[ROCm] Add MLA dual RMS norm fusion (Q, KV) pass for DeepSeek/Kimi-K2 (#39242)
Signed-off-by: Rita Brugarolas Brufau <rita.brugarolasbrufau@amd.com>
This commit is contained in:
@@ -56,6 +56,7 @@ Fusions:
|
||||
- `-cc.pass_config.fuse_norm_quant=True`*
|
||||
- `-cc.pass_config.fuse_act_quant=True`*
|
||||
- `-cc.pass_config.fuse_act_padding=True`†
|
||||
- `-cc.pass_config.fuse_mla_dual_rms_norm=True`†
|
||||
|
||||
\* These fusions are only enabled when either op is using a custom kernel, otherwise Inductor fusion is better.</br>
|
||||
† These fusions are ROCm-only and require AITER.
|
||||
|
||||
Reference in New Issue
Block a user