Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Load Qwen2.5-Omni checkpoints that name Qwen2_5OmniModel (#13087)

Daniel Han9656b589717e
diff --git a/tests/test_omni_auto_class_resolution.py b/tests/test_omni_auto_class_resolution.py
index d963536097b..6c8ed70275f 100644
--- a/tests/test_omni_auto_class_resolution.py
+++ b/tests/test_omni_auto_class_resolution.py
@@ -233,3 +233,47 @@ def test_a_normal_model_still_gets_its_hook():
model = _tiny(Normal)
model.enable_input_require_grads()
assert model.embed._forward_hooks, "the input-gradient hook must be registered"
+
+
+class _Routed(Exception):
+ pass
+
+
+def _route(tmp_path, monkeypatch, config):
+ import json
+ import unsloth.models.loader as loader
+
+ (tmp_path / "config.json").write_text(json.dumps(config.to_dict()))
+
+ def capture(*args, **kwargs):
+ raise _Routed(kwargs.get("auto_model"))
+
+ monkeypatch.setattr(loader.FastBaseModel, "from_pretrained", staticmethod(capture))
+ # Its pinned host buffer needs an accelerator; CPU CI has none.
+ monkeypatch.setattr(loader, "apply_unsloth_gradient_checkpointing", lambda value, *args: value)
+ with pytest.raises(_Routed) as routed:
+ loader.FastModel.from_pretrained(str(tmp_path), load_in_4bit = False)
+ return routed.value.args[0]
+
+
+ "architecture", ["Qwen2_5OmniModel", "Qwen2_5OmniForConditionalGeneration"]
+)
+def test_qwen2_5_omni_routes_to_a_multimodal_class(tmp_path, monkeypatch, architecture):
+ """Qwen/Qwen2.5-Omni-3B / -7B name `Qwen2_5OmniModel` and nest vision under thinker_config (#2325)."""
+ config_class = getattr(transformers, "Qwen2_5OmniConfig", None)
+ if config_class is None:
+ pytest.skip("this transformers has no Qwen2.5-Omni")
+ config = config_class()
+ config.architectures = [architecture]
+ auto_model = _route(tmp_path, monkeypatch, config)
+ assert auto_model is not transformers.AutoModelForCausalLM
+ assert resolve_model_class(auto_model, config) is not None
+
+
+def test_a_text_checkpoint_still_routes_to_causal_lm(tmp_path, monkeypatch):
+ config = transformers.Qwen2Config(
+ hidden_size = 8, num_hidden_layers = 1, num_attention_heads = 1, intermediate_size = 8
+ )
+ config.architectures = ["Qwen2ForCausalLM"]
+ assert _route(tmp_path, monkeypatch, config) is transformers.AutoModelForCausalLM
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 29e75e194d7..fd3ef675b90 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -622,6 +622,13 @@ def _built_by_unsloth_compiler(cls):
return False
+def _has_vision_config(config):
+ # Qwen2.5-Omni checkpoints name `Qwen2_5OmniModel` (no ForConditionalGeneration) and nest vision under thinker_config.
+ return any(
+ hasattr(sub, "vision_config") for sub in (config, getattr(config, "thinker_config", None))
+ )
+
+
def _resolve_omni_auto_model(
model_config,
trust_remote_code = None,
@@ -2489,9 +2496,9 @@ class FastModel(FastBaseModel):
# Keep the local checkpoint dir as tokenizer when self-sufficient; a VLM also needs local processor files, else fall back to the base repo so its cached processor loads.
_ckpt_arch = getattr(model_config, "architectures", None) or []
- _ckpt_is_vlm = any(x.endswith("ForConditionalGeneration") for x in _ckpt_arch) or hasattr(
- model_config, "vision_config"
- )
+ _ckpt_is_vlm = any(
+ x.endswith("ForConditionalGeneration") for x in _ckpt_arch
+ ) or _has_vision_config(model_config)
# T5 / BART end in ForConditionalGeneration too but ship a tokenizer, not a processor.
_ckpt_is_vlm = _ckpt_is_vlm and not _is_text_seq2seq_config(model_config)
tokenizer_name = _resolve_checkpoint_tokenizer_name(
@@ -2523,7 +2530,7 @@ class FastModel(FastBaseModel):
if architectures is None:
architectures = []
is_vlm = any(x.endswith("ForConditionalGeneration") for x in architectures)
- is_vlm = is_vlm or hasattr(model_config, "vision_config")
+ is_vlm = is_vlm or _has_vision_config(model_config)
if (
is_peft
and not text_only