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

[v0.18.0][BugFix] Fix Qwen3.5 MoE FC1 error under high concurrency when dp>1 (#8395)

Browse Source 
### What this PR does / why we need it?
GDN Attention uses FIA's query_start_loc (padded), which may cause
conv1d update errors under high concurrency when dp > 1, and this PR is
to make GDN use its own query_start_loc (unpadded).

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
- vLLM version: v0.18.0

Signed-off-by: Wangbingjie <wangbj1207@126.com>
This commit is contained in:
wangbj127
2026-04-20 10:26:19 +08:00
committed by GitHub
parent b992b11545
commit e6ba5a88f7
3 changed files with 18 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
vllm_ascend/patch/worker/patch_qwen3_5.py
View File

@@ -33,7 +33,7 @@ from vllm_ascend.ascend_forward_context import _EXTRA_CTX
from vllm_ascend.attention.utils import maybe_save_kv_layer_to_connector
from vllm_ascend.ops.triton.fla.sigmoid_gating import fused_sigmoid_gating_delta_rule_update
from vllm_ascend.ops.triton.fused_gdn_gating import fused_gdn_gating_patch
from vllm_ascend.utils import enable_sp, vllm_version_is
from vllm_ascend.utils import vllm_version_is
def to_int64_tuple(t):
@@ -141,10 +141,9 @@ class AscendQwen3_5GatedDeltaNet(Qwen3_5GatedDeltaNet):
num_actual_tokens = attn_metadata.num_actual_tokens
num_accepted_tokens = attn_metadata.num_accepted_tokens
if not enable_sp():
mixed_qkv = mixed_qkv[:num_actual_tokens]
b = b[:num_actual_tokens]
a = a[:num_actual_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))
@@ -315,20 +314,11 @@ class AscendQwen3_5GatedDeltaNet(Qwen3_5GatedDeltaNet):
)
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)
if not enable_sp():
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)[:num_actual_tokens]
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
elif spec_sequence_masks is not None:
if not enable_sp():
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)[:num_actual_tokens]
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
if not enable_sp():
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)[:num_actual_tokens]
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
maybe_save_kv_layer_to_connector("", [])

24
vllm_ascend/patch/worker/patch_qwen3_next.py
View File

@@ -33,7 +33,7 @@ from vllm_ascend.attention.utils import maybe_save_kv_layer_to_connector
from vllm_ascend.ops.triton.fla.fused_qkvzba_split_reshape import fused_qkvzba_split_reshape_cat
from vllm_ascend.ops.triton.fused_gdn_gating import fused_gdn_gating_patch
from vllm_ascend.patch.worker.patch_qwen3_5 import to_int64_tuple
from vllm_ascend.utils import enable_sp, vllm_version_is
from vllm_ascend.utils import vllm_version_is
class AscendQwen3Next_GatedDeltaNet(Qwen3NextGatedDeltaNet):
@@ -131,10 +131,9 @@ class AscendQwen3Next_GatedDeltaNet(Qwen3NextGatedDeltaNet):
num_actual_tokens = attn_metadata.num_actual_tokens
num_accepted_tokens = attn_metadata.num_accepted_tokens
if not enable_sp():
mixed_qkv = mixed_qkv[:num_actual_tokens]
b = b[:num_actual_tokens]
a = a[:num_actual_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))
@@ -294,20 +293,11 @@ class AscendQwen3Next_GatedDeltaNet(Qwen3NextGatedDeltaNet):
)
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)
if not enable_sp():
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)[:num_actual_tokens]
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
elif spec_sequence_masks is not None:
if not enable_sp():
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)[:num_actual_tokens]
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
if not enable_sp():
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)[:num_actual_tokens]
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
Qwen3NextGatedDeltaNet.forward = AscendQwen3Next_GatedDeltaNet.forward

5
vllm_ascend/worker/model_runner_v1.py
View File

@@ -2201,7 +2201,7 @@ class NPUModelRunner(GPUModelRunner):
if self._has_gdn:
attn_group = self.attn_groups[kv_cache_gid][0]
builder = attn_group.get_metadata_builder(0)
if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
if isinstance(builder, GDNAttentionMetadataBuilder):
cm.query_start_loc_cpu = self.gdn_query_start_loc.cpu[: num_reqs_padded + 1]
cm.query_start_loc = self.gdn_query_start_loc.gpu[: num_reqs_padded + 1]
@@ -2399,6 +2399,9 @@ class NPUModelRunner(GPUModelRunner):
num_reqs_padded = self._pad_query_start_loc_for_fia(
num_tokens_padded, num_reqs_padded, num_reqs, cudagraph_runtime_mode, batch_desc.num_reqs
)
if self._has_gdn:
self.gdn_query_start_loc.np[1 : num_reqs_padded + 1] = cum_num_tokens
self.gdn_query_start_loc.copy_to_gpu()
pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
attn_metadata, _ = self._build_attention_metadata(