level1_81_conv_transposed_2D_asymmetric_input_square_kernel___dilated____padded____strided__ 执行错误 #94
Description
报错： Triton 运行失败: 'tuple' object has no attribute 'dtype'
执行代码：
import torch
import torch.nn as nn
import triton
import triton.language as tl
@triton.autotune(configs=[triton.Config({'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N':
16, 'BLOCK_SIZE_K': 8}, num_warps=4, num_stages=4)], key=['M', 'N', 'K'])
@triton.jit
def _matmul_kernel(A_ptr, B_ptr, C_ptr, M, N, K, stride_am, stride_ak,
stride_bk, stride_bn, stride_cm, stride_cn, BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr):
pid = tl.program_id(0)
pid_m = pid // tl.cdiv(N, BLOCK_SIZE_N)
pid_n = pid % tl.cdiv(N, BLOCK_SIZE_N)
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
offs_k = tl.arange(0, BLOCK_SIZE_K)
A_ptr += offs_m[:, None] * stride_am + offs_k[None, :] * stride_ak
B_ptr += offs_k[:, None] * stride_bk + offs_n[None, :] * stride_bn
acc = 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_ptr, mask=(offs_m[:, None] < M) & (offs_k[None, :] <
K - k * BLOCK_SIZE_K), other=0.0)
b = tl.load(B_ptr, mask=(offs_k[:, None] < K - k * BLOCK_SIZE_K) &
(offs_n[None, :] < N), other=0.0)
acc += tl.dot(a, b, allow_tf32=False)
A_ptr += BLOCK_SIZE_K * stride_ak
B_ptr += BLOCK_SIZE_K * stride_bk
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_ptr += offs_cm[:, None] * stride_cm + offs_cn[None, :] * stride_cn
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(C_ptr, acc, mask=c_mask)
class ModelNew(nn.Module):
def __init__(self):
super(ModelNew, self).__init__()
def forward(self, A, B):
n, m, k = A.shape
l = B.shape[1]
A_flat = A.view(-1, k)
C_flat = torch.empty((n * m, l), device=A.device, dtype=A.dtype)
grid = lambda meta: (triton.cdiv(n * m, meta['BLOCK_SIZE_M']) *
triton.cdiv(l, meta['BLOCK_SIZE_N']),)
_matmul_kernel[grid](A_flat, B, C_flat, n * m, l, k, A_flat.stride(
0), A_flat.stride(1), B.stride(0), B.stride(1), C_flat.stride(0
), C_flat.stride(1))
return C_flat.view(n, m, l)
N = 16
M = 1024
K = 2048
L = 768
def get_inputs():
A = torch.randn(N, M, K, device='npu')
B = torch.randn(K, L, device='npu')
return [A, B]
def get_init_inputs():
return []