Skip to content

[FA]: squeeze Z and H into one axis to align with XeTLA #2618

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

Merged
quintinwang5 merged 2 commits into from
Nov 6, 2024
Merged

[FA]: squeeze Z and H into one axis to align with XeTLA #2618

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c2080b6
[FA]: squeeze Z and H into one axis to align with XeTLA
quintinwang5 Nov 4, 2024
6fbf4fb
Fix format
quintinwang5 Nov 4, 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
5 changes: 3 additions & 2 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 @@ -78,10 +78,11 @@ def _attn_fwd(Q, K, V, sm_scale, M, Out, #
start_m = tl.program_id(2)
off_z = tl.program_id(0)
off_h = tl.program_id(1)
qvk_offset = off_z.to(tl.int64) * stride_qz + off_h.to(tl.int64) * stride_qh
if N_CTX <= 512:
start_m = tl.program_id(0)
off_z = tl.program_id(2)
qvk_offset = off_z.to(tl.int64) * stride_qz + off_h.to(tl.int64) * stride_qh
qvk_offset = off_z.to(tl.int64) * stride_qh

# block pointers
Q_block_ptr = tl.make_block_ptr(
Expand Down Expand Up @@ -181,7 +182,7 @@ def forward(q, k, v, causal, sm_scale):
grid = lambda args: (q.shape[0], q.shape[1], triton.cdiv(q.shape[2], args['BLOCK_M']))
n_ctx = q.shape[2]
if n_ctx <= 512:
grid = lambda args: (triton.cdiv(q.shape[2], args['BLOCK_M']), q.shape[1], q.shape[0])
grid = lambda args: (triton.cdiv(q.shape[2], args['BLOCK_M']), 1, q.shape[0] * q.shape[1])
M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)

if os.getenv('TRITON_INTEL_ADVANCED_PATH', '0') == '0':
Expand Down