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

【main】【bugfix】fix: restrict default MLAPO activation to Decode nodes only (#6451)

Browse Source 
### What this PR does / why we need it?
There is an issue with the current default logic for MLAPO (MLA Policy
Optimization). By design, MLAPO should only be enabled by default on
Decode (D) nodes. However, in hybrid (collocated prefill and decode)
scenarios, the strategy is erroneously activated during the Prefill
stage.
This PR corrects the default behavior to ensure that MLAPO is
exclusively enabled for the Decoding phase. This prevents unexpected
policy interference during Prefill and ensures optimal performance in
hybrid deployment environments.
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.14.1
- vLLM main:
dc917cceb8

---------

Signed-off-by: fems14 <1804143737@qq.com>
This commit is contained in:
fems14
2026-01-31 22:44:56 +08:00
committed by GitHub
parent ef02d20086
commit 775fbc4cd2
2 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/attention/mla_v1.py
View File

@@ -22,7 +22,7 @@ from vllm_ascend.attention.utils import (
AscendCommonAttentionMetadata, AscendCommonAttentionMetadata,
ascend_chunked_prefill_workspace_size, ascend_chunked_prefill_workspace_size,
enable_cp, enable_cp,
enabling_malpo, enabling_mlapo,
maybe_save_kv_layer_to_connector, maybe_save_kv_layer_to_connector,
split_decodes_and_prefills, split_decodes_and_prefills,
trans_rope_weight, trans_rope_weight,
@@ -710,7 +710,7 @@ class AscendMLAImpl(MLAAttentionImpl):
self.ring_mla_mask_size = 512 self.ring_mla_mask_size = 512
self.speculative_config = self.vllm_config.speculative_config self.speculative_config = self.vllm_config.speculative_config
self.enable_mlapo = enabling_malpo(self.vllm_config) self.enable_mlapo = enabling_mlapo(self.vllm_config)
self.is_kv_producer = ( self.is_kv_producer = (
self.vllm_config.kv_transfer_config is not None and self.vllm_config.kv_transfer_config.is_kv_producer self.vllm_config.kv_transfer_config is not None and self.vllm_config.kv_transfer_config.is_kv_producer

8
vllm_ascend/attention/utils.py
View File

@@ -305,6 +305,10 @@ def transdata(nd_mat, block_size: tuple = (16, 16)):
return nz_mat return nz_mat
def enabling_malpo(vllm_config: VllmConfig) -> bool: def enabling_mlapo(vllm_config: VllmConfig) -> bool:
is_decode_instance = vllm_config.kv_transfer_config is not None and vllm_config.kv_transfer_config.is_kv_consumer is_decode_instance = (
vllm_config.kv_transfer_config is not None
and vllm_config.kv_transfer_config.is_kv_consumer
and not vllm_config.kv_transfer_config.is_kv_producer
)
return bool(envs.VLLM_ASCEND_ENABLE_MLAPO and is_decode_instance) return bool(envs.VLLM_ASCEND_ENABLE_MLAPO and is_decode_instance)