[New model][Multimodal] Add DeepSeek-V4-Flash-Vision-Exp support (#54566)
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
@@ -465,3 +465,34 @@ def test_draft_model_arch_config(
|
||||
_assert_model_config_methods(
|
||||
model_config, expected, check_head_size=check_head_size
|
||||
)
|
||||
|
||||
|
||||
def test_deepseek_v4_convertor_splits_vision_architecture():
|
||||
"""The vision checkpoint shares model_type/architectures with the
|
||||
text-only DeepSeek-V4; the convertor routes it to the VL wrapper class
|
||||
by rewriting hf_config.architectures (model-class resolution reads the
|
||||
raw attribute)."""
|
||||
from vllm.transformers_utils.configs.deepseek_v4 import DeepseekV4Config
|
||||
from vllm.transformers_utils.model_arch_config_convertor import (
|
||||
DeepseekV4ModelArchConfigConvertor,
|
||||
)
|
||||
|
||||
text_cfg = DeepseekV4Config(architectures=["DeepseekV4ForCausalLM"])
|
||||
conv = DeepseekV4ModelArchConfigConvertor(text_cfg, text_cfg)
|
||||
assert conv.get_architectures() == ["DeepseekV4ForCausalLM"]
|
||||
assert text_cfg.architectures == ["DeepseekV4ForCausalLM"]
|
||||
|
||||
vl_cfg = DeepseekV4Config(
|
||||
architectures=["DeepseekV4ForCausalLM"], vision_n_layers=32
|
||||
)
|
||||
conv = DeepseekV4ModelArchConfigConvertor(vl_cfg, vl_cfg)
|
||||
assert conv.get_architectures() == ["DeepseekV4ForConditionalGeneration"]
|
||||
assert vl_cfg.architectures == ["DeepseekV4ForConditionalGeneration"]
|
||||
|
||||
# Speculative-draft configs (set by SpeculativeConfig.hf_config_override)
|
||||
# keep their own architecture even with a vision tower in the config.
|
||||
for draft_arch in ("DSparkDraftModel", "DeepSeekV4MTPModel"):
|
||||
draft_cfg = DeepseekV4Config(architectures=[draft_arch], vision_n_layers=32)
|
||||
conv = DeepseekV4ModelArchConfigConvertor(draft_cfg, draft_cfg)
|
||||
assert conv.get_architectures() == [draft_arch]
|
||||
assert draft_cfg.architectures == [draft_arch]
|
||||
|
||||
Reference in New Issue
Block a user