[V1][Spec Decode] Apply torch.compile & cudagraph to EAGLE3 (vllm-project#17504)

Signed-off-by: qizixi <qizixi@meta.com>

Author: zixi-qi

vllm/model_executor/models/llama_eagle3.py

@@ -6,7 +6,8 @@
 import torch.nn as nn
 from transformers import LlamaConfig
 
-from vllm.config import ModelConfig, VllmConfig
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import QKVParallelLinear
@@ -76,17 +77,19 @@ def forward(
         return hidden_states, residual
 
 
+@support_torch_compile
 class LlamaModel(nn.Module):
 
     def __init__(
         self,
         *,
-        model_config: ModelConfig,
+        vllm_config: VllmConfig,
         start_layer_id: int = 0,
         prefix: str = "",
     ) -> None:
         super().__init__()
-        self.config = model_config.hf_config
+        self.config = vllm_config. \
+            speculative_config.draft_model_config.hf_config
         self.vocab_size = self.config.vocab_size
         self.embed_tokens = VocabParallelEmbedding(
             self.config.vocab_size,
@@ -119,8 +122,7 @@ def forward(
         hidden_states: torch.Tensor,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         input_embeds = self.embed_tokens(input_ids)
-        if (hidden_states.shape[-1] != input_embeds.shape[-1]):
-            hidden_states = self.fc(hidden_states)
+        assert hidden_states.shape[-1] == input_embeds.shape[-1]
 
         residual = None
         hidden_states, residual = self.layers[0](
@@ -169,9 +171,9 @@ class Eagle3LlamaForCausalLM(LlamaForCausalLM):
 
     def __init__(self, *, vllm_config: VllmConfig, start_layer_id: int = 0):
         nn.Module.__init__(self)
-        model_config = vllm_config.speculative_config.draft_model_config
-        self.config = model_config.hf_config
-        self.model = LlamaModel(model_config=model_config,
+        self.config = vllm_config. \
+            speculative_config.draft_model_config.hf_config
+        self.model = LlamaModel(vllm_config=vllm_config,
                                 start_layer_id=start_layer_id,
                                 prefix="model")
 
@@ -214,6 +216,13 @@ def compute_logits(
         logits_new[:, targets] = logits
         return logits_new
 
+    def combine_hidden_states(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        # combine multiple auxiliary hidden states returned by eagle3
+        return self.model.fc(hidden_states)
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         loader = AutoWeightsLoader(
             self,

vllm/v1/spec_decode/eagle.py

@@ -10,6 +10,7 @@
 from vllm.model_executor.model_loader.loader import get_model_loader
 from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 from vllm.model_executor.models import ModelRegistry
+from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
 from vllm.v1.sample.metadata import SamplingMetadata
 
@@ -39,11 +40,9 @@ def __init__(
 
         self.hidden_size = vllm_config.model_config.get_hidden_size()
 
-        # TODO: make eagle3 compatible with cudagraph
-        self.use_cuda_graph = self.method != 'eagle3' and \
-            (self.vllm_config.compilation_config.level
-             == CompilationLevel.PIECEWISE and
-             not self.vllm_config.model_config.enforce_eager)
+        self.use_cuda_graph = (self.vllm_config.compilation_config.level
+                               == CompilationLevel.PIECEWISE and
+                               not self.vllm_config.model_config.enforce_eager)
 
         self.cudagraph_batch_sizes = list(
             reversed(
@@ -90,6 +89,12 @@ def propose(
         batch_size = next_token_ids.shape[0]
         last_token_indices = cu_num_tokens[1:] - 1
 
+        if self.method == "eagle3":
+            assert isinstance(self.model, Eagle3LlamaForCausalLM)
+            target_hidden_states = self.model.combine_hidden_states(
+                target_hidden_states)
+            assert target_hidden_states.shape[-1] == self.hidden_size
+
         # Shift the input ids by one token.
         # E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
         self.input_ids[:num_tokens - 1] = target_token_ids[1:]
@@ -126,20 +131,15 @@ def propose(
         # copy inputs to buffer for cudagraph
         self.positions[:num_tokens] = target_positions
 
-        if self.method == 'eagle':
-            self.hidden_states[:num_tokens] = target_hidden_states
-            hidden_states = self.hidden_states
-        else:
-            # TODO: make eagle3 compatible with cuda graph
-            hidden_states = target_hidden_states
+        self.hidden_states[:num_tokens] = target_hidden_states
 
         with set_forward_context(attn_metadata,
                                  self.vllm_config,
                                  num_tokens=num_input_tokens):
             last_hidden_states, hidden_states = self.model(
                 input_ids=self.input_ids[:num_input_tokens],
                 positions=self.positions[:num_input_tokens],
-                hidden_states=hidden_states[:num_input_tokens],
+                hidden_states=self.hidden_states[:num_input_tokens],
             )
         sample_hidden_states = last_hidden_states[last_token_indices]
         logits = self.model.compute_logits(sample_hidden_states, None)
@@ -209,10 +209,7 @@ def propose(
             self.input_ids[:batch_size] = input_ids
             self.positions[:batch_size] = clamped_positions
 
-            if self.method == 'eagle':
-                # TODO: make eagle3 compatible with cudagraph.
-                self.hidden_states[:batch_size] = hidden_states
-                hidden_states = self.hidden_states
+            self.hidden_states[:batch_size] = hidden_states
 
             # Run the model.
             with set_forward_context(attn_metadata,
@@ -221,7 +218,7 @@ def propose(
                 last_hidden_states, hidden_states = self.model(
                     input_ids=self.input_ids[:input_batch_size],
                     positions=self.positions[:input_batch_size],
-                    hidden_states=hidden_states[:input_batch_size],
+                    hidden_states=self.hidden_states[:input_batch_size],
                 )
             hidden_states = hidden_states[:batch_size]
             logits = self.model.compute_logits(last_hidden_states[:batch_size],
@@ -314,12 +311,11 @@ def dummy_run(
     ) -> None:
         with set_forward_context(None, self.vllm_config,
                                  num_tokens=num_tokens):
-            if self.method == 'eagle':
-                self.model(
-                    input_ids=self.input_ids[:num_tokens],
-                    positions=self.positions[:num_tokens],
-                    hidden_states=self.hidden_states[:num_tokens],
-                )
+            self.model(
+                input_ids=self.input_ids[:num_tokens],
+                positions=self.positions[:num_tokens],
+                hidden_states=self.hidden_states[:num_tokens],
+            )
 
 
 # NOTE(woosuk): Currently, the below code is not used and we always use argmax