-
Notifications
You must be signed in to change notification settings - Fork 50
[Blackwell] add non-causal bwd/FA with TMA and atomic_add #603
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.
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
Conversation
|
@manman-ren has imported this pull request. If you are a Meta employee, you can view this in D85880773. |
htyu
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM.
njriasan
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM!
| for blk_idx in range(num_steps): | ||
| q = desc_q.load([(off_bh + curr_m).to(tl.int32), 0]) | ||
| qT = tl.trans(q) | ||
| # Load m before computing qk to reduce pipeline stall. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is this still relevant/required with WS?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
depending on whether the 1D load of m is in the load partition or not. We can load m first, then wait for qk.
|
Can you help run |
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
2bb886a to
6974d4d
Compare
Summary: Copied from Hongtao's TLX implementation in third_party/tlx/tutorials/blackwell-fa-ws-pipelined-persistent_test.py
Test Plan:
python run.py --op blackwell_attentions --seq-len 8192 --batch 4 --n-heads 32 --d-head 128 --only triton_tutorial_flash_v2_persistent_blackwell --bwd --force --metrics tflops
Reviewers:
Subscribers:
Tasks:
Tags: