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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Add Qwen2.5 Coder model support to registry and chat templates (#4221)
Daniel Hanfa61ee46d729
diff --git a/tests/test_model_registry.py b/tests/test_model_registry.pyindex 8552157620f..6c5aac52987 100644--- a/tests/test_model_registry.py+++ b/tests/test_model_registry.py@@ -268,3 +268,13 @@ def test_unreachable_hub_skips_instead_of_reporting_missing(monkeypatch, make_ermonkeypatch.setattr(sys.modules[__name__], "HfApi", lambda: _FakeApi(error))with pytest.raises(pytest.skip.Exception):_test_model_uploaded(["unsloth/Qwen2.5-7B"])+++def test_qwen_2_5_coder_registered_without_dynamic_quants():+ from unsloth.registry._qwen import Qwen_2_5_CoderMeta+ from unsloth.registry.registry import _register_models++ MODEL_REGISTRY.clear()+ _register_models(Qwen_2_5_CoderMeta)+ assert "unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit" in MODEL_REGISTRY+ assert not [k for k in MODEL_REGISTRY if "Coder" in k and k.endswith("unsloth-bnb-4bit")]diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.pyindex 83e4720e91b..7f9933ed96b 100644--- a/unsloth/chat_templates.py+++ b/unsloth/chat_templates.py@@ -648,6 +648,19 @@ DEFAULT_SYSTEM_MESSAGE["qwen25"] = qwen25_default_system_messageCHAT_TEMPLATES["qwen2.5"] = (qwen25_template, qwen25_template_eos_token, False, qwen25_ollama,)DEFAULT_SYSTEM_MESSAGE["qwen2.5"] = qwen25_default_system_message+qwen25_coder_ollama = _ollama_template("qwen-25-coder")+CHAT_TEMPLATES["qwen-2.5-coder"] = (qwen25_template, qwen25_template_eos_token, False, qwen25_coder_ollama,)+DEFAULT_SYSTEM_MESSAGE["qwen-2.5-coder"] = qwen25_default_system_message++CHAT_TEMPLATES["qwen-25-coder"] = (qwen25_template, qwen25_template_eos_token, False, qwen25_coder_ollama,)+DEFAULT_SYSTEM_MESSAGE["qwen-25-coder"] = qwen25_default_system_message++CHAT_TEMPLATES["qwen2.5-coder"] = (qwen25_template, qwen25_template_eos_token, False, qwen25_coder_ollama,)+DEFAULT_SYSTEM_MESSAGE["qwen2.5-coder"] = qwen25_default_system_message++CHAT_TEMPLATES["qwen25-coder"] = (qwen25_template, qwen25_template_eos_token, False, qwen25_coder_ollama,)+DEFAULT_SYSTEM_MESSAGE["qwen25-coder"] = qwen25_default_system_message+# "{{ bos_token }}"\ # Phi-4 removes BOS?# =========================================== Phi-4phi4_template = \diff --git a/unsloth/registry/REGISTRY.md b/unsloth/registry/REGISTRY.mdindex a0b3d96cade..540572939c6 100644--- a/unsloth/registry/REGISTRY.md+++ b/unsloth/registry/REGISTRY.md@@ -99,7 +99,7 @@ Prints the following (abridged) output:- [x] MistralSmall- [x] Qwen2.5- [x] Qwen2.5-VL- - [ ] Qwen2.5 Coder+ - [x] Qwen2.5 Coder- [x] QwenQwQ-32B- [x] Deepseek v3- [x] Deepseek R1diff --git a/unsloth/registry/_qwen.py b/unsloth/registry/_qwen.pyindex d754c633a96..31719cd95a1 100644--- a/unsloth/registry/_qwen.py+++ b/unsloth/registry/_qwen.py@@ -3,6 +3,7 @@ from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_IS_QWEN_2_5_REGISTERED = False_IS_QWEN_2_5_VL_REGISTERED = False_IS_QWEN_QWQ_REGISTERED = False+_IS_QWEN_2_5_CODER_REGISTERED = Falseclass QwenModelInfo(ModelInfo):@@ -33,6 +34,13 @@ class QwenQVQPreviewModelInfo(ModelInfo):return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)+class Qwen2_5CoderModelInfo(ModelInfo):+ @classmethod+ def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):+ key = f"{base_name}{version}-Coder-{size}B"+ return super().construct_model_name(base_name, version, size, quant_type, instruct_tag, key)++Qwen_2_5_Meta = ModelMeta(org = "Qwen",base_name = "Qwen",@@ -78,6 +86,16 @@ QwenQVQPreviewMeta = ModelMeta(is_multimodal = True,quant_types = [QuantType.BNB],)+Qwen_2_5_CoderMeta = ModelMeta(+ org = "Qwen",+ base_name = "Qwen",+ instruct_tags = [None, "Instruct"],+ model_version = "2.5",+ model_sizes = ["0.5", "1.5", "3", "7", "14", "32"],+ model_info_cls = Qwen2_5CoderModelInfo,+ is_multimodal = False,+ quant_types = [QuantType.NONE, QuantType.BNB], # no *-unsloth-bnb-4bit Coder uploads exist+)def register_qwen_2_5_models(include_original_model: bool = False):@@ -105,10 +123,19 @@ def register_qwen_qwq_models(include_original_model: bool = False):_IS_QWEN_QWQ_REGISTERED = True+def register_qwen_2_5_coder_models(include_original_model: bool = False):+ global _IS_QWEN_2_5_CODER_REGISTERED+ if _IS_QWEN_2_5_CODER_REGISTERED:+ return+ _register_models(Qwen_2_5_CoderMeta, include_original_model = include_original_model)+ _IS_QWEN_2_5_CODER_REGISTERED = True++def register_qwen_models(include_original_model: bool = False):register_qwen_2_5_models(include_original_model = include_original_model)register_qwen_2_5_vl_models(include_original_model = include_original_model)register_qwen_qwq_models(include_original_model = include_original_model)+ register_qwen_2_5_coder_models(include_original_model = include_original_model)if __name__ == "__main__":