[V1][Spec Decode] Small refactors to improve eagle bookkeeping performance (#18424)

Signed-off-by: qizixi <[email protected]>

Author: zixi-qi

tests/v1/spec_decode/test_eagle.py

@@ -100,8 +100,12 @@ def test_prepare_inputs():
         dtype=torch.int32,
         device=device)
 
+    # n1 + n2 + n3 - a - b -c
+    num_tokens = cu_target_query_lens[-1].item() - num_rejected_tokens.sum(
+    ).item()
+
     cu_num_tokens, token_indices = EagleProposer.prepare_inputs(
-        cu_target_query_lens, num_rejected_tokens)
+        cu_target_query_lens, num_rejected_tokens, num_tokens)
 
     assert torch.equal(cu_num_tokens, expected_cu_num_tokens)
     assert token_indices.shape[0] == expected_cu_num_tokens[-1].item()

vllm/v1/spec_decode/eagle.py

@@ -271,6 +271,7 @@ def prepare_inputs(
         cu_target_query_lens: torch.Tensor,
         # [batch_size]
         num_rejected_tokens: torch.Tensor,
+        num_tokens: int,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         # cu_target_query_lens: [0, a, a + b, a + b + c]
         # num_rejected_tokens: [n1, n2, n3]
@@ -288,18 +289,13 @@ def prepare_inputs(
 
         # [a - n1, b - n2, c - n3] ->
         # [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
-        cu_num_tokens = torch.empty_like(cu_target_query_lens)
+        cu_num_tokens = torch.zeros_like(cu_target_query_lens)
         torch.cumsum(num_tokens_per_req, dim=0, out=cu_num_tokens[1:])
-        cu_num_tokens[0] = 0
-
-        # FIXME(woosuk): Avoid synchronization.
-        num_tokens = cu_num_tokens[-1].item()
         token_indices = torch.empty(
             num_tokens,
             dtype=torch.int32,
-            device=cu_num_tokens.device,
+            device=cu_target_query_lens.device,
         )
-
         batch_size = num_rejected_tokens.shape[0]
         BLOCK_SIZE = 1024
         prepare_eagle_input_kernel[(batch_size, )](

vllm/v1/worker/gpu_model_runner.py

@@ -34,8 +34,8 @@
 from vllm.sampling_params import SamplingType
 from vllm.sequence import IntermediateTensors
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
-                        GiB_bytes, LazyLoader, cdiv, check_use_alibi,
-                        is_pin_memory_available)
+                        GiB_bytes, LazyLoader, async_tensor_h2d, cdiv,
+                        check_use_alibi, is_pin_memory_available)
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
 from vllm.v1.attention.backends.utils import CommonAttentionMetadata
 from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
@@ -281,7 +281,7 @@ def __init__(
     def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> bool:
         """
         Update the order of requests in the batch based on the attention
-        backend's needs. For example, some attention backends (namely MLA) may 
+        backend's needs. For example, some attention backends (namely MLA) may
         want to separate requests based on if the attention computation will be
         compute-bound or memory-bound.
 
@@ -1360,9 +1360,10 @@ def execute_model(
                                scheduler_output.num_scheduled_tokens[req_id])
                     next_token_id = req_state.get_token_id(seq_len)
                 next_token_ids.append(next_token_id)
-            next_token_ids = torch.tensor(next_token_ids,
-                                          dtype=torch.int32,
-                                          device=self.device)
+            next_token_ids = async_tensor_h2d(next_token_ids,
+                                              dtype=torch.int32,
+                                              target_device=self.device,
+                                              pin_memory=True)
             eagle_attn_metadata = attn_metadata[self.drafter.attn_layer_name]
 
             # NOTE: deepseek_mtp uses MLA which does not have `block_table`
@@ -1390,14 +1391,16 @@ def execute_model(
                     n + 1 - len(valid_sampled_token_ids[i]) if n > 0 else 0
                     for i, n in enumerate(num_draft_tokens)
                 ]
-                num_rejected_tokens = torch.tensor(
+                num_rejected_tokens_tensor = async_tensor_h2d(
                     num_rejected_tokens,
                     dtype=torch.int32,
-                    device=self.device,
-                )
+                    target_device=self.device,
+                    pin_memory=True)
+                num_tokens = num_scheduled_tokens - sum(num_rejected_tokens)
                 cu_num_tokens, token_indices = self.drafter.prepare_inputs(
                     eagle_attn_metadata.query_start_loc,
-                    num_rejected_tokens,
+                    num_rejected_tokens_tensor,
+                    num_tokens,
                 )
                 target_token_ids = self.input_ids[token_indices]
                 target_positions = positions[token_indices]
@@ -1408,7 +1411,6 @@ def execute_model(
                     target_hidden_states = hidden_states[token_indices]
                 target_slot_mapping = eagle_attn_metadata.slot_mapping[
                     token_indices]
-
             draft_token_ids = self.drafter.propose(
                 target_token_ids=target_token_ids,
                 target_positions=target_positions,