[feat]1.支持mtp模型 full_cuda_graph; 2.优化mtp拒绝采样

vllm/model_executor/models/deepseek_mtp.py

@@ -150,7 +150,7 @@ class DeepSeekMultiTokenPredictor(nn.Module):
                                        sampling_metadata)
         return logits
 
-#@support_torch_compile
+@support_torch_compile
 class DeepSeekMTP(nn.Module, SupportsPP):
 
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):

vllm/v1/spec_decode/eagle.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any, Optional
+import numpy as np
+
 import torch
 import torch.nn as nn
 
@@ -29,10 +32,10 @@ PADDING_SLOT_ID = -1
 class EagleProposer:
 
     def __init__(
-            self,
-            vllm_config: VllmConfig,
-            device: torch.device,
-            runner=None,
+        self,
+        vllm_config: VllmConfig,
+        device: torch.device,
+        runner=None,
     ):
         self.vllm_config = vllm_config
         self.speculative_config = vllm_config.speculative_config
@@ -56,6 +59,9 @@ class EagleProposer:
         self.use_cuda_graph = (self.vllm_config.compilation_config.level
                                == CompilationLevel.PIECEWISE and
                                not self.vllm_config.model_config.enforce_eager)
+        self.use_full_cuda_graph = (
+            self.use_cuda_graph
+            and vllm_config.compilation_config.full_cuda_graph)
         self.cudagraph_batch_sizes = list(
             reversed(
                 self.vllm_config.compilation_config.cudagraph_capture_sizes))
@@ -71,6 +77,9 @@ class EagleProposer:
             (self.max_num_tokens, self.hidden_size),
             dtype=self.dtype,
             device=device)
+        
+        # attention metadata captured in full cudagraph mode
+        self.attn_metadata_cudagraph = None
         # We need +1 here because the arange is used to set query_start_loc,
         # which has one more element than batch_size.
         self.arange = torch.arange(vllm_config.scheduler_config.max_num_seqs +
@@ -79,25 +88,25 @@ class EagleProposer:
                                    dtype=torch.int32)
 
     def propose(
-            self,
-            # [num_tokens]
-            target_token_ids: torch.Tensor,
-            # [num_tokens]
-            target_positions: torch.Tensor,
-            # [num_tokens, hidden_size]
-            target_hidden_states: torch.Tensor,
-            # [num_tokens]
-            target_slot_mapping: torch.Tensor,
-            # [batch_size]
-            next_token_ids: torch.Tensor,
-            # [batch_size + 1] starting with 0
-            cu_num_tokens: torch.Tensor,
-            # [batch_size, max_num_blocks_per_req]
-            block_table: torch.Tensor,
-            # [batch_size]
-            num_rejected_tokens: list[int],
-            # [batch_size]
-            sampling_metadata: SamplingMetadata
+        self,
+        # [num_tokens]
+        target_token_ids: torch.Tensor,
+        # [num_tokens]
+        target_positions: torch.Tensor,
+        # [num_tokens, hidden_size]
+        target_hidden_states: torch.Tensor,
+        # [num_tokens]
+        target_slot_mapping: torch.Tensor,
+        # [batch_size]
+        next_token_ids: torch.Tensor,
+        # [batch_size + 1] starting with 0
+        cu_num_tokens: torch.Tensor,
+        # [batch_size, max_num_blocks_per_req]
+        block_table: torch.Tensor,
+        # [batch_size]
+        num_rejected_tokens: list[int],
+        # [batch_size]
+        sampling_metadata: SamplingMetadata
     ) -> torch.Tensor:
         num_tokens = target_token_ids.shape[0]
         batch_size = next_token_ids.shape[0]
@@ -157,7 +166,7 @@ class EagleProposer:
             # FIXME: need to consider multiple kv_cache_groups
             attn_metadata = self.runner.attn_metadata_builders[0].build(
                 common_prefix_len=0,
-                common_attn_metadata=common_attn_metadata,
+                common_attn_metadata=common_attn_metadata
             )
         else:
             raise ValueError(f"Unsupported method: {self.method}")
@@ -168,7 +177,7 @@ class EagleProposer:
         for layer_name in self.attn_layer_names:
             per_layer_attn_metadata[layer_name] = attn_metadata
         if self.use_cuda_graph and \
-                num_tokens <= self.cudagraph_batch_sizes[-1]:
+            num_tokens <= self.cudagraph_batch_sizes[-1]:
             num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
         else:
             num_input_tokens = num_tokens
@@ -176,6 +185,38 @@ class EagleProposer:
         self.positions[:num_tokens] = target_positions
         self.hidden_states[:num_tokens] = target_hidden_states
 
+        if (self.use_full_cuda_graph
+                and num_tokens <= self.cudagraph_batch_sizes[-1]):
+            assert self.attn_metadata_cudagraph
+            if self.method in ["eagle", "eagle3"]:
+                self.attn_metadata_cudagraph.seq_lens[:batch_size] = (
+                    attn_metadata.seq_lens)
+                self.attn_metadata_cudagraph.slot_mapping[:num_tokens] = (
+                    attn_metadata.slot_mapping)
+                self.attn_metadata_cudagraph.query_start_loc[:batch_size + 1] = (
+                    attn_metadata.query_start_loc)
+                self.attn_metadata_cudagraph.block_table[:batch_size] = (
+                    attn_metadata.block_table)
+            elif self.method == "deepseek_mtp":
+                self.attn_metadata_cudagraph.num_actual_tokens = (
+                    attn_metadata.num_actual_tokens)
+                self.attn_metadata_cudagraph.query_start_loc[:batch_size + 1] = (
+                    attn_metadata.query_start_loc)
+                self.attn_metadata_cudagraph.slot_mapping[:num_tokens] = (
+                    attn_metadata.slot_mapping)
+                self.attn_metadata_cudagraph.num_decodes = (
+                    attn_metadata.num_decodes)
+                self.attn_metadata_cudagraph.num_decode_tokens = (
+                    attn_metadata.num_decode_tokens)
+                self.attn_metadata_cudagraph.num_prefills = (
+                    attn_metadata.num_prefills)
+                
+                if attn_metadata.decode is not None:
+                    self.attn_metadata_cudagraph.decode.block_table[:attn_metadata.num_decode_tokens] = (
+                        attn_metadata.decode.block_table)
+                    self.attn_metadata_cudagraph.decode.seq_lens[:attn_metadata.num_decode_tokens] = (
+                        attn_metadata.decode.seq_lens)
+
         with set_forward_context(per_layer_attn_metadata,
                                  self.vllm_config,
                                  num_tokens=num_input_tokens):
@@ -192,10 +233,14 @@ class EagleProposer:
         logits = self.model.compute_logits(sample_hidden_states, None)
         draft_token_ids = logits.argmax(dim=-1)
 
+        draft_prob = logits.softmax(dim=-1, dtype=torch.float32)
+
+        draft_probs_list = [draft_prob]
+
         # Early exit if there is only one draft token to be generated.
         if self.num_speculative_tokens == 1:
             # [batch_size, 1]
-            return draft_token_ids.view(-1, 1)
+            return draft_token_ids.view(-1, 1), draft_probs_list
 
         # TODO: Currently, MTP module released by deepseek only has
         # one layer. Adapt this code to support multiple layers once
@@ -212,7 +257,7 @@ class EagleProposer:
             hidden_states = hidden_states[last_token_indices]
 
         if self.use_cuda_graph and \
-                batch_size <= self.cudagraph_batch_sizes[-1]:
+            batch_size <= self.cudagraph_batch_sizes[-1]:
             input_batch_size = self.vllm_config.pad_for_cudagraph(batch_size)
         else:
             input_batch_size = batch_size
@@ -230,7 +275,7 @@ class EagleProposer:
                 seq_lens=(seq_lens + 1),
             )
 
-        for _ in range(self.num_speculative_tokens - 1):
+        for i in range(self.num_speculative_tokens - 1):
             # Update the inputs.
             # cast to int32 is crucial when eagle model is compiled.
             # tensor.argmax() returns int64 by default.
@@ -267,10 +312,10 @@ class EagleProposer:
             # Compute the slot mapping.
             block_numbers = clamped_positions // self.block_size
             block_ids = block_table.gather(dim=1,
-                                           index=block_numbers.view(-1, 1))
+                                        index=block_numbers.view(-1, 1))
             block_ids = block_ids.view(-1)
             attn_metadata.slot_mapping = (block_ids * self.block_size +
-                                          clamped_positions % self.block_size)
+                                        clamped_positions % self.block_size)
             # Mask out the slot mappings that exceed the max model length.
             # Otherwise, the KV cache will be inadvertently updated with the
             # padding tokens.
@@ -282,6 +327,43 @@ class EagleProposer:
             self.positions[:batch_size] = clamped_positions
             self.hidden_states[:batch_size] = hidden_states
 
+            if (self.use_full_cuda_graph
+                    and batch_size <= self.cudagraph_batch_sizes[-1]):
+                assert self.attn_metadata_cudagraph
+                if self.method in ["eagle", "eagle3"]:
+                    self.attn_metadata_cudagraph.seq_lens[:batch_size] = (
+                        attn_metadata.seq_lens)
+                    self.attn_metadata_cudagraph.slot_mapping[:batch_size] = (
+                        attn_metadata.slot_mapping)
+                    if i == 0:
+                        self.attn_metadata_cudagraph.query_start_loc[:batch_size +
+                                                                    1] = (
+                                                                        attn_metadata
+                                                                        .
+                                                                        query_start_loc
+                                                                    )
+                        self.attn_metadata_cudagraph.block_table[:batch_size] = (
+                            attn_metadata.block_table)
+                elif self.method == "deepseek_mtp":
+                    self.attn_metadata_cudagraph.num_actual_tokens = (
+                        attn_metadata.num_actual_tokens)
+                    self.attn_metadata_cudagraph.slot_mapping[:attn_metadata.num_decode_tokens] = (
+                        attn_metadata.slot_mapping)
+                    self.attn_metadata_cudagraph.num_decodes = (
+                        attn_metadata.num_decodes)
+                    self.attn_metadata_cudagraph.num_decode_tokens = (
+                        attn_metadata.num_decode_tokens)
+                    self.attn_metadata_cudagraph.num_prefills = (
+                        attn_metadata.num_prefills)
+                    self.attn_metadata_cudagraph.decode.seq_lens[:attn_metadata.num_decode_tokens] = (
+                        attn_metadata.decode.seq_lens)
+                    
+                    if i == 0:
+                        self.attn_metadata_cudagraph.query_start_loc[:batch_size + 1] = (
+                            attn_metadata.query_start_loc)
+                        self.attn_metadata_cudagraph.decode.block_table[:attn_metadata.num_decode_tokens] = (
+                            attn_metadata.decode.block_table)
+
             # Run the model.
             with set_forward_context(per_layer_attn_metadata,
                                      self.vllm_config,
@@ -305,17 +387,22 @@ class EagleProposer:
             draft_token_ids = logits.argmax(dim=-1)
             draft_token_ids_list.append(draft_token_ids)
 
+            draft_prob = logits.softmax(dim=-1, dtype=torch.float32)
+            draft_probs_list.append(draft_prob)
+
         # [batch_size, num_speculative_tokens]
         draft_token_ids = torch.stack(draft_token_ids_list, dim=1)
-        return draft_token_ids
+        draft_probs = torch.stack(draft_probs_list, dim=1).contiguous()
+
+        return draft_token_ids, draft_probs
 
     @staticmethod
     def prepare_inputs(
-            # [batch_size + 1]
-            cu_target_query_lens: torch.Tensor,
-            # [batch_size]
-            num_rejected_tokens: torch.Tensor,
-            num_tokens: int,
+        # [batch_size + 1]
+        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]
@@ -342,7 +429,7 @@ class EagleProposer:
         )
         batch_size = num_rejected_tokens.shape[0]
         BLOCK_SIZE = 1024
-        prepare_eagle_input_kernel[(batch_size,)](
+        prepare_eagle_input_kernel[(batch_size, )](
             token_indices,
             cu_target_query_lens,
             cu_num_tokens,
@@ -362,8 +449,8 @@ class EagleProposer:
                                    model_config=draft_model_config)
 
         draft_attn_layer_names = (
-                get_layers_from_vllm_config(self.vllm_config, Attention).keys() -
-                target_attn_layer_names)
+            get_layers_from_vllm_config(self.vllm_config, Attention).keys() -
+            target_attn_layer_names)
 
         self.attn_layer_names = list(draft_attn_layer_names)
 
@@ -376,8 +463,8 @@ class EagleProposer:
             target_language_model = target_model
         # share embed_tokens with the target model if needed
         if get_pp_group().world_size == 1 \
-                and self.method != "deepseek_mtp" \
-                and self.model.model.embed_tokens.weight.shape \
+            and self.method != "deepseek_mtp" \
+            and self.model.model.embed_tokens.weight.shape \
                 == target_language_model.model.embed_tokens.weight.shape:
             logger.info(
                 "Assuming the EAGLE head shares the same vocab embedding" \
@@ -402,10 +489,15 @@ class EagleProposer:
 
     @torch.inference_mode()
     def dummy_run(
-            self,
-            num_tokens: int,
+        self,
+        num_tokens: int,
+        attn_metadata: Optional[dict[str, Any]] = None,
     ) -> None:
-        with set_forward_context(None, self.vllm_config,
+        if attn_metadata is not None and self.attn_metadata_cudagraph is None:
+            self.attn_metadata_cudagraph = attn_metadata[
+                self.attn_layer_names[0]]
+        with set_forward_context(attn_metadata,
+                                 self.vllm_config,
                                  num_tokens=num_tokens):
             self.model(
                 self.input_ids[:num_tokens],
@@ -440,8 +532,8 @@ class EagleProposer:
 # FIXME(woosuk): The logic here is duplicated with the main sampling code.
 # We should refactor this to reuse the same sampling implementation.
 def compute_probs_and_sample_next_token(
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+    logits: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     if sampling_metadata.all_greedy:
         # For greedy requests, draft_probs is not used in rejection sampling.

vllm/v1/spec_decode/utils.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import msgspec
+from abc import ABC
+import torch
+
 from vllm.sampling_params import SamplingParams
 from vllm.triton_utils import tl, triton
 
@@ -39,3 +43,41 @@ def prepare_eagle_input_kernel(
             index_start + offset,
             mask=offset < num_tokens,
         )
+
+class DraftProbs(ABC):  # type: ignore[call-arg]
+    """Draft probs corresponding to in-progress sequences."""
+
+    # spec tokens probs. 
+    draft_probs: torch.Tensor
+
+    # The request id list.
+    _req_ids: list[str]
+
+    def __init__(self, draft_probs, req_ids):
+        assert len(req_ids) == len(draft_probs)
+        self.draft_probs = draft_probs
+        self._req_ids = req_ids
+
+    def update(self,
+               draft_probs: torch.Tensor,
+               tmp_req_ids: list[str]):
+        diff_req_ids = [item for item in self._req_ids if item not in tmp_req_ids]
+        index = [self._req_ids.index(req_id) for req_id in diff_req_ids]
+        self._req_ids = diff_req_ids
+        self.draft_probs = self.draft_probs[index]
+        self.draft_probs = torch.cat([self.draft_probs, draft_probs])
+        
+        self._req_ids.extend(tmp_req_ids)
+        assert len(self._req_ids) == len(self.draft_probs)
+
+    def prune(self, req_ids: list[str]):
+        new_req_ids = [req_id for req_id in self._req_ids if req_id not in req_ids]
+        if new_req_ids != self._req_ids:
+            # Batch contents changed - prune removed sequences.
+            index = [self._req_ids.index(req_id) for req_id in new_req_ids]
+            self.draft_probs = self.draft_probs[index]
+            self._req_ids = new_req_ids
+
+    def get_probs(self, req_ids: list[str]):
+        index = [self._req_ids.index(req_id) for req_id in req_ids]
+        return self.draft_probs[index]

vllm/v1/worker/gpu_model_runner.py

@@ -58,11 +58,13 @@ from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
 from vllm.v1.pool.metadata import PoolingMetadata
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.sample.rejection_sampler import RejectionSampler
+from vllm.v1.sample.rejection_sampler_mtp import MtpRejectionSampler
 from vllm.v1.sample.sampler import Sampler
 from vllm.v1.spec_decode.eagle import EagleProposer
 from vllm.v1.spec_decode.medusa import MedusaProposer
 from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
+from vllm.v1.spec_decode.utils import DraftProbs
 from vllm.v1.utils import bind_kv_cache
 from vllm.v1.worker.block_table import BlockTable
 from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
@@ -191,7 +193,12 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             else:
                 raise ValueError("Unknown speculative decoding method: "
                                  f"{self.speculative_config.method}")
-            self.rejection_sampler = RejectionSampler()
+            
+            self.use_mtp = self.speculative_config.method == "deepseek_mtp"
+            if not self.use_mtp:
+                self.rejection_sampler = RejectionSampler()
+            else:
+                self.rejection_sampler = MtpRejectionSampler()
 
         # Request states.
         self.requests: dict[str, CachedRequestState] = {}
@@ -318,6 +325,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         # from the KV cache of `shared_kv_cache_layers[layer_name]`.
         self.shared_kv_cache_layers: dict[str, str] = {}
 
+        self.draft_probs : Optional[DraftProbs] = None
+
     def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
         """
         Update the order of requests in the batch based on the attention
@@ -377,6 +386,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         for req_id in scheduler_output.finished_req_ids:
             self.input_batch.remove_request(req_id)
 
+        # prune draft probs of finished requests
+        if self.use_mtp and self.draft_probs is not None and len(scheduler_output.finished_req_ids) > 0:
+            self.draft_probs.prune(list(scheduler_output.finished_req_ids))
+
         # Free the cached encoder outputs.
         for req_id, input_id in scheduler_output.free_encoder_input_ids:
             encoder_outputs = self.encoder_cache.get(req_id)
@@ -534,6 +547,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             # Add spec_token_ids to token_ids_cpu.
             spec_token_ids = (
                 scheduler_output.scheduled_spec_decode_tokens.get(req_id, ()))
+            
             if spec_token_ids:
                 num_spec_tokens = len(spec_token_ids)
                 start_index = self.input_batch.num_tokens_no_spec[req_index]
@@ -1451,7 +1465,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             target_logits = logits[spec_decode_metadata.target_logits_indices]
             output_token_ids = self.rejection_sampler(
                 spec_decode_metadata,
-                None,  # draft_probs
+                self.draft_probs.get_probs(self.input_batch.req_ids) \
+                    if self.draft_probs is not None else None,  # draft_probs
                 target_logits,
                 bonus_token_ids,
                 sampling_metadata,
@@ -1536,7 +1551,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             # Speculative decoding is not enabled.
             spec_token_ids = None
         else:
-            spec_token_ids = self.propose_draft_token_ids(
+            spec_token_ids, draft_probs = self.propose_draft_token_ids(
                 scheduler_output,
                 valid_sampled_token_ids,
                 sampling_metadata,
@@ -1547,6 +1562,15 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 attn_metadata,
             )
 
+            if self.use_mtp:
+                if self.draft_probs is None:
+                    self.draft_probs = DraftProbs(
+                    draft_probs, self.input_batch.req_ids)
+                else:
+                    self.draft_probs.update(draft_probs, self.input_batch.req_ids)
+
+            spec_token_ids = spec_token_ids.tolist()
+
         # Clear KVConnector state after all KVs are generated.
         if has_kv_transfer_group():
             get_kv_transfer_group().clear_connector_metadata()
@@ -1563,7 +1587,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             pooler_output=[],
             finished_sending=finished_sending,
             finished_recving=finished_recving,
-            num_nans_in_logits=num_nans_in_logits,
+            num_nans_in_logits=num_nans_in_logits
         )
 
     def propose_draft_token_ids(
@@ -1675,7 +1699,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                     target_hidden_states = hidden_states[token_indices]
                 target_slot_mapping = eagle_attn_metadata.slot_mapping[
                     token_indices]
-            draft_token_ids = self.drafter.propose(
+            spec_token_ids, draft_probs = self.drafter.propose(
                 target_token_ids=target_token_ids,
                 target_positions=target_positions,
                 target_hidden_states=target_hidden_states,
@@ -1686,8 +1710,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 sampling_metadata=sampling_metadata,
                 num_rejected_tokens=num_rejected_tokens
             )
-            spec_token_ids = draft_token_ids.tolist()
-        return spec_token_ids
+            
+        return spec_token_ids, draft_probs
 
     def kv_connector_no_forward(
             self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput:
@@ -2076,7 +2100,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
 
             if self.speculative_config and self.speculative_config.use_eagle():
                 assert isinstance(self.drafter, EagleProposer)
-                self.drafter.dummy_run(num_tokens)
+                self.drafter.dummy_run(num_tokens, attn_metadata)
 
         # This is necessary to avoid blocking DP.
         # For dummy runs, we typically skip EPLB since we don't have any real
@@ -2143,10 +2167,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 draft_token_ids, self.device)
 
             num_tokens = sum(len(ids) for ids in draft_token_ids)
-            # draft_probs = torch.randn(
-            #     num_tokens, logits.shape[-1], device=self.device,
-            #     dtype=logits.dtype)
-            draft_probs = None
+            draft_probs = torch.randn(
+                num_tokens, logits.shape[-1], device=self.device,
+                dtype=logits.dtype)
+            # draft_probs = None
             target_logits = torch.randn(num_tokens,
                                         logits.shape[-1],
                                         device=self.device,