revert is_marlin_format check (#1316)

Author: CSY-ModelCloud

gptqmodel/quantization/config.py

@@ -33,6 +33,7 @@
 
 FORMAT_FIELD_CODE = "format"
 FORMAT_FIELD_JSON = "checkpoint_format"
+FORMAT_FIELD_COMPAT_MARLIN = "is_marlin_format"
 QUANT_METHOD_FIELD = "quant_method"
 PACK_DTYPE_FIELD = "pack_dtype"
 QUANT_CONFIG_FILENAME = "quantize_config.json"
@@ -184,6 +185,8 @@ class QuantizeConfig():
     # pending used field
     adapter: Optional[Union[Dict[str, Any], Lora]] = field(default=None)
 
+    is_marlin_format: bool = False
+
     def __post_init__(self):
         fields_info = fields(self)
 
@@ -351,6 +354,8 @@ def from_quant_config(cls, quantize_cfg, format: str = None):
                     raise ValueError(f"QuantizeConfig: Unknown quantization method: `{val}`.")
                 else:
                     normalized[QUANT_METHOD_FIELD] = val
+            elif key == FORMAT_FIELD_COMPAT_MARLIN and val:
+                normalized[FORMAT_FIELD_CODE] = FORMAT.MARLIN
             elif key in field_names:
                 normalized[key] = val
             else:

tests/test_quant_formats.py

@@ -49,9 +49,9 @@ def setUpClass(self):
 
     @parameterized.expand(
         [
-            # (QUANT_METHOD.GPTQ, BACKEND.AUTO, False, FORMAT.GPTQ, 8),
+            (QUANT_METHOD.GPTQ, BACKEND.AUTO, False, FORMAT.GPTQ, 8),
             (QUANT_METHOD.GPTQ, BACKEND.EXLLAMA_V2, True, FORMAT.GPTQ_V2, 4),
-            # (QUANT_METHOD.GPTQ, BACKEND.EXLLAMA_V2, False, FORMAT.GPTQ, 4),
+            (QUANT_METHOD.GPTQ, BACKEND.EXLLAMA_V2, False, FORMAT.GPTQ, 4),
         ]
     )
     def test_quantize(self, method: QUANT_METHOD, backend: BACKEND, sym: bool, format: FORMAT, bits: int):
@@ -115,12 +115,13 @@ def test_quantize(self, method: QUANT_METHOD, backend: BACKEND, sym: bool, forma
             if not sym and format == FORMAT.GPTQ or format == FORMAT.IPEX:
                 return
 
-            # test compat: 1) with simple dict type
+            # test compat: 1) with simple dict type 2) is_marlin_format
             compat_quantize_config = {
                 "bits": bits,
                 "group_size": 128,
                 "sym": sym,
                 "desc_act": False if format == FORMAT.MARLIN else True,
+                "is_marlin_format": backend == BACKEND.MARLIN,
             }
 
             model = GPTQModel.load(