Enable ModelOpt Llama4 fp8 checkpoint deployment (#20419)

Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>
This commit is contained in:
Zhiyu
2025-07-11 23:07:16 -07:00
committed by GitHub
parent 5de8d9f111
commit 4afe687a82
5 changed files with 501 additions and 35 deletions
@@ -762,6 +762,10 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
modelopt_scale_names = [
".self_attn.k_proj.k_scale", ".self_attn.v_proj.v_scale"
]
# Also support qkv_proj scale parameters (from stacked parameter processing)
qkv_proj_scale_names = [
".self_attn.qkv_proj.k_scale", ".self_attn.qkv_proj.v_scale"
]
for scale_name in possible_scale_names:
if name.endswith(scale_name):
if any(mo_scale_name in name
@@ -769,6 +773,12 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
remapped_name = name.replace(
f".self_attn.{scale_name[1]}_proj{scale_name}",
f".self_attn.attn{scale_name}")
elif any(qkv_scale_name in name
for qkv_scale_name in qkv_proj_scale_names):
# Handle qkv_proj scale parameters
remapped_name = name.replace(
f".self_attn.qkv_proj{scale_name}",
f".self_attn.attn{scale_name}")
else:
remapped_name = name.replace(scale_name, f".attn{scale_name}")
if remapped_name not in params_dict: