Datasets:
File size: 4,117 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 | import os
import torch
import pytest
import tile_kernels
from tile_kernels.testing.bench import dtype_to_str, make_param_id
from tile_kernels.testing.generator import generate_topk_idx, generate_hidden_sizes, generate_moe_params
from tile_kernels.testing.numeric import assert_equal, count_bytes
# Disable TileLang prints
os.environ['TILELANG_PRINT_ON_COMPILATION'] = '0'
def generate_test_data(params):
hidden = params['hidden']
with_weights = params['with_weights']
in_dtype = params['in_dtype']
out_dtype = params['out_dtype']
with_sf = params['with_sf']
num_experts = params['num_experts']
num_ep_ranks = params['num_ep_ranks']
num_topk = params['num_topk']
topk_idx = generate_topk_idx(params)
num_tokens = topk_idx.shape[0]
num_expanded_tokens = num_tokens * num_topk
expanded = torch.randn((num_expanded_tokens, hidden), dtype=in_dtype, device='cuda')
_, _, _, token_topk_to_pos, _, _, _, _ = tile_kernels.moe.get_fused_mapping(topk_idx, num_experts, 0, 1)
topk_weights = torch.rand((num_tokens, num_topk), dtype=torch.float32, device='cuda') if with_weights else None
if out_dtype == torch.float8_e4m3fn:
sf = torch.randn((1,), dtype=torch.float32, device='cuda')
else:
sf = None
if with_sf:
x_sf = torch.randn((num_expanded_tokens,), dtype=torch.float32, device='cuda')
else:
x_sf = None
fp8_format = 'e4m3' if out_dtype == torch.float8_e4m3fn else ''
x_input = (expanded, x_sf) if x_sf is not None else expanded
return (expanded, token_topk_to_pos, topk_weights, sf, x_sf, fp8_format, x_input, num_tokens)
def generate_test_params(is_benchmark: bool) -> list[dict]:
params = [
{**moe, 'hidden': hidden, 'with_weights': with_weights,
'in_dtype': in_dtype, 'out_dtype': out_dtype, 'with_sf': with_sf}
for moe in generate_moe_params(is_benchmark=is_benchmark)
for hidden in generate_hidden_sizes(256)
for with_weights in (True, False)
for in_dtype in (torch.float32, torch.bfloat16)
for out_dtype in (in_dtype, torch.float8_e4m3fn)
for with_sf in (True, False)
]
if is_benchmark:
params = [p for p in params if p['num_topk'] == 6 and p['with_weights']]
return params
@pytest.mark.parametrize('params', generate_test_params(is_benchmark=False), ids=make_param_id)
def test_reduce_fused(params):
(expanded, token_topk_to_pos, topk_weights, sf, x_sf, fp8_format, x_input,
_) = generate_test_data(params)
# Test correctness: tile_kernels kernel
func = lambda: tile_kernels.moe.reduce_fused(
x_input, topk_weights, token_topk_to_pos, fp8_format, sf, None
)
r_tk = func()
# Test correctness: torch reference
r_ref = tile_kernels.torch.reduce_fused(
x_input, topk_weights, token_topk_to_pos, fp8_format, sf
)
assert_equal(r_tk, r_ref)
@pytest.mark.benchmark
@pytest.mark.parametrize('params', generate_test_params(is_benchmark=True), ids=make_param_id)
def test_reduce_fused_benchmark(benchmark_timer, benchmark_record, params):
hidden = params['hidden']
out_dtype = params['out_dtype']
(expanded, token_topk_to_pos, topk_weights, sf, x_sf, fp8_format, x_input,
num_tokens) = generate_test_data(params)
in_dtype = params['in_dtype']
func = lambda: tile_kernels.moe.reduce_fused(
x_input, topk_weights, token_topk_to_pos, fp8_format, sf, None
)
r_tk = func()
num_bytes = count_bytes(token_topk_to_pos, x_sf, r_tk)
num_bytes += torch.count_nonzero(token_topk_to_pos != -1).item() * hidden * (torch.finfo(in_dtype).bits // 8)
if topk_weights is not None:
num_bytes += count_bytes(topk_weights)
t_us = benchmark_timer(func)
bandwidth_gbs = num_bytes / t_us / 1e3
params.pop('num_send_tokens')
benchmark_record(
kernel='reduce_fused',
operation='fwd',
params={'num_tokens': num_tokens, **params, 'in_dtype': dtype_to_str(in_dtype), 'out_dtype': dtype_to_str(out_dtype)},
time_us=t_us,
bandwidth_gbs=bandwidth_gbs,
)
|