[Spec-Decode] Support piecewise cudagraphs for Eagle head (#25109)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: Benjamin Chislett <chislett.ben@gmail.com>

vllm/config/compilation.py

@@ -50,11 +50,14 @@ class CUDAGraphMode(enum.Enum):
     def mixed_mode(self) -> "CUDAGraphMode":
         return CUDAGraphMode(self.value[1]) if self.separate_routine() else self
 
+    def has_mode(self, mode: "CUDAGraphMode") -> bool:
+        assert not mode.separate_routine()
+        if self.separate_routine():
+            return mode.value in self.value
+        return self == mode
+
     def requires_piecewise_compilation(self) -> bool:
-        return (
-            self.decode_mode() == CUDAGraphMode.PIECEWISE
-            or self.mixed_mode() == CUDAGraphMode.PIECEWISE
-        )
+        return self.has_mode(CUDAGraphMode.PIECEWISE)
 
     def max_cudagraph_mode(self) -> "CUDAGraphMode":
         return CUDAGraphMode(max(self.value)) if self.separate_routine() else self

vllm/forward_context.py

@@ -283,6 +283,12 @@ def set_forward_context(
             vllm_config.parallel_config, num_tokens or 0, num_tokens_across_dp
         )
 
+    # Convenience: if cudagraph is used and num_tokens is given, we can just
+    # create a batch descriptor here if not given (there's no harm since if it
+    # doesn't match in the wrapper it'll fall through).
+    if cudagraph_runtime_mode != CUDAGraphMode.NONE and num_tokens is not None:
+        batch_descriptor = batch_descriptor or BatchDescriptor(num_tokens=num_tokens)
+
     forward_context = create_forward_context(
         attn_metadata,
         vllm_config,

vllm/model_executor/models/deepseek_mtp.py

@@ -7,6 +7,7 @@ import torch
 import torch.nn as nn
 from transformers import PretrainedConfig
 
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import VllmConfig
 from vllm.model_executor.layers.fused_moe import FusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -162,6 +163,7 @@ class DeepSeekMultiTokenPredictor(nn.Module):
         return logits
 
 
+@support_torch_compile
 class DeepSeekMTP(nn.Module, SupportsPP):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()

vllm/model_executor/models/llama_eagle3.py

@@ -8,6 +8,7 @@ import torch
 import torch.nn as nn
 from transformers import LlamaConfig
 
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import VllmConfig, get_current_vllm_config
 from vllm.logger import init_logger
 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -21,6 +22,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 )
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.llama import LlamaDecoderLayer, LlamaForCausalLM
+from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import NestedTensors
 
 from .utils import AutoWeightsLoader, maybe_prefix
@@ -119,6 +121,15 @@ class LlamaDecoderLayer(LlamaDecoderLayer):
         return hidden_states, residual
 
 
+@support_torch_compile(
+    # torch.compile is disabled for multimodal EAGLE3 models due to constraint
+    # violations with dynamic shapes during tensor concatenation operations.
+    # See: https://github.com/vllm-project/vllm/pull/22872/files#r2362028132
+    # Non-multimodal EAGLE3 models can still use torch.compile safely.
+    enable_if=lambda vllm_config: not MULTIMODAL_REGISTRY.supports_multimodal_inputs(
+        vllm_config.model_config
+    ),
+)
 class LlamaModel(nn.Module):
     def __init__(
         self,

vllm/v1/spec_decode/eagle.py

@@ -9,7 +9,12 @@ import numpy as np
 import torch
 import torch.nn as nn
 
-from vllm.config import CompilationLevel, VllmConfig, get_layers_from_vllm_config
+from vllm.config import (
+    CompilationLevel,
+    CUDAGraphMode,
+    VllmConfig,
+    get_layers_from_vllm_config,
+)
 from vllm.distributed.parallel_state import get_pp_group
 from vllm.forward_context import set_forward_context
 from vllm.logger import init_logger
@@ -80,12 +85,25 @@ class EagleProposer:
         self.attn_layer_names: list[str] = []
         self.indexer_layer_names: list[str] = []
 
-        self.use_cuda_graph = (
-            not current_platform.is_xpu()
-            and self.vllm_config.compilation_config.level == CompilationLevel.PIECEWISE
-            and not self.vllm_config.model_config.enforce_eager
-            and not self.speculative_config.enforce_eager
-        )
+        self.use_cuda_graph = False
+
+        compilation_config = self.vllm_config.compilation_config
+        if compilation_config.level == CompilationLevel.PIECEWISE:
+            cudagraph_mode = compilation_config.cudagraph_mode
+            if cudagraph_mode != CUDAGraphMode.NONE and not cudagraph_mode.has_mode(
+                CUDAGraphMode.PIECEWISE
+            ):
+                logger.warning(
+                    "Currently the eagle proposer only supports cudagraph_mode "
+                    "PIECEWISE, if you want the drafter to use cuda graphs, "
+                    "please set compilation_config.cudagraph_mode to PIECEWISE "
+                    "or FULL_AND_PIECEWISE"
+                )
+            self.use_cuda_graph = (
+                cudagraph_mode.has_mode(CUDAGraphMode.PIECEWISE)
+                and not self.speculative_config.enforce_eager
+            )
+
         self.cudagraph_batch_sizes = (
             list(reversed(self.vllm_config.compilation_config.cudagraph_capture_sizes))
             if self.use_cuda_graph
@@ -239,12 +257,15 @@ class EagleProposer:
         per_layer_attn_metadata = {}
         for layer_name in self.attn_layer_names:
             per_layer_attn_metadata[layer_name] = attn_metadata
+
         for layer_name in self.indexer_layer_names:
             assert draft_indexer_metadata is not None
             per_layer_attn_metadata[layer_name] = draft_indexer_metadata
 
+        cudagraph_runtime_mode = CUDAGraphMode.NONE
         if self.use_cuda_graph and num_tokens <= self.cudagraph_batch_sizes[-1]:
             num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
+            cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE
         else:
             num_input_tokens = num_tokens
         # copy inputs to buffer for cudagraph
@@ -267,7 +288,10 @@ class EagleProposer:
             inputs_embeds = None
 
         with set_forward_context(
-            per_layer_attn_metadata, self.vllm_config, num_tokens=num_input_tokens
+            per_layer_attn_metadata,
+            self.vllm_config,
+            num_tokens=num_input_tokens,
+            cudagraph_runtime_mode=cudagraph_runtime_mode,
         ):
             ret_hidden_states = self.model(
                 input_ids=input_ids,
@@ -326,8 +350,10 @@ class EagleProposer:
 
         if self.use_cuda_graph and batch_size <= self.cudagraph_batch_sizes[-1]:
             input_batch_size = self.vllm_config.pad_for_cudagraph(batch_size)
+            cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE
         else:
             input_batch_size = batch_size
+            cudagraph_runtime_mode = CUDAGraphMode.NONE
 
         common_attn_metadata.num_actual_tokens = batch_size
         common_attn_metadata.max_query_len = 1
@@ -424,7 +450,10 @@ class EagleProposer:
 
             # Run the model.
             with set_forward_context(
-                per_layer_attn_metadata, self.vllm_config, num_tokens=input_batch_size
+                per_layer_attn_metadata,
+                self.vllm_config,
+                num_tokens=input_batch_size,
+                cudagraph_runtime_mode=cudagraph_runtime_mode,
             ):
                 ret_hidden_states = self.model(
                     input_ids=input_ids,
@@ -731,11 +760,16 @@ class EagleProposer:
 
             if self.use_cuda_graph and num_tokens <= self.cudagraph_batch_sizes[-1]:
                 num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
+                cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE
             else:
                 num_input_tokens = num_tokens
+                cudagraph_runtime_mode = CUDAGraphMode.NONE
             # Run the model.
             with set_forward_context(
-                per_layer_attn_metadata, self.vllm_config, num_tokens=num_input_tokens
+                per_layer_attn_metadata,
+                self.vllm_config,
+                num_tokens=num_input_tokens,
+                cudagraph_runtime_mode=cudagraph_runtime_mode,
             ):
                 last_hidden_states, hidden_states = self.model(
                     input_ids=self.input_ids[:num_input_tokens],
@@ -1015,8 +1049,19 @@ class EagleProposer:
     def dummy_run(
         self,
         num_tokens: int,
+        use_cudagraphs=True,
     ) -> None:
-        with set_forward_context(None, self.vllm_config, num_tokens=num_tokens):
+        if use_cudagraphs and num_tokens <= self.cudagraph_batch_sizes[-1]:
+            num_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
+
+        with set_forward_context(
+            None,
+            self.vllm_config,
+            num_tokens=num_tokens,
+            cudagraph_runtime_mode=CUDAGraphMode.PIECEWISE
+            if use_cudagraphs
+            else CUDAGraphMode.NONE,
+        ):
             if self.supports_mm_inputs:
                 input_ids = None
                 inputs_embeds = self.inputs_embeds[:num_tokens]

vllm/v1/worker/gpu_model_runner.py

@@ -3441,7 +3441,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
 
             if self.speculative_config and self.speculative_config.use_eagle():
                 assert isinstance(self.drafter, EagleProposer)
-                self.drafter.dummy_run(num_tokens)
+                use_cudagraphs = cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
+                self.drafter.dummy_run(num_tokens, use_cudagraphs=use_cudagraphs)
 
         # This is necessary to avoid blocking DP.
         # For dummy runs, we typically skip EPLB since we don't have any real