[Observers] group size + channel wise + per token

src/compressed_tensors/quantization/lifecycle/forward.py

@@ -13,9 +13,13 @@
 # limitations under the License.
 
 from functools import wraps
+from math import ceil
 
 import torch
-from compressed_tensors.quantization.quant_args import QuantizationArgs
+from compressed_tensors.quantization.quant_args import (
+    QuantizationArgs,
+    QuantizationStrategy,
+)
 from compressed_tensors.quantization.quant_config import QuantizationStatus
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
 from torch.nn import Module
@@ -32,10 +36,9 @@ def quantize(
     q_min: torch.Tensor,
     q_max: torch.Tensor,
 ) -> torch.Tensor:
+
     return torch.clamp(
-        torch.round(
-            x / scale + zero_point,
-        ),
+        torch.round(x / scale + zero_point),
         q_min,
         q_max,
     )
@@ -57,12 +60,81 @@ def fake_quantize(
     zero_point: torch.Tensor,
     args: QuantizationArgs,
 ) -> torch.Tensor:
+    """
+    Fake quantize the input tensor x depending on the group_size.
+    if group_size is greater than 0, then q/dq by groups. The groups
+    must be divisible by the column size
+    if group_size is -1, then channel wise q/dq. THe input scale and
+    zero_points are reshaped to support vectorization (Assumes 1 is
+    the channel dimension)
+
+    :param x: Input tensor
+    :param scale: scale tensor
+    :param zero_point: zero point tensor
+    :param args: quantization args that contain group_size info
+    :return: fake quantized tensor
+
+    """
     bit_range = 2**args.num_bits
     max_q = torch.tensor(bit_range / 2 - 1, device=x.device)
     min_q = torch.tensor(-bit_range / 2, device=x.device)
-    Q = torch.zeros_like(x)
-    Q = quantize(x, scale, zero_point, min_q, max_q)
-    return dequantize(Q, scale, zero_point)
+
+    group_size = args.group_size
+
+    # group
+    if args.strategy == QuantizationStrategy.GROUP:
+
+        DQ = torch.zeros_like(x)
+
+        # TODO: vectorize the for loop
+        # TODO: fix genetric assumption about the tensor size for computing group
+
+        # TODO: make validation step for inputs
+
+        while scale.ndim < 2:
+            scale = scale.unsqueeze(1)
+            zero_point = zero_point.unsqueeze(1)
+
+        columns = x.shape[1]
+        if columns >= group_size:
+            assert columns % group_size == 0
+        for i in range(ceil(columns / group_size)):
+
+            # scale.shape should be [nchan, ndim]
+            # sc.shape should be [nchan, 1] after unsqueeze
+            sc = scale[:, i].unsqueeze(1)
+            zp = zero_point[:, i].unsqueeze(1)
+
+            idx = i * group_size
+            Q = quantize(x[:, idx : (idx + group_size)], sc, zp, min_q, max_q)
+            DQ[:, idx : (idx + group_size)] = dequantize(Q, sc, zp)
+
+    # channel-wise
+    elif args.strategy == QuantizationStrategy.CHANNEL:  # group_size == -1
+        # before: scale shape = [channel_size]
+        # after: scale shape = [1, channel_size]
+        scale = scale.unsqueeze(0)
+        zero_point = zero_point.unsqueeze(0)
+
+        Q = quantize(x, scale, zero_point, min_q, max_q)
+        DQ = dequantize(Q, scale, zero_point)
+
+    # per-token
+    elif args.strategy == QuantizationStrategy.TOKEN:
+        # before: scale shape = [channel_size]
+        # after: scale shape = [channel_size, 1]
+
+        scale = scale.unsqueeze(1)
+        zero_point = zero_point.unsqueeze(1)
+
+        Q = quantize(x, scale, zero_point, min_q, max_q)
+        DQ = dequantize(Q, scale, zero_point)
+
+    else:
+        Q = quantize(x, scale, zero_point, min_q, max_q)
+        DQ = dequantize(Q, scale, zero_point)
+
+    return DQ
 
 
 def wrap_module_forward_quantized(module: Module, scheme: QuantizationScheme):
@@ -138,5 +210,4 @@ def _maybe_calibrate_or_quantize(
             device = next(module.parameters()).device
             scale.data = updated_scale.to(device)
             zero_point.data = updated_zero_point.to(device)
-
     return fake_quantize(value, scale, zero_point, args)

src/compressed_tensors/quantization/observers/base.py

@@ -14,7 +14,11 @@
 
 from typing import Optional, Tuple
 
-from compressed_tensors.quantization.quant_args import QuantizationArgs
+import torch
+from compressed_tensors.quantization.quant_args import (
+    QuantizationArgs,
+    QuantizationStrategy,
+)
 from compressed_tensors.registry.registry import RegistryMixin
 from torch import FloatTensor, IntTensor, Tensor
 from torch.nn import Module
@@ -52,6 +56,12 @@ def calculate_qparams(self, observed: Tensor) -> Tuple[FloatTensor, IntTensor]:
         """
         raise NotImplementedError(f"{self.__class__} must implement calculate_qparams")
 
+    def post_calculate_qparams(self) -> None:
+        """
+        Run any logic specific to its observers after running calculate_qparams
+        """
+        ...
+
     def get_qparams(
         self, observed: Optional[Tensor] = None
     ) -> Tuple[FloatTensor, IntTensor]:
@@ -64,6 +74,57 @@ def get_qparams(
         :return: tuple of scale and zero point based on last observed value
         """
         if observed is not None:
-            # re-calcualte scale and zero point, update the stored value
-            self._scale, self._zero_point = self.calculate_qparams(observed)
+            group_size = self.quantization_args.group_size
+
+            if self.quantization_args.strategy == QuantizationStrategy.TENSOR:
+
+                # re-calculate scale and zero point, update the stored value
+                self._scale, self._zero_point = self.calculate_qparams(observed)
+
+            elif self.quantization_args.strategy == QuantizationStrategy.GROUP:
+                columns = observed.shape[1]
+                scales, zero_points = [], []
+                for i in range(0, columns, self.quantization_args.group_size):
+                    scale, zero_point = self.get_qparams_along_dim(
+                        observed[:, i : (i + group_size)],
+                        0,
+                    )
+                    scales.append(scale)
+                    zero_points.append(zero_point)
+
+                self._scale = torch.stack(scales, dim=1)
+                self._zero_point = torch.stack(zero_points, dim=1)
+
+            elif self.quantization_args.strategy == QuantizationStrategy.CHANNEL:
+                # assume observed is transposed, because its the output, hence use dim 0
+                self._scale, self._zero_point = self.get_qparams_along_dim(observed, 0)
+
+            elif self.quantization_args.strategy == QuantizationStrategy.TOKEN:
+
+                # use dim 1, assume the obsersed.shape = [batch, token, hidden]
+                # should be batch, token
+
+                self._scale, self._zero_point = self.get_qparams_along_dim(
+                    observed, dim=1
+                )
+
         return self._scale, self._zero_point
+
+    def get_qparams_along_dim(self, observed, dim: int):
+        # TODO: add documentation that specifies the shape must
+        #   be padded with 1-dims so the scales are along the right channel
+        # TODO: generalize the logic for reduce_dims
+        scales, zero_points = [], []
+
+        # TODO: make a more generic way to get the channel
+        num_dims = observed.shape[dim]
+
+        for dim_idx in range(num_dims):
+            scale, zero_point = self.calculate_qparams(
+                observed.select(dim=dim, index=dim_idx)
+            )
+
+            scales.append(scale)
+            zero_points.append(zero_point)
+        # breakpoint()
+        return torch.stack(scales), torch.stack(zero_points)

src/compressed_tensors/quantization/quant_args.py

@@ -15,7 +15,7 @@
 from enum import Enum
 from typing import Any, Dict, Optional
 
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, validator
 
 
 __all__ = ["QuantizationType", "QuantizationStrategy", "QuantizationArgs"]
@@ -39,6 +39,7 @@ class QuantizationStrategy(str, Enum):
     CHANNEL = "channel"
     GROUP = "group"
     BLOCK = "block"
+    TOKEN = "token"
 
 
 class QuantizationArgs(BaseModel):
@@ -63,7 +64,7 @@ class QuantizationArgs(BaseModel):
     num_bits: int = 8
     type: QuantizationType = QuantizationType.INT
     symmetric: bool = True
-    strategy: QuantizationStrategy = QuantizationStrategy.TENSOR
+    strategy: Optional[QuantizationStrategy] = None
     group_size: Optional[int] = None
     block_structure: Optional[str] = None
     dynamic: bool = False
@@ -94,3 +95,32 @@ def get_observer(self):
             self.observer = "memoryless"
 
         return Observer.load_from_registry(self.observer, quantization_args=self)
+
+    @validator("strategy", pre=True)
+    def validate_strategy(cls, value, values):
+        group_size = values.get("group_size")
+
+        # use group_size to determinine strategy if not given explicity
+        if group_size is not None and value is None:
+            if group_size > 0:
+                return QuantizationStrategy.GROUP
+
+            elif group_size == -1:
+                return QuantizationStrategy.CHANNEL
+
+            else:
+                raise ValueError(
+                    f"group_size={group_size} with strategy {value} is invald. "
+                    "group_size > 0 for strategy='group' and "
+                    "group_size = -1 for 'channel'"
+                )
+        # breakpoint()
+        group_size = 128
+        if value == QuantizationStrategy.GROUP:
+            if group_size is None:
+                raise ValueError(f"strategy {value} requires group_size to be set.")
+
+        if value is None:
+            return QuantizationStrategy.TENSOR
+
+        return value

src/compressed_tensors/quantization/utils/helpers.py

@@ -108,6 +108,7 @@ def calculate_compression_ratio(model: Module) -> float:
             compressed_bits = uncompressed_bits
             if is_module_quantized(submodule):
                 compressed_bits = submodule.quantization_scheme.weights.num_bits
+
             num_weights = parameter.numel()
             total_compressed += compressed_bits * num_weights
             total_uncompressed += uncompressed_bits * num_weights