[wip] make scale shape 2d and match qdata shape in NVFP4Tensor

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
ghstack-source-id: 9dd2ad6
ghstack-comment-id: 3357503258
Pull-Request: #3108

Author: vkuzo

test/prototype/mx_formats/test_nvfp4_tensor.py

@@ -21,6 +21,7 @@
     per_tensor_amax_to_scale,
     unpack_uint4,
 )
+from torchao.prototype.mx_formats.utils import ceil_div
 from torchao.quantization.utils import compute_error
 from torchao.testing.utils import skip_if_rocm
 from torchao.utils import (
@@ -525,3 +526,64 @@ def test_nvfp4_to_copy():
     assert x.act_quant_kwargs == y.act_quant_kwargs
     assert x.dtype == torch.float32
     assert y.dtype == torch.bfloat16
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+@pytest.mark.skipif(
+    not torch_version_at_least("2.8.0"), reason="NVFP4 requires PyTorch 2.8+"
+)
+@pytest.mark.parametrize("transpose", [False, True])
+@pytest.mark.parametrize("use_triton_kernel", [False, True])
+@pytest.mark.parametrize("is_swizzled_scales", [False, True])
+@pytest.mark.parametrize(
+    "mk",
+    (
+        (128, 64),
+        (128 + 16, 64),
+        (128, 64 + 16),
+        (128 + 16, 64 + 16),
+    ),
+)
+def test_scale_shape_matches_qdata(
+    transpose, use_triton_kernel, is_swizzled_scales, mk
+):
+    if use_triton_kernel and not is_sm_at_least_100():
+        pytest.skip("CUDA capability >= 10.0 required for nvfp4 triton kernel")
+    if use_triton_kernel and not is_swizzled_scales:
+        pytest.skip("triton kernel requires swizzled scales")
+
+    M, K = mk
+
+    block_size = 16
+
+    x_hp = torch.randn(M, K, device="cuda")
+    x = NVFP4Tensor.to_nvfp4(
+        x_hp, is_swizzled_scales=is_swizzled_scales, use_triton_kernel=use_triton_kernel
+    )
+
+    m_dim, k_dim = 0, 1
+    if transpose:
+        x_hp = x_hp.t()
+        x = x.t()
+        m_dim, k_dim = 1, 0
+
+    orig_m = x_hp.shape[m_dim]
+    expected_padded_m = orig_m
+    if is_swizzled_scales:
+        # in swizzled nvfp4, a 128x128 data unpacked / 128x64 data packed maps to a 32x16 scale tile
+        expected_padded_m = ceil_div(orig_m, 128) * 32
+    actual_padded_m = x._scale_e4m3.shape[m_dim]
+    assert expected_padded_m == actual_padded_m, (
+        f"incompatible padded shape for dim {m_dim}: {expected_padded_m=}, {actual_padded_m=}, {x.shape}, {x._scale_e4m3.shape}"
+    )
+
+    orig_k = x_hp.shape[k_dim]
+    expected_padded_k = orig_k // block_size
+    if is_swizzled_scales:
+        # in swizzled nvfp4, a 128x128 data unpacked / 128x64 data packed maps to a 32x16 scale tile
+        expected_padded_k = ceil_div(orig_k // block_size, 4) * 16
+    actual_padded_k = x._scale_e4m3.shape[k_dim]
+
+    assert expected_padded_k == actual_padded_k, (
+        f"incompatible padded shape for dim {k_dim}: {expected_padded_k}, {actual_padded_k=}, {x.shape}, {x._scale_e4m3.shape}"
+    )

torchao/prototype/mx_formats/nvfp4_tensor.py

@@ -24,7 +24,11 @@
     tensor_size_fp4x2_to_hp,
     tensor_size_hp_to_fp4x2,
 )
-from torchao.prototype.mx_formats.utils import from_blocked, to_blocked
+from torchao.prototype.mx_formats.utils import (
+    from_blocked,
+    hp_data_dims_to_swizzled_scale_dims_nvfp4,
+    to_blocked,
+)
 from torchao.quantization.quantize_.common import (
     QuantizeTensorKwargs,
 )
@@ -170,6 +174,9 @@ def to_nvfp4(
         Returns:
             NVFP4Tensor: Quantized tensor in NVFP4 format
         """
+        assert len(data_hp.shape) == 2, "unsupported"
+        M, K = data_hp.shape[0], data_hp.shape[1]
+
         if use_triton_kernel:
             assert is_swizzled_scales, "Triton kernel only supports swizzled scales"
             assert data_hp.shape[1] % 16 == 0, (
@@ -181,12 +188,19 @@ def to_nvfp4(
                 data_hp, block_size, per_tensor_scale
             )
             if is_swizzled_scales:
-                M, K = data_hp.shape[0], data_hp.shape[1]
                 scale_shape = (M, K // block_size)
                 blockwise_scales = to_blocked(
                     blockwise_scales.view(scale_shape)
                 ).flatten()
 
+        if is_swizzled_scales:
+            scale_M, scale_K = hp_data_dims_to_swizzled_scale_dims_nvfp4(M, K)
+        else:
+            # a 1x16 unpacked or 1x8 packed qdata tile corresponds to 1
+            # scale element
+            scale_M, scale_K = M, K // block_size
+        blockwise_scales = blockwise_scales.view(scale_M, scale_K)
+
         return NVFP4Tensor(
             data_lp,
             blockwise_scales,
@@ -239,13 +253,13 @@ def get_hp_scales(self) -> torch.Tensor:
         is_transposed = self.qdata.stride(0) < self.qdata.stride(1)
         if is_transposed:
             M, K = self.shape[1], self.shape[0]
+            scale_e4m3 = self._scale_e4m3.t()
         else:
             M, K = self.shape[0], self.shape[1]
+            scale_e4m3 = self._scale_e4m3
 
         if self._is_swizzled_scales:
-            scale_e4m3 = from_blocked(self._scale_e4m3, M, K // self._block_size)
-        else:
-            scale_e4m3 = self._scale_e4m3
+            scale_e4m3 = from_blocked(scale_e4m3, M, K // self._block_size)
 
         return (
             scale_e4m3.to(self._orig_dtype)
@@ -369,6 +383,11 @@ def nvfp4_slice(func, types, args, kwargs):
 
     M, K = x.shape[0], x.shape[1]
 
+    # The scale manipulations below assume a flattened scale. For now, we
+    # flatten the scale, go through the calculations below, and then reshape
+    # it back to the format which matches the shape of `qdata`.
+    # TODO(future PR): update this
+
     if x._is_swizzled_scales:
         scale_rows = M
         scale_cols = K // x._block_size
@@ -407,7 +426,9 @@ def nvfp4_slice(func, types, args, kwargs):
                 else None
             )
 
-            sliced_scale = aten.slice.Tensor(x._scale_e4m3, 0, start_idx, end_idx, 1)
+            sliced_scale = aten.slice.Tensor(
+                x._scale_e4m3.flatten(), 0, start_idx, end_idx, 1
+            )
             sliced_data = aten.slice.Tensor(x.qdata, 0, start, end, step)
 
         elif dim == 1:
@@ -462,7 +483,7 @@ def nvfp4_slice(func, types, args, kwargs):
                     row_start = row_block * elements_per_row_block
                     col_start = row_start + start_col_block * elements_per_block
                     col_end = row_start + end_col_block * elements_per_block
-                    slices_to_extract.append(x._scale_e4m3[col_start:col_end])
+                    slices_to_extract.append(x._scale_e4m3.flatten()[col_start:col_end])
 
                 # Concatenate all the slices
                 sliced_scale = torch.cat(slices_to_extract, dim=0)
@@ -515,6 +536,19 @@ def nvfp4_slice(func, types, args, kwargs):
 
         sliced_scale = sliced_scale.flatten()
 
+    # reshape at the end
+    sliced_M = sliced_data.shape[0]
+    # multiply by 2 to convert from bytes to num_elements
+    sliced_K = sliced_data.shape[1] * 2
+    if x._is_swizzled_scales:
+        scale_M, scale_K = hp_data_dims_to_swizzled_scale_dims_nvfp4(sliced_M, sliced_K)
+    else:
+        # a 1x16 unpacked or 1x8 packed qdata tile corresponds to 1
+        # scale element
+        scale_M = sliced_M
+        scale_K = sliced_K // x._block_size
+    sliced_scale = sliced_scale.view(scale_M, scale_K)
+
     # Create result tensor
     result = NVFP4Tensor(
         sliced_data,
@@ -537,7 +571,7 @@ def nvfp4_t(func, types, args, kwargs):
     old = args[0]
     new = NVFP4Tensor(
         old.qdata.t(),
-        old._scale_e4m3,
+        old._scale_e4m3.t(),
         old._block_size,
         old._orig_dtype,
         old._per_tensor_scale,
@@ -576,7 +610,9 @@ def _addmm_nvfp4_dispatch(
     The only difference is whether bias is None or not.
     """
     assert a.qdata.is_contiguous()
+    assert a._scale_e4m3.is_contiguous()
     assert b.qdata.t().is_contiguous()
+    assert b._scale_e4m3.t().is_contiguous()
     assert a._block_size == 16, f"NVFP4 requires block_size=16, got {a._block_size}"
     assert b._block_size == 16, f"NVFP4 requires block_size=16, got {b._block_size}"
 
@@ -591,9 +627,9 @@ def _addmm_nvfp4_dispatch(
         a_scale_blocked = to_blocked(a_scale)
 
     if b._is_swizzled_scales:
-        b_scale_blocked = b._scale_e4m3  # Already swizzled
+        b_scale_blocked = b._scale_e4m3.t()  # Already swizzled
     else:
-        b_scale = b._scale_e4m3.view(N, K // b._block_size)
+        b_scale = b._scale_e4m3.t().view(N, K // b._block_size)
         b_scale_blocked = to_blocked(b_scale)
 
     # Merge double quant scales into 1 scale for Scale_In^D

torchao/prototype/mx_formats/utils.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the root directory of this source tree.
 
+from typing import Tuple
+
 import torch
 from torch.distributed._tensor import DTensor
 
@@ -99,6 +101,22 @@ def from_blocked(
     return padded[:original_rows, :original_cols]
 
 
+def hp_data_dims_to_swizzled_scale_dims_nvfp4(
+    hp_data_M,
+    hp_data_K,
+) -> Tuple[int, int]:
+    """
+    Given the `M` and `K` dimensions of a high precision contiguous tensor,
+    returns a 2d tuple of the dims of the swizzled nvfp4 scale corresponding to
+    that tensor.
+    """
+    # a 128x64 unpacked or 128x32 packed qdata tile corresponds
+    # to a swizzled 32x16 scale tile
+    scale_M = ceil_div(hp_data_M, 128) * 32
+    scale_K = ceil_div(hp_data_K, 64) * 16
+    return scale_M, scale_K
+
+
 def _to_blocked_single(scales: Tensor) -> Tensor:
     """Assume that we have a 128x4 block of scales in K Major order