EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Support --enable-llama4-multimodal (#5254)

Browse Source
This commit is contained in:
Cheng Wan
2025-04-11 04:24:14 -04:00
committed by GitHub
parent aee62d744b
commit 038bc5d521
7 changed files with 32 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
python/sglang/srt/configs/model_config.py
View File

@@ -43,10 +43,12 @@ class ModelConfig:
context_length: Optional[int] = None,
model_override_args: Optional[str] = None,
is_embedding: Optional[bool] = None,
enable_multimodal: Optional[bool] = None,
dtype: str = "auto",
quantization: Optional[str] = None,
override_config_file: Optional[str] = None,
) -> None:
self.model_path = model_path
self.revision = revision
self.quantization = quantization
@@ -70,14 +72,28 @@ class ModelConfig:
self.hf_text_config, "attention_chunk_size", None
)
if enable_multimodal is None:
if self.hf_config.architectures == "Llama4ForConditionalGeneration":
enable_multimodal = False
else:
enable_multimodal = True
# Check model type
self.is_generation = is_generation_model(
self.hf_config.architectures, is_embedding
)
self.is_multimodal = is_multimodal_model(self.hf_config.architectures)
self.is_multimodal_gen = is_multimodal_gen_model(self.hf_config.architectures)
self.is_image_gen = is_image_gen_model(self.hf_config.architectures)
self.is_audio_model = is_audio_model(self.hf_config.architectures)
self.is_multimodal = enable_multimodal and is_multimodal_model(
self.hf_config.architectures
)
self.is_multimodal_gen = enable_multimodal and is_multimodal_gen_model(
self.hf_config.architectures
)
self.is_image_gen = enable_multimodal and is_image_gen_model(
self.hf_config.architectures
)
self.is_audio_model = enable_multimodal and is_audio_model(
self.hf_config.architectures
)
self.is_encoder_decoder = is_encoder_decoder_model(self.hf_config.architectures)
self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)