[Model] Add Clef-Flash support (#61308)

Signed-off-by: Tianyu Guo <guoty@inferact.ai>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Tianyu Guo
2026-10-13 05:46:37 +00:00
committed by GitHub
co-authored by Codex Isotr0py
parent 0d2689e0df
commit 380638afbb
11 changed files with 580 additions and 361 deletions
@@ -647,6 +647,27 @@ class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):
_strip_mimo_v2_attention_chunk_size(hf_config, hf_text_config)
class Qwen3_5ModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
hf_config: PreTrainedConfig,
hf_text_config: PreTrainedConfig,
revision: str | None = None,
):
if getattr(hf_config, "architectures", None) == [
"Qwen3_5ForConditionalGeneration"
]:
model_path = hf_config.name_or_path
if model_path and file_or_path_exists(
model_path, "joint_head_config.json", revision
):
hf_config.architectures = ["ClefFlashForJointSchema"]
hf_config.joint_head_config = get_hf_file_to_dict(
"joint_head_config.json", model_path, revision
)
super().__init__(hf_config, hf_text_config, revision)
class Qwen3OmniMoeModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
@@ -850,6 +871,7 @@ MODEL_ARCH_CONFIG_CONVERTORS = {
"nemotron-nas": NemotronNasModelArchConfigConvertor,
"bailing_hybrid_v3_mtp": BailingHybridV3MTPModelArchConfigConvertor,
"pangu_ultra_moe_mtp": PanguUltraMoeMTPModelArchConfigConvertor,
"qwen3_5": Qwen3_5ModelArchConfigConvertor,
"qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor,
"qwen3_omni_moe": Qwen3OmniMoeModelArchConfigConvertor,
"qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor,