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

[BugFix][0.18.0] Fix quant_bias missing in w8a8_static when flashcomm1 is enabled for GLM-5 (#8304)

Browse Source 
### What this PR does / why we need it?
PR #8220 in v0.18.0

In a previous PR #7843 , the o_proj layer of GLM-5 was reverted to TP
(Tensor Parallel) splitting when flashcomm1 was enabled. However, this
was a temporary workaround and did not address the root cause of the
precision issues observed in the o_proj layer under flashcomm1.

I am working on a definitive fix for this issue. Currently, a clear bug
has been identified in
880e20fdde/vllm_ascend/quantization/methods/w8a8_static.py (L124):
during quantized matrix multiplication, quant_bias is not added if
tp_rank > 0. In the flashcomm1 scenario, all ranks actually require the
addition of quant_bias, meaning tp_rank=0 should be passed to ensure the
bias is applied correctly.

This PR aims to resolve this logic error and fix the underlying
precision issue.

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

### How was this patch tested?
glm5 e2e test

---------

Signed-off-by: zjks98 <zhangjiakang4@huawei.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: triomino <15924998+triomino@users.noreply.github.com>
Co-authored-by: zjks98 <zhangjiakang4@huawei.com>
This commit is contained in:
aipaes
2026-04-17 22:46:36 +08:00
committed by GitHub
parent b72ade9acd
commit 0954fd0912
3 changed files with 5 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm_ascend/attention/sfa_v1.py
View File

@@ -445,13 +445,6 @@ class AscendSFAImpl(MLAAttentionImpl):
# Enable layer sharding via DSA-CP on the P node in the PD-disaggregated setup.
self.enable_dsa_cp_with_layer_shard = enable_dsa_cp_with_layer_shard()
# Improves glm5 accuracy after enabling dsa-cp in scenarios with strict accuracy requirements,
# especially for customized cases, at the cost of performance degradation due to extra communication.
self.enable_dsa_cp_strict_accuracy = (
self.enable_dsa_cp_with_layer_shard
and self.vllm_config.model_config.hf_config.model_type in ["glm_moe_dsa"]
)
# use original TP o_proj weight in PD mix stage, and full gather
# for o_proj weight for prefill stage.
self.enable_dsa_cp_with_o_proj_tp = enable_dsa_cp_with_o_proj_tp()
@@ -1246,16 +1239,6 @@ class AscendSFAImpl(MLAAttentionImpl):
return result
attn_output = result
if self.enable_dsa_cp_strict_accuracy:
send = (
attn_output.view(-1, self.tp_size, self.num_heads * self.v_head_dim)
.permute(1, 0, 2)
.reshape(-1, self.num_heads * self.v_head_dim)
)
attn_output = torch.empty_like(send)
torch.distributed.all_to_all_single(attn_output, send, group=get_tp_group().device_group)
output[...] = self.o_proj(attn_output)[0]
maybe_save_kv_layer_to_connector(layer_name, list(kv_cache))