diff --git a/vllm/model_executor/models/qwen3_5.py b/vllm/model_executor/models/qwen3_5.py index 43b9004638..b00b495868 100644 --- a/vllm/model_executor/models/qwen3_5.py +++ b/vllm/model_executor/models/qwen3_5.py @@ -30,6 +30,7 @@ from collections.abc import Callable, Iterable import torch from torch import nn +from vllm._aiter_ops import rocm_aiter_ops from vllm.compilation.decorators import support_torch_compile from vllm.config import VllmConfig from vllm.distributed import ( @@ -314,6 +315,15 @@ class Qwen3_5Model(Qwen3NextModel): num_experts = ( self.config.num_experts if hasattr(self.config, "num_experts") else 0 ) + from vllm.config import get_current_vllm_config + + from .qwen3_next import _is_shared_expert_fse_compatible + + is_fse = ( + rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() + and _is_shared_expert_fse_compatible(get_current_vllm_config().quant_config) + ) + for name, loaded_weight in weights: if "rotary_emb.inv_freq" in name: continue @@ -327,6 +337,15 @@ class Qwen3_5Model(Qwen3NextModel): if name is None: continue + # FSE: remap shared_expert weights to fused expert slot + if is_fse and "mlp.shared_expert." in name: + name = name.replace( + "mlp.shared_expert.", + f"mlp.experts.{num_experts}.", + ) + is_fused_expert = False + expert_params_mapping = self.get_expert_mapping() + for param_name, weight_name, shard_id in stacked_params_mapping: if "experts.gate_up_proj" in name or "experts.down_proj" in name: is_fused_expert = True diff --git a/vllm/model_executor/models/qwen3_next.py b/vllm/model_executor/models/qwen3_next.py index acce2a7679..2c7667a416 100644 --- a/vllm/model_executor/models/qwen3_next.py +++ b/vllm/model_executor/models/qwen3_next.py @@ -88,6 +88,26 @@ logger = init_logger(__name__) KVCache = tuple[torch.Tensor, torch.Tensor] +def _is_shared_expert_fse_compatible(quant_config) -> bool: + """Check if shared expert can be fused with routed experts. + + FSE requires that shared and routed expert weights use the same + quantization format. Returns False when the shared expert is + excluded from quantization (e.g. float32 shared in an MXFP4 model) + or has a different quant spec than routed experts. + """ + if quant_config is None: + return True + # Quark stores its full config dict in quant_config.quant_config + raw_config = getattr(quant_config, "quant_config", None) + if not isinstance(raw_config, dict): + return True + exclude = raw_config.get("exclude", []) + if not exclude: + return True + return not any("shared_expert." in str(e) for e in exclude) + + class Qwen3NextSparseMoeBlock(nn.Module): def __init__(self, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -142,10 +162,15 @@ class Qwen3NextSparseMoeBlock(nn.Module): prefix=f"{prefix}.shared_expert_gate", ) - if ( - rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() - or config.shared_expert_intermediate_size <= 0 - ): + _fse_requested = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() + _fse_enabled = _fse_requested and _is_shared_expert_fse_compatible(quant_config) + if _fse_requested and not _fse_enabled: + logger.warning( + "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS is enabled but " + "shared expert has a different quantization spec than routed " + "experts. Falling back to non-fused shared expert path." + ) + if _fse_enabled or config.shared_expert_intermediate_size <= 0: self.shared_expert = None else: self.shared_expert = Qwen3NextMLP(