GPTQ Lite implementation

modelopt/torch/quantization/config.py

@@ -234,6 +234,7 @@
     "algorithm": "max",
 }
 
+
 INT4_AWQ_CFG = {
     "quant_cfg": {
         "*weight_quantizer": {
@@ -1189,6 +1190,44 @@ class SVDQuantConfig(QuantizeAlgorithmConfig):
     )
 
 
+class GPTQLiteConfig(QuantizeAlgorithmConfig):
+    """The config for GPTQ lite.
+
+    GPTQ lite is a variant of GPTQ that does not exactly follow the official GPTQ implementation.
+
+    GPTQ lite does not perform sequential quantization of layers. This means that the updated
+    activations are not used to process the next layer.
+
+    The default values are taken from the official GPTQ implementation:
+    https://github.com/IST-DASLab/FP-Quant/blob/d2e3092f968262c4de5fb050e1aef568a280dadd/src/quantization/gptq.py#L35
+
+    Note: This feature is currently experimental and may not translate to improved accuracy as expected.
+
+
+    """
+
+    method: Literal["gptq_lite"] = ModeloptField("gptq_lite")
+    percdamp: float | None = ModeloptField(
+        default=0.01,
+        gt=0.0,
+        le=1.0,
+        title="Percentage damping factor.",
+        description="The percentage of average Hessian diagonal used for damping.",
+    )
+    block_size: int | None = ModeloptField(
+        default=128,
+        title="Block size for GPTQ weight update.",
+        description="""The block size for GPTQ weight update, which must be a multiple of the
+        group_size used in the quantization.""",
+    )
+    hessian_state_path: str | None = ModeloptField(
+        default=None,
+        title="Path to the Hessian state file.",
+        description="""The path to the Hessian state file. If hessian path exists, we load from
+         hessian file instead of recomputing them.""",
+    )
+
+
 QuantizeQuantCfgType = dict[
     str | Callable,
     QuantizerAttributeConfig

modelopt/torch/quantization/mode.py

@@ -37,6 +37,7 @@
     AWQFullCalibConfig,
     AWQLiteCalibConfig,
     CompressConfig,
+    GPTQLiteConfig,
     MaxCalibConfig,
     MseCalibConfig,
     QuantizeAlgoCfgType,
@@ -55,7 +56,7 @@
     restore_svdquant_model,
     update_quantize_metadata,
 )
-from .model_calib import awq, max_calibrate, mse_calibrate, smoothquant, svdquant
+from .model_calib import awq, gptq_lite, max_calibrate, mse_calibrate, smoothquant, svdquant
 
 __all__ = ["BaseCalibrateModeDescriptor"]
 
@@ -439,3 +440,15 @@ def config_class(self) -> type[QuantizeAlgorithmConfig]:
     def restore(self) -> RestoreEntrypoint:
         """The mode's entrypoint for restoring a model."""
         return restore_svdquant_model
+
+
+@CalibrateModeRegistry.register_mode
+class GPTQLiteModeDescriptor(BaseCalibrateModeDescriptor):
+    """Mode for GPTQ calibration algorithm."""
+
+    @property
+    def config_class(self) -> type[QuantizeAlgorithmConfig]:
+        """Specifies the config class for the mode."""
+        return GPTQLiteConfig
+
+    _calib_func = gptq_lite