Skip to content

Do not pass l_vec and o to mha_forward in h100_bench #92

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
acforvs wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Do not pass l_vec and o to mha_forward in h100_bench #92

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1ec071d
Remove l_vec from mha_forward in h100_bench.py
acforvs Feb 19, 2025
b95a021
remove o from mha_forward in h100_bench
acforvs Feb 19, 2025
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
11 changes: 4 additions & 7 deletions kernels/attn/h100/h100_bench.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,9 +37,6 @@ def benchmark_attention(configurations):
kg = torch.zeros_like(k, requires_grad=False, dtype=torch.float).contiguous()
vg = torch.zeros_like(v, requires_grad=False, dtype=torch.float).contiguous()

l_vec = torch.zeros(q.shape[0], q.shape[1], q.shape[2], 1, device=q.device, dtype=torch.float, requires_grad=False).contiguous()
d_vec = torch.zeros(q.shape[0], q.shape[1], q.shape[2], 1, device=q.device, dtype=torch.float, requires_grad=False).contiguous()

# Prepare for timing forward pass
start_events_fwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
end_events_fwd = [torch.cuda.Event(enable_timing=True) for _ in range(10)]
Expand All @@ -49,12 +46,12 @@ def benchmark_attention(configurations):

# Warmup for forward pass
for _ in range(10):
tk.mha_forward(q, k, v, o, l_vec, causal)
tk.mha_forward(q, k, v, causal)

# Time the forward pass
for i in range(10):
start_events_fwd[i].record()
tk.mha_forward(q, k, v, o, l_vec, causal)
_, l_vec = tk.mha_forward(q, k, v, causal)
end_events_fwd[i].record()

torch.cuda.synchronize()
Expand All @@ -77,12 +74,12 @@ def benchmark_attention(configurations):

# Warmup for backward pass
for _ in range(10):
qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, d_vec, grad_output, causal)
qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, grad_output, causal)

# Time the backward pass
for i in range(10):
start_events_bwd[i].record()
qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, d_vec, grad_output, causal)
qg, kg, vg = tk.mha_backward(q, k, v, o, l_vec, grad_output, causal)
end_events_bwd[i].record()

torch.cuda.synchronize()
Expand Down