leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 26 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/dense_blockscaled_gemm_persistent.py
+3 -3
View File
@@ -615,7 +615,7 @@ class Sm100BlockScaledPersistentDenseGemmKernel:
ab_empty_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage]
acc_full_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage]
acc_empty_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage]
tmem_dealloc_mbar_ptr: cutlass.Int64
tmem_dealloc_mbar: cutlass.Int64
tmem_holding_buf: cutlass.Int32
# (EPI_TILE_M, EPI_TILE_N, STAGE)
sC: cute.struct.Align[
@@ -794,11 +794,11 @@ class Sm100BlockScaledPersistentDenseGemmKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_alloc_barrier,
allocator_warp_id=self.epilog_warp_id[0],
is_two_cta=use_2cta_instrs,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar_ptr,
two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar.ptr,
)
# Cluster arrive after barrier init