Skip to content

Remove workaround for torch.nn.functional.scaled_dot_product_attention #2827

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

Closed
anmyachev wants to merge 1 commit into from
Closed

Remove workaround for torch.nn.functional.scaled_dot_product_attention #2827

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
9 changes: 2 additions & 7 deletions benchmarks/triton_kernels_benchmark/flash_attention_fwd_benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -256,13 +256,8 @@ def benchmark(Z, H, N_CTX, D_HEAD, CAUSAL, provider):

elif provider == 'triton':
triton_fn = lambda: forward(q, k, v, CAUSAL, sm_scale)
if benchmark_suit.USE_IPEX_OPTION:
torch_fn = lambda: torch.nn.functional.scaled_dot_product_attention(
q, k, v, attn_mask=None, dropout_p=0.0, is_causal=CAUSAL, scale=sm_scale).to(torch.float32)
else:
# FIXME: use torch sdpa for result check after https://github.com/intel/intel-xpu-backend-for-triton/issues/2042 fixed
torch_fn = lambda: torch.nn.functional.scaled_dot_product_attention(q.cpu(), k.cpu(), v.cpu(
), attn_mask=None, dropout_p=0.0, is_causal=CAUSAL, scale=sm_scale).to(torch.float32)
torch_fn = lambda: torch.nn.functional.scaled_dot_product_attention(
q, k, v, attn_mask=None, dropout_p=0.0, is_causal=CAUSAL, scale=sm_scale).to(torch.float32)
atol = 1e-1 if N_CTX == 16384 else 1e-2
benchmark_suit.assert_close(triton_fn(), torch_fn(), atol=atol, rtol=1e-3, err_msg='triton to torch')
_, min_ms, max_ms, mean, cv = benchmark_suit.do_bench(triton_fn, n_warmup=10, n_repeat=10, quantiles=quantiles,
Expand Down
Loading