feat(quanto): introduce qtype

This implies a lot of modifications but is functionally equivalent.

Author: dacorvo

bench/generation/benchmark.py

@@ -7,7 +7,7 @@
 from tqdm.auto import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, GenerationConfig
 
-from quanto import Calibration, freeze, quantize
+from quanto import Calibration, freeze, qint8, quantize
 
 
 CALIBRATION_PROMPT = "It was a bright cold day in April, and the clocks were striking thirteen."
@@ -165,8 +165,8 @@ def main():
         if args.quantization in ("w8a8", "w8a16"):
             print("quantizing")
             start = time.time()
-            weights = torch.int8
-            activations = None if "a16" in args.quantization else torch.int8
+            weights = qint8
+            activations = None if "a16" in args.quantization else qint8
             quantize(model, weights=weights, activations=activations)
             if activations is not None:
                 print("Calibrating")

examples/nlp/text-classification/sst2/quantize_sst2_model.py

@@ -8,7 +8,7 @@
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
 from transformers.pipelines.pt_utils import KeyDataset
 
-from quanto import Calibration, freeze, quantize
+from quanto import Calibration, freeze, qint8, quantize
 
 
 def evaluate_model(model, tokenizer, dataset, device, batch_size):
@@ -22,7 +22,7 @@ def evaluate_model(model, tokenizer, dataset, device, batch_size):
 
 
 def keyword_to_itype(k):
-    return {"none": None, "int8": torch.int8}[k]
+    return {"none": None, "int8": qint8}[k]
 
 
 def main():

examples/nlp/text-generation/quantize_causal_lm_model.py

@@ -5,7 +5,7 @@
 from datasets import load_dataset
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-from quanto import Calibration, freeze, quantize
+from quanto import Calibration, freeze, qfloat8_e4m3fn, qfloat8_e5m2, qint8, quantize
 
 
 @torch.no_grad()
@@ -51,7 +51,7 @@ def evaluate_model(model, tokenizer, dataset, device, batch_size, samples=None,
 
 
 def keyword_to_itype(k):
-    return {"none": None, "int8": torch.int8, "fp8_e5m2": torch.float8_e5m2, "fp8_e4m3": torch.float8_e4m3fn}[k]
+    return {"none": None, "int8": qint8, "fp8_e5m2": qfloat8_e5m2, "fp8_e4m3": qfloat8_e4m3fn}[k]
 
 
 def main():

examples/vision/image-classification/mnist/quantize_mnist_model.py

@@ -7,7 +7,7 @@
 from torchvision import datasets, transforms
 from transformers import AutoModel
 
-from quanto import Calibration, QTensor, freeze, int4, quantize
+from quanto import Calibration, QTensor, freeze, qint4, qint8, quantize
 
 
 def test(model, device, test_loader):
@@ -60,7 +60,7 @@ def train(log_interval, model, device, train_loader, optimizer, epoch):
 
 
 def keyword_to_itype(k):
-    return {"none": None, "int4": int4, "int8": torch.int8}[k]
+    return {"none": None, "int4": qint4, "int8": qint8}[k]
 
 
 def main():

quanto/nn/qlinear.py

@@ -2,7 +2,7 @@
 
 import torch
 
-from ..tensor import QBitsTensor, QTensor, absmax_scale, qbitsdtype
+from ..tensor import QBitsTensor, QTensor, absmax_scale, qint2, qint4, qint8, qtype
 from .qmodule import QModuleMixin, register_qmodule
 
 
@@ -11,12 +11,12 @@
 
 @register_qmodule(torch.nn.Linear)
 class QLinear(QModuleMixin, torch.nn.Linear):
-    def __init__(self, *args, weights: torch.dtype = torch.int8, **kwargs):
+    def __init__(self, *args, weights: qtype = qint8, **kwargs):
         super().__init__(*args, **kwargs)
         self.weights = weights
 
     @classmethod
-    def from_module(cls, module, weights=torch.int8, activations: Optional[torch.dtype] = None):
+    def from_module(cls, module, weights=qint8, activations: Optional[qtype] = None):
         qmodule = cls(
             module.in_features,
             module.out_features,
@@ -36,12 +36,12 @@ def qweight(self):
         if isinstance(self.weight, QTensor):
             return self.weight
         # Quantize the weights per-axis
-        if isinstance(self.weights, torch.dtype):
+        if self.weights == qint8:
             wscale = absmax_scale(self.weight, axis=0)
-            return QTensor.quantize(self.weight, itype=self.weights, scale=wscale)
-        elif isinstance(self.weights, qbitsdtype):
-            return QBitsTensor.quantize(self.weight, itype=self.weights, axis=0)
-        raise ValueError("Invalid quantized weights type")
+            return QTensor.quantize(self.weight, qtype=self.weights, scale=wscale)
+        elif self.weights in (qint2, qint4):
+            return QBitsTensor.quantize(self.weight, qtype=self.weights, axis=0)
+        raise ValueError(f"Invalid quantized weights type {self.weights}")
 
     def qforward(self, input: torch.Tensor) -> torch.Tensor:
         if self.activations is not None and not isinstance(input, QTensor):

quanto/nn/qmodule.py

@@ -94,9 +94,9 @@ def qforward(self, input: torch.Tensor) -> torch.Tensor:
 
     def forward(self, input: torch.Tensor) -> torch.Tensor:
         def maybe_requantize(t, scale):
-            if t.itype == self.activations and t.axis is None:
+            if t.qtype == self.activations and t.axis is None:
                 return t
-            return QTensor.quantize(t.dequantize(), itype=self.activations, scale=scale)
+            return QTensor.quantize(t.dequantize(), qtype=self.activations, scale=scale)
 
         if self.activations is not None and isinstance(input, QTensor):
             input = maybe_requantize(input, self.input_scale)

quanto/tensor/__init__.py

@@ -1 +1,2 @@
 from .core import *
+from .qtype import *