[Model]Support Step-3.7-Flash (#43859)

Signed-off-by: luotingdan <luotingdan@stepfun.com>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Co-authored-by: luotingdan <luotingdan@stepfun.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Yu Huang <yuhuang@nvidia.com>
Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
ltd0924
2026-05-28 17:01:48 -07:00
committed by GitHub
co-authored by luotingdan Isotr0py Yu Huang Jee Jee Li
parent 325a1ec4fb
commit b690b2bb67
15 changed files with 698 additions and 27 deletions
+53 -1
View File
@@ -637,6 +637,54 @@ class Step3p5Model(nn.Module):
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
(
f".moe.experts.{base_layer}w13_weight_scale_2",
".moe.gate_proj.weight_scale_2",
"w1",
),
(
f".moe.experts.{base_layer}w13_weight_scale_2",
".moe.up_proj.weight_scale_2",
"w3",
),
(
f".moe.experts.{base_layer}w2_weight_scale_2",
".moe.down_proj.weight_scale_2",
"w2",
),
(
f".moe.experts.{base_layer}w13_weight_scale",
".moe.gate_proj.weight_scale",
"w1",
),
(
f".moe.experts.{base_layer}w13_weight_scale",
".moe.up_proj.weight_scale",
"w3",
),
(
f".moe.experts.{base_layer}w2_weight_scale",
".moe.down_proj.weight_scale",
"w2",
),
# Required due to the Step3 HF model's packed expert format:
# input scales are stored as moe.{gate,up,down}_proj.input_scale
# rather than the standard per-expert format handled generically.
(
f".moe.experts.{base_layer}w13_input_scale",
".moe.gate_proj.input_scale",
"w1",
),
(
f".moe.experts.{base_layer}w13_input_scale",
".moe.up_proj.input_scale",
"w3",
),
(
f".moe.experts.{base_layer}w2_input_scale",
".moe.down_proj.input_scale",
"w2",
),
]
# New per-expert format: .moe.experts.E.gate_proj.weight_packed [out, in]
@@ -756,7 +804,11 @@ class Step3p5Model(nn.Module):
# Per-tensor global scales (e.g. weight_global_scale)
# have shape [1] in compressed-tensors NVFP4 checkpoints.
# Expand to per-expert before the iteration loop.
if (
if loaded_weight.ndim == 0:
loaded_weight = loaded_weight.unsqueeze(0).expand(
moe_expert_num
)
elif (
loaded_weight.shape[0] == 1
and loaded_weight.shape[0] != moe_expert_num
):