[Perf] fuse qk rmsnorm rope gate for qwen3.5 (#44176)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
Jiangyun Zhu
2026-06-09 22:12:17 +08:00
committed by GitHub
parent dc10e467a9
commit 7a89b72564
4 changed files with 418 additions and 12 deletions
+56 -11
View File
@@ -28,6 +28,7 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoE,
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.fused_qk_norm_rope import fused_qk_rmsnorm_rope_gate
from vllm.model_executor.layers.layernorm import (
GemmaRMSNorm as Qwen3NextRMSNorm,
)
@@ -58,6 +59,7 @@ from vllm.model_executor.model_loader.weight_utils import (
)
from vllm.model_executor.models.qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP
from vllm.model_executor.models.utils import sequence_parallel_chunk
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.qwen3_next import Qwen3NextConfig
@@ -283,13 +285,50 @@ class Qwen3NextAttention(nn.Module):
self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.k_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)
def forward(
# Fuse the gated split + QK-RMSNorm + (partial) NeoX RoPE + gate copy.
# TODO: support MRoPE
mm_config = model_config.multimodal_config if model_config else None
text_only = mm_config is None or mm_config.language_model_only
self.use_fused_qk_norm_rope_gate = (
self.attn_output_gate
and getattr(self.rotary_emb, "is_neox_style", False)
and current_platform.is_cuda()
and text_only
)
def _project_qkv_gate(
self,
qkv: torch.Tensor,
positions: torch.Tensor,
output: torch.Tensor,
hidden_states: torch.Tensor,
):
qkv, _ = self.qkv_proj(hidden_states)
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor | None]:
"""Return post-norm, post-RoPE (q, k, v) and the pre-sigmoid gate.
Dispatches between the fused Triton kernel and the eager
split + QK-RMSNorm + RoPE path. ``gate`` is ``None`` when output
gating is disabled.
"""
if self.use_fused_qk_norm_rope_gate:
q_gate, k, v = qkv.split(
[self.q_size * 2, self.kv_size, self.kv_size], dim=-1
)
# mRoPE passes positions as (3, n_tokens) for T/H/W. Fusion is only
# enabled text-only, where the three rows are identical, so taking
# the T row is exact. (1D positions pass through.)
pos = positions[0] if positions.ndim == 2 else positions
q, k, gate = fused_qk_rmsnorm_rope_gate(
q_gate,
k,
self.q_norm.weight.float() + 1.0,
self.k_norm.weight.float() + 1.0,
self.rotary_emb.cos_sin_cache,
pos,
self.q_norm.variance_epsilon,
self.num_heads,
self.num_kv_heads,
self.head_dim,
self.rotary_emb.rotary_dim,
)
return q, k, v, gate
if self.attn_output_gate:
q_gate, k, v = qkv.split(
@@ -302,6 +341,7 @@ class Qwen3NextAttention(nn.Module):
gate = gate.reshape(*orig_shape, -1)
else:
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
gate = None
q = self.q_norm(q.view(-1, self.num_heads, self.head_dim)).view(
-1, self.num_heads * self.head_dim
@@ -309,15 +349,20 @@ class Qwen3NextAttention(nn.Module):
k = self.k_norm(k.view(-1, self.num_kv_heads, self.head_dim)).view(
-1, self.num_kv_heads * self.head_dim
)
q, k = self.rotary_emb(positions, q, k)
return q, k, v, gate
def forward(
self,
positions: torch.Tensor,
output: torch.Tensor,
hidden_states: torch.Tensor,
):
qkv, _ = self.qkv_proj(hidden_states)
q, k, v, gate = self._project_qkv_gate(qkv, positions)
attn_output = self.attn(q, k, v)
if self.attn_output_gate:
gate = torch.sigmoid(gate)
attn_output = attn_output * gate
if gate is not None:
attn_output = attn_output * torch.sigmoid(gate)
output[:], _ = self.o_proj(attn_output)