[fsdp] fix: Fix Qwen3 MoE FSDP weight sync for vLLM rollout in Transformers 5

verl/utils/model.py

@@ -19,6 +19,7 @@
 import os
 import re
 import warnings
+from collections.abc import Iterable
 from dataclasses import dataclass
 from typing import Optional
 
@@ -262,6 +263,34 @@ def convert_weight_keys(state_dict: dict[str, torch.Tensor], model: PreTrainedMo
     return original_weights
 
 
+def iter_vllm_compatible_moe_params(name: str, tensor: torch.Tensor) -> Iterable[tuple[str, torch.Tensor]]:
+    """Expand Transformers 5 packed MoE expert tensors to vLLM checkpoint keys.
+
+    Transformers 5 stores Qwen-style MoE experts as packed 3D parameters:
+    ``mlp.experts.gate_up_proj`` with shape
+    ``[num_experts, 2 * intermediate_size, hidden_size]`` and
+    ``mlp.experts.down_proj`` with shape
+    ``[num_experts, hidden_size, intermediate_size]``. vLLM's Qwen MoE reload
+    path still accepts the original per-expert checkpoint keys during live
+    weight sync, so stream those keys without materializing a full dict.
+    """
+    if name.endswith(".mlp.experts.gate_up_proj") and tensor.dim() == 3:
+        gate, up = tensor.chunk(2, dim=1)
+        base = name.removesuffix(".gate_up_proj")
+        for expert_id in range(tensor.size(0)):
+            yield f"{base}.{expert_id}.gate_proj.weight", gate[expert_id].contiguous()
+            yield f"{base}.{expert_id}.up_proj.weight", up[expert_id].contiguous()
+        return
+
+    if name.endswith(".mlp.experts.down_proj") and tensor.dim() == 3:
+        base = name.removesuffix(".down_proj")
+        for expert_id in range(tensor.size(0)):
+            yield f"{base}.{expert_id}.down_proj.weight", tensor[expert_id].contiguous()
+        return
+
+    yield name, tensor
+
+
 def check_exclude_modules(config, key: str) -> bool:
     """
     A helper method to check if the passed module's key name matches any of the exclude modules in the adapter_config.

verl/workers/engine/fsdp/transformer_impl.py

@@ -59,7 +59,7 @@
     offload_fsdp_optimizer,
     replace_lora_wrapper,
 )
-from verl.utils.model import convert_weight_keys, extract_multi_modal_inputs
+from verl.utils.model import convert_weight_keys, extract_multi_modal_inputs, iter_vllm_compatible_moe_params
 from verl.utils.py_functional import convert_to_regular_types
 from verl.utils.torch_functional import logprobs_from_logits
 from verl.utils.ulysses import (
@@ -858,15 +858,16 @@ def get_per_tensor_param(self, layered_summon=False, base_sync_done=False, **kwa
         else:
             device = get_device_id()  # used when fsdp2 set cpu_offload_policy
             # TODO: cast fp32 to bf16 to reduce weight sync overhead, need more fine-grained control, e.g MoE gate
-            per_tensor_param = (
-                (
-                    name,
-                    param.to(device, non_blocking=True).full_tensor().to(torch.bfloat16, non_blocking=True)
-                    if isinstance(param, DTensor)
-                    else param,
-                )
-                for name, param in params.items()
-            )
+            def param_generator():
+                for name, param in params.items():
+                    tensor = (
+                        param.to(device, non_blocking=True).full_tensor().to(torch.bfloat16, non_blocking=True)
+                        if isinstance(param, DTensor)
+                        else param
+                    )
+                    yield from iter_vllm_compatible_moe_params(name, tensor)
+
+            per_tensor_param = param_generator()
 
         if self._qat_enabled:
             from verl.utils.qat.quantizer import QATQuantizer