[DCP][Bugfix][CI] Fix accuracy issue of DCP when using FLASH_ATTN_MLA (#30309)
Signed-off-by: FENP <yuanyongjie.yyj@antgroup.com>
This commit is contained in:
@@ -123,8 +123,11 @@ class CPTestSettings:
|
||||
|
||||
CP_TEXT_GENERATION_MODELS = {
|
||||
"deepseek-ai/DeepSeek-V2-Lite-Chat": [
|
||||
CPTestSettings.detailed(dcp_multipliers=[1]),
|
||||
CPTestSettings.detailed(
|
||||
dcp_multipliers=[0.5, 1], cp_kv_cache_interleave_size=64
|
||||
dcp_multipliers=[0.5],
|
||||
cp_kv_cache_interleave_size=64,
|
||||
attn_backend="FLASHMLA",
|
||||
),
|
||||
],
|
||||
"Qwen/Qwen2.5-1.5B-Instruct": [
|
||||
|
||||
Reference in New Issue
Block a user