mirror of
https://github.com/deepseek-ai/DeepGEMM
synced 2025-04-09 16:14:03 +00:00
156 lines
7.5 KiB
Python
156 lines
7.5 KiB
Python
import pytest
|
|
import random
|
|
import torch
|
|
from typing import Tuple
|
|
|
|
import deep_gemm
|
|
from deep_gemm import bench_kineto, calc_diff, ceil_div, get_col_major_tma_aligned_tensor
|
|
|
|
# Set seeds and TF32 flags up front
|
|
torch.backends.cuda.matmul.allow_tf32 = True
|
|
torch.backends.cudnn.allow_tf32 = True
|
|
torch.manual_seed(0)
|
|
random.seed(0)
|
|
|
|
def per_token_cast_to_fp8(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
assert x.dim() == 2 and x.size(1) % 128 == 0
|
|
m, n = x.shape
|
|
x_view = x.view(m, -1, 128)
|
|
x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4)
|
|
return (x_view * (448.0 / x_amax.unsqueeze(2))).to(torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1)
|
|
|
|
def per_block_cast_to_fp8(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
assert x.dim() == 2
|
|
m, n = x.shape
|
|
x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128), dtype=x.dtype, device=x.device)
|
|
x_padded[:m, :n] = x
|
|
x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
|
|
x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
|
|
x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
|
|
return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
|
|
|
|
def construct(m: int, k: int, n: int) -> \
|
|
Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor], torch.Tensor, torch.Tensor]:
|
|
x = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
|
|
y = torch.randn((n, k), device='cuda', dtype=torch.bfloat16)
|
|
out = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
|
|
ref_out = x @ y.t()
|
|
|
|
x_fp8, y_fp8 = per_token_cast_to_fp8(x), per_block_cast_to_fp8(y)
|
|
# Transpose earlier so that the testing will not trigger transposing kernels
|
|
x_fp8 = (x_fp8[0], get_col_major_tma_aligned_tensor(x_fp8[1]))
|
|
return x_fp8, y_fp8, out, ref_out
|
|
|
|
def construct_grouped(num_groups: int, m: int, k: int, n: int, is_masked: bool) -> \
|
|
Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor], torch.Tensor, torch.Tensor]:
|
|
x = torch.randn((num_groups, m, k), device='cuda', dtype=torch.bfloat16)
|
|
y = torch.randn((num_groups, n, k), device='cuda', dtype=torch.bfloat16)
|
|
out = torch.empty((num_groups, m, n), device='cuda', dtype=torch.bfloat16)
|
|
ref_out = torch.einsum('gmk,gnk->gmn', x, y)
|
|
|
|
assert m % 4 == 0, f'TMA alignment error: {m}'
|
|
x_fp8 = (
|
|
torch.empty_like(x, dtype=torch.float8_e4m3fn),
|
|
torch.empty((num_groups, m, k // 128), device='cuda', dtype=torch.float)
|
|
)
|
|
y_fp8 = (
|
|
torch.empty_like(y, dtype=torch.float8_e4m3fn),
|
|
torch.empty((num_groups, (n + 127) // 128, k // 128), device='cuda', dtype=torch.float)
|
|
)
|
|
for i in range(num_groups):
|
|
x_fp8[0][i], x_fp8[1][i] = per_token_cast_to_fp8(x[i])
|
|
y_fp8[0][i], y_fp8[1][i] = per_block_cast_to_fp8(y[i])
|
|
|
|
# For non-masked input, we must merge the group and M dims
|
|
if not is_masked:
|
|
x_fp8 = (x_fp8[0].view(-1, k), per_token_cast_to_fp8(x.view(-1, k))[1])
|
|
out, ref_out = out.view(-1, n), ref_out.view(-1, n)
|
|
|
|
# Transpose earlier so that the testing will not trigger transposing kernels
|
|
x_fp8 = (x_fp8[0], get_col_major_tma_aligned_tensor(x_fp8[1]))
|
|
return x_fp8, y_fp8, out, ref_out
|
|
|
|
@pytest.mark.parametrize("m", [64, 128, 4096])
|
|
@pytest.mark.parametrize("k,n", [(7168, 2112), (1536, 24576), (512, 32768), (16384, 7168), (7168, 4096), (2048, 7168)])
|
|
def test_gemm(m, k, n):
|
|
"""Test single GEMM with various dimensions."""
|
|
x_fp8, y_fp8, out, ref_out = construct(m, k, n)
|
|
deep_gemm.gemm_fp8_fp8_bf16_nt(x_fp8, y_fp8, out)
|
|
diff = calc_diff(out, ref_out)
|
|
assert diff < 0.001, f'GEMM mismatch: m={m}, k={k}, n={n}, diff={diff:.5f}'
|
|
|
|
# (Optional) performance timing
|
|
def test_func():
|
|
# Construct new tensors every time to avoid L2 cache effects
|
|
x_fp8_local, y_fp8_local, out_local, _ = construct(m, k, n)
|
|
deep_gemm.gemm_fp8_fp8_bf16_nt(x_fp8_local, y_fp8_local, out_local)
|
|
|
|
t = bench_kineto(test_func, 'fp8_gemm', suppress_kineto_output=True)
|
|
print(f'[test_gemm] (m={m:5}, n={n:5}, k={k:5}): {t * 1e6:6.0f} us | '
|
|
f'{2 * m * n * k / t / 1e12:5.2f} TFLOPS')
|
|
|
|
@pytest.mark.parametrize("num_groups,m,k,n", [
|
|
(4, 8192, 7168, 4096),
|
|
(4, 8192, 2048, 7168),
|
|
(8, 4096, 7168, 4096),
|
|
(8, 4096, 2048, 7168)
|
|
])
|
|
def test_m_grouped_gemm_contiguous(num_groups, m, k, n):
|
|
"""Test grouped GEMM with merged group+M dimension."""
|
|
x_fp8, y_fp8, out, ref_out = construct_grouped(num_groups, m, k, n, is_masked=False)
|
|
m_indices = torch.arange(0, num_groups, device='cuda', dtype=torch.int)
|
|
m_indices = m_indices.unsqueeze(-1).expand(num_groups, m).contiguous().view(-1)
|
|
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(x_fp8, y_fp8, out, m_indices)
|
|
diff = calc_diff(out, ref_out)
|
|
assert diff < 0.001, f'Grouped GEMM mismatch: m={m*num_groups}, k={k}, n={n}, diff={diff:.5f}'
|
|
|
|
# (Optional) performance timing
|
|
def test_func():
|
|
x_fp8_local, y_fp8_local, out_local, _ = construct_grouped(num_groups, m, k, n, is_masked=False)
|
|
m_indices_local = torch.arange(0, num_groups, device='cuda', dtype=torch.int)
|
|
m_indices_local = m_indices_local.unsqueeze(-1).expand(num_groups, m).contiguous().view(-1)
|
|
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(x_fp8_local, y_fp8_local, out_local, m_indices_local)
|
|
|
|
t = bench_kineto(test_func, 'fp8_gemm', suppress_kineto_output=True)
|
|
print(f'[test_m_grouped_gemm_contiguous] (groups={num_groups}, m={m}, n={n}, k={k}): '
|
|
f'{t * 1e6:6.0f} us | {2 * num_groups * m * n * k / t / 1e12:5.2f} TFLOPS')
|
|
|
|
@pytest.mark.parametrize("num_groups,m", [
|
|
(1, 1024),
|
|
(2, 512),
|
|
(4, 256),
|
|
])
|
|
@pytest.mark.parametrize("k,n", [
|
|
(7168, 4096),
|
|
(2048, 7168),
|
|
])
|
|
def test_m_grouped_gemm_masked(num_groups, m, k, n):
|
|
"""Test grouped GEMM where each group can have a different 'masked' M dimension."""
|
|
# Test correctness with random partial masks
|
|
masked_m_candidates = list(filter(lambda candidate: candidate <= m, (64, 128, 192, 256, 320, 384)))
|
|
for _ in range(10):
|
|
x_fp8, y_fp8, out, ref_out = construct_grouped(num_groups, m, k, n, is_masked=True)
|
|
masked_m = torch.empty((num_groups,), device='cuda', dtype=torch.int)
|
|
for j in range(num_groups):
|
|
masked_m[j] = random.choice(masked_m_candidates)
|
|
# We pick an expected M that covers the average usage
|
|
expected_m = min(int(masked_m.float().mean()) + 1, m)
|
|
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8, y_fp8, out, masked_m, expected_m)
|
|
|
|
for j in range(num_groups):
|
|
diff = calc_diff(out[j, :masked_m[j].item()], ref_out[j, :masked_m[j].item()])
|
|
assert diff < 0.001, (
|
|
f'Grouped Masked GEMM mismatch: m={m}, k={k}, n={n}, group={j}, '
|
|
f'masked_m={masked_m[j]}, groups={num_groups}, diff={diff:.5f}'
|
|
)
|
|
|
|
# Test performance with full dimension (no actual mask)
|
|
def test_func():
|
|
x_fp8_local, y_fp8_local, out_local, _ = construct_grouped(num_groups, m, k, n, is_masked=True)
|
|
masked_m = torch.ones((num_groups,), device='cuda', dtype=torch.int) * m
|
|
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8_local, y_fp8_local, out_local, masked_m, m)
|
|
|
|
t = bench_kineto(test_func, 'fp8_gemm', suppress_kineto_output=True)
|
|
print(f'[test_m_grouped_gemm_masked] (groups={num_groups}, m={m}, n={n}, k={k}): '
|
|
f'{t * 1e6:6.0f} us | {2 * num_groups * m * n * k / t / 1e12:5.2f} TFLOPS')
|