[Model][Frontend] Add Ling 3.0 Flash BF16, MTP, and parser support (#51045)

Signed-off-by: zexplorerhj <zhjoneson@163.com>
This commit is contained in:
zexplorerhj
2026-08-06 00:38:28 +08:00
committed by GitHub
parent 66b3c0e61f
commit d4da0c55af
15 changed files with 2070 additions and 1 deletions
@@ -289,6 +289,7 @@ class ModelArchConfigConvertorBase:
"pangu_ultra_moe_mtp",
"bailing_hybrid",
"bailing_hybrid_mtp",
"bailing_hybrid_v3_mtp",
):
# check is deepseek_v4 model
if hasattr(self.hf_text_config, "compress_ratios"):
@@ -571,6 +572,12 @@ class BailingHybridMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
return getattr(self.hf_text_config, "num_nextn_predict_layers", 0)
class BailingHybridV3MTPModelArchConfigConvertor(
BailingHybridMTPModelArchConfigConvertor
):
pass
class Qwen3_5MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def get_num_hidden_layers(self) -> int:
return getattr(self.hf_text_config, "mtp_num_hidden_layers", 0)
@@ -693,6 +700,7 @@ MODEL_ARCH_CONFIG_CONVERTORS = {
"moss_audio": MossAudioModelArchConfigConvertor,
"mpt": MPTModelArchConfigConvertor,
"nemotron-nas": NemotronNasModelArchConfigConvertor,
"bailing_hybrid_v3_mtp": BailingHybridV3MTPModelArchConfigConvertor,
"pangu_ultra_moe_mtp": PanguUltraMoeMTPModelArchConfigConvertor,
"qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor,
"qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor,