chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

support llama4 eagle3 (#6985)

Browse Source 
Co-authored-by: shuaills <shishuaiuoe@gmail.com>
Co-authored-by: Shenggui Li <somerlee.9@gmail.com>
Co-authored-by: Yingyi Huang <yingyihuang2000@outlook.com>
Co-authored-by: yizhang2077 <1109276519@qq.com>
This commit is contained in:
lukec
2025-07-01 13:34:10 +08:00
committed by GitHub
parent 637bfee448
commit 886d344964
7 changed files with 138 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
python/sglang/srt/model_loader/loader.py
View File

@@ -124,6 +124,9 @@ def _get_quantization_config(
quant_config = get_quant_config(
model_config, load_config, packed_modules_mapping
)
# (yizhang2077) workaround for nvidia/Llama-4-Maverick-17B-128E-Eagle3
if quant_config is None:
return None
major, minor = get_device_capability()
if major is not None and minor is not None:

11
python/sglang/srt/model_loader/weight_utils.py
View File

@@ -209,6 +209,17 @@ def get_quant_config(
config["adapter_name_or_path"] = model_name_or_path
elif model_config.quantization == "modelopt":
if config["producer"]["name"] == "modelopt":
# (yizhang2077) workaround for nvidia/Llama-4-Maverick-17B-128E-Eagle3
if config["quantization"]["quant_algo"] is None:
if (
model_config.hf_config.architectures[0]
!= "LlamaForCausalLMEagle3"
):
raise ValueError(
f"Invalid quant_config, quantization method: {model_config.quantization},"
f"hf architectures: {model_config.hf_config.architectures[0]}. "
)
return None
if "FP4" in config["quantization"]["quant_algo"]:
return ModelOptFp4Config.from_config(config)
else: