DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-06-26 23:15:49 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Weight gradient kernels for dense and MoE models (#95)

Browse Source 
* Init weight gradient kernels.

* Support unaligned n,k and gmem stride

* Update docs

* Several cleanups

* Remove restrictions on N

* Add stride(0) assertions

---------

Co-authored-by: Chenggang Zhao <chenggangz@deepseek.com>
This commit is contained in:
Zhean Xu
2025-05-14 14:47:58 +08:00
committed by GitHub
parent d75b218b7b
commit 04278f6dee
12 changed files with 911 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
tests/test_core.py
View File

@@ -5,7 +5,7 @@ print(f'NVRTC version: {nvrtc.nvrtcVersion()[1:]}')
import random
import torch
from typing import Tuple
from typing import List, Tuple
import deep_gemm
from deep_gemm import bench_kineto, calc_diff, ceil_div, get_col_major_tma_aligned_tensor
@@ -13,11 +13,14 @@ from deep_gemm.jit_kernels.utils import get_m_alignment_for_contiguous_layout
def per_token_cast_to_fp8(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
assert x.dim() == 2 and x.size(1) % 128 == 0
assert x.dim() == 2
m, n = x.shape
pad_size = (128 - (n % 128)) % 128
x = torch.nn.functional.pad(x, (0, pad_size), value=0) if pad_size > 0 else x
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)
fp8_data = (x_view * (448.0 / x_amax.unsqueeze(2))).to(torch.float8_e4m3fn)
return fp8_data.view(m, n + pad_size)[:, :n], (x_amax / 448.0).view(m, -1)
def per_block_cast_to_fp8(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
@@ -94,10 +97,74 @@ def construct_masked_grouped(num_groups: int, m: int, k: int, n: int) -> \
return x_fp8, y_fp8, out, ref_out
def construct_wgrad(m: int, k: int, n: int) -> \
Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, 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)
residual = torch.randn((m, n), device='cuda', dtype=torch.float) * 10
out = residual.clone()
ref_out = residual + (x.float() @ y.float().t())
x_fp8 = per_token_cast_to_fp8(x)
y_fp8 = per_token_cast_to_fp8(y)
# NOTES: please do inplace add on the `out` later
return x_fp8, y_fp8, residual, out, ref_out
def construct_k_grouped_wgrad(m: int, n: int, k_sizes: List[int]) -> \
Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor], torch.Tensor, torch.Tensor, List[int]]:
num_groups, total_k = len(k_sizes), sum(k_sizes)
x_flat = torch.empty((m * total_k,), device='cuda', dtype=torch.bfloat16)
y_flat = torch.empty((n * total_k,), device='cuda', dtype=torch.bfloat16)
out = torch.zeros((num_groups, m, n), device='cuda', dtype=torch.float)
ref_out = torch.zeros((num_groups, m, n), device='cuda', dtype=torch.float)
# Fill tensors with data and compute reference output
x_offset, y_offset = 0, 0
for idx, k in enumerate(k_sizes):
x_chunk = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
y_chunk = torch.randn((n, k), device='cuda', dtype=torch.bfloat16)
x_flat[x_offset:x_offset + m * k].copy_(x_chunk.flatten())
y_flat[y_offset:y_offset + n * k].copy_(y_chunk.flatten())
ref_out[idx] = x_chunk.float() @ y_chunk.float().t()
x_offset += m * k
y_offset += n * k
x_fp8_flat = torch.empty_like(x_flat, dtype=torch.float8_e4m3fn)
y_fp8_flat = torch.empty_like(y_flat, dtype=torch.float8_e4m3fn)
total_scale_factors = sum((k + 127) // 128 for k in k_sizes)
x_scales = torch.empty((total_scale_factors, m), device='cuda', dtype=torch.float)
y_scales = torch.empty((total_scale_factors, n), device='cuda', dtype=torch.float)
# Cast to FP8 and prepare scale factors
x_offset, y_offset, scale_offset = 0, 0, 0
for k in k_sizes:
x_fp8_chunk, x_scale_chunk = per_token_cast_to_fp8(x_flat[x_offset:x_offset + m * k].view(m, k))
y_fp8_chunk, y_scale_chunk = per_token_cast_to_fp8(y_flat[y_offset:y_offset + n * k].view(n, k))
x_fp8_flat[x_offset:x_offset + m * k].copy_(x_fp8_chunk.flatten())
y_fp8_flat[y_offset:y_offset + n * k].copy_(y_fp8_chunk.flatten())
num_scales = (k + 127) // 128
x_scales[scale_offset:scale_offset + num_scales].copy_(x_scale_chunk.T)
y_scales[scale_offset:scale_offset + num_scales].copy_(y_scale_chunk.T)
x_offset += m * k
y_offset += n * k
scale_offset += num_scales
return (x_fp8_flat, x_scales), (y_fp8_flat, y_scales), out, ref_out, k_sizes
def test_gemm() -> None:
print('Testing GEMM:')
for m in (64, 128, 4096):
for k, n in [(7168, 2112), (1536, 24576), (512, 32768), (16384, 7168), (7168, 4096), (2048, 7168)]:
for k, n in [(576, 7168), (7168, 2112), (1536, 24576), (512, 32768), (16384, 7168), (7168, 4096), (2048, 7168)]:
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)
@@ -175,6 +242,62 @@ def test_m_grouped_gemm_masked() -> None:
print()
def test_wgrad_gemm():
print('Testing weight gradient GEMM:')
for k in (4096, 8192):
for m, n in ((7168, 2112), (1536, 24576), (512, 32768), (16384, 7168), (7168, 4096), (2048, 7168)):
# Test correctness
x_fp8, y_fp8, residual, out, ref_out = construct_wgrad(m, k, n)
deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt(x_fp8, y_fp8, out)
diff = calc_diff(out, ref_out)
assert diff < 0.001, f'{m=}, {k=}, {n=}, {diff:.5f}'
# Construct new tensors only once to avoid L2 cache acceleration (creating them puts them in L2)
x_fp8, y_fp8, residual, out, ref_out = construct_wgrad(m, k, n)
# noinspection PyShadowingNames
def test_func():
deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt(x_fp8, y_fp8, out)
t = bench_kineto(test_func, 'fp8_wgrad_gemm', suppress_kineto_output=True)
print(f' > Performance (m={m:5}, n={n:5}, k={k:5}): {t * 1e6:4.0f} us | '
f'throughput: {2 * m * n * k / t / 1e12:4.0f} TFLOPS, '
f'{(m * k + k * n + m * n * 2) / 1e9 / t:4.0f} GB/s')
print()
def test_k_grouped_wgrad_gemm():
print('Testing grouped weight gradient GEMM:')
for num_groups, base_k in ((4, 4096), (4, 8192), (8, 4096)):
for m, n in ((7168, 4096), (2048, 7168)):
# Vary k sizes around base_k
k_sizes = [base_k + random.randint(-1, 1) * 128 for _ in range(num_groups - 1)]
k_sizes.append(base_k * num_groups - sum(k_sizes))
# Test correctness
x_fp8, y_fp8, out, ref_out, k_sizes = construct_k_grouped_wgrad(m, n, k_sizes)
deep_gemm.k_grouped_wgrad_gemm_fp8_fp8_fp32_nt(x_fp8, y_fp8, out, k_sizes)
for idx in range(num_groups):
diff = calc_diff(out[idx], ref_out[idx])
assert diff < 0.001, f'{num_groups=}, {m=}, {n=}, k={k_sizes[idx]}, batch={idx}, {diff:.5f}'
# Construct new tensors to avoid L2 cache acceleration
x_fp8, y_fp8, out, ref_out, k_sizes = construct_k_grouped_wgrad(m, n, k_sizes)
total_k = sum(k_sizes)
def test_func():
deep_gemm.k_grouped_wgrad_gemm_fp8_fp8_fp32_nt(x_fp8, y_fp8, out, k_sizes)
t = bench_kineto(test_func, 'fp8_wgrad_gemm', suppress_kineto_output=True, with_multiple_kernels=True) * num_groups
print(f' > Performance ({num_groups=}, m={m:5}, n={n:5}, avg_k={total_k//num_groups:5}): {t * 1e6:4.0f} us | '
f'throughput: {2 * num_groups * m * n * (total_k/num_groups) / t / 1e12:4.0f} TFLOPS, '
f'{(m * total_k + n * total_k + num_groups * m * n * 2) / 1e9 / t:4.0f} GB/s')
print()
if __name__ == '__main__':
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
@@ -187,3 +310,6 @@ if __name__ == '__main__':
test_gemm()
test_m_grouped_gemm_contiguous()
test_m_grouped_gemm_masked()
test_wgrad_gemm()
test_k_grouped_wgrad_gemm()