[Model] Support Cloudflare/clef-omni on /v1/systemone (#61064)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Jiangyun Zhu
2026-10-12 21:27:34 +08:00
committed by GitHub
co-authored by Claude Opus 5.5 Isotr0py
parent 314ab8cffd
commit 092c35dd15
20 changed files with 1592 additions and 69 deletions
@@ -17,6 +17,10 @@ from vllm.transformers_utils.config import (
ConfigFormat,
get_safetensors_params_metadata,
)
from vllm.transformers_utils.repo_utils import (
file_or_path_exists,
get_hf_file_to_dict,
)
from vllm.utils.torch_utils import common_broadcastable_dtype
logger = init_logger(__name__)
@@ -643,6 +647,31 @@ class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):
_strip_mimo_v2_attention_chunk_size(hf_config, hf_text_config)
class Qwen3OmniMoeModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
hf_config: PreTrainedConfig,
hf_text_config: PreTrainedConfig,
revision: str | None = None,
):
# Clef-Omni keeps Qwen3-Omni's config and adds a joint schema head;
# route to its wrapper class when the head's config file is present.
# Mutate hf_config (like the deepseek_v4/mimo_v2 rewrites above)
# because model-class resolution reads hf_config.architectures.
if getattr(hf_config, "architectures", None) == [
"Qwen3OmniMoeForConditionalGeneration"
]:
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 = ["ClefOmniForJointSchema"]
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 MimoV2MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
@@ -822,6 +851,7 @@ MODEL_ARCH_CONFIG_CONVERTORS = {
"bailing_hybrid_v3_mtp": BailingHybridV3MTPModelArchConfigConvertor,
"pangu_ultra_moe_mtp": PanguUltraMoeMTPModelArchConfigConvertor,
"qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor,
"qwen3_omni_moe": Qwen3OmniMoeModelArchConfigConvertor,
"qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor,
"qwen4_exp_mtp": Qwen4ExpMTPModelArchConfigConvertor,
"RefinedWeb": FalconModelArchConfigConvertor,