Skip to content

Fix incorrect MLAPO weight release in PD mixex scenarios. #4774

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wangxiyuan merged 3 commits into from
Dec 8, 2025
Merged

Fix incorrect MLAPO weight release in PD mixex scenarios. #4774

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1db0135
Fix incorrect MLAPO weight release in PD mixex scenarios.
ZYang6263 Dec 8, 2025
4d7bac1
Merge branch 'main' into pr-mlapo-fix
ZYang6263 Dec 8, 2025
e1a3e70
Merge branch 'main' into pr-mlapo-fix
wangxiyuan Dec 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 5 additions & 4 deletions vllm_ascend/attention/sfa_v1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -470,7 +470,7 @@ def get_and_maybe_dequant_weights(layer: LinearBase):
if self.fused_qkv_a_proj is None or not isinstance(
quant_method, AscendW8A8LinearMethod):
reasons.append(
"Currently mlapo only supports W8A8 quantization in MLA scenario."
"Currently mlapo only supports W8A8 quantization in SFA scenario."
"Some layers in your model are not quantized with W8A8,"
"thus mlapo is disabled for these layers.")
if self.enable_sfa_cp:
Expand Down Expand Up @@ -597,8 +597,6 @@ def _process_weights_for_fused_mlapo(self, act_dtype: torch.dtype):
q_a_proj_wt = self.fused_qkv_a_proj.weight.data[
..., :self.q_lora_rank].contiguous()

self.fused_qkv_a_proj.weight = None

kv_a_proj_wt = kv_a_proj_wt.t().contiguous()
kv_a_proj_wt = trans_rope_weight(kv_a_proj_wt, self.qk_rope_head_dim)
kv_a_proj_wt = kv_a_proj_wt.t().contiguous()
Expand Down Expand Up @@ -673,9 +671,12 @@ def _process_weights_for_fused_mlapo(self, act_dtype: torch.dtype):
self.ctkv_scale = torch.tensor([1], dtype=act_dtype, device=device)
self.q_nope_scale = torch.tensor([1], dtype=act_dtype, device=device)

if self.vllm_config.kv_transfer_config is not None:
if self.vllm_config.kv_transfer_config is not None and \
self.vllm_config.kv_transfer_config.is_kv_consumer:
self.fused_qkv_a_proj.weight = None
self.fused_qkv_a_proj.deq_scale = None
self.fused_qkv_a_proj.quant_bias = None
self.q_proj.weight = None
self.q_proj.deq_scale = None
self.q_proj.quant_bias = None
torch.npu.empty_cache()
Comment on lines +674 to 682
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Dec 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

critical

The current logic for releasing MLAPO weights only covers the case where the node is a KV consumer in a Prefill-Decode (PD) mixed scenario. This correctly fixes the original bug but introduces a memory leak in non-PD (standalone) scenarios.

Previously, self.fused_qkv_a_proj.weight was released unconditionally after its data was processed. With this change, it is no longer released in non-PD scenarios, as self.vllm_config.kv_transfer_config would be None.

The logic should be updated to release the processed weights in both non-PD scenarios and on the consumer side of PD scenarios. This ensures memory is freed correctly in all configurations.

Suggested change
if self.vllm_config.kv_transfer_config is not None and \
self.vllm_config.kv_transfer_config.is_kv_consumer:
self.fused_qkv_a_proj.weight = None
self.fused_qkv_a_proj.deq_scale = None
self.fused_qkv_a_proj.quant_bias = None
self.q_proj.weight = None
self.q_proj.deq_scale = None
self.q_proj.quant_bias = None
torch.npu.empty_cache()
if self.vllm_config.kv_transfer_config is None or \
self.vllm_config.kv_transfer_config.is_kv_consumer:
self.fused_qkv_a_proj.weight = None
self.fused_qkv_a_proj.deq_scale = None
self.fused_qkv_a_proj.quant_bias = None
self.q_proj.weight = None
self.q_proj.deq_scale = None
self.q_proj.quant_bias = None
torch.npu.empty_cache()

Expand Down
Loading