biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

D1: TMA producer uses acquire_and_advance + commit (no wait_and_advance)

Browse Source
This commit is contained in:
biondizzle
2026-05-24 06:38:15 +00:00
parent 3bf1e62b58
commit 583c509bcd
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
dsv4/kernels/attention/fmha.py
View File

@@ -221,15 +221,15 @@ class FmhaKernel:
# Load Q[k_sub] → sQ
qh = qp.acquire_and_advance()
cute.copy(tma_q, tAgQ[(None, Int32(k_sub))], tAsQ[(None, qh.index)], tma_bar_ptr=qh.barrier)
qh.commit()
# Load K[k_sub] → sK
kvh = kvp.acquire_and_advance()
cute.copy(tma_k, tBgK[(None, Int32(k_sub))], tBsK[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
qh = qp.wait_and_advance(); qh.release()
kvh = kvp.wait_and_advance(); pk = cutlass.Boolean(1)
kvh.commit()
# Load V[kt] → sV
kvh = kvp.acquire_and_advance(pk)
kvh = kvp.acquire_and_advance()
cute.copy(tma_v, tVgV[(None, Int32(kt))], tVsV[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
kvh = kvp.wait_and_advance(); pk = cutlass.Boolean(1)
kvh.commit()
qp.tail()
kvp.tail()
else: