Skip to content

fix(dyt): put the row count on grid dimension 0 in the forward launch #1607

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

Open
bhargavchintam wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(dyt): put the row count on grid dimension 0 in the forward launch #1607

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 4 additions & 3 deletions src/liger_kernel/ops/dyt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,9 +32,9 @@
)
@triton.jit
def _dyt_fwd_kernel(X, Y, Alpha, Gamma, Beta, HAVE_BETA: tl.constexpr, N: tl.constexpr, BLOCK_N: tl.constexpr):
col = tl.cast(tl.program_id(0), tl.int64) * BLOCK_N + tl.arange(0, BLOCK_N)
col = tl.cast(tl.program_id(1), tl.int64) * BLOCK_N + tl.arange(0, BLOCK_N)
mask = col < N
row_id = tl.cast(tl.program_id(1), tl.int64)
row_id = tl.cast(tl.program_id(0), tl.int64)

X += row_id * N
Y += row_id * N
Expand Down Expand Up @@ -107,7 +107,8 @@ def liger_dyt_fwd(x, alpha, gamma, beta):

y = torch.empty_like(x)

grid = lambda meta: (triton.cdiv(N, meta["BLOCK_N"]), M)
# CUDA limits grid dimensions 1 and 2 to 65535 programs, so the row count has to go on dimension 0.
grid = lambda meta: (M, triton.cdiv(N, meta["BLOCK_N"]))
_dyt_fwd_kernel[grid](
x,
y,
Expand Down
30 changes: 30 additions & 0 deletions test/transformers/test_dyt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,6 +100,36 @@ def test_liger_dyt_correctness(B, T, hidden_size, beta, init_alpha, dtype, atol,
)


@pytest.mark.parametrize("beta", [False, True])
def test_liger_dyt_more_rows_than_grid_dimension_limit(beta):
# CUDA limits grid dimensions 1 and 2 to 65535 programs, and this input has 65538 rows.
B, T, hidden_size = 2, 32769, 8
_input = torch.randn(B, T, hidden_size, device=device, dtype=torch.float32)
x1 = _input.clone().requires_grad_(True)
x2 = _input.clone().requires_grad_(True)

torch_dyt = TorchDyT(hidden_size=hidden_size, beta=beta).to(device)
liger_dyt = LigerDyT(hidden_size=hidden_size, beta=beta).to(device)

torch_output = torch_dyt(x1)
liger_output = liger_dyt(x2)
assert_verbose_allclose(torch_output, liger_output, rtol=1e-5, atol=1e-5, extra_info="[output]")

grad_output = torch.randn_like(_input)
torch_output.backward(grad_output)
liger_output.backward(grad_output)
assert_verbose_allclose(x1.grad, x2.grad, rtol=1e-5, atol=1e-5, extra_info="[input.grad]")

# These gradients are sums over all the rows, so they carry more fp32 rounding error.
rtol, atol = 1e-3, 1e-3
assert_verbose_allclose(torch_dyt.alpha.grad, liger_dyt.alpha.grad, rtol=rtol, atol=atol, extra_info="[alpha.grad]")
assert_verbose_allclose(torch_dyt.gamma.grad, liger_dyt.gamma.grad, rtol=rtol, atol=atol, extra_info="[gamma.grad]")
if beta:
assert_verbose_allclose(
torch_dyt.beta.grad, liger_dyt.beta.grad, rtol=rtol, atol=atol, extra_info="[beta.grad]"
)


@pytest.mark.parametrize("beta", [False, True])
@pytest.mark.parametrize(
"B, T, hidden_size",
Expand Down