[quantization] Support override in GPTQ

This commit supports override option in GPTQ.

TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>

Author: mhs4670go

test/quantization/algorithm/test_gptq.py

@@ -19,6 +19,7 @@
 import torch
 
 from tico.quantization import convert, prepare
+from tico.quantization.algorithm.gptq.quantizer import GPTQQuantizer
 from tico.quantization.algorithm.gptq.utils import SensitivityCalibrator
 from tico.quantization.config.gptq import GPTQConfig
 from tico.quantization.config.ptq import PTQConfig
@@ -48,6 +49,23 @@ def get_zero_inputs(self):
         return (torch.zeros(1, 2048),), {}
 
 
+class SmallLinear(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.m = torch.nn.ModuleList()
+        for _ in range(3):
+            self.m.append(torch.nn.Linear(16, 16))
+
+    def forward(self, x):
+        z = self.m[0](x)
+        z = self.m[1](z)
+        z = self.m[2](z)
+        return z
+
+    def get_example_inputs(self):
+        return (torch.randn(1, 16),), {}
+
+
 class NormConv2D(torch.nn.Module):
     def __init__(self):
         super().__init__()
@@ -278,6 +296,87 @@ def get_example_inputs(self):
 
 
 class GPTQTest(unittest.TestCase):
+    def test_gptq_config_validate_weight_bits_overrides(self):
+        conf = GPTQConfig(weight_bits=4, weight_bits_overrides={"m.1": 8})
+        conf.validate()
+
+    def test_gptq_config_validate_rejects_non_positive_weight_bits_override(self):
+        conf = GPTQConfig(weight_bits=4, weight_bits_overrides={"m.1": 0})
+        with self.assertRaises(ValueError):
+            conf.validate()
+
+    def test_resolve_weight_bits_priority(self):
+        quantizer = GPTQQuantizer(
+            GPTQConfig(
+                weight_bits=4,
+                weight_bits_overrides={
+                    "proj": 5,
+                    "layer.proj": 6,
+                    "model.layers.0.layer.proj": 8,
+                },
+            )
+        )
+
+        assert isinstance(quantizer.config, GPTQConfig)
+        self.assertEqual(
+            quantizer._resolve_weight_bits(
+                quantizer.config,
+                full_module_name="model.layers.0.layer.proj",
+                local_module_name="layer.proj",
+            ),
+            8,
+        )
+        self.assertEqual(
+            quantizer._resolve_weight_bits(
+                quantizer.config,
+                full_module_name="model.layers.1.layer.proj",
+                local_module_name="layer.proj",
+            ),
+            6,
+        )
+        self.assertEqual(
+            quantizer._resolve_weight_bits(
+                quantizer.config,
+                full_module_name="model.layers.2.other.proj",
+                local_module_name="other.proj",
+            ),
+            5,
+        )
+        self.assertEqual(
+            quantizer._resolve_weight_bits(
+                quantizer.config,
+                full_module_name="model.layers.2.other.up_proj",
+                local_module_name="other.up_proj",
+            ),
+            4,
+        )
+
+    @torch.inference_mode()
+    def test_weight_bits_overrides_are_applied_per_module(self):
+        q_m = SmallLinear()
+        q_m.eval()
+        ori_m = q_m
+
+        q_m = prepare(
+            q_m,
+            GPTQConfig(
+                show_progress=False,
+                weight_bits=4,
+                weight_bits_overrides={
+                    "m.1": 8,
+                },
+            ),
+        )
+        for _ in range(8):
+            args, kwargs = ori_m.get_example_inputs()
+            q_m(*args, **kwargs)
+        convert(q_m, inplace=True)
+
+        self.assertTrue(hasattr(q_m, "quantizers"))
+        self.assertEqual(q_m.quantizers["model.layers.0.m.0"].maxq.item(), 15)
+        self.assertEqual(q_m.quantizers["model.layers.0.m.1"].maxq.item(), 255)
+        self.assertEqual(q_m.quantizers["model.layers.0.m.2"].maxq.item(), 15)
+
     @unittest.skipIf(
         not IS_INTERNAL_TEST, "Internal test — run only if --include-internal is set"
     )

tico/quantization/algorithm/gptq/quantizer.py

@@ -60,6 +60,31 @@ def __init__(self, config: GPTQConfig):
         self._orig_layer_forward: Optional[Callable[..., Any]] = None
         self._first_layer_ref: Optional[torch.nn.Module] = None
 
+    def _resolve_weight_bits(
+        self,
+        gptq_conf: GPTQConfig,
+        *,
+        full_module_name: str,
+        local_module_name: str,
+    ) -> int:
+        """Resolve the effective bit-width for a quantized submodule."""
+        if full_module_name in gptq_conf.weight_bits_overrides:
+            return gptq_conf.weight_bits_overrides[full_module_name]
+
+        if local_module_name in gptq_conf.weight_bits_overrides:
+            return gptq_conf.weight_bits_overrides[local_module_name]
+
+        suffix_matches = [
+            bits
+            for pattern, bits in gptq_conf.weight_bits_overrides.items()
+            if full_module_name.endswith(f".{pattern}")
+        ]
+
+        if suffix_matches:
+            return suffix_matches[-1]
+
+        return gptq_conf.weight_bits
+
     @torch.no_grad()
     def prepare(
         self,
@@ -220,18 +245,22 @@ def convert(self, model):
                 gptq: Dict[str, GPTQ] = {}
                 for name in subset:
                     gptq[name] = GPTQ(subset[name])
+                    full_module_name = module_name[subset[name]]
+                    weight_bits = self._resolve_weight_bits(
+                        gptq_conf,
+                        full_module_name=full_module_name,
+                        local_module_name=name,
+                    )
                     if (
                         gptq_conf.sensitivity is not None
                         and isinstance(gptq_conf.sensitivity, dict)
-                        and module_name[subset[name]] in gptq_conf.sensitivity
+                        and full_module_name in gptq_conf.sensitivity
                     ):
-                        cur_sensitivity = gptq_conf.sensitivity[
-                            module_name[subset[name]]
-                        ]
+                        cur_sensitivity = gptq_conf.sensitivity[full_module_name]
                     else:
                         cur_sensitivity = None
                     gptq[name].quantizer.configure(
-                        bits=gptq_conf.weight_bits,
+                        bits=weight_bits,
                         perchannel=gptq_conf.perchannel,
                         sym=gptq_conf.symmetric,
                         mse=gptq_conf.mse,

tico/quantization/config/gptq.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 
 import torch
 
@@ -23,6 +23,21 @@
 class GPTQConfig(BaseConfig):
     """
     Configuration for GPTQ weight quantization.
+
+    Attributes
+    ----------
+    weight_bits : int
+        Default bit-width applied to quantized weights.
+    weight_bits_overrides : dict[str, int]
+        Optional per-module bit-width overrides.
+
+        Supported keys are matched in the following order:
+          1) Full module name, for example `model.layers.0.self_attn.o_proj`
+          2) Layer-local module name, for example `self_attn.o_proj`
+          3) Full-name suffix, for example `self_attn.o_proj` or `down_proj`
+
+        This makes it possible to keep a default bit-width for most modules
+        while selectively increasing precision for specific projections.
     """
 
     # general
@@ -31,6 +46,7 @@ class GPTQConfig(BaseConfig):
 
     # quantizer.configure params (weight quantization spec)
     weight_bits: int = 8
+    weight_bits_overrides: dict[str, int] = field(default_factory=dict)
     perchannel: bool = True
     symmetric: bool = False
     mse: str | None = None
@@ -49,6 +65,11 @@ def name(self) -> str:
     def validate(self) -> None:
         if self.weight_bits <= 0:
             raise ValueError(f"weight_bits must be positive. got {self.weight_bits}")
+        for module_name, bits in self.weight_bits_overrides.items():
+            if bits <= 0:
+                raise ValueError(
+                    f"weight_bits_overrides[{module_name!r}] must be positive. got {bits}"
+                )
         if self.groupsize != -1 and self.groupsize <= 0:
             raise ValueError(f"groupsize must be -1 or positive. got {self.groupsize}")
         if not (0.0 < self.percdamp <= 1.0):