leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

v4.5 dev update. (#3153)

Browse Source
This commit is contained in:
Junkai-Wu
2026-04-07 12:16:05 -04:00
committed by GitHub
parent 418d38a5de
commit a221da7ccf
265 changed files with 4913 additions and 1478 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/python/CuTeDSL/blackwell/fmha.py
+2 -2
View File
@@ -998,7 +998,7 @@ class BlackwellFusedMultiHeadAttentionForward:
# Alloc tmem buffer
tmem_alloc_cols = Int32(self.tmem_alloc_cols)
cute.arch.alloc_tmem(tmem_alloc_cols, storage.tmem_holding_buf)
cute.arch.alloc_tmem(tmem_alloc_cols, storage.tmem_holding_buf.ptr)
self.tmem_alloc_barrier.arrive_and_wait()
tile_sched = fmha_utils.create_fmha_static_tile_scheduler(
tile_sched_params, cute.arch.block_idx(), cute.arch.grid_dim()
@@ -1260,7 +1260,7 @@ class BlackwellFusedMultiHeadAttentionForward:
tmem_ptr = cute.arch.retrieve_tmem_ptr(
Float32,
alignment=16,
ptr_to_buffer_holding_addr=storage.tmem_holding_buf,
ptr_to_buffer_holding_addr=storage.tmem_holding_buf.ptr,
)
cute.arch.dealloc_tmem(tmem_ptr, tmem_alloc_cols)