Add support for int4 weight-only QAT (pytorch#383)

* Add cachemask variant for fake_quantize_affine

Summary: In QAT, we often wish to filter out the gradients
corresponding to values outside the expected quantization
range, for example:

```
q = _quantize_affine_no_dtype_cast(...)
dq = _dequantize_affine_no_dtype_check(...)
mask = torch.logical_and((q >= quant_min), (q <= quant_max))

grad = grad * mask
```

The existing `fake_quantize_affine` returns the dequantized
values only, so callers do not have access to this mask.
This commit adds the variant to this op that returns both
the dequantized values and the mask, similar to
`fake_quantize_per_tensor_affine_cachemask` in core.

Test Plan:
python test/quantization/test_quant_primitives.py -k test_fake_quantize_affine_cachemask

* Add support for int4 weight-only QAT

Summary: This commit adds support for int4 weight-only QAT,
which simulates the numerics of the existing
Int4WeightOnlyQuantizer. The main motivation for this is to
provide an end-to-end path for running QAT and lowering to
the efficient int4 tinygemm cuda kernel. To enable this,
we have to add new fake quantization primitives to match
the numerics of the tinygemm kernel, and this required
refactoring existing quant primitives to skip dtype casting.

Test Plan:
python test/quantization/test_qat.py -k test_qat_4w_linear

Reviewers: jerryzh168, msaroufim

Subscribers: jerryzh168, msaroufim, HDCharles, supriyar

test/quantization/test_qat.py

@@ -18,31 +18,41 @@
     fake_quantize_per_channel_group,
     fake_quantize_per_token,
 )
-from torchao.quantization.utils import get_group_qparams_symmetric
+from torchao.quantization.quant_primitives import (
+    fake_quantize_affine,
+    ZeroPointDomain,
+)
+from torchao.quantization.utils import (
+    get_group_qparams_symmetric,
+    get_groupwise_affine_qparams,
+    groupwise_affine_quantize_tensor,
+)
 from torchao.utils import TORCH_VERSION_AFTER_2_4
 
 
 # TODO: put this in a common test utils file
+_CUDA_IS_AVAILABLE = torch.cuda.is_available()
+
 class Sub(torch.nn.Module):
     def __init__(self):
         super().__init__()
-        self.linear = torch.nn.Linear(32, 32, bias=False).to(torch.float)
+        self.linear = torch.nn.Linear(256, 256, bias=False).to(torch.float)
 
     def example_inputs(self):
-        return (torch.randn(1, 32).to(torch.float),)
+        return (torch.randn(1, 256).to(torch.float),)
 
     def forward(self, x):
         return self.linear(x)
 
 class M(torch.nn.Module):
     def __init__(self):
         super().__init__()
-        self.linear1 = torch.nn.Linear(64, 32, bias=False).to(torch.float)
+        self.linear1 = torch.nn.Linear(512, 256, bias=False).to(torch.float)
         self.sub = Sub()
-        self.linear2 = torch.nn.Linear(32, 64, bias=False).to(torch.float)
+        self.linear2 = torch.nn.Linear(256, 512, bias=False).to(torch.float)
 
     def example_inputs(self):
-        return (torch.randn(1, 64).to(torch.float),)
+        return (torch.randn(1, 512).to(torch.float),)
 
     def forward(self, x):
         x = self.linear1(x)
@@ -111,23 +121,46 @@ def test_fake_quantize_per_token(self):
 
     def _set_ptq_weight(
         self,
-        ptq_linear: "Int8DynActInt4WeightLinear",
-        fp32_weight: torch.Tensor,
-        group_size: int,
+        ptq_linear: torch.nn.Module,
+        qat_linear: torch.nn.Module,
     ):
         """
         Set the weight to the quantized version of the given fp32 weights,
         for making linear outputs comparable with QAT.
         """
+        from torchao.quantization.GPTQ import (
+            Int8DynActInt4WeightLinear,
+            WeightOnlyInt4Linear,
+        )
+        from torchao.quantization.prototype.qat import (
+            Int8DynActInt4WeightQATLinear,
+            Int4WeightOnlyQATLinear,
+        )
         n_bit = 4
         (qmin, qmax) = self._get_qmin_qmax(n_bit)
-        (s, zp) = get_group_qparams_symmetric(fp32_weight, n_bit, group_size)
-        q_weight = torch.ops.quantized_decomposed.quantize_per_channel_group(
-            fp32_weight, s, zp, qmin, qmax, torch.int8, group_size,
-        )
-        ptq_linear.weight = q_weight
-        ptq_linear.scales = s
-        ptq_linear.zeros = zp
+        if isinstance(ptq_linear, Int8DynActInt4WeightLinear):
+            assert isinstance(qat_linear, Int8DynActInt4WeightQATLinear)
+            fp32_weight = qat_linear.weight
+            group_size = qat_linear.groupsize
+            (s, zp) = get_group_qparams_symmetric(fp32_weight, n_bit, group_size)
+            q_weight = torch.ops.quantized_decomposed.quantize_per_channel_group(
+                fp32_weight, s, zp, qmin, qmax, torch.int8, group_size,
+            )
+            ptq_linear.weight = q_weight
+            ptq_linear.scales = s
+            ptq_linear.zeros = zp
+        elif isinstance(ptq_linear, WeightOnlyInt4Linear):
+            assert isinstance(qat_linear, Int4WeightOnlyQATLinear)
+            (q_weight, scales_and_zeros) = groupwise_affine_quantize_tensor(
+                qat_linear.weight, n_bit, qat_linear.groupsize,
+            )
+            q_weight = torch.ops.aten._convert_weight_to_int4pack(
+                q_weight.to("cuda"), qat_linear.inner_k_tiles,
+            )
+            ptq_linear.weight = q_weight
+            ptq_linear.scales_and_zeros = scales_and_zeros
+        else:
+            raise ValueError("Unknown ptq_linear type: %s" % type(ptq_linear))
 
     @unittest.skipIf(not TORCH_VERSION_AFTER_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_8da4w_linear(self):
@@ -144,7 +177,7 @@ def test_qat_8da4w_linear(self):
         )
 
         # Force the weights to be the same
-        self._set_ptq_weight(ptq_linear, qat_linear.weight, group_size)
+        self._set_ptq_weight(ptq_linear, qat_linear)
 
         # Compare linear values
         torch.manual_seed(self.SEED)
@@ -280,7 +313,7 @@ def test_qat_8da4w_quantizer_disable_fake_quant_backward(self):
         loss_fn1 = torch.nn.CrossEntropyLoss()
         loss_fn2 = torch.nn.CrossEntropyLoss()
         example_inputs = nn_model.example_inputs()
-        target = torch.randn(1, 64).float()
+        target = torch.randn(1, 512).float()
         output1 = nn_model(*example_inputs)
         output2 = qat_model(*example_inputs)
         torch.testing.assert_close(output1, output2, atol=0, rtol=0)
@@ -322,6 +355,130 @@ def test_qat_generic_fake_quantize(self):
         torch.testing.assert_close(py_out, ao_out, atol=0, rtol=0)
         torch.testing.assert_close(py_input.grad, ao_input.grad, atol=0, rtol=0)
 
+    def _assert_close_4w(self, val, ref):
+        # Note: for int4 weight-only quantization, we do not expect exact match
+        # because torch._weight_int4pack_mm and torch.mm do not match exactly.
+        # Here we use the same error bar as PyTorch core to determine closeness:
+        # https://github.com/pytorch/pytorch/blob/6079c5091091d872b8dafbaa4e31a5b6194647ad/test/test_linalg.py#L6079
+        mean_err = ((val - ref) / ref).mean().abs()
+        print(mean_err)
+        self.assertTrue(mean_err < 0.05)
+
+    @unittest.skipIf(not TORCH_VERSION_AFTER_2_4, "skipping when torch version is 2.4 or lower")
+    @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
+    def test_qat_4w_primitives(self):
+        n_bit = 4
+        group_size = 32
+        inner_k_tiles = 8
+        scales_precision = torch.bfloat16
+        device = torch.device("cuda")
+        dtype = torch.bfloat16
+        torch.manual_seed(self.SEED)
+        x = torch.randn(100, 256, dtype=dtype, device=device)
+        weight = torch.randn(512, 256, dtype=dtype, device=device)
+
+        # PTQ
+        (q_weight, scales_and_zeros) = groupwise_affine_quantize_tensor(
+            weight, n_bit, group_size, scales_precision,
+        )
+        q_weight = torch.ops.aten._convert_weight_to_int4pack(
+            q_weight.to(device), inner_k_tiles,
+        )
+        ptq_out = torch.ops.aten._weight_int4pack_mm(
+            x, q_weight, group_size, scales_and_zeros
+        )
+
+        # QAT
+        block_size = (1, group_size)
+        quant_min = 0
+        quant_max = 2 ** n_bit - 1
+        scales, zero_points = get_groupwise_affine_qparams(
+            weight, n_bit, group_size, scales_precision,
+        )
+        w_fq = fake_quantize_affine(
+            weight,
+            block_size,
+            scales,
+            zero_points,
+            torch.int32,
+            quant_min,
+            quant_max,
+            zero_point_domain = ZeroPointDomain.FLOAT,
+        )
+        qat_out = torch.nn.functional.linear(x, w_fq)
+
+        self._assert_close_4w(qat_out, ptq_out)
+
+    @unittest.skipIf(not TORCH_VERSION_AFTER_2_4, "skipping when torch version is 2.4 or lower")
+    @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
+    def test_qat_4w_linear(self):
+        from torchao.quantization.prototype.qat import Int4WeightOnlyQATLinear
+        from torchao.quantization.GPTQ import WeightOnlyInt4Linear
+
+        group_size = 128
+        device = torch.device("cuda")
+        dtype = torch.bfloat16
+        torch.manual_seed(self.SEED)
+        qat_linear = Int4WeightOnlyQATLinear(
+            256, 688, bias=False, groupsize=group_size, device=device,
+        )
+        ptq_linear = WeightOnlyInt4Linear(
+            256, 688, bias=False, groupsize=group_size, device=device,
+        )
+
+        # Force the weights to be the same
+        self._set_ptq_weight(ptq_linear, qat_linear)
+
+        # Compare linear values
+        torch.manual_seed(self.SEED)
+        x = torch.randn(100, 256, dtype=dtype, device=device)
+        x2 = copy.deepcopy(x)
+        qat_out = qat_linear(x)
+        ptq_out = ptq_linear(x2)
+        self._assert_close_4w(qat_out, ptq_out)
+
+    @unittest.skipIf(not TORCH_VERSION_AFTER_2_4, "skipping when torch version is 2.4 or lower")
+    @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
+    def test_qat_4w_quantizer(self):
+        from torchao.quantization.prototype.qat import Int4WeightOnlyQATQuantizer
+        from torchao.quantization.GPTQ import Int4WeightOnlyQuantizer
+
+        group_size = 32
+        inner_k_tiles = 8
+        device = torch.device("cuda")
+        dtype = torch.bfloat16
+        torch.manual_seed(self.SEED)
+        m = M().to(device).to(dtype)
+        m2 = copy.deepcopy(m)
+        qat_quantizer = Int4WeightOnlyQATQuantizer(
+            groupsize=group_size, inner_k_tiles=inner_k_tiles,
+        )
+        ptq_quantizer = Int4WeightOnlyQuantizer(
+            groupsize=group_size, inner_k_tiles=inner_k_tiles,
+        )
+        qat_model = qat_quantizer.prepare(m)
+        ptq_model = ptq_quantizer.quantize(m2)
+
+        # Compare model values
+        torch.manual_seed(self.SEED)
+        x = [i.to(device).to(dtype) for i in m.example_inputs()]
+        x2 = copy.deepcopy(x)
+        qat_out = qat_model(*x)
+        ptq_out = ptq_model(*x2)
+        self._assert_close_4w(qat_out, ptq_out)
+
+        # Convert QAT model and compare model values
+        converted_model = qat_quantizer.convert(qat_model)
+        converted_out = converted_model(*x)
+        torch.testing.assert_close(converted_out, ptq_out, atol=0, rtol=0)
+
+        # Compare converted state dict
+        ptq_state_dict = ptq_model.state_dict()
+        converted_state_dict = converted_model.state_dict()
+        self.assertEqual(ptq_state_dict.keys(), converted_state_dict.keys())
+        for k in ptq_state_dict.keys():
+            torch.testing.assert_close(ptq_state_dict[k], converted_state_dict[k], atol=0, rtol=0)
+
 
 if __name__ == "__main__":
     unittest.main()