vllm.model_executor.layers.quantization.utils.fp8_utils ¶

def _maybe_pad_fp8_weight(weight: torch.Tensor) -> torch.Tensor:
    """Pad the weight tensor. This is an optimization on ROCm platform, which
    can benefit from tensors located far enough from one another in memory"""
    if (envs.VLLM_ROCM_FP8_PADDING and current_platform.is_rocm()
            and weight.stride(-1) == 1
            and (weight.stride(-2) * weight.element_size()) % 512 == 0):
        num_pad = 256 // weight.element_size()
        import torch.nn.functional as F
        weight = F.pad(weight, (0, num_pad), "constant", 0)[..., :-num_pad]
        torch.cuda.empty_cache()
    return weight

@triton.jit
def _per_token_group_quant_fp8(
    # Pointers to inputs and output
    y_ptr,
    y_q_ptr,
    y_s_ptr,
    group_size,
    # Num columns of y
    y_num_columns,
    y_row_stride,
    # Avoid to divide zero
    eps,
    # Information for float8
    fp8_min,
    fp8_max,
    use_ue8m0: tl.constexpr,
    # Meta-parameters
    BLOCK: tl.constexpr,
):
    """A Triton-accelerated function to perform per-token-group
    quantization on a tensor.
    This function converts the tensor values into float8 values.
    """
    groups_per_row = y_num_columns // group_size

    # Map the program id to the row of X and Y it should compute.
    g_id = tl.program_id(0)
    row = g_id // groups_per_row
    row_g_id = g_id % groups_per_row

    # Ensure offset calculations use int64 to prevent overflow
    y_ptr_offset = (row.to(tl.int64) * y_row_stride) + (row_g_id.to(tl.int64) *
                                                        group_size)
    y_ptr += y_ptr_offset

    y_q_ptr_offset = g_id.to(tl.int64) * group_size
    y_q_ptr += y_q_ptr_offset
    y_s_ptr += g_id

    cols = tl.arange(0, BLOCK)  # N <= BLOCK
    mask = cols < group_size

    y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
    # Quant
    _absmax = tl.maximum(tl.max(tl.abs(y)), eps)
    scale_raw = _absmax / fp8_max
    y_s = tl.math.exp2(tl.ceil(tl.log2(scale_raw))) if use_ue8m0 else scale_raw
    y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)

    tl.store(y_q_ptr + cols, y_q, mask=mask)
    tl.store(y_s_ptr, y_s)

@triton.jit
def _per_token_group_quant_fp8_colmajor(
    # Pointers to inputs and output
    y_ptr,
    y_q_ptr,
    y_s_ptr,
    group_size,
    # Num columns of y
    y_num_columns,
    y_row_stride,
    # Stride from one column to the next of y_s
    y_s_col_stride,
    # Avoid to divide zero
    eps,
    # Information for float8
    fp8_min,
    fp8_max,
    use_ue8m0: tl.constexpr,
    # Meta-parameters
    BLOCK: tl.constexpr,
):
    """A Triton-accelerated function to perform per-token-group
    quantization on a tensor.
    This function converts the tensor values into float8 values.
    """
    groups_per_row = y_num_columns // group_size

    # Map the program id to the row of X and Y it should compute.
    g_id = tl.program_id(0)
    row = g_id // groups_per_row
    row_g_id = g_id % groups_per_row

    # Ensure offset calculations use int64 to prevent overflow
    y_ptr_offset = (row.to(tl.int64) * y_row_stride) + (row_g_id.to(tl.int64) *
                                                        group_size)
    y_ptr += y_ptr_offset

    y_q_ptr_offset = g_id.to(tl.int64) * group_size
    y_q_ptr += y_q_ptr_offset

    # Convert g_id the flattened block coordinate to 2D so we can index
    # into the output y_scales matrix
    blocks_per_row = y_num_columns // group_size
    scale_col = g_id % blocks_per_row
    scale_row = g_id // blocks_per_row
    # Ensure offset calculation uses int64 for y_s_ptr
    y_s_ptr_offset = (scale_col.to(tl.int64) * y_s_col_stride) + scale_row.to(
        tl.int64)
    y_s_ptr += y_s_ptr_offset

    cols = tl.arange(0, BLOCK)  # group_size <= BLOCK
    mask = cols < group_size

    y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
    # Quant
    _absmax = tl.maximum(tl.max(tl.abs(y)), eps)
    scale_raw = _absmax / fp8_max
    y_s = tl.math.exp2(tl.ceil(tl.log2(scale_raw))) if use_ue8m0 else scale_raw
    y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)

    tl.store(y_q_ptr + cols, y_q, mask=mask)
    tl.store(y_s_ptr, y_s)

@triton.jit
def _w8a8_block_fp8_matmul(
    # Pointers to inputs and output
    A,
    B,
    C,
    As,
    Bs,
    # Shape for matmul
    M,
    N,
    K,
    # Block size for block-wise quantization
    group_n,
    group_k,
    # Stride for inputs and output
    stride_am,
    stride_ak,
    stride_bk,
    stride_bn,
    stride_cm,
    stride_cn,
    stride_As_m,
    stride_As_k,
    stride_Bs_k,
    stride_Bs_n,
    # Meta-parameters
    BLOCK_SIZE_M: tl.constexpr,
    BLOCK_SIZE_N: tl.constexpr,
    BLOCK_SIZE_K: tl.constexpr,
    GROUP_SIZE_M: tl.constexpr,
):
    """Triton-accelerated function used to perform linear operations (dot
    product) on input tensors `A` and `B` with block-wise quantization, and
    store the result in output tensor `C`.
    """

    pid = tl.program_id(axis=0)
    num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
    num_pid_in_group = GROUP_SIZE_M * num_pid_n
    group_id = pid // num_pid_in_group
    first_pid_m = group_id * GROUP_SIZE_M
    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
    pid_m = first_pid_m + (pid % group_size_m)
    pid_n = (pid % num_pid_in_group) // group_size_m

    offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
    offs_k = tl.arange(0, BLOCK_SIZE_K)
    a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
    b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)

    As_ptrs = As + offs_am * stride_As_m
    offs_bsn = offs_bn // group_n
    Bs_ptrs = Bs + offs_bsn * stride_Bs_n

    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
        a = tl.load(a_ptrs,
                    mask=offs_k[None, :] < K - k * BLOCK_SIZE_K,
                    other=0.0)
        b = tl.load(b_ptrs,
                    mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
                    other=0.0)

        k_start = k * BLOCK_SIZE_K
        offs_ks = k_start // group_k
        a_s = tl.load(As_ptrs + offs_ks * stride_As_k)
        b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k)

        accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
        a_ptrs += BLOCK_SIZE_K * stride_ak
        b_ptrs += BLOCK_SIZE_K * stride_bk

    if C.dtype.element_ty == tl.bfloat16:
        c = accumulator.to(tl.bfloat16)
    elif C.dtype.element_ty == tl.float16:
        c = accumulator.to(tl.float16)
    else:
        c = accumulator.to(tl.float32)

    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
    tl.store(c_ptrs, c, mask=c_mask)

def apply_fp8_block_linear(layer: torch.nn.Module, input: torch.Tensor,
                           bias: Optional[torch.Tensor],
                           cutlass_block_fp8_supported: bool,
                           use_aiter_and_is_supported: bool) -> torch.Tensor:
    """Apply block-wise FP8 linear operation."""
    assert layer.weight_block_size is not None

    return torch.ops.vllm.apply_w8a8_block_fp8_linear(
        input=input,
        weight=layer.weight,
        block_size=layer.weight_block_size,
        weight_scale=layer.weight_scale,
        input_scale=layer.input_scale,
        bias=bias,
        cutlass_block_fp8_supported=cutlass_block_fp8_supported,
        use_aiter_and_is_supported=use_aiter_and_is_supported,
    )

def apply_w8a8_block_fp8_linear(
    input: torch.Tensor,
    weight: torch.Tensor,
    block_size: list[int],
    weight_scale: torch.Tensor,
    input_scale: Optional[torch.Tensor] = None,
    bias: Optional[torch.Tensor] = None,
    cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
    use_aiter_and_is_supported: bool = False,
) -> torch.Tensor:
    assert input_scale is None
    # View input as 2D matrix for fp8 methods
    input_2d = input.view(-1, input.shape[-1])
    output_shape = [*input.shape[:-1], weight.shape[0]]
    output_dtype = input.dtype

    if should_use_deepgemm_for_fp8_linear(output_dtype, weight):

        input_2d = input.view(-1, input.shape[-1])
        output_shape = [*input.shape[:-1], weight.shape[0]]

        q_input, x_scale = per_token_group_quant_fp8(
            input_2d,
            block_size[1],
            column_major_scales=True,
        )
        output = torch.empty((q_input.shape[0], weight.shape[0]),
                             dtype=torch.bfloat16,
                             device=q_input.device)
        fp8_gemm_nt((q_input, x_scale), (weight, weight_scale), output)
        if bias is not None:
            output += bias
        return output.to(dtype=output_dtype).view(*output_shape)

    w8a8_blockscale_func = dispatch_w8a8_blockscale_func(
        cutlass_block_fp8_supported, use_aiter_and_is_supported)
    if cutlass_block_fp8_supported:
        num_pad = 0
        if current_platform.is_device_capability(90):
            # pad first dimension to be divisible by 4 due to
            # cutlass blockwise gemm limitation for hopper
            num_pad = 4 - (input_2d.shape[0] % 4)
            if num_pad > 0:
                input_2d = torch.nn.functional.pad(input_2d,
                                                   (0, 0, 0, num_pad),
                                                   "constant", 0)
        q_input, x_scale = per_token_group_quant_fp8(input_2d,
                                                     block_size[1],
                                                     column_major_scales=True)
        output = w8a8_blockscale_func(q_input, weight, x_scale, weight_scale,
                                      block_size, input.dtype)
        if num_pad > 0:
            output = output[:-num_pad]
    else:
        if use_aiter_and_is_supported:
            q_input, x_scale = aiter_per1x128_quant(
                input_2d.contiguous(), quant_dtype=rocm_aiter.dtypes.fp8)
        else:
            q_input, x_scale = per_token_group_quant_fp8(
                input_2d, block_size[1], column_major_scales=False)

        output = w8a8_blockscale_func(q_input, weight, x_scale, weight_scale,
                                      block_size, input.dtype)

    if bias is not None:
        output = output + bias
    return output.to(dtype=input.dtype).view(*output_shape)

def apply_w8a8_block_fp8_linear_fake(
    input: torch.Tensor,
    weight: torch.Tensor,
    block_size: list[int],
    weight_scale: torch.Tensor,
    input_scale: Optional[torch.Tensor] = None,
    bias: Optional[torch.Tensor] = None,
    cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
    use_aiter_and_is_supported: bool = False,
) -> torch.Tensor:
    output_shape = [*input.shape[:-1], weight.shape[0]]
    return torch.empty(output_shape, dtype=input.dtype, device=input.device)

def block_quant_to_tensor_quant(
    x_q_block: torch.Tensor,
    x_s: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    """This function converts block-wise quantization to tensor-wise
    quantization. The inputs are block-wise quantization tensor `x_q_block`,
    block-wise quantization scale and the block size.
    The outputs are tensor-wise quantization tensor and tensor-wise
    quantization scale. Note only float8 is supported for now.
    """
    x_dq_block = group_broadcast(x_q_block, x_s)
    x_q_tensor, scale = input_to_float8(x_dq_block, dtype=x_q_block.dtype)
    return x_q_tensor, scale

def check_aiter_fp8_linear_support() -> bool:
    """AITER is only supported on ROCm and only for FP8_FNUZ
    and at the moment are MI300 series"""
    return (current_platform.is_rocm() and envs.VLLM_ROCM_USE_AITER
            and envs.VLLM_ROCM_USE_AITER_LINEAR
            and current_platform.is_fp8_fnuz())

def create_fp8_input_scale(
        output_partition_sizes: list[int],
        weight_loader: Optional[Callable]) -> torch.nn.Parameter:
    """Create input scale parameter for static activation quantization."""
    from vllm.model_executor.parameter import PerTensorScaleParameter

    scale = PerTensorScaleParameter(data=torch.empty(
        len(output_partition_sizes), dtype=torch.float32),
                                    weight_loader=weight_loader)
    scale[:] = torch.finfo(torch.float32).min
    return scale

def create_fp8_scale_parameter(
        parameter_type: torch.nn.Parameter, output_partition_sizes: list[int],
        input_size_per_partition: int, block_size: Optional[list[int]],
        weight_loader: Optional[Callable]) -> torch.nn.Parameter:
    """Create scale parameter based on quantization strategy."""
    if parameter_type == ChannelQuantScaleParameter:
        scale = parameter_type(data=torch.empty(
            (sum(output_partition_sizes), 1), dtype=torch.float32),
                               output_dim=0,
                               weight_loader=weight_loader)
    elif parameter_type == BlockQuantScaleParameter:
        assert block_size is not None
        block_n, block_k = block_size[0], block_size[1]
        output_size_per_partition = sum(output_partition_sizes)
        scale = parameter_type(
            data=torch.empty(
                (output_size_per_partition + block_n - 1) // block_n,
                (input_size_per_partition + block_k - 1) // block_k,
                dtype=torch.float32,
            ),
            input_dim=1,
            output_dim=0,
            weight_loader=weight_loader,
        )
    elif parameter_type == PerTensorScaleParameter:
        scale = parameter_type(data=torch.empty(len(output_partition_sizes),
                                                dtype=torch.float32),
                               weight_loader=weight_loader)
    else:
        raise ValueError(f"Unknown parameter type: {parameter_type}")

    scale[:] = torch.finfo(torch.float32).min
    return scale

def create_fp8_weight_parameter(
        output_size_per_partition: int, input_size_per_partition: int,
        weight_loader: Optional[Callable]) -> torch.nn.Parameter:
    """Create FP8 weight parameter."""
    from vllm.model_executor.parameter import ModelWeightParameter

    return ModelWeightParameter(data=torch.empty(output_size_per_partition,
                                                 input_size_per_partition,
                                                 dtype=torch.float8_e4m3fn),
                                input_dim=1,
                                output_dim=0,
                                weight_loader=weight_loader)

def cutlass_scaled_mm(
    A: torch.Tensor,
    B: torch.Tensor,
    As: torch.Tensor,
    Bs: torch.Tensor,
    block_size: list[int],
    output_dtype: torch.dtype = torch.float16,
) -> torch.Tensor:
    return ops.cutlass_scaled_mm(
        A,
        B.T,
        out_dtype=output_dtype,
        scale_a=As,
        # SM90 block FP8 requires row-major scale_b, which we do ahead of time
        scale_b=Bs if block_size is not None
        and current_platform.is_device_capability(90) else Bs.T)

def dispatch_w8a8_blockscale_func(
    use_cutlass: bool, use_aiter_and_is_supported: bool
) -> Callable[[
        torch.Tensor,
        torch.Tensor,
        torch.Tensor,
        torch.Tensor,
        list[int],
        torch.dtype,
], torch.Tensor]:
    if use_cutlass:
        return cutlass_scaled_mm
    if (use_aiter_and_is_supported):
        return torch.ops.vllm.rocm_aiter_gemm_w8a8_blockscale
    return w8a8_block_fp8_matmul

def expert_weight_is_col_major(x: torch.Tensor) -> bool:
    assert x.dim() == 3
    b, m, n = x.shape
    return x.stride(0) == m * n and x.stride(1) == 1 and x.stride(2) == m

@functools.lru_cache
def get_w8a8_block_fp8_configs(N: int, K: int, block_n: int,
                               block_k: int) -> Optional[dict[int, Any]]:
    """
    Return optimized configurations for the w8a8 block fp8 kernel.
    The return value will be a dictionary that maps an irregular grid of
    batch sizes to configurations of the w8a8 block fp8 kernel. To evaluate the
    kernel on a given batch size bs, the closest batch size in the grid should
    be picked and the associated configuration chosen to invoke the kernel.
    """

    # First look up if an optimized configuration is available in the configs
    # directory
    device_name = current_platform.get_device_name().replace(" ", "_")
    json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n},{block_k}].json"  # noqa: E501

    config_file_path = os.path.join(
        os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
    if os.path.exists(config_file_path):
        with open(config_file_path) as f:
            logger.info(
                "Using configuration from %s for W8A8 Block FP8 kernel.",
                config_file_path,
            )
            # If a configuration has been found, return it
            return {int(key): val for key, val in json.load(f).items()}

    # If no optimized configuration is available, we will use the default
    # configuration
    logger.warning(
        "Using default W8A8 Block FP8 kernel config. Performance might "
        "be sub-optimal! Config file not found at %s",
        config_file_path,
    )
    return None

def input_to_float8(
        x: torch.Tensor,
        dtype: Optional[torch.dtype] = None
) -> tuple[torch.Tensor, torch.Tensor]:
    """This function quantizes input values to float8 values "
    "with tensor-wise quantization."""
    dtype = current_platform.fp8_dtype() if dtype is None else dtype
    finfo = torch.finfo(dtype)
    min_val, max_val = x.aminmax()
    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
    scale = finfo.max / amax
    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
    return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()

def is_fp8(x: Union[torch.dtype, torch.Tensor]) -> bool:
    if isinstance(x, torch.Tensor):
        x = x.dtype
    return x == torch.float8_e4m3fn or x == torch.float8_e4m3fnuz

def maybe_post_process_fp8_weight_block(layer: torch.nn.Module,
                                        cutlass_block_fp8_supported: bool):
    assert layer.weight_block_size is not None

    from vllm.utils.deep_gemm import (is_deep_gemm_e8m0_used,
                                      should_use_deepgemm_for_fp8_linear)

    # On Blackwell or Hopper, if E8M0 for DeepGemm is used, we need to
    # requantize the weight and input to the specific scale
    # at the same time.
    should_use_deepgemm = should_use_deepgemm_for_fp8_linear(
        layer.orig_dtype, layer.weight)
    if is_deep_gemm_e8m0_used() and should_use_deepgemm:
        block_sz = tuple(layer.weight_block_size)
        requant_weight_ue8m0_inplace(layer.weight.data,
                                     layer.weight_scale.data, block_sz)
    # SM90 Block FP8 CUTLASS requires row-major weight scales
    elif (current_platform.is_device_capability(90)
          and cutlass_block_fp8_supported and not should_use_deepgemm):
        layer.weight_scale = torch.nn.Parameter(
            layer.weight_scale.data.T.contiguous(), requires_grad=False)

def per_token_group_quant_fp8(
    x: torch.Tensor,
    group_size: int,
    eps: float = 1e-10,
    dtype: Optional[torch.dtype] = None,
    column_major_scales: bool = False,
    out_q: Optional[torch.Tensor] = None,
    use_ue8m0: Optional[bool] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Function to perform per-token-group quantization on an input tensor `x`.
    It converts the tensor values into signed float8 values and returns the
    quantized tensor along with the scaling factor used for quantization.
    Args:
        x: The input tensor with ndim >= 2.
        group_size: The group size used for quantization.
        eps: The minimum to avoid dividing zero.
        dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
        is supported for now.
        column_major_scales: Outputs scales in column major.
        out_q: Optional output tensor. If not provided, function will create.
    Returns:
        tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
        scaling factor.
    """
    if use_ue8m0 is None:
        use_ue8m0 = is_deep_gemm_e8m0_used()
    dtype = current_platform.fp8_dtype() if dtype is None else dtype
    assert (x.shape[-1] % group_size == 0), (
        f"the last dimension of `x` {x.shape[-1]} must be divisible "
        f"by `group_size` {group_size}")
    assert x.stride(-1) == 1, "`x` groups must be contiguous"

    finfo = torch.finfo(dtype)
    fp8_min = finfo.min
    fp8_max = finfo.max

    assert out_q is None or out_q.shape == x.shape
    x_q = out_q
    if x_q is None:
        x_q = torch.empty_like(x, device=x.device, dtype=dtype)

    # Allocate the scale tensor in either row- or column-major format.
    if column_major_scales:
        shape = (x.shape[-1] // group_size, ) + x.shape[:-1]
        x_s = torch.empty(shape, device=x.device,
                          dtype=torch.float32).permute(-1, -2)
    else:
        shape = x.shape[:-1] + (x.shape[-1] // group_size, )
        x_s = torch.empty(shape, device=x.device, dtype=torch.float32)

    # prefer CUDA kernel if available
    # TODO(bnell): this causes some fp8 moe test to fail.
    if current_platform.is_cuda() and x.is_contiguous():
        torch.ops._C.per_token_group_fp8_quant(x, x_q, x_s, group_size, eps,
                                               fp8_min, fp8_max, use_ue8m0)
        return x_q, x_s

    # TRITON FALLBACK
    M = x.numel() // group_size
    N = group_size
    BLOCK = triton.next_power_of_2(N)
    # heuristics for number of warps
    num_warps = min(max(BLOCK // 256, 1), 8)
    num_stages = 1
    if column_major_scales:
        _per_token_group_quant_fp8_colmajor[(M, )](
            x,
            x_q,
            x_s,
            group_size,
            x.shape[1],
            x.stride(0),
            x_s.stride(1),
            eps,
            fp8_min=fp8_min,
            fp8_max=fp8_max,
            use_ue8m0=use_ue8m0,
            BLOCK=BLOCK,
            num_warps=num_warps,
            num_stages=num_stages,
        )
    else:
        _per_token_group_quant_fp8[(M, )](
            x,
            x_q,
            x_s,
            group_size,
            x.shape[1],
            x.stride(0),
            eps,
            fp8_min=fp8_min,
            fp8_max=fp8_max,
            use_ue8m0=use_ue8m0,
            BLOCK=BLOCK,
            num_warps=num_warps,
            num_stages=num_stages,
        )

    return x_q, x_s

def process_fp8_weight_block_strategy(
    weight: torch.Tensor,
    weight_scale: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Process weights for block-wise quantization strategy."""
    from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
        normalize_e4m3fn_to_e4m3fnuz)

    if current_platform.is_fp8_fnuz():
        weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
            weight=weight, weight_scale=weight_scale)

    weight = _maybe_pad_fp8_weight(weight)
    return weight, weight_scale

def process_fp8_weight_channel_strategy(
    weight: torch.Tensor,
    weight_scale: torch.Tensor,
    input_scale: Optional[torch.Tensor] = None
) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
    """Process weights for channel-wise quantization strategy."""
    from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
        normalize_e4m3fn_to_e4m3fnuz)

    if current_platform.is_fp8_fnuz():
        weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
            weight=weight, weight_scale=weight_scale, input_scale=input_scale)

    return weight, weight_scale, input_scale

def process_fp8_weight_tensor_strategy(
    weight: torch.Tensor,
    weight_scale: torch.Tensor,
    logical_widths: list[int],
    input_scale: Optional[torch.Tensor] = None
) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
    """Process weights for tensor-wise quantization strategy."""
    from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
        normalize_e4m3fn_to_e4m3fnuz, requantize_with_max_scale)

    if current_platform.is_fp8_fnuz():
        weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
            weight=weight, weight_scale=weight_scale, input_scale=input_scale)

    # Requantize with max scale
    weight_scale, weight = requantize_with_max_scale(
        weight=weight,
        weight_scale=weight_scale,
        logical_widths=logical_widths,
    )

    weight = _maybe_pad_fp8_weight(weight)
    return weight, weight_scale, input_scale