Function dequant_kernel_dim0

quantize/triton_utils/kernels.py:83–123

dequant the quantized tensor to fp tensor B is of shape (M/(32//bits), N) int32 C is of shape (M, N) float16

(
    b_ptr, c_ptr,
    M, N,
    bits, maxq,
    stride_bk, stride_bn,
    stride_cm, stride_cn,
    BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr
)

)
@triton.jit
def dequant_kernel_dim0(
    b_ptr, c_ptr,
    M, N,
    bits, maxq,
    stride_bk, stride_bn,
    stride_cm, stride_cn,
    BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr
):
    """
    dequant the quantized tensor to fp tensor
    B is of shape (M/(32//bits), N) int32
    C is of shape (M, N) float16
    """

    bits_per_feature = 32 // bits

    pid = tl.program_id(axis=0)
    num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)

    pid_m = pid // num_pid_n
    pid_n = pid % num_pid_n

    offs_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    offs_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)


    b_ptrs = b_ptr + ((offs_am[:, None] // bits_per_feature) * stride_bk + offs_bn[None, :] * stride_bn)

    shifter = (offs_am[:, None] % bits_per_feature) * bits



    b = tl.load(b_ptrs)
    b = (b >> shifter) & maxq
  
    c = b

    c_ptrs = c_ptr + stride_cm * offs_am[:, None] + stride_cn * offs_bn[None, :]
    c_mask = (offs_am[:, None] < M) & (offs_bn[None, :] < N)
    tl.store(c_ptrs, c, mask=c_mask)

@custom_autotune.autotune(
    configs=[