[Performance][Spec Decode] Optimize ngram lookup performance (vllm-pr…

…oject#9333)

vllm/spec_decode/ngram_worker.py

@@ -67,9 +67,16 @@ def sampler_output(
                 execute_model_req.seq_group_metadata_list):
             seq_data = next(iter(seq_group_metadata.seq_data.values()))
 
+            seq_len = seq_data.get_len()
+            # When seq_len is less than 3072 (3K), we use CPU to perform
+            # the ngram match. Otherwise, we use the device specified in
+            # the model config (normally GPU). 3072 is a rough threshold
+            # based on profiling on H100, and it can be adjusted based
+            # on the actual performance on different hardware.
+            cur_device = "cpu" if seq_len < 3072 else self.device
             input_ids = torch.as_tensor(seq_data.get_token_ids(),
                                         dtype=torch.long,
-                                        device=self.device)
+                                        device=cur_device)
             input_length = seq_data.get_len()
 
             for ngram_size in range(
@@ -91,17 +98,15 @@ def sampler_output(
                 # first_match includes "values" (bool), indicating whether
                 # the match is found, and "indices", indicating the index
                 # of the first match.
-                # Note that "first_match.values.item()" triggers GPU-CPU
-                # sync so it is a bit inefficient, but we have not found
-                # a better way to do this.
                 first_match = matches.max(dim=-1)
                 if first_match.values.item():
                     proposal_start_idx = first_match.indices.add_(ngram_size)
                     spec_indices = (
                         proposal_start_idx).repeat(sample_len) + torch.arange(
-                            sample_len, device=self.device)
+                            sample_len, device=cur_device)
                     spec_indices.clamp_(max=input_ids.shape[-1] - 1)
-                    res = input_ids.gather(dim=-1, index=spec_indices)
+                    res = input_ids.gather(dim=-1,
+                                           index=spec_indices).to(self.device)
                     token_id_list.append(res)
                     token_prob_list.append(
                         torch.nn.functional.one_hot(