6 +103,7 @@ class Sampler(nn.Module):
             
         #     logits = scores_processed
         #     # logits = logits_processor_stage1(logits)
+        print("logits.shape:", logits.shape)
         if logits.dim() == 1 or (logits.size(0) != sampling_metadata.top_k.size(0)):
             batch_size = sampling_metadata.top_k.size(0)
             # 1. 构造假的采样结果 (batch_size,)