Support expert parallel in Transformers backend (#26162)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>

docs/models/supported_models.md

@@ -32,8 +32,9 @@ If the Transformers model implementation follows all the steps in [writing a cus
 - All the features listed in the [compatibility matrix](../features/README.md#feature-x-feature)
 - Any combination of the following vLLM parallelisation schemes:
     - Data parallel
-    - Pipeline parallel
     - Tensor parallel
+    - Expert parallel
+    - Pipeline parallel
 
 Checking if the modeling backend is Transformers is as simple as:
 

vllm/model_executor/models/transformers_moe.py

@@ -22,6 +22,7 @@ import torch.nn as nn
 
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config.utils import getattr_iter
+from vllm.distributed import get_dp_group, get_ep_group
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.model_executor.custom_op import CustomOp
 from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -40,42 +41,54 @@ class TransformersFusedMoE(FusedMoE):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self._top_k_index: torch.Tensor = None
+        self._topk_ids: torch.Tensor = None
 
         def custom_routing_function(hidden_states, gating_output, topk,
                                     renormalize):
-            """Return `top_k_weights` from `gating_output` and the
-            `top_k_index` we stored in the layer earlier."""
-            return gating_output, self._top_k_index
+            """Return `topk_weights` from `gating_output` and the
+            `topk_ids` we stored in the layer earlier."""
+            topk_weights = gating_output
+            topk_ids = self._topk_ids
+            # Handle all gather in expert parallel
+            if topk_ids.size(0) != hidden_states.size(0):
+                dp_metadata = get_forward_context().dp_metadata
+                sizes = dp_metadata.get_chunk_sizes_across_dp_rank()
+                is_sp = self.is_sequence_parallel
+                dist_group = get_ep_group() if is_sp else get_dp_group()
+                assert sizes[dist_group.rank_in_group] == topk_ids.shape[0]
+                topk_ids, = dist_group.all_gatherv([topk_ids], 0, sizes)
+            return topk_weights, topk_ids
 
         self.custom_routing_function = custom_routing_function
 
-    def forward(self, hidden_states: torch.Tensor, top_k_index: torch.Tensor,
-                top_k_weights: torch.Tensor, **kwargs: Any) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor, topk_ids: torch.Tensor,
+                topk_weights: torch.Tensor, **kwargs: Any) -> torch.Tensor:
         """In Transformers `experts.forward` will have this signature.
 
         We discard any extra kwargs because we cannot use them here."""
-        return torch.ops.vllm.transformers_moe_forward(hidden_states,
-                                                       top_k_index,
-                                                       top_k_weights,
-                                                       self.layer_name)
+        return torch.ops.vllm.transformers_moe_forward(
+            hidden_states,
+            topk_ids.to(torch.int32),
+            topk_weights.to(torch.float32),
+            self.layer_name,
+        )
 
 
 def transformers_moe_forward(hidden_states: torch.Tensor,
-                             top_k_index: torch.Tensor,
-                             top_k_weights: torch.Tensor,
+                             topk_ids: torch.Tensor,
+                             topk_weights: torch.Tensor,
                              layer_name: str) -> torch.Tensor:
-    """Store the `top_k_index` in the layer and call the actual forward."""
+    """Store the `topk_ids` in the layer and call the actual forward."""
     forward_context: ForwardContext = get_forward_context()
     self = forward_context.no_compile_layers[layer_name]
-    self._top_k_index = top_k_index
+    self._topk_ids = topk_ids
     # Clone hidden_states because it will be mutated in-place in FusedMoE
-    return self.forward_impl(hidden_states.clone(), top_k_weights)
+    return self.forward_impl(hidden_states.clone(), topk_weights)
 
 
 def transformers_moe_forward_fake(hidden_states: torch.Tensor,
-                                  top_k_index: torch.Tensor,
-                                  top_k_weights: torch.Tensor,
+                                  topk_ids: torch.Tensor,
+                                  topk_weights: torch.Tensor,
                                   layer_name: str) -> torch.Tensor:
     return torch.empty_like(hidden_states)
 
@@ -96,9 +109,6 @@ class TransformersMoEBase(TransformersBase):
         self.check_version("4.57.0.dev0", "MoE models support")
         super().__init__(vllm_config=vllm_config, prefix=prefix)
 
-        if self.parallel_config.enable_expert_parallel:
-            raise NotImplementedError(
-                "Transformers backend does not support expert parallel yet.")
         if self.parallel_config.enable_eplb:
             raise NotImplementedError(
                 "Transformers backend does not support expert parallel load "