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

Revert "[Refactor] Unify full-graph parameter update logic (#6041)" (#6227)

Browse Source 
This reverts commit 8966a99710.

It breaks the test
`tests/e2e/singlecard/spec_decode/test_mtp_eagle_correctness.py::test_deepseek_mtp_correctness[True-FULL_DECODE_ONLY-2-wemaster/deepseek_mtp_main_random_bf16]`

- vLLM version: v0.14.0
- vLLM main:
d68209402d
This commit is contained in:
wangxiyuan
2026-01-25 15:25:38 +08:00
committed by GitHub
parent 7799c4ca3b
commit 95649344aa
10 changed files with 415 additions and 420 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/distributed/kv_transfer/kv_pool/cpu_offload/cpu_offload_connector.py
View File

@@ -416,7 +416,7 @@ def get_kv_cache_spec(vllm_config: VllmConfig) -> dict[str, KVCacheSpec]:
kv_cache_spec: dict[str, KVCacheSpec] = {}
attn_layers = get_layers_from_vllm_config(vllm_config, AttentionLayerBase)
# NOTE: Must process Attention/MLAAttention before MambaBase to maintain
# ordering expected by graph parameter update logic in attention backends.
# ordering expected by acl_graph.py's _update_attn_fia_params.
mamba_layers: dict[str, MambaBase] = {}
for layer_name, attn_module in attn_layers.items():
if isinstance(attn_module, Attention):