[Bugfix] Refactor to support DP parallel in R3 (#32306)

Signed-off-by: xhx1022 <1737006628@qq.com>
Co-authored-by: arlenxu <arlenxu@tencent.com>

vllm/model_executor/layers/fused_moe/routed_experts_capturer.py

@@ -17,8 +17,9 @@ from unittest.mock import patch
 import numpy as np
 import torch
 
-from vllm.config import ModelConfig
+from vllm.config import VllmConfig
 from vllm.distributed import get_tensor_model_parallel_rank
+from vllm.forward_context import get_forward_context
 
 logger = logging.getLogger(__name__)
 
@@ -89,7 +90,6 @@ class RoutedExpertsCapturer:
         self._shm: shared_memory.SharedMemory | None = None
         self._host_buffer_view: np.ndarray | None = None
         self._lock_file: str | None = None
-        self._shm_name: str | None = None
 
     @classmethod
     def create(cls) -> RoutedExpertsCapturer:
@@ -110,8 +110,7 @@ class RoutedExpertsCapturer:
         self,
         max_num_batched_tokens: int,
         max_num_kv_tokens: int,
-        model_config: ModelConfig,
-        instance_id: str,
+        vllm_config: VllmConfig,
     ) -> None:
         """
         Initialize the device buffer and optionally shared memory buffer.
@@ -119,14 +118,13 @@ class RoutedExpertsCapturer:
         Args:
             max_num_batched_tokens: Maximum number of tokens in a batch.
             max_num_kv_tokens: Maximum number of KV tokens for shared memory.
-            model_config: Model configuration containing layer and expert info.
-            instance_id: Unique identifier for the shared memory buffer.
+            vllm_config: vllm configuration containing layer and expert info.
         """
 
         if self._device_buffer is not None:
             raise RuntimeError("Device buffer has already been initialized")
 
-        hf_config = model_config.hf_text_config
+        hf_config = vllm_config.model_config.hf_text_config
         num_layers = hf_config.num_hidden_layers
         num_experts_per_tok = hf_config.num_experts_per_tok
 
@@ -136,6 +134,7 @@ class RoutedExpertsCapturer:
             dtype=torch.int32,
             device="cuda",
         )
+        self.dp_rank = vllm_config.parallel_config.data_parallel_rank
 
         if get_tensor_model_parallel_rank() != 0:
             return
@@ -143,19 +142,19 @@ class RoutedExpertsCapturer:
         # Initialize shared memory
         shape = (max_num_kv_tokens, num_layers, num_experts_per_tok)
         buffer_size = int(np.prod(shape)) * np.dtype(np.int32).itemsize
-
-        self._lock_file = f"{_LOCK_FILE_PREFIX}_{instance_id}.lock"
-        self._shm_name = f"{_BUFFER_PREFIX}_{instance_id}"
+        instance_id = vllm_config.instance_id
+        self._lock_file = f"{_LOCK_FILE_PREFIX}_{instance_id}_{self.dp_rank}.lock"
+        shm_name = f"{_BUFFER_PREFIX}_{instance_id}_{self.dp_rank}"
 
         self._shm = _create_or_attach_shared_memory(
-            self._shm_name, buffer_size, self._lock_file
+            shm_name, buffer_size, self._lock_file
         )
         self._host_buffer_view = np.ndarray(shape, dtype=np.int32, buffer=self._shm.buf)
         self._host_buffer_view.fill(0)
 
         logger.debug(
             "Created shared memory buffer '%s' with shape %s",
-            self._shm.name,
+            shm_name,
             shape,
         )
 
@@ -170,11 +169,24 @@ class RoutedExpertsCapturer:
         if self._device_buffer is None:
             raise RuntimeError("Buffer not initialized. Call init_buffer() first.")
 
+        ctx = get_forward_context()
+        if ctx.dp_metadata is None:  # single dp
+            start_loc = 0
+            end_loc = topk_ids.shape[0]
+            token_num_per_dp = topk_ids.shape[0]
+        else:  # multi dp
+            token_num_per_dp = ctx.dp_metadata.num_tokens_across_dp_cpu[self.dp_rank]
+            cumsum = torch.cumsum(ctx.dp_metadata.num_tokens_across_dp_cpu, dim=0)
+            assert cumsum[-1] == topk_ids.shape[0]
+            end_loc = cumsum[self.dp_rank]
+            start_loc = end_loc - token_num_per_dp
+
         if layer_id >= self._device_buffer.shape[1]:
             return
 
-        batch_size = topk_ids.shape[0]
-        self._device_buffer[:batch_size, layer_id, :] = topk_ids
+        self._device_buffer[:token_num_per_dp, layer_id, :] = topk_ids[
+            start_loc:end_loc, :
+        ]
 
     def clear_buffer(self) -> None:
         """Clear the device buffer."""
@@ -254,30 +266,30 @@ class RoutedExpertsReader:
     def attach_buffer(
         self,
         max_num_kv_tokens: int,
-        model_config: ModelConfig,
-        instance_id: str,
+        vllm_config: VllmConfig,
     ) -> None:
         """
         Attach to an existing shared memory buffer.
 
         Args:
             max_num_kv_tokens: Maximum number of KV tokens.
-            model_config: Model configuration.
-            instance_id: Unique identifier for the shared memory buffer.
+            vllm_config: vllm configuration.
         """
         if self._shm is not None:
             logger.warning("Already attached to shared memory buffer.")
             return  # Already attached
 
-        hf_config = model_config.hf_text_config
+        hf_config = vllm_config.model_config.hf_text_config
         shape = (
             max_num_kv_tokens,
             hf_config.num_hidden_layers,
             hf_config.num_experts_per_tok,
         )
 
-        self._lock_file = f"{_LOCK_FILE_PREFIX}_{instance_id}.lock"
-        shm_name = f"{_BUFFER_PREFIX}_{instance_id}"
+        self.dp_rank = vllm_config.parallel_config.data_parallel_rank
+        instance_id = vllm_config.instance_id
+        self._lock_file = f"{_LOCK_FILE_PREFIX}_{instance_id}_{self.dp_rank}.lock"
+        shm_name = f"{_BUFFER_PREFIX}_{instance_id}_{self.dp_rank}"
 
         with _file_lock(self._lock_file, mode="rb+"):
             # Avoid resource_tracker registering the shared memory

vllm/v1/core/sched/scheduler.py

@@ -245,8 +245,7 @@ class Scheduler(SchedulerInterface):
 
             self.routed_experts_reader.attach_buffer(
                 max_num_kv_tokens=self.max_num_kv_tokens,
-                model_config=self.vllm_config.model_config,
-                instance_id=self.vllm_config.instance_id,
+                vllm_config=self.vllm_config,
             )
 
     def schedule(self) -> SchedulerOutput:

vllm/v1/worker/gpu_model_runner.py

@@ -5675,12 +5675,10 @@ class GPUModelRunner(
             self.kv_cache_config.num_blocks // len(self.kv_cache_config.kv_cache_groups)
             + 1
         ) * block_size
-
         routed_experts_capturer.init_buffer(
             max_num_batched_tokens=self.scheduler_config.max_num_batched_tokens,
             max_num_kv_tokens=self.max_num_kv_tokens,
-            model_config=self.model_config,
-            instance_id=self.vllm_config.instance_id,
+            vllm_config=self.vllm_config,
         )
 
     def may_add_encoder_only_layers_to_kv_cache_config(self) -> None: