[Bugfix] Fuse Qwen3.5 in_qkvz_proj forwarding with LoRA enabled (#37912)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Isotr0py
2026-05-10 10:59:02 +00:00
committed by GitHub
parent 0a309b5ee9
commit 48698b1b9b
4 changed files with 84 additions and 111 deletions
@@ -263,7 +263,6 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
config: Qwen3NextConfig,
vllm_config: VllmConfig,
prefix: str = "",
create_in_proj_qkvz: bool = True,
gqa_interleaved_layout=False,
) -> None:
super().__init__()
@@ -323,32 +322,14 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
# we need to create qkvz_proj adaptively here.
# When create_in_proj_qkvz is False (e.g. LoRA enabled in Qwen3.5),
# in_proj_qkv and in_proj_z are created separately instead.
self.has_lora_projections = not create_in_proj_qkvz
if create_in_proj_qkvz:
self.in_proj_qkvz = self.create_qkvz_proj(
hidden_size=self.hidden_size,
key_dim=self.key_dim,
value_dim=self.value_dim,
quant_config=quant_config,
prefix=f"{prefix}.in_proj_qkvz",
)
else:
# LoRA case (Qwen3.5 only): keep q/k/v and z as separate modules
# so that LoRA adapters can be applied independently.
self.in_proj_qkv = MergedColumnParallelLinear(
input_size=self.hidden_size,
output_sizes=[self.key_dim, self.key_dim, self.value_dim],
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.in_proj_qkv",
)
self.in_proj_z = ColumnParallelLinear(
input_size=self.hidden_size,
output_size=self.value_dim,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.in_proj_z",
)
self.in_proj_qkvz = self.create_qkvz_proj(
hidden_size=self.hidden_size,
key_dim=self.key_dim,
value_dim=self.value_dim,
quant_config=quant_config,
prefix=f"{prefix}.in_proj_qkvz",
)
# ba_proj doesn't support blockwise fp8 quantization.
# Qwen3-Next and Qwen3.5 have different in_proj_ba checkpoint
# layouts, so we use a factory method to create the projection.
@@ -707,7 +688,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
):
"""ROCm forward using AITER Triton fused projection+attention when
available, otherwise falling back to the generic CUDA path."""
if not self.has_lora_projections and GDN_AITER_TRITON_AVAILABLE:
if GDN_AITER_TRITON_AVAILABLE:
num_tokens = hidden_states.size(0)
projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states)
projected_states_ba, _ = self.in_proj_ba(hidden_states)
@@ -752,37 +733,27 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
# ============================================================
# Part 1: Input Projection
# ============================================================
if self.has_lora_projections:
# LoRA path (Qwen3.5 only): separate in_proj_qkv and in_proj_z
mixed_qkv, _ = self.in_proj_qkv(hidden_states)
ba, _ = self.in_proj_ba(hidden_states)
z, _ = self.in_proj_z(hidden_states)
mixed_qkvz, _ = self.in_proj_qkvz(hidden_states)
ba, _ = self.in_proj_ba(hidden_states)
if self.gqa_interleaved_layout:
# Qwen3-Next: unpack the interleaved GQA layout
query, key, value, z, b, a = self.fix_query_key_value_ordering(
mixed_qkvz, ba
)
query, key, value = map(
lambda x: rearrange(x, "l p d -> l (p d)"), (query, key, value)
)
mixed_qkv = torch.cat((query, key, value), dim=-1)
else:
# Qwen3.5: weights are already in [q, k, v, z] and [b, a] order
qkv_size = (self.key_dim * 2 + self.value_dim) // self.tp_size
z_size = self.value_dim // self.tp_size
mixed_qkv, z = mixed_qkvz.split([qkv_size, z_size], dim=-1)
z = z.reshape(z.size(0), -1, self.head_v_dim)
b, a = ba.chunk(2, dim=-1)
b = b.contiguous()
a = a.contiguous()
else:
mixed_qkvz, _ = self.in_proj_qkvz(hidden_states)
ba, _ = self.in_proj_ba(hidden_states)
if self.gqa_interleaved_layout:
# Qwen3-Next: unpack the interleaved GQA layout
query, key, value, z, b, a = self.fix_query_key_value_ordering(
mixed_qkvz, ba
)
query, key, value = map(
lambda x: rearrange(x, "l p d -> l (p d)"), (query, key, value)
)
mixed_qkv = torch.cat((query, key, value), dim=-1)
else:
# Qwen3.5: weights are already in [q, k, v, z] and [b, a] order
qkv_size = (self.key_dim * 2 + self.value_dim) // self.tp_size
z_size = self.value_dim // self.tp_size
mixed_qkv, z = mixed_qkvz.split([qkv_size, z_size], dim=-1)
z = z.reshape(z.size(0), -1, self.head_v_dim)
b, a = ba.chunk(2, dim=-1)
b = b.contiguous()
a = a.contiguous()
# ============================================================
# Part 2: Core Attention (Custom Op)
@@ -822,8 +793,6 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
"""
num_tokens = hidden_states.size(0)
assert not self.has_lora_projections, "lora isn't supported on XPU."
# ============================================================
# Part 1: Input Projection
# ============================================================