Only fuse if flash_attn_fuse_mlp is True
This commit is contained in:
@@ -385,6 +385,7 @@ def load_model(
|
||||
**model_kwargs,
|
||||
)
|
||||
|
||||
if cfg.flash_attn_fuse_mlp:
|
||||
LOG.info("Mixtral MoE: Replacing experts with SwiGLU")
|
||||
replace_mixtral_mlp_with_swiglu(model)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user