Add a way to do power of 2 scaling

stack-info: PR: #2256, branch: drisspg/stack/57

Author: drisspg

test/dtypes/test_affine_quantized_float.py

@@ -44,6 +44,7 @@
     choose_qparams_affine,
 )
 from torchao.utils import (
+    TORCH_VERSION_AT_LEAST_2_8,
     is_sm_at_least_89,
     is_sm_at_least_90,
 )
@@ -345,6 +346,63 @@ def test_mm_float8dq(self, in_features, out_features, leading_shape, bias: bool)
         error = compute_error(ref_output, quant_output)
         assert error > 20, f"Quantization error is too high got a SQNR of {error}"
 
+    @unittest.skipIf(not torch.cuda.is_available(), "Need CUDA available")
+    @unittest.skipIf(
+        not is_sm_at_least_89(), "Requires GPU with compute capability >= 8.9"
+    )
+    @common_utils.parametrize(
+        "granularity", [PerTensor(), PerRow()] if is_sm_at_least_90() else [PerTensor()]
+    )
+    @unittest.skipIf(
+        not TORCH_VERSION_AT_LEAST_2_8, "Requires PyTorch 2.8+ with e8m0 support"
+    )
+    def test_fp8_e8m0_scale_dtype(self, granularity):
+        """Test float8 quantization with e8m0 scale dtype on PyTorch 2.8+"""
+        device = "cuda"
+        dtype = torch.bfloat16
+        in_features, out_features = 256, 512
+
+        # Create model
+        model = ToyLinearModel(in_features, out_features).to(device).to(dtype)
+        quant_model = copy.deepcopy(model)
+
+        # Create config with e8m0 scale dtype
+        config = Float8DynamicActivationFloat8WeightConfig(
+            granularity=granularity, scale_dtype=torch.float8_e8m0fnu
+        )
+
+        # Quantize the model
+        quantize_(quant_model, config)
+
+        # Verify that the scale dtype is correctly set
+        for layer_name in ["linear1", "linear2"]:
+            layer = getattr(quant_model, layer_name)
+            weight_impl = layer.weight.original_weight_tensor.tensor_impl
+
+            # All though we specify w/ e8m0 we still cast to fp32
+            self.assertEqual(weight_impl.scale.dtype, torch.float32)
+
+            # Verify scale is power of 2 (requirement for e8m0)
+            scale_values = weight_impl.scale.float()
+            log2_scales = torch.log2(scale_values)
+            self.assertTrue(
+                torch.allclose(log2_scales, torch.round(log2_scales), atol=0),
+                "e8m0 scales should be powers of 2",
+            )
+
+        # Test forward pass
+        input_tensor = torch.randn(32, in_features, device=device, dtype=dtype)
+
+        with torch.no_grad():
+            output = model(input_tensor)
+            output_quant = quant_model(input_tensor)
+
+        # Verify output shape and that computation completes without error
+        expected_shape = (32, in_features)  # ToyLinearModel returns to original size
+        self.assertEqual(output.shape, expected_shape)
+        error = compute_error(output, output_quant)
+        assert error > 20, f"Quantization error is too high got a SQNR of {error}"
+
 
 common_utils.instantiate_parametrized_tests(TestAffineQuantizedFloat8Compile)
 

torchao/dtypes/affine_quantized_tensor.py

@@ -463,7 +463,10 @@ def from_hp_to_floatx(
             original_shape = input_float.shape
             input_float = _layout.pre_process(input_float)
             scale = choose_qparams_affine_float8(
-                input_float, float8_dtype=target_dtype, block_size=block_size
+                input_float,
+                float8_dtype=target_dtype,
+                block_size=block_size,
+                scale_dtype=scale_dtype,
             )
             data = quantize_affine_float8(input_float, scale, target_dtype)
             data, scale, zero_point = _layout.post_process(

torchao/quantization/quant_api.py

@@ -1407,7 +1407,7 @@ def _float8_weight_only_quant_tensor(weight, config):
         input_float=weight,
         block_size=block_size,
         target_dtype=config.weight_dtype,
-        scale_dtype=None,
+        scale_dtype=torch.float32,
         _layout=Float8Layout(mm_config=None),
     )
     return new_weight
@@ -1561,6 +1561,7 @@ class Float8DynamicActivationFloat8WeightConfig(AOBaseConfig):
             only PerTensor and PerRow are supported.
         mm_config (Float8MMConfig): Configuration for the matrix multiplication. Default uses fast accumulation.
         set_inductor_config (bool): if True, adjusts `torchinductor` settings to recommended values.
+        scale_dtype: By default we set to fp32, if a user is on 12.8 and sets it to e8m0 we well ensure power of 2 scaling
 
     """
 
@@ -1571,6 +1572,7 @@ class Float8DynamicActivationFloat8WeightConfig(AOBaseConfig):
     ] = None
     mm_config: Optional[Float8MMConfig] = None
     set_inductor_config: bool = True
+    scale_dtype: torch.dtype = torch.float32
 
     def __post_init__(self):
         if self.mm_config is None:
@@ -1591,6 +1593,7 @@ def _float8_dynamic_activation_float8_weight_quantize_tensor(weight, config):
     weight_dtype = config.weight_dtype
     granularity = config.granularity
     mm_config = config.mm_config
+    scale_dtype = config.scale_dtype
 
     activation_granularity, weight_granularity = _normalize_granularity(granularity)
 
@@ -1610,7 +1613,7 @@ def _float8_dynamic_activation_float8_weight_quantize_tensor(weight, config):
         input_float=weight,
         block_size=block_size,
         target_dtype=weight_dtype,
-        scale_dtype=torch.float32,
+        scale_dtype=scale_dtype,
         _layout=Float8Layout(mm_config=mm_config),
     )
 

torchao/quantization/quant_primitives.py

@@ -2002,7 +2002,12 @@ def choose_qparams_affine_float8(
         ]
         scale = scale.reshape(output_shape)
 
-    return scale.to(dtype=scale_dtype)
+    if scale_dtype is not torch.float32:
+        # Shielding for Version > 2.8
+        assert scale_dtype is torch.float8_e8m0fnu, "Only float8_e8m0fnuz is supported"
+        scale = torch.exp2(torch.round(torch.log2(scale)))
+
+    return scale.to(dtype=torch.float32)
 
 
 def quantize_affine_float8(