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.accelerator.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: tl.constexpr,
    fp8_max: tl.constexpr,
    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
    # Use multiply-by-reciprocal instead of division to match PyTorch's
    # tensor/scalar division precision (GPU fast-division for constexpr
    # divisors can introduce 1-ULP error that flips FP8 quantization at
    # representable-value boundaries).
    _absmax = tl.maximum(tl.max(tl.abs(y)), eps)
    scale_raw = _absmax * (1.0 / 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: tl.constexpr,
    fp8_max: tl.constexpr,
    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 * (1.0 / 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 _silu_mul_per_token_group_quant_fp8_colmajor(
    y_ptr,  # [M, N]
    y_q_ptr,  # [M, N // 2]
    y_s_ptr,  # [M, (N // 2) // GROUP_SIZE]
    M,  # num tokens
    N,  # intermediate size
    # Stride
    y_s_col_stride: tl.int64,
    # Information for float8
    eps,
    clamp_limit,
    alpha,
    beta,
    fp8_min: tl.constexpr,
    fp8_max: tl.constexpr,
    use_ue8m0: tl.constexpr,
    HAS_CLAMP: tl.constexpr,
    # Meta-parameters
    GROUP_SIZE: tl.constexpr,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
):
    # TODO(varun) : Add expert_ids so we may early-exit no-op thread blocks.
    """
    Each thread block (BLOCK_N) computes [BLOCK_M, GROUP_SIZE] act-mul outputs. Then
    the thread block quantizes the [BLOCK_M, GROUP_SIZE] block of values and fills
    the outputs tensors at the right positions.
    """

    pid_m = tl.program_id(0)
    pid_n = tl.program_id(1)
    N_2 = N // 2

    m_offset = pid_m.to(tl.int64) * BLOCK_M
    n_offset = pid_n.to(tl.int64) * BLOCK_N
    if m_offset >= M:
        return

    offs_n = tl.arange(0, BLOCK_N).to(tl.int64)
    offs_m = tl.arange(0, BLOCK_M).to(tl.int64)

    base_y_ptr = y_ptr + m_offset * N + n_offset

    act_in_ptrs = base_y_ptr + offs_m[:, None] * N + offs_n[None, :]

    act_in = tl.load(act_in_ptrs)
    mul_in = tl.load(act_in_ptrs + N_2)

    # silu & mul — match C++ silu_and_mul: clamp in fp32 then store back to the
    # input dtype, run silu in fp32 then narrow, and do the mul at input
    # precision so HAS_CLAMP True/False share the same multiplication path.
    if HAS_CLAMP:
        act_in = tl.minimum(act_in.to(tl.float32), clamp_limit).to(
            y_ptr.dtype.element_ty
        )
        mul_in = tl.clamp(mul_in.to(tl.float32), -clamp_limit, clamp_limit).to(
            y_ptr.dtype.element_ty
        )
    # Unified gated activation: silu == swigluoai with alpha=1, beta=0.
    #   glu = gate * sigmoid(alpha * gate); y = (up + beta) * glu
    # Keep glu/up at input precision (narrow before the mul) so the alpha=1,
    # beta=0 defaults match the C++ silu_and_mul path bit-for-bit.
    act_in = act_in.to(tl.float32)
    glu = (act_in / (1.0 + tl.exp(-act_in * alpha))).to(y_ptr.dtype.element_ty)
    up = (mul_in.to(tl.float32) + beta).to(y_ptr.dtype.element_ty)
    y = (glu * up).to(tl.float32)

    # quant
    _absmax = tl.maximum(tl.max(tl.abs(y), axis=1), eps)
    scale_raw = _absmax * (1.0 / fp8_max)
    y_s = tl.math.exp2(tl.ceil(tl.log2(scale_raw))) if use_ue8m0 else scale_raw
    y_s = tl.reshape(y_s, (BLOCK_M, 1))
    y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)

    # store y_q
    base_y_q_ptr = y_q_ptr + m_offset * N_2 + n_offset
    y_q_ptrs = base_y_q_ptr + offs_m[:, None] * N_2 + offs_n[None, :]
    tl.store(y_q_ptrs, y_q)

    # store y_s
    group_id = n_offset // GROUP_SIZE
    base_y_s_ptr = y_s_ptr + group_id * y_s_col_stride + m_offset
    y_s_ptrs = base_y_s_ptr + offs_m
    y_s = tl.reshape(y_s, (BLOCK_M,))
    tl.store(y_s_ptrs, y_s)

def _upcast_e8m0_to_fp32(scale: torch.Tensor) -> torch.Tensor:
    """Upcast E8M0 (exponent-only) scale to float32.

    E8M0 stores only the 8-bit biased exponent (bias=127). To convert
    to float32 we place those 8 bits into the exponent field of an
    IEEE-754 float32 (bits 23-30) with sign=0 and mantissa=0.
    """
    exp_bits = scale.view(torch.uint8).to(torch.int32)
    fp32_bits = exp_bits << 23
    return fp32_bits.view(torch.float32)

@triton.jit
def _w8a8_triton_block_scaled_mm(
    # 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 create_fp8_input_scale(
    output_partition_sizes: list[int], weight_loader: Callable | None
) -> 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: list[int] | None,
    weight_loader: Callable | None,
    scale_dtype: torch.dtype | None = None,
) -> torch.nn.Parameter:
    """Create scale parameter based on quantization strategy."""
    dtype = scale_dtype if scale_dtype is not None else torch.float32
    if parameter_type == ChannelQuantScaleParameter:
        scale = parameter_type(
            data=torch.empty((sum(output_partition_sizes), 1), dtype=dtype),
            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=dtype,
            ),
            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=dtype),
            weight_loader=weight_loader,
        )
    else:
        raise ValueError(f"Unknown parameter type: {parameter_type}")

    if dtype == torch.float32:
        scale[:] = torch.finfo(torch.float32).min
    set_weight_attrs(scale, {"scale_type": "weight_scale"})
    return scale

def create_fp8_weight_parameter(
    output_size_per_partition: int,
    input_size_per_partition: int,
    weight_loader: Callable | None,
) -> 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,
    )

@functools.lru_cache
def get_w8a8_block_fp8_configs(
    N: int, K: int, block_n: int, block_k: int
) -> dict[int, Any] | None:
    """
    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 = get_device_name_as_file_name()
    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: torch.dtype | None = 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 per_token_group_quant_fp8(
    x: torch.Tensor,
    group_size: int,
    eps: float = 1e-10,
    dtype: torch.dtype | None = None,
    column_major_scales: bool = False,
    tma_aligned_scales: bool = False,
    out_q: torch.Tensor | None = None,
    use_ue8m0: bool | None = 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 dtype of output tensor. Note that only `torch.float8_e4m3fn`
        is supported for now.
        column_major_scales: Outputs scales in column major.
        tma_aligned_scales: Outputs scales in TMA-aligned layout.
        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"

    fp8_min, fp8_max = get_fp8_min_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(x.shape, device=x.device, dtype=dtype)

    # Allocate the scale tensor in either row- or column-major format.
    if column_major_scales:
        if tma_aligned_scales:
            m = x.shape[-2]
            sf_k = x.shape[-1] // group_size
            tma_aligned_m = get_tma_aligned_size(m, 4)
            shape = x.shape[:-2] + (m, sf_k)
            stride = (
                (1, tma_aligned_m)
                if x.dim() == 2
                else (tma_aligned_m * sf_k, 1, tma_aligned_m)
            )
            x_s = torch.empty_strided(
                shape, stride, device=x.device, dtype=torch.float32
            )
        else:
            shape = x.shape[:-2] + (x.shape[-1] // group_size, x.shape[-2])
            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/XPU kernel if available
    # TODO(bnell): this causes some fp8 moe test to fail.
    if (
        current_platform.is_cuda_alike() or current_platform.is_xpu()
    ) 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,
            column_major_scales,
            tma_aligned_scales,
        )
        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 per_token_group_quant_fp8_packed_for_deepgemm(
    x: torch.Tensor,
    group_size: int,
    eps: float = 1e-10,
    use_ue8m0: bool | None = None,
    out_q: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    """FP8 per-token-group quantization for DeepGEMM.

    Returns:
        (x_q, x_s_packed)
            x_q: FP8 activations, same shape as `x`.
            x_s_packed: Int32 tensor with logical shape
                        [mn, ceil(num_groups_per_row / 4)], laid out with
                        TMA-aligned stride along the packed-K dimension
    """
    if use_ue8m0 is None:
        use_ue8m0 = is_deep_gemm_e8m0_used()
    # for DeepGEMM UE8M0-packed layout we *require* UE8M0 scales.
    assert use_ue8m0, (
        "per_token_group_quant_fp8_packed_for_deepgemm requires UE8M0 scales."
    )

    dtype = current_platform.fp8_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"

    fp8_min, fp8_max = get_fp8_min_max()

    # compute DeepGEMM-style packed scale tensor shape.
    hidden_dim = x.shape[-1]
    mn = x.numel() // hidden_dim
    num_groups_per_row = hidden_dim // group_size
    k_num_packed_sf_k = (num_groups_per_row + 3) // 4
    tma_aligned_mn = ((mn + 3) // 4) * 4

    x_s_packed = torch.empty_strided(
        (mn, k_num_packed_sf_k),
        (1, tma_aligned_mn),
        device=x.device,
        dtype=torch.int32,
    )

    # Native kernel (libtorch stable); used with DeepGEMM on CUDA and
    # available on ROCm for the same packed UE8M0 scale layout.
    assert current_platform.is_cuda_alike(), (
        "per_token_group_quant_fp8_packed_for_deepgemm requires a CUDA or ROCm GPU."
    )

    x_contiguous = x.contiguous()
    if out_q is not None:
        x_q_local = out_q
    else:
        x_q_local = torch.empty_like(x_contiguous, device=x.device, dtype=dtype)

    torch.ops._C.per_token_group_fp8_quant_packed(
        x_contiguous,
        x_q_local,
        x_s_packed,
        group_size,
        eps,
        fp8_min,
        fp8_max,
    )

    # return a tensor with the original logical shape.
    x_q = x_q_local.view_as(x)
    return x_q, x_s_packed

def process_fp8_input_tensor_strategy_moe(
    w13_input_scale: torch.Tensor,
    w2_input_scale: torch.Tensor,
    enable_eplb: bool,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Process moe input scales for tensor-wise quantization strategy."""

    if not all_close_1d(w13_input_scale) or not all_close_1d(w2_input_scale):
        logger.info_once(
            "Found input_scales that are not equal for "
            "fp8 MoE layer. Using the maximum across experts "
            "for each layer."
        )

    return (
        amax_for_moe_activation_quant(w13_input_scale, enable_eplb),
        amax_for_moe_activation_quant(w2_input_scale, enable_eplb),
    )

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() and weight.dtype == torch.float8_e4m3fn:
        if weight_scale.dtype == torch.float8_e8m0fnu:
            # UE8M0 scales: e8m0 stores exponent-only values (2^(exp-127)),
            # so doubling the dequant scale == incrementing the exponent byte
            # by 1. Convert the OCP E4M3 weight bytes to FNUZ in place by
            # reinterpreting and patching the NaN sentinel (-128 in int8),
            # then double the UE8M0 exponent so the dequantized magnitudes
            # match.
            weight_as_int8 = weight.view(torch.int8)
            ROCM_FP8_NAN_AS_INT = -128
            weight_as_int8[weight_as_int8 == ROCM_FP8_NAN_AS_INT] = 0
            weight = weight_as_int8.view(torch.float8_e4m3fnuz)
            exp_bytes = weight_scale.view(torch.uint8)
            weight_scale = (
                (exp_bytes.to(torch.int16) + 1)
                .clamp(max=254)
                .to(torch.uint8)
                .view(torch.float8_e8m0fnu)
            )
        else:
            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: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]:
    """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() and weight.dtype == torch.float8_e4m3fn:
        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: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]:
    """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() and weight.dtype == torch.float8_e4m3fn:
        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

def process_fp8_weight_tensor_strategy_moe(
    weight: torch.Tensor,
    weight_scales: torch.Tensor,
    shard_size: int,
    num_experts: int,
    is_act_and_mul: bool = True,
) -> tuple[torch.Tensor, torch.Tensor]:
    """Process moe weights for tensor-wise quantization strategy."""
    max_scales = weight_scales.max(dim=1).values

    # For w1 case (i.e. not w13): there is already just one scale per expert.
    if not is_act_and_mul:
        assert weight_scales.shape[1] == 1
        # One scale per expert
        assert max_scales.shape == (num_experts,)
        return weight, max_scales

    # For w13 case (common): require single scale for w13 per expert, but
    # on disk there is a scale for w1 and w3. Use the max to requantize.
    for expert_id in range(num_experts):
        start = 0
        for shard_id in range(2):
            dq_weight = per_tensor_dequantize(
                weight[expert_id][start : start + shard_size, :],
                weight_scales[expert_id][shard_id],
            )
            weight[expert_id][start : start + shard_size, :], _ = ops.scaled_fp8_quant(
                dq_weight, max_scales[expert_id]
            )
            start += shard_size
    return weight, max_scales