Datasets:
File size: 4,064 Bytes
167e081 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 | import os
import pytest
import torch
import tile_kernels
from tile_kernels.testing.numeric import assert_equal, count_bytes
from tile_kernels.testing.bench import dtype_to_str, make_param_id
from tile_kernels.testing.generator import generate_hidden_sizes, generate_num_tokens
# Disable TileLang prints
os.environ['TILELANG_PRINT_ON_COMPILATION'] = '0'
def twice_stride(w):
# Make a 2D tensor's leading dim twice strided
twice_w = w.new_empty((w.shape[0], w.shape[1] * 2))
ret = torch.chunk(twice_w, 2, dim=1)[0]
ret[:] = w
assert not ret.is_contiguous()
return ret
def generate_test_data_transpose(params):
num_tokens = params['num_tokens']
hidden = params['hidden']
dtype = params['dtype']
x = torch.randn((num_tokens, hidden), dtype=torch.bfloat16, device='cuda')
if dtype == torch.float8_e4m3fn:
x = x.to(torch.float8_e4m3fn)
if num_tokens > 0:
x = twice_stride(x)
return (x,)
def generate_test_data_batched_transpose(params):
num_tokens = params['num_tokens']
hidden = params['hidden']
num_experts = params['num_experts']
dtype = params['dtype']
x = torch.randn((num_experts, num_tokens, hidden), dtype=torch.bfloat16, device='cuda')
if dtype == torch.float8_e4m3fn:
x = x.to(torch.float8_e4m3fn)
return (x,)
def generate_test_params_transpose(is_benchmark: bool) -> list[dict]:
return [
{'num_tokens': t, 'hidden': hidden_size, 'dtype': dtype}
for t in generate_num_tokens(64, is_benchmark=is_benchmark)
for hidden_size in generate_hidden_sizes()
for dtype in (torch.float8_e4m3fn, torch.bfloat16)
]
def generate_test_params_batched_transpose(is_benchmark: bool) -> list[dict]:
return [
{'num_tokens': num_tokens, 'hidden': hidden_size, 'num_experts': num_experts, 'dtype': dtype}
for num_tokens in generate_num_tokens(64, is_benchmark=is_benchmark)
for hidden_size in generate_hidden_sizes()
for num_experts in (8, 32)
for dtype in (torch.float8_e4m3fn, torch.bfloat16, torch.float32)
]
@pytest.mark.parametrize('params', generate_test_params_transpose(is_benchmark=False), ids=make_param_id)
def test_transpose(params):
num_tokens = params['num_tokens']
(x,) = generate_test_data_transpose(params)
y = tile_kernels.transpose.transpose(x)
if num_tokens == 0:
return
y_ref = x.T.contiguous()
assert_equal(y, y_ref)
@pytest.mark.benchmark
@pytest.mark.parametrize('params', generate_test_params_transpose(is_benchmark=True), ids=make_param_id)
def test_transpose_benchmark(benchmark_timer, benchmark_record, params):
(x,) = generate_test_data_transpose(params)
num_bytes = count_bytes(x, tile_kernels.transpose.transpose(x))
t_us = benchmark_timer(lambda: tile_kernels.transpose.transpose(x))
benchmark_record(
kernel='transpose',
operation='fwd',
params={**params, 'dtype': dtype_to_str(params['dtype'])},
time_us=t_us,
bandwidth_gbs=num_bytes / t_us / 1e3,
)
@pytest.mark.parametrize('params', generate_test_params_batched_transpose(is_benchmark=False), ids=make_param_id)
def test_batched_transpose(params):
(x,) = generate_test_data_batched_transpose(params)
y = tile_kernels.transpose.batched_transpose(x)
y_ref = torch.transpose(x, 1, 2).contiguous()
assert_equal(y, y_ref)
@pytest.mark.benchmark
@pytest.mark.parametrize('params', generate_test_params_batched_transpose(is_benchmark=True), ids=make_param_id)
def test_batched_transpose_benchmark(benchmark_timer, benchmark_record, params):
(x,) = generate_test_data_batched_transpose(params)
num_bytes = count_bytes(x, tile_kernels.transpose.batched_transpose(x))
t_us = benchmark_timer(lambda: tile_kernels.transpose.batched_transpose(x))
benchmark_record(
kernel='batched_transpose',
operation='fwd',
params={**params, 'dtype': dtype_to_str(params['dtype'])},
time_us=t_us,
bandwidth_gbs=num_bytes / t_us / 1e3,
)
|