Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Kernel] Add CUTLASS sparse support, heuristics, and torch operators #10340

Open
wants to merge 21 commits into
base: main
Choose a base branch
from
+7,454 −596
Open

[Kernel] Add CUTLASS sparse support, heuristics, and torch operators #10340

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
21 commits
Select commit Hold shift + click to select a range
5d51361
Add cutlass 2:4 infrastructure
Faraz9877 Oct 22, 2024
17f5b96
Update with test code
Faraz9877 Oct 28, 2024
471a03c
Clean up a bit; both fp8 and int8 working
Faraz9877 Oct 30, 2024
0b332fb
Add fp16 and bf16 support to sparse cutlass mm
Faraz9877 Oct 30, 2024
ccadad0
Add multiprocessing for kernel sweep benchmarking
Faraz9877 Nov 8, 2024
807737c
Add multi-GPU
Faraz9877 Nov 8, 2024
04c19a5
Add cutlass_scaled_sparse_mm op
Faraz9877 Nov 14, 2024
2a85c5a
Clean up
Faraz9877 Nov 14, 2024
1b381c9
Update code
Faraz9877 Nov 14, 2024
4e31076
Update code
Faraz9877 Nov 14, 2024
13fccf4
Clean up the benchmarking
Faraz9877 Nov 14, 2024
b345cc8
Clean up the cutlass benchmarking
Faraz9877 Nov 14, 2024
2d03e1d
Fix cmake errors
Faraz9877 Nov 14, 2024
e9439cc
Fix the cmake TAG
Faraz9877 Nov 14, 2024
6870093
Fix batch size and zeros issue
Faraz9877 Nov 18, 2024
8d7b0df
Enable other datatypes
Faraz9877 Nov 18, 2024
cdb213d
Add the heuristics for fp8
Faraz9877 Nov 19, 2024
246c8ab
Add cherry-picked heuristic for Llama3 8B model
Faraz9877 Nov 20, 2024
4bc043a
Add cherry-picked configs for all data types
Faraz9877 Dec 2, 2024
6773df8
Add scaled_sparse_azp op
Faraz9877 Dec 2, 2024
07defd0
Revert fp8 kernels to best performing version at: 246c8ab
Faraz9877 Dec 3, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Add cherry-picked configs for all data types
  • Loading branch information
@Faraz9877
Faraz9877 committed Dec 2, 2024
commit 4bc043a006993bca057dc5018099934c37368ef5
40 changes: 16 additions & 24 deletions benchmarks/cutlass_benchmarks/sparse_mm/bench_v1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,15 +36,14 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
# Create tensors
b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k)
aT = a.t()
bT = b.t()
scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)

out = ops.cutlass_scaled_sparse_mm(b_compressed, e, aT, scale_b, scale_a, torch.bfloat16)
out_ref = ops.cutlass_scaled_mm(a, bT, scale_a, scale_b, torch.bfloat16)
out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)

if not torch.allclose(out.t(), out_ref):
if not torch.allclose(out, out_ref):
print("Incorrect result")
exit()

Expand Down Expand Up @@ -96,16 +95,15 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
# Create tensors
b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k)
aT = a.t()
bT = b
scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)

out = ops.cutlass_scaled_sparse_mm(b_compressed, e, aT, scale_b, scale_a, torch.bfloat16)
out_ref = ops.cutlass_scaled_mm(a, bT, scale_a, scale_b, torch.bfloat16)
out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)

if not torch.allclose(out, out_ref, rtol=1e-2, atol=1e-2):
print(f"Incorrect result for {m}, {k}, {n}")
if not torch.allclose(out, out_ref):
print("Incorrect result")
exit()

timers = []
Expand All @@ -114,7 +112,7 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
timers.append(
bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
bT.to(dtype=torch.bfloat16, device="cuda")))
b.to(dtype=torch.bfloat16, device="cuda")))

# pytorch impl: bf16 output, without fp8 fast accum
timers.append(
Expand All @@ -123,7 +121,7 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
"pytorch_fp8_fp8_bf16_scaled_mm",
torch._scaled_mm,
a,
bT,
b,
scale_a=scale_a,
scale_b=scale_b,
out_dtype=torch.bfloat16))
Expand All @@ -135,7 +133,7 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
"pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
torch._scaled_mm,
a,
bT,
b,
scale_a=scale_a,
scale_b=scale_b,
out_dtype=torch.bfloat16,
Expand All @@ -148,7 +146,7 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
"pytorch_fp8_fp8_fp16_scaled_mm",
torch._scaled_mm,
a,
bT,
b,
scale_a=scale_a,
scale_b=scale_b,
out_dtype=torch.float16))
Expand All @@ -160,7 +158,7 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
"pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
torch._scaled_mm,
a,
bT,
b,
scale_a=scale_a,
scale_b=scale_b,
out_dtype=torch.float16,
Expand All @@ -183,26 +181,19 @@ def bench_fp16(dtype: torch.dtype, m: int, k: int, n: int, label: str,
sub_label: str) -> Iterable[TMeasurement]:
assert dtype == torch.float16

m, k, n = 1, 128, 256

# Create tensors
b_compressed, e, a, b = make_rand_sparse_tensors(torch.float16, m, n, k)
aT = a.t()
bT = b.t()
scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)

out = ops.cutlass_scaled_sparse_mm(b_compressed, e, aT, scale_b, scale_a, torch.bfloat16)
out_ref = to_bf16(a@bT)
out_ref = to_bf16(a@b)

if not torch.allclose(out.t(), out_ref, rtol=1e-2, atol=1e-2):
if not torch.allclose(out, out_ref, rtol=1e-2, atol=1e-2):
print("Incorrect result")
print(out.t())
print(out_ref)
exit()
else:
print("Correct result")

timers = []

Expand Down Expand Up @@ -269,16 +260,17 @@ def bench_bf16(dtype: torch.dtype, m: int, k: int, n: int, label: str,
# Create tensors
b_compressed, e, a, b = make_rand_sparse_tensors(torch.bfloat16, m, n, k)
aT = a.t()
bT = b.t()
scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)

out = ops.cutlass_scaled_sparse_mm(b_compressed, e, aT, scale_b, scale_a, torch.bfloat16)
out_ref = to_bf16(a@bT)
out_ref = to_bf16(a@b)

if not torch.allclose(out.t(), out_ref):
if not torch.allclose(out, out_ref, rtol=1e-1, atol=1e-1):
print("Incorrect result")
print(out)
print(out_ref)
exit()

timers = []
Expand Down
Loading