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.

Let AWQ checkpoints train with LoRA (#13584)

Daniel Han8e68a99263f6
diff --git a/tests/test_gptq_packed_lora.py b/tests/test_gptq_packed_lora.py
index 362441e13ff..b531f8f07b6 100644
--- a/tests/test_gptq_packed_lora.py
+++ b/tests/test_gptq_packed_lora.py
@@ -96,13 +96,35 @@ def test_gptq_backend_left_alone(monkeypatch):
gptq_trainable_quantization_config(_config({**_GPTQ8, "backend": "gptq_torch"}), None)
is None
)
- assert (
- gptq_trainable_quantization_config(_config({"quant_method": "awq", "bits": 4}), None)
- is None
- )
assert (
gptq_trainable_quantization_config(_config({"quant_method": "bitsandbytes"}), None) is None
)
assert gptq_trainable_quantization_config(_config(None), None) is None
monkeypatch.setattr(transformers.utils, "is_gptqmodel_available", lambda: False)
assert gptq_trainable_quantization_config(_config(dict(_GPTQ8)), None) is None
+
+
+_AWQ4 = {"quant_method": "awq", "bits": 4, "group_size": 64, "zero_point": True, "version": "gemm"}
+
+
+def test_awq_checkpoint_gets_trainable_backend(monkeypatch):
+ from transformers.quantizers.auto import AutoHfQuantizer
+
+ monkeypatch.setattr(transformers.utils, "is_gptqmodel_available", lambda: True)
+ cfg = gptq_trainable_quantization_config(_config(dict(_AWQ4)), None)
+ assert type(cfg).__name__ == "AwqConfig" and cfg.backend == "auto_trainable"
+ merged = AutoHfQuantizer.merge_quantization_configs(dict(_AWQ4), cfg)
+ assert (merged.bits, merged.group_size, merged.zero_point) == (4, 64, True)
+ assert merged.format == "gemm" and merged.backend == "auto_trainable"
+ # Configs saved by transformers 4.x (Qwen/Qwen3-8B-AWQ) carry the legacy default backend.
+ for legacy in ("autoawq", "llm-awq"):
+ legacy_cfg = gptq_trainable_quantization_config(_config({**_AWQ4, "backend": legacy}), None)
+ assert legacy_cfg is not None and legacy_cfg.backend == "auto_trainable"
+
+
+def test_awq_backend_left_alone(monkeypatch):
+ monkeypatch.setattr(transformers.utils, "is_gptqmodel_available", lambda: True)
+ assert gptq_trainable_quantization_config(_config(dict(_AWQ4)), object()) is None
+ assert gptq_trainable_quantization_config(_config({**_AWQ4, "backend": "gemm"}), None) is None
+ monkeypatch.setattr(transformers.utils, "is_gptqmodel_available", lambda: False)
+ assert gptq_trainable_quantization_config(_config(dict(_AWQ4)), None) is None
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index e69da4a678a..e89486c203a 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -3364,7 +3364,7 @@ def sync_load_when_quantizing(quantization_config, model_config):
def gptq_trainable_quantization_config(model_config, user_quantization_config):
- """GPTQConfig asking gptqmodel for a trainable kernel on a GPTQ checkpoint, else None.
+ """GPTQConfig / AwqConfig asking gptqmodel for a trainable kernel on a GPTQ or AWQ checkpoint, else None.
gptqmodel's default kernels (Marlin / ExLlama) raise NotImplementedError on model.train(); only
`backend` is a loading attribute, so the checkpoint's own bits / group_size still apply.
@@ -3374,17 +3374,31 @@ def gptq_trainable_quantization_config(model_config, user_quantization_config):
qc = getattr(model_config, "quantization_config", None)
if qc is not None and not isinstance(qc, dict):
qc = qc.to_dict()
- if not qc or str(qc.get("quant_method", "")).lower() != "gptq":
+ quant_method = str((qc or {}).get("quant_method", "")).lower()
+ if quant_method not in ("gptq", "awq"):
return None
- if qc.get("backend") not in (None, "auto"):
+ # transformers 4.x wrote "autoawq" / "llm-awq" into AWQ configs (Qwen3-8B-AWQ); AwqConfig reads both as auto.
+ default_backends = (
+ (None, "auto", "autoawq", "llm-awq") if quant_method == "awq" else (None, "auto")
+ )
+ if qc.get("backend") not in default_backends:
return None
try:
- from transformers import GPTQConfig
+ from transformers import AwqConfig, GPTQConfig
from transformers.utils import is_gptqmodel_available
except ImportError:
return None
if not is_gptqmodel_available() or "backend" not in inspect.signature(GPTQConfig).parameters:
return None
+ if quant_method == "awq":
+ # transformers < 5 AwqConfig has no auto_trainable backend (autoawq era).
+ try:
+ from transformers.utils.quantization_config import AwqBackend
+ except ImportError:
+ return None
+ if not hasattr(AwqBackend, "AUTO_TRAINABLE"):
+ return None
+ return AwqConfig(bits = qc.get("bits", 4), backend = "auto_trainable")
return GPTQConfig(bits = qc["bits"], backend = "auto_trainable")