leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 29 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/mixed_input_gemm/grouped_mixed_input_gemm.py
+3 -3
View File
@@ -624,7 +624,7 @@ class GroupedMixedInputGemmKernel:
tile_info_empty_mbar_ptr: cute.struct.MemRange[
cutlass.Int64, self.num_tile_info_stage
]
tmem_dealloc_mbar_ptr: cutlass.Int64
tmem_dealloc_mbar: cutlass.Int64
tmem_holding_buf: cutlass.Int32
self.shared_storage = SharedStorage
@@ -824,11 +824,11 @@ class GroupedMixedInputGemmKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_ptr_sync_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
examples/python/CuTeDSL/blackwell/mixed_input_gemm/grouped_mixed_input_gemm_acc_scale.py
+3 -3
View File
@@ -518,7 +518,7 @@ class GroupedMixedInputGemmAccScaleKernel:
tile_info_empty_mbar_ptr: cute.struct.MemRange[
cutlass.Int64, self.num_tile_info_stage
]
tmem_dealloc_mbar_ptr: cutlass.Int64
tmem_dealloc_mbar: cutlass.Int64
tmem_holding_buf: cutlass.Int32
self.shared_storage = SharedStorage
@@ -708,11 +708,11 @@ class GroupedMixedInputGemmAccScaleKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_ptr_sync_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
examples/python/CuTeDSL/blackwell/mixed_input_gemm/mixed_input_gemm.py
+3 -3
View File
@@ -618,7 +618,7 @@ class MixedInputGemmKernel:
]
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
# Tensor buffers
# (EPI_TILE_M, EPI_TILE_N, STAGE)
@@ -820,11 +820,11 @@ class MixedInputGemmKernel:
# Tensor memory dealloc barrier init
tmem = utils.TmemAllocator(
storage.tmem_holding_buf,
storage.tmem_holding_buf.ptr,
barrier_for_retrieve=self.tmem_ptr_sync_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