Add all fbgemm kernel Tensors into Int4WeightOnlyConfig and Float8DynamicActivationInt4WeightConfig

Summary:
we will
* deprecate FbgemmConfig since it's a single kernel (later).
* we'd like to categorize things to derived dtype + packed format, e.g. int4 preshuffled, float8 plain
* Added PackingFormat that has preshuffled, plain and _legacy for legacy implementation

Test Plan:
python test/quantization/quantize_/workflows/int4/test_int4_tensor.py
python test/quantization/quantize_/workflows/int4/test_int4_preshuffled_tensor.py
python test/quantization/quantize_/workflows/float8/test_float8_tensor.py

Reviewers:

Subscribers:

Tasks:

Tags:

stack-info: PR: #2474, branch: jerryzh168/stack/10

Author: jerryzh168

test/integration/test_serialization_bc.py

@@ -17,13 +17,13 @@
 from torchao.utils import _is_fbgemm_genai_gpu_available, is_sm_at_least_90
 
 _MODEL_NAMES = [
-    "torchao-testing/opt-125m-float8dq-row-fbgemm",
+    "torchao-testing/opt-125m-float8dq-row-0.13-dev",
+    "torchao-testing/opt-125m-int4wo-preshuffled-0.13-dev",
 ]
 
 
 @unittest.skipIf(not torch.cuda.is_available(), "Need CUDA available")
 @unittest.skipIf(not is_sm_at_least_90(), "Nedd sm90+")
-@unittest.skip("temporary skip since we have some refactor next")
 class TestSerializationBC(TestCase):
     """Test we can still load and run serialized model in previous AO versions
     we commit to have BC for 3 pytorch releases

test/quantization/quantize_/workflows/int4/test_int4_preshuffled_tensor.py

@@ -15,9 +15,9 @@
     run_tests,
 )
 
-from torchao.float8.config import e4m3_dtype
 from torchao.quantization import (
-    FbgemmConfig,
+    Float8ActivationInt4WeightConfig,
+    Int4WeightOnlyConfig,
     quantize_,
 )
 from torchao.quantization.utils import compute_error
@@ -27,44 +27,15 @@
     is_sm_at_least_90,
 )
 
-if TORCH_VERSION_AT_LEAST_2_8:
-    BF16_ACT_CONFIG = FbgemmConfig(
-        input_dtype=torch.bfloat16,
-        weight_dtype=torch.int4,
-        output_dtype=torch.bfloat16,
-        block_size=[1, 128],
-        preshuffle=True,
-    )
-
-    BF16_ACT_BMM_CONFIG = FbgemmConfig(
-        input_dtype=torch.bfloat16,
-        weight_dtype=torch.int4,
-        output_dtype=torch.bfloat16,
-        block_size=[1, 1, 128],
-        preshuffle=True,
-    )
-
-    FP8_ACT_CONFIG = FbgemmConfig(
-        input_dtype=e4m3_dtype,
-        weight_dtype=torch.int4,
-        output_dtype=torch.bfloat16,
-        block_size=[1, 128],
-        preshuffle=True,
-    )
-
-    FP8_ACT_BMM_CONFIG = FbgemmConfig(
-        input_dtype=e4m3_dtype,
-        weight_dtype=torch.int4,
-        output_dtype=torch.bfloat16,
-        block_size=[1, 1, 128],
-        preshuffle=True,
-    )
-
-else:
-    BF16_ACT_CONFIG = None
-    BF16_ACT_BMM_CONFIG = None
-    FP8_ACT_CONFIG = None
-    FP8_ACT_BMM_CONFIG = None
+BF16_ACT_CONFIG = Int4WeightOnlyConfig(
+    group_size=128,
+    packing_format="preshuffled",
+)
+
+FP8_ACT_CONFIG = Float8ActivationInt4WeightConfig(
+    group_size=128,
+    packing_format="preshuffled",
+)
 
 
 @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_8, "Need pytorch 2.8+")
@@ -90,7 +61,7 @@ def test_linear(self, config):
 
     # Note: this order will error out: `Got bad cuda status: an illegal memory access was encountered at line: 449`
     # @parametrize("bmm_config", [BF16_ACT_BMM_CONFIG, FP8_ACT_BMM_CONFIG])
-    @parametrize("bmm_config", [FP8_ACT_BMM_CONFIG, BF16_ACT_BMM_CONFIG])
+    @parametrize("bmm_config", [FP8_ACT_CONFIG, BF16_ACT_CONFIG])
     def test_bmm(self, bmm_config):
         class M(torch.nn.Module):
             def __init__(self, weight):

test/dtypes/test_fbgemm_int4.py renamed to test/quantization/quantize_/workflows/int4/test_int4_tensor.py

@@ -13,7 +13,7 @@
 )
 
 from torchao.quantization import (
-    FbgemmConfig,
+    Int4WeightOnlyConfig,
     quantize_,
 )
 from torchao.quantization.utils import compute_error
@@ -26,19 +26,11 @@
 @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_8, "Need pytorch 2.8+")
 @unittest.skipIf(not torch.cuda.is_available(), "Need CUDA available")
 @unittest.skipIf(not is_sm_at_least_90(), "Nedd sm90+")
-class TestFbgemmInt4Tensor(TestCase):
+class TestInt4Tensor(TestCase):
     def setUp(self):
-        self.config = FbgemmConfig(
-            input_dtype=torch.bfloat16,
-            weight_dtype=torch.int4,
-            output_dtype=torch.bfloat16,
-            block_size=[1, 128],
-        )
-        self.bmm_config = FbgemmConfig(
-            input_dtype=torch.bfloat16,
-            weight_dtype=torch.int4,
-            output_dtype=torch.bfloat16,
-            block_size=[1, 1, 128],
+        self.config = Int4WeightOnlyConfig(
+            group_size=128,
+            packing_format="plain",
         )
         self.GPU_DEVICES = ["cuda"] if torch.cuda.is_available() else []
 
@@ -68,13 +60,9 @@ def test_slice(self):
         quantize_(dummy, self.config)
         weight1 = dummy.weight.narrow(0, 0, 64)
         weight2 = dummy.weight.narrow(1, 0, 128)
-        self.assertEqual(
-            weight1.packed_weight, dummy.weight.packed_weight.narrow(0, 0, 64)
-        )
+        self.assertEqual(weight1._data, dummy.weight._data.narrow(0, 0, 64))
         self.assertEqual(weight1.scale, dummy.weight.scale.narrow(1, 0, 64))
-        self.assertEqual(
-            weight2.packed_weight, dummy.weight.packed_weight.narrow(1, 0, 64)
-        )
+        self.assertEqual(weight2._data, dummy.weight._data.narrow(1, 0, 64))
         self.assertEqual(weight2.scale, dummy.weight.scale.narrow(0, 0, 1))
 
         # check for sliced weight, before and after float8 quantization
@@ -100,12 +88,10 @@ def test_slice_and_copy_(self):
         param = l.weight
         param_data = param.data
         param_data = param_data.narrow(0, 0, 512)
-        assert (
-            param.data.packed_weight.data_ptr() == param_data.packed_weight.data_ptr()
-        )
+        assert param.data._data.data_ptr() == param_data._data.data_ptr()
         assert param.data.scale.data_ptr() == param_data.scale.data_ptr()
         assert param.data.zero_point.data_ptr() == param_data.zero_point.data_ptr()
-        orig_value = param.data.packed_weight[0][0].item()
+        orig_value = param.data._data[0][0].item()
 
         # dummy_l has random input (shouldn't be 0)
         dummy_l = torch.nn.Linear(1024, 1024).to("cuda").to(torch.bfloat16)
@@ -116,7 +102,7 @@ def test_slice_and_copy_(self):
         param_data.copy_(quantized)
 
         # making sure param.data is updated
-        assert param.data.packed_weight[0][0] != orig_value
+        assert param.data._data[0][0] != orig_value
 
     def test_bmm(self):
         class M(torch.nn.Module):
@@ -135,7 +121,7 @@ def forward(self, x):
         original = m(input)
         # we need to transpose the weight first for bmm
         m.weight = torch.nn.Parameter(m.weight.transpose(1, 2).contiguous())
-        quantize_(m, self.bmm_config, filter_fn=lambda x, fqn: True)
+        quantize_(m, self.config, filter_fn=lambda x, fqn: True)
         quantized = m(input)
         self.assertTrue(compute_error(original, quantized) > 18)
 

torchao/dtypes/__init__.py

@@ -9,7 +9,6 @@
     to_affine_quantized_intx_static,
 )
 from .fbgemm_fp8_tensor import FbgemmFp8Tensor, to_fbgemm_fp8
-from .fbgemm_int4_tensor import FbgemmInt4Tensor, to_fbgemm_int4
 from .floatx import (
     CutlassSemiSparseLayout,
     Float8Layout,
@@ -64,8 +63,6 @@
     "PackedLinearInt8DynamicActivationIntxWeightLayout",
     "to_affine_quantized_packed_linear_int8_dynamic_activation_intx_weight",
     "Int4XPULayout",
-    "to_fbgemm_int4",
-    "FbgemmInt4Tensor",
     "to_fbgemm_fp8",
     "FbgemmFp8Tensor",
     "Int8DynamicActInt4WeightCPULayout",

torchao/quantization/__init__.py

@@ -44,6 +44,7 @@
 from .quant_api import (
     CutlassInt4PackedLayout,
     FbgemmConfig,
+    Float8ActivationInt4WeightConfig,
     Float8DynamicActivationFloat8SemiSparseWeightConfig,
     Float8DynamicActivationFloat8WeightConfig,
     Float8MMConfig,
@@ -90,6 +91,7 @@
 from .quantize_.workflows import (
     Float8Tensor,
     Int4PreshuffledTensor,
+    Int4Tensor,
 )
 from .smoothquant import (
     SmoothFakeDynamicallyQuantizedLinear,
@@ -141,6 +143,7 @@
     "Int8DynamicActivationInt8WeightConfig",
     "Int8DynamicActivationIntxWeightConfig",
     "Int4WeightOnlyConfig",
+    "Float8ActivationInt4WeightConfig",
     "Int8WeightOnlyConfig",
     "Float8WeightOnlyConfig",
     "Float8DynamicActivationFloat8WeightConfig",
@@ -154,6 +157,7 @@
     "ModuleFqnToConfig",
     "FbgemmConfig",
     # tensor subclasses
+    "Int4Tensor",
     "Int4PreshuffledTensor",
     "Float8Tensor",
     # smooth quant - subject to change

torchao/quantization/quant_api.py

@@ -49,7 +49,6 @@
     to_affine_quantized_floatx_static,
     to_affine_quantized_intx,
     to_fbgemm_fp8,
-    to_fbgemm_int4,
     to_marlinqqq_quantized_intx,
 )
 from torchao.dtypes.uintx.packed_linear_int8_dynamic_activation_intx_weight_layout import (
@@ -79,6 +78,7 @@
 from torchao.quantization.quantize_.workflows import (
     Float8Tensor,
     Int4PreshuffledTensor,
+    Int4Tensor,
 )
 from torchao.quantization.transform_module import (
     _QUANTIZE_CONFIG_HANDLER,
@@ -1115,6 +1115,8 @@ class Int4WeightOnlyConfig(AOBaseConfig):
         `zero_point_domain`: data type of zeros points, choices are [ZeroPointDomain.FLOAT, ZeroPointDomain.INT, ZeroPointDomain.NONE]
         `set_inductor_config`: if True, adjusts `torchinductor` settings to recommended values.
         `preserve_zero`: whether to preserve zero, default is None. Will be set to True if zero_point_domain is ZeroPointDomain.INT
+        `packing_format`: the packing format for int4 tensor, this is the new structure we plan to use
+           existing layout based structure will use `_legacy`
     """
 
     group_size: int = 128
@@ -1123,6 +1125,9 @@ class Int4WeightOnlyConfig(AOBaseConfig):
     zero_point_domain: Optional[ZeroPointDomain] = ZeroPointDomain.NONE
     set_inductor_config: bool = True
     preserve_zero: Optional[bool] = None
+    # since not all tensors are migrated to the new structure yet,
+    # we use `_legacy' to represent the previous layout
+    packing_format: PackingFormat = "_legacy"
 
 
 # for BC
@@ -1140,15 +1145,33 @@ def _int4_weight_only_quantize_tensor(weight, config):
     layout = config.layout
     use_hqq = config.use_hqq
     zero_point_domain = config.zero_point_domain
+    packing_format = config.packing_format
 
     if weight.shape[-1] % group_size != 0:
         logger.info(
             f"Skipping quantizing weight with int4 weight only quantization because the shape of weight {weight.shape} is not compatible with group_size {group_size}"
         )
         return weight
 
+    block_size = tuple([1 for _ in range(weight.ndim - 1)] + [group_size])
+
+    if packing_format == "preshuffled":
+        new_weight = Int4PreshuffledTensor.from_float(
+            weight,
+            block_size,
+            activation_dtype=torch.bfloat16,
+        )
+        return new_weight
+    elif packing_format == "plain":
+        new_weight = Int4Tensor.from_float(
+            weight,
+            block_size,
+        )
+        return new_weight
+
+    assert packing_format == "_legacy", f"Unsupported packing_format: {packing_format}"
+
     mapping_type = MappingType.ASYMMETRIC
-    block_size = tuple([1 for _ in range(weight.dim() - 1)] + [group_size])
     target_dtype = torch.int32
     quant_min = 0
     quant_max = 15
@@ -1220,6 +1243,38 @@ def _int4_weight_only_transform(
     return module
 
 
+@dataclass
+class Float8ActivationInt4WeightConfig(AOBaseConfig):
+    group_size: int = 128
+    packing_format: PackingFormat = "preshuffled"
+
+
+@register_quantize_module_handler(Float8ActivationInt4WeightConfig)
+def _float8_activation_int4_weight_transform(
+    module: torch.nn.Module, config: Float8ActivationInt4WeightConfig
+) -> torch.nn.Module:
+    assert hasattr(module, "weight"), (
+        "applying int8 weight only quant requires module to have weight attribute"
+        + " but {module} does not have one"
+    )
+    group_size = config.group_size
+    packing_format = config.packing_format
+
+    assert packing_format == "preshuffled", (
+        f"only preshuffled packing_format supported right now, got: {packing_format}"
+    )
+    weight = module.weight
+    block_size = tuple([1 for _ in range(weight.ndim - 1)] + [group_size])
+    new_weight = Int4PreshuffledTensor.from_float(
+        module.weight,
+        block_size,
+        activation_dtype=torch.float8_e4m3fn,
+    )
+    module.weight = torch.nn.Parameter(new_weight, requires_grad=False)
+    module.extra_repr = types.MethodType(_linear_extra_repr, module)
+    return module
+
+
 @dataclass
 class Int8WeightOnlyConfig(AOBaseConfig):
     """
@@ -1685,6 +1740,7 @@ def _float8_dynamic_activation_float8_weight_quantize_tensor(weight, config):
         # TODO(future PR): this should really throw an exception instead of silently
         # not doing what the user asked
         return weight
+
     if isinstance(weight_granularity, PerRow):
         assert weight.dtype == torch.bfloat16, (
             "PerRow quantization only works for bfloat16 precision input weight"
@@ -2125,7 +2181,7 @@ def _(module: torch.nn.Module, config: FbgemmConfig) -> torch.nn.Module:
                 activation_dtype=torch.bfloat16,
             )
         else:
-            weight = to_fbgemm_int4(
+            weight = Int4Tensor.from_float(
                 module.weight,
                 config.block_size,
             )

torchao/quantization/quantize_/workflows/__init__.py

@@ -4,8 +4,16 @@
 from .int4.int4_preshuffled_tensor import (
     Int4PreshuffledTensor,
 )
+from .int4.int4_tensor import (
+    Int4Tensor,
+)
+from .packing_format import (
+    PackingFormat,
+)
 
 __all__ = [
+    "Int4Tensor",
     "Int4PreshuffledTensor",
     "Float8Tensor",
+    "PackingFormat",
 ]

torchao/quantization/quantize_/workflows/int4/__init__.py

@@ -0,0 +1,7 @@
+from .int4_preshuffled_tensor import Int4PreshuffledTensor
+from .int4_tensor import Int4Tensor
+
+__all__ = [
+    "Int4PreshuffledTensor",
+    "Int4Tensor",
+]

torchao/quantization/quantize_/workflows/int4/int4_preshuffled_tensor.py

@@ -194,7 +194,7 @@ def from_float(
         if quantize_int4_preshuffle is None:
             raise ImportError("Requires fbgemm-gpu-genai >= 1.2.0")
 
-        assert all(x == 1 for x in block_size[:-1]), (
+        assert all(x == 1 for x in block_size[:-1]) and block_size[-1] != 1, (
             "Only groupwise quant is supported right now"
         )
         group_size = block_size[-1]