EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[pref] qwen3_next add triton ops : fused_sigmoid_gating_delta_rule_update (#4818)

Browse Source 
### What this PR does / why we need it?
qwen3_next add fused_sigmoid_gating_delta_rule_update op which fused
fused_gdn_gating+fused_recurrent_gated_delta_rule

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
This commit is contained in:
XiaoxinWang
2025-12-19 16:34:11 +08:00
committed by GitHub
parent 118b0ed346
commit 0cc3fc357f
5 changed files with 539 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_ascend/patch/__init__.py
View File

@@ -285,3 +285,15 @@
# Future Plan:
# Remove this patch when vLLM support these operators.
#
# ** 15. File: worker/patch_qwen3_next.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.model_executor.models.qwen3_next.Qwen3NextGatedDeltaNet._forward_core`
# Why:
# triton ops fused_recurrent_gated_delta_rule and fused_gdn_gating in vLLM perform not good on NPU.
# How：
# add a new fused triton ops in vLLM with ascend implementation.
# Related PR (if no, explain why):
# https://github.com/vllm-project/vllm/pull/30860
# Future Plan:
# Remove this patch when vLLM support these operators.
#

1
vllm_ascend/patch/worker/__init__.py
View File

@@ -35,3 +35,4 @@ import vllm_ascend.patch.worker.patch_rope # noqa
import vllm_ascend.patch.worker.patch_qwen3_next # noqa
import vllm_ascend.patch.worker.patch_qwen3_next_mtp # noqa
import vllm_ascend.patch.worker.patch_rejection_sampler # noqa
import vllm_ascend.patch.worker.patch_qwen3_next # noqa

236
vllm_ascend/patch/worker/patch_qwen3_next.py
View File

@@ -18,14 +18,23 @@
import torch
from einops import rearrange
from torch import nn
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.config import CUDAGraphMode
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.fla.ops import (
chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.models.qwen3_next import Qwen3NextGatedDeltaNet
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
causal_conv1d_fn, causal_conv1d_update)
from vllm.model_executor.models.qwen3_next import (Qwen3NextGatedDeltaNet,
fused_gdn_gating)
from vllm.triton_utils import triton
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
from vllm_ascend.ops.triton.fla.fused_qkvzba_split_reshape import \
fused_qkvzba_split_reshape_cat
from vllm_ascend.ops.triton.fla.sigmoid_gating import \
fused_sigmoid_gating_delta_rule_update
class AscendQwen3Next_GatedDeltaNet(nn.Module, MambaBase):
@@ -101,5 +110,230 @@ class AscendQwen3Next_GatedDeltaNet(nn.Module, MambaBase):
core_attn_out = rearrange(core_attn_out, "... h d -> ... (h d)")
output[:num_tokens], _ = self.out_proj(core_attn_out)
def _forward_core(
self,
mixed_qkv: torch.Tensor,
b: torch.Tensor,
a: torch.Tensor,
core_attn_out: torch.Tensor,
):
"""
Core attention computation (called by custom op).
"""
forward_context = get_forward_context()
attn_metadata: AttentionMetadata = forward_context.attn_metadata
if attn_metadata is None:
# V1 profile run
return
assert isinstance(attn_metadata, dict)
attn_metadata = attn_metadata[self.prefix]
assert isinstance(attn_metadata, GDNAttentionMetadata)
has_initial_state = attn_metadata.has_initial_state
spec_query_start_loc = attn_metadata.spec_query_start_loc
non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
spec_sequence_masks = attn_metadata.spec_sequence_masks
spec_token_indx = attn_metadata.spec_token_indx
non_spec_token_indx = attn_metadata.non_spec_token_indx
spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor # noqa: E501
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor # noqa: E501
self_kv_cache = self.kv_cache[forward_context.virtual_engine]
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens
num_accepted_tokens = attn_metadata.num_accepted_tokens
mixed_qkv = mixed_qkv[:num_actual_tokens]
b = b[:num_actual_tokens]
a = a[:num_actual_tokens]
# 1. Convolution sequence transformation
conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0),
self.conv1d.weight.size(2))
if spec_sequence_masks is not None:
if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
mixed_qkv_spec = mixed_qkv
mixed_qkv_non_spec = None
else:
mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
mixed_qkv_non_spec = mixed_qkv.index_select(
0, non_spec_token_indx)
else:
mixed_qkv_spec = None
mixed_qkv_non_spec = mixed_qkv
# 1.1: Process the multi-query part
if spec_sequence_masks is not None:
mixed_qkv_spec = causal_conv1d_update(
mixed_qkv_spec,
conv_state,
conv_weights,
self.conv1d.bias,
self.activation,
conv_state_indices=spec_state_indices_tensor[:, 0]
[:attn_metadata.num_spec_decodes],
num_accepted_tokens=num_accepted_tokens,
query_start_loc=spec_query_start_loc,
max_query_len=spec_state_indices_tensor.size(-1),
validate_data=False,
)
# 1.2: Process the remaining part
if attn_metadata.num_prefills > 0:
if mixed_qkv_non_spec is not None:
mixed_qkv_non_spec_T = mixed_qkv_non_spec.transpose(0, 1)
# - "cache_indices" updates the conv_state cache in positions
# pointed to by "state_indices_tensor"
mixed_qkv_non_spec = causal_conv1d_fn(
mixed_qkv_non_spec_T,
conv_weights,
self.conv1d.bias,
activation=self.activation,
conv_states=conv_state,
has_initial_state=has_initial_state,
cache_indices=non_spec_state_indices_tensor,
query_start_loc=non_spec_query_start_loc,
metadata=attn_metadata,
).transpose(0, 1)
elif attn_metadata.num_decodes > 0:
mixed_qkv_non_spec = causal_conv1d_update(
mixed_qkv_non_spec,
conv_state,
conv_weights,
self.conv1d.bias,
self.activation,
conv_state_indices=
non_spec_state_indices_tensor[:attn_metadata.
num_actual_tokens],
validate_data=True,
)
else:
mixed_qkv_non_spec = None
query_spec, key_spec, value_spec = self.rearrange_mixed_qkv(
mixed_qkv_spec)
query_non_spec, key_non_spec, value_non_spec = self.rearrange_mixed_qkv(
mixed_qkv_non_spec)
if attn_metadata.num_prefills > 0 or spec_sequence_masks is not None:
g, beta = fused_gdn_gating(self.A_log, a, b, self.dt_bias)
if spec_sequence_masks is not None:
if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
g_spec = g
beta_spec = beta
g_non_spec = None
beta_non_spec = None
else:
g_spec = g.index_select(1, spec_token_indx)
beta_spec = beta.index_select(1, spec_token_indx)
g_non_spec = g.index_select(1, non_spec_token_indx)
beta_non_spec = beta.index_select(1, non_spec_token_indx)
else:
g_spec = None
beta_spec = None
g_non_spec = g
beta_non_spec = beta
# 2. Recurrent attention
# 2.1: Process the multi-query part
if spec_sequence_masks is not None:
core_attn_out_spec, last_recurrent_state = fused_recurrent_gated_delta_rule(
q=query_spec,
k=key_spec,
v=value_spec,
g=g_spec,
beta=beta_spec,
initial_state=ssm_state,
inplace_final_state=True,
cu_seqlens=spec_query_start_loc[:attn_metadata.
num_spec_decodes + 1],
ssm_state_indices=spec_state_indices_tensor,
num_accepted_tokens=num_accepted_tokens,
use_qk_l2norm_in_kernel=True,
)
else:
core_attn_out_spec, last_recurrent_state = None, None
# 2.2: Process the remaining part
if attn_metadata.num_prefills > 0:
initial_state = ssm_state[
non_spec_state_indices_tensor].contiguous()
initial_state[~has_initial_state, ...] = 0
(
core_attn_out_non_spec,
last_recurrent_state,
) = chunk_gated_delta_rule(
q=query_non_spec,
k=key_non_spec,
v=value_non_spec,
g=g_non_spec,
beta=beta_non_spec,
initial_state=initial_state,
output_final_state=True,
cu_seqlens=non_spec_query_start_loc,
head_first=False,
use_qk_l2norm_in_kernel=True,
)
# Init cache
ssm_state[
non_spec_state_indices_tensor] = last_recurrent_state.to(
ssm_state.dtype)
elif attn_metadata.num_decodes > 0:
core_attn_out_non_spec, last_recurrent_state = (
fused_recurrent_gated_delta_rule(
q=query_non_spec,
k=key_non_spec,
v=value_non_spec,
g=g_non_spec,
beta=beta_non_spec,
initial_state=ssm_state,
inplace_final_state=True,
cu_seqlens=non_spec_query_start_loc[:attn_metadata.
num_decodes + 1],
ssm_state_indices=non_spec_state_indices_tensor,
use_qk_l2norm_in_kernel=True,
))
else:
core_attn_out_non_spec, last_recurrent_state = None, None
elif attn_metadata.num_decodes > 0:
core_attn_out_non_spec = fused_sigmoid_gating_delta_rule_update(
A_log=self.A_log.contiguous(),
dt_bias=self.dt_bias.contiguous(),
q=query_non_spec.contiguous(),
k=key_non_spec.contiguous(),
v=value_non_spec.contiguous(),
a=a.contiguous(),
b=b.contiguous(),
initial_state_source=ssm_state,
initial_state_indices=non_spec_state_indices_tensor,
cu_seqlens=non_spec_query_start_loc,
use_qk_l2norm_in_kernel=True,
softplus_beta=1.0,
softplus_threshold=20.0,
)
# 3. Merge core attention output
if spec_sequence_masks is not None and core_attn_out_non_spec is not None:
merged_out = torch.empty(
(1, num_actual_tokens, *core_attn_out_spec.shape[2:]),
dtype=core_attn_out_non_spec.dtype,
device=core_attn_out_non_spec.device,
)
merged_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
merged_out.index_copy_(1, non_spec_token_indx,
core_attn_out_non_spec)
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
elif spec_sequence_masks is not None:
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(
0)
Qwen3NextGatedDeltaNet.forward = AscendQwen3Next_GatedDeltaNet.forward
Qwen3NextGatedDeltaNet._forward_core = AscendQwen3Next_GatedDeltaNet._forward_core