paddlepaddle--paddle/test/legacy_test/test_fused_act_dequant_op.py

# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import unittest

import numpy as np

import paddle


def ceil_div(x: int, y: int) -> int:
    return (x + y - 1) // y


def align(x: int, y: int) -> int:
    return ceil_div(x, y) * y


def get_tma_aligned_size(x: int, element_size: int) -> int:
    """
    Align x to TMA-required size.
    Args:
        x: size in elements
        element_size: size of each element in bytes
    Returns:
        Aligned size in elements
    """
    kNumTMAAlignmentBytes = 16
    assert kNumTMAAlignmentBytes % element_size == 0
    return align(x, kNumTMAAlignmentBytes // element_size)


def ceil_to_ue8m0_paddle(x: paddle.Tensor):
    """
    x > 0
    return 2 ^ ceil(log2(x))
    """
    # log2(x)
    log2_x = paddle.log(x) / paddle.log(paddle.to_tensor(2.0, dtype=x.dtype))
    # ceil
    ceil_log2_x = paddle.ceil(log2_x)
    # 2^k
    return paddle.pow(paddle.to_tensor(2.0, dtype=x.dtype), ceil_log2_x)


def _get_mn_major_tma_aligned_packed_ue8m0_tensor_torch_impl(
    x: paddle.Tensor,
):
    assert x.dtype == paddle.float and x.dim() in (2, 3)

    ue8m0_tensor = (x.view(paddle.int) >> 23).to(paddle.uint8)

    mn, k = x.shape[-2], x.shape[-1]
    remove_dim = False

    if x.dim() == 2:
        x, remove_dim = x.unsqueeze(0), True
    b = x.shape[0]

    aligned_mn = get_tma_aligned_size(mn, 4)
    aligned_k = align(k, 4)

    padded = paddle.zeros(
        (b, aligned_mn, aligned_k), device=x.device, dtype=paddle.uint8
    )
    padded[:, :mn, :k] = ue8m0_tensor

    padded = (
        padded.view(-1)
        .view(dtype=paddle.int)
        .view(b, aligned_mn, aligned_k // 4)
    )

    transposed = paddle.zeros(
        (b, aligned_k // 4, aligned_mn), device=x.device, dtype=paddle.int
    ).mT
    transposed[:, :, :] = padded

    aligned_x = transposed[:, :mn, :]

    return aligned_x.squeeze(0) if remove_dim else aligned_x


def transform_scale_ue8m0(sf, mn, weight_block_size=None):
    get_mn_major_tma_aligned_packed_ue8m0_tensor = (
        _get_mn_major_tma_aligned_packed_ue8m0_tensor_torch_impl
    )
    if weight_block_size:
        assert weight_block_size == [128, 128]
        sf = sf.index_select(-2, paddle.arange(mn, device=sf.device) // 128)
    sf = get_mn_major_tma_aligned_packed_ue8m0_tensor(sf)
    return sf


def quant_ref(x_scale_fp32, mn, weight_block_size=None):
    x_scale_fp32_ = ceil_to_ue8m0_paddle(x_scale_fp32)
    ref_e8m0_scale = transform_scale_ue8m0(
        x_scale_fp32_, mn=mn, weight_block_size=weight_block_size
    )
    return ref_e8m0_scale


class TestActQuantDequant(unittest.TestCase):
    """Test cases for activation quantization and dequantization functions."""

    def setUp(self):
        """Set up test fixtures before each test method."""
        paddle.set_default_dtype('float32')

    def act_quant(
        self, x: paddle.Tensor, block_size: int = 128
    ) -> tuple[paddle.Tensor, paddle.Tensor]:
        """
        Quantize activation tensor to float8_e4m3fn format.

        Args:
            x: Input tensor to quantize
            block_size: Block size for quantization (default: 128)

        Returns:
            Tuple of (quantized_tensor, scale_factors)
        """
        self.assertTrue(x.is_contiguous(), "Input tensor must be contiguous")
        self.assertEqual(
            x.shape[-1] % block_size,
            0,
            f"Last dimension size must be divisible by block_size (block_size={block_size})",
        )

        # Convert to float32 for computation
        x_float = x.astype('float32')

        # Reshape to process blocks
        original_shape = x.shape
        # Reshape to (..., num_blocks, block_size)
        new_shape = [
            *original_shape[:-1],
            original_shape[-1] // block_size,
            block_size,
        ]
        x_reshaped = x_float.reshape(new_shape)

        # Compute scaling factors: max(abs(x)) / 448.0 for each block
        abs_x = paddle.abs(x_reshaped)
        max_vals = paddle.max(
            abs_x, axis=-1, keepdim=False
        )  # (..., num_blocks)
        s = max_vals / 448.0

        # Expand scaling factors to match x_reshaped shape for division
        s_expanded = s.unsqueeze(-1)  # (..., num_blocks, 1)

        # Quantize: x / s
        y_reshaped = x_reshaped / s_expanded

        # Reshape back to original shape
        y_float = y_reshaped.reshape(original_shape)

        # Convert to target dtype
        y = y_float.astype('float8_e4m3fn')

        return y, s

    def dequant_ref(
        self, x: paddle.Tensor, s: paddle.Tensor, block_size: int = 128
    ) -> paddle.Tensor:
        """
        Reference implementation for dequantizing activation tensor.

        Args:
            x: Quantized tensor
            s: Scale factors
            block_size: Block size used in quantization (default: 128)

        Returns:
            Dequantized tensor
        """
        self.assertTrue(
            x.is_contiguous() and s.is_contiguous(),
            "Input tensors must be contiguous",
        )
        self.assertEqual(x.dim(), 2, "Input tensor x must have 2 dimensions")
        self.assertEqual(s.dim(), 2, "Input tensor s must have 2 dimensions")

        M, N = x.shape

        # Convert to float32 for computation
        x_float = x.astype('float32')

        # Check if s needs to be expanded to match x shape
        if s.shape[1] == N // block_size:
            # s has shape (M, N//block_size), need to expand to (M, N)
            # Reshape s to (M, N//block_size, 1) then repeat along last dimension
            s_expanded = s.unsqueeze(-1)  # (M, N//block_size, 1)
            s_expanded = paddle.tile(
                s_expanded, [1, 1, block_size]
            )  # (M, N//block_size, block_size)
            s_expanded = s_expanded.reshape([M, N])  # (M, N)
        else:
            # s already has shape (M, N)
            s_expanded = s

        # Dequantize: x * s
        y = x_float * s_expanded

        # Convert to default dtype
        y = y.astype(paddle.get_default_dtype())

        return y

    def test_act_quant_basic_functionality(self):
        """Test basic functionality of act_quant function."""
        # Test with simple case
        x = paddle.randn([4, 256]).astype("bfloat16")
        x = paddle.clip(x, min=-10, max=10)

        x_fp8, scale = self.act_quant(x, block_size=128)

        # Check output shapes
        self.assertEqual(x_fp8.shape, x.shape)
        self.assertEqual(scale.shape, [4, 2])  # 256 // 128 = 2

        # Check output dtypes
        self.assertEqual(x_fp8.dtype, paddle.float8_e4m3fn)
        self.assertEqual(scale.dtype, paddle.float32)

    def test_act_dequant_consistency_small(self):
        """Test consistency between reference and fused implementations with small tensors."""
        test_cases = [
            (512, 7168),
            (2048, 7168),
            (4096, 7168),
        ]

        for height, width in test_cases:
            with self.subTest(height=height, width=width):
                self._test_single_case(height, width)

    def test_act_dequant_consistency_various_sizes(self):
        """Test with various tensor sizes."""
        test_cases = [
            (128, 256),
            (256, 512),
            (1024, 2048),
        ]

        for height, width in test_cases:
            with self.subTest(height=height, width=width):
                self._test_single_case(height, width)

    def _test_single_case(
        self, height: int, width: int, rtol: float = 1e-2, atol: float = 1e-2
    ):
        """
        Test a single case with given dimensions.

        Args:
            height: Tensor height
            width: Tensor width
            rtol: Relative tolerance for comparison
            atol: Absolute tolerance for comparison
        """
        # Generate test data
        x = paddle.clip(
            paddle.randn([height, width]).astype("bfloat16"), min=-50, max=50
        )

        # Perform quantization
        x_fp8, scale = self.act_quant(x)

        # Get results from both implementations
        if hasattr(paddle.incubate.nn.functional, 'fused_act_dequant'):
            dequant_result_fused = (
                paddle.incubate.nn.functional.fused_act_dequant(x_fp8, scale)
            )
        else:
            # Skip fused test if not available
            self.skipTest(
                "fused_act_dequant not available in this Paddle version"
            )

        dequant_result_ref = self.dequant_ref(x_fp8, scale)

        # Convert to numpy for comparison
        fused_np = dequant_result_fused.astype("float32").numpy()
        ref_np = dequant_result_ref.astype("float32").numpy()

        # Check for NaN values
        nan_cnt_fused = np.sum(np.isnan(fused_np))
        nan_cnt_ref = np.sum(np.isnan(ref_np))

        self.assertEqual(
            nan_cnt_fused,
            0,
            f"Fused result contains {nan_cnt_fused} NaN values",
        )
        self.assertEqual(
            nan_cnt_ref,
            0,
            f"Reference result contains {nan_cnt_ref} NaN values",
        )

        # Compare results
        try:
            np.testing.assert_allclose(fused_np, ref_np, rtol=rtol, atol=atol)
        except AssertionError as e:
            self.fail(
                f"Results don't match for shape [{height}, {width}]: {e!s}"
            )

    def test_ue8m0_support(self):
        """Test ue8m0 support in fused_act_dequant."""
        if not hasattr(paddle.incubate.nn.functional, 'fused_act_dequant'):
            self.skipTest(
                "fused_act_dequant not available in this Paddle version"
            )

        height, width = 4096, 7168
        x = paddle.clip(
            paddle.randn([height, width]).astype("bfloat16"), min=-50, max=50
        )
        x_fp8, scale = paddle.incubate.nn.functional.fp8_quant_blockwise(
            x, quant_method="1x128", output_scale_transpose=False
        )

        # 1. Align scale to ue8m0 values (2^k) in float32
        scale_aligned_fp32 = ceil_to_ue8m0_paddle(scale)

        # 2. Pack aligned scale to ue8m0 format (int32)
        scale_packed_int32 = transform_scale_ue8m0(
            scale_aligned_fp32, mn=height
        )

        # 3. Run fused_act_dequant with aligned float32 scale
        out_fp32 = paddle.incubate.nn.functional.fused_act_dequant(
            x_fp8, scale_aligned_fp32
        )

        # 4. Run fused_act_dequant with packed int32 scale
        out_ue8m0 = paddle.incubate.nn.functional.fused_act_dequant(
            x_fp8, scale_packed_int32
        )

        # 5. Compare
        out_fp32_np = out_fp32.numpy()
        out_ue8m0_np = out_ue8m0.numpy()

        np.testing.assert_allclose(out_fp32_np, out_ue8m0_np, rtol=0, atol=0)

    def test_invalid_inputs(self):
        """Test error handling for invalid inputs."""
        # Test non-divisible block size
        x = paddle.randn([4, 255]).astype(
            "bfloat16"
        )  # 255 is not divisible by 128

        with self.assertRaises(AssertionError):
            self.act_quant(x, block_size=128)


if __name__ == '__main__':
    # Configure test runner
    unittest.main(verbosity=2)