qwen3 moe supports npu_add_rms_norm_quant op by default, update op with norm bias

Signed-off-by: huangdong2022 <huangdong51@huawei.com>

Author: huangdong2022

vllm_ascend/ascend_forward_context.py

@@ -147,12 +147,14 @@ def set_ascend_forward_context(
         # Once the necessary conditions are met, support for MOE models will also be added.
         from vllm_ascend.quantization.quant_config import AscendQuantConfig
         addrmsnorm_quant_fusion_enabled = isinstance(vllm_config.quant_config, AscendQuantConfig) and \
-            vllm_config.model_config.hf_config.model_type in ["llama", "qwen2", "qwen3"] and \
+            vllm_config.model_config.hf_config.model_type in ["llama", "qwen2", "qwen3", "qwen3_moe"] and \
             forward_context.layer_idx is not None
         if addrmsnorm_quant_fusion_enabled:
             forward_context.model_instance = model_instance
             forward_context.num_hidden_layers = vllm_config.model_config.hf_config.num_hidden_layers
             forward_context.fusion_linear = "gate_up_dense" if forward_context.layer_idx == 0 else "qkv_dense"
+            if vllm_config.model_config.hf_config.model_type == "qwen3_moe":
+                forward_context.fusion_linear = "gate_moe" if forward_context.layer_idx == 0 else "qkv_moe"
         forward_context.addrmsnorm_quant_fusion_enabled = addrmsnorm_quant_fusion_enabled
 
         if num_tokens is None and attn_metadata is not None:

vllm_ascend/ops/layernorm.py

@@ -15,9 +15,10 @@
 # This file is a part of the vllm-ascend project.
 #
 
-from typing import Optional, Tuple, Union, cast
+from typing import Optional, Tuple, Union
 
 import torch
+from vllm.config import get_current_vllm_config
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.layernorm import RMSNorm
 
@@ -27,6 +28,7 @@ def _addrmsnorm_forward_oot(
     x: torch.Tensor,
     residual: torch.Tensor,
     layer: Optional[torch.nn.Module] = None,
+    bias: Optional[torch.nn.Parameter] = None,
 ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
     import torch_npu
 
@@ -39,6 +41,7 @@ def _addrmsnorm_forward_oot(
             self.weight,
             layer.aclnn_input_scale,
             layer.aclnn_input_offset,
+            beta=bias,
             epsilon=self.variance_epsilon)
     else:
         if is_310p():
@@ -50,12 +53,31 @@ def _addrmsnorm_forward_oot(
         else:
             x, _, residual = torch_npu.npu_add_rms_norm(
                 x, residual, self.weight, self.variance_epsilon)
+        if bias is not None:
+            x.add_(bias)
     torch.ops.vllm.maybe_wait_prefetch_done(x)
     return x, residual
 
 
 class AscendRMSNorm(RMSNorm):
 
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-6,
+        var_hidden_size: Optional[int] = None,
+        has_weight: bool = True,
+        dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        super().__init__(hidden_size, eps, var_hidden_size, has_weight, dtype)
+        vllm_config = get_current_vllm_config()
+        self.bias = None
+        # m4
+        if vllm_config is not None and vllm_config.quant_config is not None and \
+                any("norm.bias" in name for name in vllm_config.quant_config.quant_description.keys()):
+            self.bias = torch.nn.Parameter(torch.zeros(hidden_size),
+                                           requires_grad=False)
+
     def forward_oot(
         self,
         x: torch.Tensor,
@@ -67,10 +89,13 @@ def forward_oot(
             residual = torch.ops.vllm.maybe_chunk_residual(x, residual)
             assert x.size(0) == residual.size(0)
             x, residual = _addrmsnorm_forward_oot(
-                self, x, residual, self.next_need_quant_fusion_linear)
+                self, x, residual, self.next_need_quant_fusion_linear,
+                self.bias)
             return x, residual
         x, residual = torch_npu.npu_rms_norm(x, self.weight,
                                              self.variance_epsilon)
+        if self.bias is not None:
+            x.add_(self.bias)
         return x
 
     @property
@@ -100,33 +125,15 @@ def next_need_quant_fusion_linear(self):
             # does not need to be repeated
             if not forward_context.prefetch_mlp_enabled:
                 forward_context.layer_idx += 1
+        elif fusion_linear == "qkv_moe":
+            next_linear = model_instance.model.layers[
+                layer_idx].self_attn.qkv_proj
+            forward_context.fusion_linear = "gate_moe"
+        elif fusion_linear == "gate_moe":
+            forward_context.fusion_linear = "qkv_moe"
+            forward_context.layer_idx += 1
         from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
         if next_linear is not None and \
             not isinstance(next_linear.quant_method.quant_method, AscendW8A8LinearMethod):
             next_linear = None
         return next_linear
-
-
-class AscendQuantRMSNorm(AscendRMSNorm):
-
-    def __init__(
-        self,
-        hidden_size: int,
-        eps: float = 1e-6,
-        var_hidden_size: Optional[int] = None,
-        has_weight: bool = True,
-        dtype: Optional[torch.dtype] = None,
-    ) -> None:
-        super().__init__(hidden_size, eps, var_hidden_size, has_weight, dtype)
-        self.bias = torch.nn.Parameter(torch.zeros(hidden_size),
-                                       requires_grad=False)
-
-    def forward_oot(
-        self,
-        x: torch.Tensor,
-        residual: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
-        if residual is not None:
-            x, residual = super().forward_oot(x, residual)
-            return x.add_(self.bias), residual
-        return cast(torch.Tensor, super().forward_oot(x)).add_(self.bias)

vllm_ascend/utils.py

@@ -500,7 +500,7 @@ def register_ascend_customop(vllm_config: Optional[VllmConfig] = None):
     from vllm_ascend.ops.activation import AscendQuickGELU, AscendSiluAndMul
     from vllm_ascend.ops.common_fused_moe import (AscendFusedMoE,
                                                   AscendSharedFusedMoE)
-    from vllm_ascend.ops.layernorm import AscendQuantRMSNorm, AscendRMSNorm
+    from vllm_ascend.ops.layernorm import AscendRMSNorm
     from vllm_ascend.ops.linear import (AscendColumnParallelLinear,
                                         AscendMergedColumnParallelLinear,
                                         AscendQKVParallelLinear,
@@ -530,11 +530,6 @@ def register_ascend_customop(vllm_config: Optional[VllmConfig] = None):
         "MultiHeadLatentAttention": AscendMultiHeadLatentAttention,
     }
 
-    if vllm_config is not None and \
-        vllm_config.quant_config is not None and \
-        any("norm.bias" in name for name in vllm_config.quant_config.quant_description.keys()):
-        REGISTERED_ASCEND_OPS["RMSNorm"] = AscendQuantRMSNorm
-
     for name, op_cls in REGISTERED_ASCEND_OPS.items():
         CustomOp.register_oot(_decorated_op_cls=op_cls, name=name)