add fused transpose and non-transpose kernel and use it for grad output (#1497)

Author: danielvegamyhre

torchao/prototype/float8nocompile/__init__.py

@@ -16,8 +16,7 @@
     ToFP8ColumnMajor,
     ToFP8ColumnMajorT,
     ToFP8RowAndColumnMajor,
-    ToFP8RowMajor,
-    ToFP8RowMajorT,
+    ToFP8RowMajorTAndNonT,
 )
 from torchao.prototype.float8nocompile.kernels.fp8_dynamic_tensorwise import (
     KernelAlgorithm,
@@ -138,12 +137,14 @@ def backward(ctx, grad_output):
         input_fp8_col_major, weight_hp = ctx.saved_tensors
 
         # cast grad output to float8_e5m2 for backward
-        grad_output_fp8_row_major = ToFP8RowMajor.apply(
-            grad_output,
-            ctx.config.cast_config_grad_output.target_dtype,
-            ctx.linear_mm_config,
-            GemmInputRole.GRAD_OUTPUT,
-            ctx.kernel_algo,
+        grad_output_fp8_row_major, grad_output_t_row_major = (
+            ToFP8RowMajorTAndNonT.apply(
+                grad_output,
+                ctx.config.cast_config_grad_output.target_dtype,
+                ctx.linear_mm_config,
+                GemmInputRole.GRAD_OUTPUT,
+                ctx.kernel_algo,
+            )
         )
 
         # grad_input = grad_output @ weight
@@ -159,12 +160,6 @@ def backward(ctx, grad_output):
         # grad_weight = grad_output_t @ input
         # apparently this variant is slightly faster than `grad_weight_t = input_t @ grad_output`
         # source: https://github.com/pytorch/ao/blob/fe5f11b2c58b452e01ba9ec7359629928b143619/torchao/float8/float8_linear.py#L84-L85
-        grad_output_t_row_major = ToFP8RowMajorT.apply(
-            grad_output,
-            ctx.config.cast_config_grad_output.target_dtype,
-            ctx.linear_mm_config,
-            GemmInputRole.GRAD_OUTPUT,
-            ctx.kernel_algo,
-        )
         grad_weight = torch.mm(grad_output_t_row_major, input_fp8_col_major)
+
         return grad_input, grad_weight, None, None, None

torchao/prototype/float8nocompile/float8nocompile_linear.py

@@ -10,17 +10,15 @@
 
 import torch
 
-from torchao.float8.float8_tensor import (
-    GemmInputRole,
-    LinearMMConfig,
-)
+from torchao.float8.float8_tensor import GemmInputRole, LinearMMConfig
 from torchao.prototype.float8nocompile.kernels.fp8_dynamic_tensorwise import (
     KernelAlgorithm,
     hp_to_fp8_col_major,
     hp_to_fp8_col_major_t,
     hp_to_fp8_row_and_col_major,
     hp_to_fp8_row_major,
     hp_to_fp8_row_major_t,
+    hp_to_fp8_row_major_t_and_non_t,
 )
 
 
@@ -172,3 +170,33 @@ def forward(
     @staticmethod
     def backward(ctx, g):
         return g, None, None, None, None
+
+
+class ToFP8RowMajorTAndNonT(torch.autograd.Function):
+    """
+    A differentiable conversion to fp8.
+    * forward: convert from high precision to float8 and produces both row-major (transposed) and row-major (non-transposed) outputs
+    * backward: pass the gradient without changes
+    """
+
+    @staticmethod
+    def forward(
+        ctx,
+        tensor: torch.Tensor,
+        float8_dtype: torch.dtype,
+        linear_mm_config: LinearMMConfig,
+        gemm_input_role: GemmInputRole,
+        kernel_algo: KernelAlgorithm = KernelAlgorithm.ATOMIC_MAX,
+    ):
+        fp8_row_major, fp8_row_major_t = hp_to_fp8_row_major_t_and_non_t(
+            tensor,
+            float8_dtype,
+            linear_mm_config,
+            gemm_input_role,
+            algo=kernel_algo,
+        )
+        return fp8_row_major, fp8_row_major_t
+
+    @staticmethod
+    def backward(ctx, g):
+        return g, None, None, None, None

torchao/prototype/float8nocompile/float8nocompile_scaling_utils.py

@@ -305,6 +305,82 @@ def _to_fp8_row_and_col_major(
     tl.store(col_major_out_ptr + col_major_offs, fp8_vals, mask=mask)
 
 
+@triton.autotune(
+    configs=kernel_configs_2D,
+    key=["num_elements"],
+)
+@triton.jit
+def _to_fp8_row_major_t_and_non_t(
+    input_ptr,
+    row_major_out_ptr,
+    row_major_t_out_ptr,
+    scale_ptr,
+    num_elements: int,
+    fp8_dtype_min: float,
+    fp8_dtype_max: float,
+    input_num_rows: int,
+    input_num_cols: int,
+    input_stride_row: int,
+    input_stride_col: int,
+    row_major_out_stride_row: int,
+    row_major_out_stride_col: int,
+    row_major_t_out_stride_row: int,
+    row_major_t_out_stride_col: int,
+    input_dtype: tl.constexpr,
+    output_dtype: tl.constexpr,
+    BLOCK_SIZE_ROWS: tl.constexpr,
+    BLOCK_SIZE_COLS: tl.constexpr,
+    EPS: tl.constexpr,
+):
+    """
+    Reads a row-major, high precision input tensor and writes 2 output tensors:
+    1) fp8 row major tensor (transposed)
+    2) fp8 row major tensor
+    """
+    block_row_id = tl.program_id(axis=0)
+    block_col_id = tl.program_id(axis=1)
+
+    # load scaling factor
+    scale = tl.load(scale_ptr).to(tl.float32)
+
+    # load block of input tensor
+    block_row_start = block_row_id * BLOCK_SIZE_ROWS
+    block_col_start = block_col_id * BLOCK_SIZE_COLS
+    block_row_offs = block_row_start + tl.arange(0, BLOCK_SIZE_ROWS)
+    block_col_offs = block_col_start + tl.arange(0, BLOCK_SIZE_COLS)
+    input_offs = (
+        block_row_offs[:, None] * input_stride_row
+        + block_col_offs[None, :] * input_stride_col
+    )
+    mask = (block_row_offs[:, None] < input_num_rows) & (
+        block_col_offs[None, :] < input_num_cols
+    )
+    vals = tl.load(input_ptr + input_offs, mask=mask).to(input_dtype)
+
+    # perform conversion
+    vals = vals * scale
+    fp8_vals = tl.clamp(vals, min=fp8_dtype_min, max=fp8_dtype_max).to(output_dtype)
+
+    # write row-major output
+    row_major_offs = (
+        block_row_offs[:, None] * row_major_out_stride_row
+        + block_col_offs[None, :] * row_major_out_stride_col
+    )
+    tl.store(row_major_out_ptr + row_major_offs, fp8_vals, mask=mask)
+
+    # write tranposed row-major output
+    row_major_t_num_rows = input_num_cols
+    row_major_t_num_cols = input_num_rows
+    row_major_t_offs = (
+        block_col_offs[:, None] * row_major_t_out_stride_row
+        + block_row_offs[None, :] * row_major_t_out_stride_col
+    )
+    mask = (block_row_offs[:, None] < row_major_t_num_rows) & (
+        block_col_offs[None, :] < row_major_t_num_cols
+    )
+    tl.store(row_major_t_out_ptr + row_major_t_offs, fp8_vals.trans(1, 0), mask=mask)
+
+
 @triton.autotune(configs=kernel_configs_1D, key=["num_elements"])
 @triton.jit
 def _amax_atomic(
@@ -701,6 +777,88 @@ def hp_to_fp8_row_and_col_major(
     return fp8_tensor_row_major, fp8_tensor_col_major
 
 
+def hp_to_fp8_row_major_t_and_non_t(
+    hp_tensor: torch.Tensor,
+    fp8_dtype: torch.dtype,
+    linear_mm_config: LinearMMConfig,
+    gemm_input_role: GemmInputRole = GemmInputRole.INPUT,
+    algo: KernelAlgorithm = KernelAlgorithm.ATOMIC_MAX,
+) -> Float8Tensor:
+    assert hp_tensor.is_contiguous(), "input tensor must be contiguous"
+
+    tl_input_dtype = FP8_DTYPE_MAP[hp_tensor.dtype]
+    tl_output_dtype = FP8_DTYPE_MAP[fp8_dtype]
+
+    fp8_dtype_min = torch.finfo(fp8_dtype).min
+    fp8_dtype_max = torch.finfo(fp8_dtype).max
+
+    # compute scaling factor for tensor
+    scale = _hp_tensor_to_scale(
+        hp_tensor,
+        tl_input_dtype,
+        fp8_dtype_max,
+        algo,
+    )
+
+    # perform fp8 conversion
+    input_num_rows, input_num_cols = hp_tensor.shape
+    transposed_num_rows, transposed_num_cols = input_num_cols, input_num_rows
+    num_elements = hp_tensor.numel()
+
+    # preallocate necessary output tensors
+    fp8_output_row_major = torch.empty(
+        (input_num_rows, input_num_cols), dtype=fp8_dtype, device=hp_tensor.device
+    )
+    fp8_output_row_major_t = torch.empty(
+        (transposed_num_rows, transposed_num_cols),
+        dtype=fp8_dtype,
+        device=hp_tensor.device,
+    )
+
+    # launch triton kernel to perform conversion
+    grid = lambda meta: (
+        triton.cdiv(input_num_rows, meta["BLOCK_SIZE_ROWS"]),
+        triton.cdiv(input_num_cols, meta["BLOCK_SIZE_COLS"]),
+    )
+    _to_fp8_row_major_t_and_non_t[grid](
+        hp_tensor,
+        fp8_output_row_major,
+        fp8_output_row_major_t,
+        scale,
+        num_elements,
+        fp8_dtype_min,
+        fp8_dtype_max,
+        input_num_rows,
+        input_num_cols,
+        hp_tensor.stride(0),
+        hp_tensor.stride(1),
+        fp8_output_row_major.stride(0),
+        fp8_output_row_major.stride(1),
+        fp8_output_row_major_t.stride(0),
+        fp8_output_row_major_t.stride(1),
+        input_dtype=tl_input_dtype,
+        output_dtype=tl_output_dtype,
+        EPS=EPS,
+    )
+
+    # wrap outputs in Float8Tensors
+    fp8_tensor_row_major = Float8Tensor(
+        fp8_output_row_major,
+        scale,
+        orig_dtype=hp_tensor.dtype,
+        linear_mm_config=linear_mm_config,
+        gemm_input_role=gemm_input_role,
+    )
+    fp8_tensor_row_major_t = Float8Tensor(
+        fp8_output_row_major_t,
+        scale,
+        orig_dtype=hp_tensor.dtype,
+        linear_mm_config=linear_mm_config,
+        gemm_input_role=gemm_input_role,
+    )
+    return fp8_tensor_row_major, fp8_tensor_row_major_t
+
+
 def _hp_tensor_to_scale(
     hp_tensor: torch.Tensor,
     tl_input_dtype: tl.core.dtype,

torchao/prototype/float8nocompile/kernels/__init__.py

@@ -11,6 +11,7 @@
     hp_to_fp8_row_and_col_major,
     hp_to_fp8_row_major,
     hp_to_fp8_row_major_t,
+    hp_to_fp8_row_major_t_and_non_t,
 )
 
 
@@ -335,3 +336,77 @@ def test_fp8_hp_to_fp8_row_and_col_major(
             torch.float8_e4m3fn,
             LinearMMConfig(),
         )
+
+
+@pytest.mark.parametrize(
+    "algo",
+    [KernelAlgorithm.REDUCTION, KernelAlgorithm.ATOMIC_MAX],
+)
+@pytest.mark.parametrize(
+    "input_shape",
+    [(2, 4), (32, 16), (512, 512)],
+)
+def test_fp8_hp_to_fp8_row_major_t_and_non_t(
+    input_shape: tuple[int, int], algo: KernelAlgorithm
+):
+    assert torch.cuda.is_available()
+    device = "cuda"
+    input_bf16 = torch.randn(input_shape, dtype=torch.bfloat16, device=device)
+    x_bf16 = input_bf16.clone().detach().to(device)
+    y_bf16 = input_bf16.clone().detach().to(device)
+
+    # production implementation
+    x_fp8_row_major = hp_tensor_to_float8_dynamic(
+        x_bf16,
+        torch.float8_e4m3fn,
+        LinearMMConfig(),
+    )
+    x_fp8_row_major_t = x_fp8_row_major.t().contiguous()
+
+    # float8nocompile triton implementation
+    y_fp8_row_major, y_fp8_row_major_t = hp_to_fp8_row_major_t_and_non_t(
+        y_bf16,
+        torch.float8_e4m3fn,
+        LinearMMConfig(),
+        algo=algo,
+    )
+
+    # check scales
+    assert torch.eq(x_fp8_row_major._scale, y_fp8_row_major._scale)
+    assert torch.eq(x_fp8_row_major_t._scale, y_fp8_row_major_t._scale)
+
+    # check data
+    assert torch.all(torch.eq(x_fp8_row_major._data, y_fp8_row_major._data))
+    assert torch.all(torch.eq(x_fp8_row_major_t._data, y_fp8_row_major_t._data))
+
+    # check shapes
+    assert x_fp8_row_major.shape == y_fp8_row_major.shape
+    assert x_fp8_row_major_t.shape == y_fp8_row_major_t.shape
+
+    # check strides
+    assert x_fp8_row_major.stride() == y_fp8_row_major.stride()
+    assert x_fp8_row_major_t.stride() == y_fp8_row_major_t.stride()
+
+    # check memory layout
+    assert is_row_major(x_fp8_row_major.stride())
+    assert is_row_major(y_fp8_row_major.stride())
+    assert is_row_major(x_fp8_row_major_t.stride())
+    assert is_row_major(y_fp8_row_major_t.stride())
+
+    # check underlying memory layout
+    assert (
+        x_fp8_row_major._data.storage().tolist()
+        == y_fp8_row_major._data.storage().tolist()
+    )
+    assert (
+        x_fp8_row_major_t._data.storage().tolist()
+        == y_fp8_row_major_t._data.storage().tolist()
+    )
+
+    # assert that error is raised when input tensor is not contiguous
+    with pytest.raises(AssertionError, match="tensor must be contiguous"):
+        hp_to_fp8_row_major_t_and_non_t(
+            y_bf16.t(),  # transpose so tensor memory layout is no longer contiguous
+            torch.float8_e4m3fn,
+            LinearMMConfig(),
+        )