biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make various updates and fixes: (#164)

Browse Source 
- Add BF16 support for SM90 and SM100
- Refactor Python APIs
- Other fixes and code refactoring
This commit is contained in:
Ray Wang
2025-08-15 18:32:35 +08:00
committed by GitHub
parent 3254b758e2
commit f85ec649d7
34 changed files with 2293 additions and 495 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
tests/generators.py
View File

@@ -59,6 +59,7 @@ def get_out_dtype() -> tuple:
def get_major_ab(freeze_a: bool) -> tuple:
# TODO: test other major-ness for SM90 BF16 GEMMs
if get_arch_major() == 9:
return ((MajorTypeAB.KMajor, MajorTypeAB.KMajor), )
if freeze_a:
@@ -70,15 +71,15 @@ def get_major_ab(freeze_a: bool) -> tuple:
def enumerate_normal(use_bf16: bool = False) -> Generator:
for kernel_type in get_kernel_types(use_bf16):
for m in (128, 4096):
for n, k in [(2112, 7168), (24576, 1536), (32768, 512), (7168, 16384), (4096, 7168), (7168, 2048), (129280, 7168)]:
for n, k in [(2112, 7168), (24576, 1536), (32768, 512), (7168, 16384), (4096, 7168), (7168, 2048)]:
for major_a, major_b in get_major_ab(False):
for out_dtype in get_out_dtype():
for accumulate in (False, ) if out_dtype == torch.bfloat16 or not kernel_type.is_1d1d() else (False, True):
for accumulate in (False, ) if out_dtype == torch.bfloat16 or kernel_type.is_1d2d() else (False, True):
yield kernel_type, m, n, k, major_a, major_b, accumulate, out_dtype
def enumerate_m_grouped_contiguous() -> Generator:
for kernel_type in get_kernel_types():
def enumerate_m_grouped_contiguous(use_bf16: bool = False) -> Generator:
for kernel_type in get_kernel_types(use_bf16):
for num_groups, expected_m_per_group, n, k in ((4, 8192, 4096, 7168), (4, 8192, 7168, 2048), (8, 4096, 4096, 7168), (8, 4096, 7168, 2048)):
for major_a, major_b in get_major_ab(True):
yield kernel_type, num_groups, expected_m_per_group, n, k, major_a, major_b
@@ -106,15 +107,12 @@ def enumerate_k_grouped_contiguous():
def enumerate_sf_layout():
for with_transpose in (True, False):
for mn in (4096, 4097, 8192):
for k in (128, 7168, 7296):
for num_groups in (1, 2, 4) if with_transpose else (1, ):
if num_groups > 1 and (mn * ceil_div(k, 128)) % 4 != 0:
continue
if not with_transpose and mn % 4 != 0:
continue
yield mn, k, with_transpose, num_groups
for use_ue8m0 in (False, True):
for with_transpose in (True, False):
for mn in (4096, 4097, 8192):
for k in (128, 7168, 7296):
for num_groups in (1, 2, 4):
yield mn, k, with_transpose, use_ue8m0, num_groups
def enumerate_k_grouped_sf_layout():
@@ -126,6 +124,13 @@ def enumerate_k_grouped_sf_layout():
yield mn, ks, num_groups
def enumerate_transpose():
for mn in (64, 4096, 16384):
for delta in (0, 101, 202, 303):
for k in (128, 1024, 4096, 9984, 16384):
yield mn + delta, k
def generate_normal(m: int, n: int, k: int,
major_a: MajorTypeAB, major_b: MajorTypeAB,
accumulate: bool, out_dtype: torch.dtype,
@@ -149,8 +154,8 @@ def generate_normal(m: int, n: int, k: int,
def generate_m_grouped_contiguous(num_groups: int, expected_m_per_group: int, n: int, k: int,
major_a: MajorTypeAB, major_b: MajorTypeAB, use_ue8m0: bool) -> \
Tuple[int, Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor], torch.Tensor, torch.Tensor, torch.Tensor]:
major_a: MajorTypeAB, major_b: MajorTypeAB,
use_ue8m0: bool = False, use_bf16: bool = False):
actual_ms = [int(expected_m_per_group * random.uniform(0.7, 1.3)) for _ in range(num_groups)]
aligned_ms = [align(actual_m, get_mk_alignment_for_contiguous_layout()) for actual_m in actual_ms]
m = sum(aligned_ms)
@@ -171,6 +176,10 @@ def generate_m_grouped_contiguous(num_groups: int, expected_m_per_group: int, n:
start = aligned_end
ref_d = torch.where((m_indices == -1).unsqueeze(1), torch.zeros_like(ref_d), ref_d)
if use_bf16:
b = b if major_b.is_k_major() else b.mT.contiguous().mT
return m, a, b, m_indices, d, ref_d
assert major_a.is_k_major()
a_fp8 = per_token_cast_to_fp8(a, use_ue8m0=use_ue8m0)
b_fp8 = (torch.empty_like(b, dtype=torch.float8_e4m3fn),
@@ -181,24 +190,27 @@ def generate_m_grouped_contiguous(num_groups: int, expected_m_per_group: int, n:
return m, a_fp8, b_fp8, m_indices, d, ref_d
def generate_m_grouped_masked(num_groups: int, max_m: int, expected_m_per_group: int, n: int, k: int, use_ue8m0: bool) -> \
Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor], torch.Tensor, torch.Tensor, torch.Tensor]:
def generate_m_grouped_masked(num_groups: int, max_m: int, expected_m_per_group: int, n: int, k: int,
use_ue8m0: bool = False, use_bf16: bool = False):
a = torch.randn((num_groups, max_m, k), device='cuda', dtype=torch.bfloat16)
b = torch.randn((num_groups, n, k), device='cuda', dtype=torch.bfloat16)
d = torch.empty((num_groups, max_m, n), device='cuda', dtype=torch.bfloat16)
ref_d = torch.einsum('gmk,gnk->gmn', a, b)
masked_m = torch.empty((num_groups, ), device='cuda', dtype=torch.int)
for j in range(num_groups):
masked_m[j] = int(expected_m_per_group * random.uniform(0.7, 1.3))
assert masked_m.amax().item() <= max_m
if use_bf16:
return a, b, masked_m, d, ref_d
a_fp8 = (torch.empty_like(a, dtype=torch.float8_e4m3fn), torch.empty((num_groups, max_m, ceil_div(k, 128)), device='cuda', dtype=torch.float))
b_fp8 = (torch.empty_like(b, dtype=torch.float8_e4m3fn), torch.empty((num_groups, ceil_div(n, 128), ceil_div(k, 128)), device='cuda', dtype=torch.float))
for i in range(num_groups):
a_fp8[0][i], a_fp8[1][i] = per_token_cast_to_fp8(a[i], use_ue8m0=use_ue8m0)
b_fp8[0][i], b_fp8[1][i] = per_block_cast_to_fp8(b[i], use_ue8m0=use_ue8m0)
masked_m = torch.empty((num_groups, ), device='cuda', dtype=torch.int)
for j in range(num_groups):
masked_m[j] = int(expected_m_per_group * random.uniform(0.7, 1.3))
assert masked_m.amax().item() <= max_m
return a_fp8, b_fp8, masked_m, d, ref_d

125
tests/test_bf16.py Normal file
View File

@@ -0,0 +1,125 @@
import torch
import random
import deep_gemm
from deep_gemm.testing import (
bench_kineto,
calc_diff, count_bytes
)
from generators import (
enumerate_normal, enumerate_m_grouped_contiguous, enumerate_m_grouped_masked, generate_normal,
generate_m_grouped_contiguous, generate_m_grouped_masked
)
def test_gemm() -> None:
print('Testing GEMM:')
for _, m, n, k, major_a, major_b, accumulate, out_dtype in enumerate_normal(use_bf16=True):
major_opt = 'N' if major_a.is_k_major() else 'T'
major_opt += 'T' if major_b.is_k_major() else 'N'
out_opt = 'FP32' if out_dtype == torch.float else 'BF16'
acc_opt = f'acc={int(accumulate)}'
for test_alias in (False, True):
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, use_bf16=True)
func_name = f'bf16_gemm_{major_opt.lower() if test_alias else "nt"}'
if test_alias:
a = a if major_a.is_k_major() else a.T
b = b if major_b.is_k_major() else b.T
assert a.is_contiguous() and b.is_contiguous()
getattr(deep_gemm, func_name)(a, b, d, c=c)
diff = calc_diff(d, ref_d)
assert diff < 0.0001, (f'{m=}, {n=}, {k=}, {major_opt=}, {accumulate=}, {out_dtype=}, '
f'{diff:.5f}, alias={test_alias}')
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, use_bf16=True)
cublas_t = 0
t = bench_kineto(lambda: deep_gemm.bf16_gemm_nt(a, b, d, c=c), 'bf16_gemm', suppress_kineto_output=True)
if accumulate == 0 and out_dtype == torch.bfloat16:
# noinspection PyBroadException
try:
cublas_t = bench_kineto(lambda: a @ b.T, 'nvjet', suppress_kineto_output=True)
except Exception:
pass
print(f' > Perf (m={m:5}, n={n:5}, k={k:5}, layout={major_opt}, {out_opt}, {acc_opt}): '
f'{t * 1e6:4.0f} us | '
f'{2 * m * n * k / t / 1e12:4.0f} TFLOPS | '
f'{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s | '
f'{cublas_t / t:.2f}x cuBLAS')
print()
def test_m_grouped_gemm_contiguous() -> None:
print('Testing m-grouped contiguous GEMM:')
for _, num_groups, expected_m_per_group, n, k, major_a, major_b in enumerate_m_grouped_contiguous(use_bf16=True):
major_opt = 'N' if major_a.is_k_major() else 'T'
major_opt += 'T' if major_b.is_k_major() else 'N'
for test_alias in (False, True):
m, a, b, m_indices, d, ref_d = generate_m_grouped_contiguous(num_groups, expected_m_per_group, n, k, major_a, major_b, use_bf16=True)
func_name = f"m_grouped_bf16_gemm_{(major_opt.lower() if test_alias else 'nt')}_contiguous"
if test_alias:
assert major_a.is_k_major()
b = b if major_b.is_k_major() else b.mT
assert a[0].is_contiguous() and b[0].is_contiguous()
getattr(deep_gemm, func_name)(a, b, d, m_indices)
d = torch.where((m_indices == -1).unsqueeze(1), torch.zeros_like(d), d)
diff = calc_diff(d, ref_d)
assert diff < 0.001, f'{m=}, {n=}, {k=}, {major_opt}, {kernel_opt}, {diff:.5f}, alias={test_alias}'
m, a, b, m_indices, d, ref_d = generate_m_grouped_contiguous(num_groups, expected_m_per_group, n, k, major_a, major_b, use_bf16=True)
# noinspection PyShadowingNames
def test_func():
deep_gemm.m_grouped_bf16_gemm_nt_contiguous(a, b, d, m_indices)
t = bench_kineto(test_func, 'bf16_gemm', suppress_kineto_output=True)
print(f' > Perf ({num_groups=}, m={m:5}, n={n:5}, k={k:5}, layout={major_opt}): '
f'{t * 1e6:4.0f} us | '
f'{2 * m * n * k / t / 1e12:4.0f} TFLOPS | '
f'{count_bytes(a, b, d) / 1e9 / t:4.0f} GB/s')
print()
def test_m_grouped_gemm_masked() -> None:
print('Testing m-grouped masked GEMM:')
# TODO: when the actual `m` is greater than `expected_m_per_group`, efficiency may significantly decrease.
for _, num_groups, max_m, expected_m_per_group, n, k in enumerate_m_grouped_masked():
# Test correctness
for i in range(10):
a, b, masked_m, d, ref_d = generate_m_grouped_masked(num_groups, max_m, expected_m_per_group, n, k, use_bf16=True)
deep_gemm.m_grouped_bf16_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group)
for j in range(num_groups):
diff = calc_diff(d[j, :masked_m[j].item()], ref_d[j, :masked_m[j].item()])
assert diff < 0.001, f'{m=}, {n=}, {k=}, {j=}, masked_m={masked_m[j]}, {num_groups=}, {diff:.5f}'
# Construct full cases
a, b, masked_m, d, ref_d = generate_m_grouped_masked(num_groups, max_m, expected_m_per_group, n, k, use_bf16=True)
# noinspection PyShadowingNames
def test_func():
deep_gemm.m_grouped_bf16_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group)
# Test performance with fixed shapes
valid_m = masked_m.sum().item()
t = bench_kineto(test_func, 'bf16_gemm', suppress_kineto_output=True)
print(f' > Perf ({num_groups=}, expected_m_per_group={expected_m_per_group:4}, n={n:4}, k={k:4}): '
f'{t * 1e6:4.0f} us | '
f'{2 * valid_m * n * k / t / 1e12:4.0f} TFLOPS | '
f'{(count_bytes(a, d) * valid_m / (max_m * num_groups) + count_bytes(b)) / 1e9 / t:4.0f} GB/s')
print()
if __name__ == '__main__':
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.manual_seed(0)
random.seed(0)
print('Library path:')
print(f' > {deep_gemm.__path__}\n')
test_gemm()
test_m_grouped_gemm_contiguous()
test_m_grouped_gemm_masked()

4
tests/test_fp8.py
View File

@@ -105,7 +105,7 @@ def test_m_grouped_gemm_masked() -> None:
# Test correctness
for i in range(10):
a, b, masked_m, d, ref_d = generate_m_grouped_masked(num_groups, max_m, expected_m_per_group, n, k, use_ue8m0=use_ue8m0)
deep_gemm.fp8_m_grouped_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
deep_gemm.m_grouped_fp8_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
for j in range(num_groups):
diff = calc_diff(d[j, :masked_m[j].item()], ref_d[j, :masked_m[j].item()])
assert diff < 0.001, f'{max_m=}, {n=}, {k=}, {j=}, masked_m={masked_m[j]}, {kernel_opt}, {num_groups=}, {diff:.5f}'
@@ -115,7 +115,7 @@ def test_m_grouped_gemm_masked() -> None:
# noinspection PyShadowingNames
def test_func():
deep_gemm.fp8_m_grouped_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
deep_gemm.m_grouped_fp8_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
# Test performance with fixed shapes
valid_m = masked_m.sum().item()

46
tests/test_layout.py
View File

@@ -1,16 +1,18 @@
import time
import torch
import random
from deep_gemm.testing import bench_kineto, count_bytes
from deep_gemm.testing import bench_kineto, count_bytes, calc_diff
from deep_gemm.utils import (
align, ceil_div,
per_token_cast_to_fp8, per_channel_cast_to_fp8,
get_tma_aligned_size,
get_mn_major_tma_aligned_tensor,
get_mn_major_tma_aligned_packed_ue8m0_tensor,
get_k_grouped_mn_major_tma_aligned_packed_ue8m0_tensor
)
from generators import (
enumerate_transpose,
enumerate_sf_layout,
enumerate_k_grouped_sf_layout
)
@@ -43,29 +45,39 @@ def get_mn_major_tma_aligned_packed_ue8m0_tensor_torch_impl(x: torch.Tensor) ->
def test_sf_layout_kernels() -> None:
print('Testing SF layout kernels:')
for mn, k, with_transpose, num_groups in enumerate_sf_layout():
for mn, k, with_transpose, use_ue8m0, num_groups in enumerate_sf_layout():
x = torch.randn((num_groups * mn, k), dtype=torch.bfloat16, device='cuda')
x, fp32_sf = per_token_cast_to_fp8(x, use_ue8m0=True)
x, fp32_sf = per_token_cast_to_fp8(x, use_ue8m0=use_ue8m0)
fp32_sf = fp32_sf if num_groups == 1 else fp32_sf.view(num_groups, mn, -1)
fp32_sf = fp32_sf if with_transpose else fp32_sf.transpose(-1, -2).contiguous().transpose(-1, -2)
# Correctness
packed_sf = get_mn_major_tma_aligned_packed_ue8m0_tensor(fp32_sf)
ref_packed_sf = get_mn_major_tma_aligned_packed_ue8m0_tensor_torch_impl(fp32_sf)
assert torch.equal(packed_sf, ref_packed_sf), f'{mn=}, {k=}, {with_transpose=}, {num_groups=}'
assert packed_sf.shape == ref_packed_sf.shape
assert all([packed_sf.stride(i) == ref_packed_sf.stride(i) for i in range(packed_sf.dim())])
# Test launch overhead
launch_start_t = time.time_ns()
get_mn_major_tma_aligned_packed_ue8m0_tensor(fp32_sf)
launch_end_t = time.time_ns()
if use_ue8m0:
impl, name = get_mn_major_tma_aligned_packed_ue8m0_tensor, 'pack_fp32_into_ue8m0'
packed_sf = get_mn_major_tma_aligned_packed_ue8m0_tensor(fp32_sf)
ref_packed_sf = get_mn_major_tma_aligned_packed_ue8m0_tensor_torch_impl(fp32_sf)
assert torch.equal(packed_sf, ref_packed_sf), f'{mn=}, {k=}, {with_transpose=}, {num_groups=}'
assert packed_sf.shape == ref_packed_sf.shape
assert all([packed_sf.stride(i) == ref_packed_sf.stride(i) for i in range(packed_sf.dim())])
else:
impl, name = get_mn_major_tma_aligned_tensor, 'transpose'
transposed_sf = get_mn_major_tma_aligned_tensor(fp32_sf)
tma_aligned_mn, sf_k = get_tma_aligned_size(mn, fp32_sf.element_size()), ceil_div(k, 128)
if num_groups > 1:
assert transposed_sf.size(0) == num_groups
assert transposed_sf.stride(0) == tma_aligned_mn * sf_k
assert transposed_sf.shape[-2:] == (mn, sf_k)
assert transposed_sf.stride()[-2:] == (1, tma_aligned_mn)
assert torch.equal(fp32_sf, transposed_sf)
# Performance
t = bench_kineto(lambda: get_mn_major_tma_aligned_packed_ue8m0_tensor(fp32_sf), 'pack_fp32_into_ue8m0')
print(f' > Perf ({num_groups=:2}, {mn=:5}, {k=:5}, transpose={int(with_transpose)}): '
f'launch {(launch_end_t - launch_start_t) / 1e3:3.0f} us | {t * 1e6:4.0f} us | '
f'{count_bytes(fp32_sf, packed_sf) / 1e9 / t:4.0f} GB/s')
try:
t = bench_kineto(lambda: impl(fp32_sf), name)
except AssertionError as e:
# Some cases may fallback to PyTorch impl
t = 0
print(f' > Perf ({num_groups=:2}, {mn=:5}, {k=:5}, transpose={int(with_transpose)}, use_ue8m0={int(use_ue8m0)}): '
f'{t * 1e6:4.0f} us | {count_bytes(fp32_sf, impl(fp32_sf)) / 1e9 / t if t else 0:4.0f} GB/s')
print()

15
tests/test_lazy_init.py Normal file
View File

@@ -0,0 +1,15 @@
import torch
import torch.multiprocessing as mp
import deep_gemm
def main(local_rank: int):
torch.cuda.set_device(local_rank)
if __name__ == '__main__':
procs = [mp.Process(target=main, args=(i, ), ) for i in range(8)]
for p in procs:
p.start()
for p in procs:
p.join()