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

v4.1 release update v2. (#2481)

Browse Source
This commit is contained in:
Junkai-Wu
2025-07-22 10:03:55 +08:00
committed by GitHub
parent 9baa06dd57
commit fd6cfe1ed0
179 changed files with 7878 additions and 1286 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/41_fused_multi_head_attention/fused_multihead_attention_fixed_seqlen.cu
View File

@@ -55,7 +55,7 @@
```
In practice, and for numerical stability reasons,
we also substract the maximum so far (`mi`) before doing
we also subtract the maximum so far (`mi`) before doing
the exponential. When we encounter new keys, the maximum
used to compute O so far (`m_prime`) can differ from the
current maximum, so we update O before accumulating with

2
examples/41_fused_multi_head_attention/fused_multihead_attention_variable_seqlen.cu
View File

@@ -55,7 +55,7 @@
```
In practice, and for numerical stability reasons,
we also substract the maximum so far (`mi`) before doing
we also subtract the maximum so far (`mi`) before doing
the exponential. When we encounter new keys, the maximum
used to compute O so far (`m_prime`) can differ from the
current maximum, so we update O before accumulating with

4
examples/41_fused_multi_head_attention/gemm/find_default_mma.h
View File

@@ -31,7 +31,7 @@
/*! \file
\brief Cutlass provides helper template functions to figure out the right
datastructures to instanciate to run a GEMM with various parameters (see
datastructures to instantiate to run a GEMM with various parameters (see
`cutlass/gemm/threadblock/default_mma.h`). However, due to template
instantiation priority rules, it will only create an MmaMultiStage with
kStages=3 (otherwise creates an MmePipelined - which is not compatible with
@@ -83,7 +83,7 @@ template <
typename InstructionShape,
/// Number of stages used in the pipelined mainloop
int Stages,
/// Operation perfomed by GEMM
/// Operation performed by GEMM
typename Operator,
typename Enable_ = void>
struct FindDefaultMma {

2
examples/41_fused_multi_head_attention/gemm/mma_from_smem.h
View File

@@ -522,7 +522,7 @@ class MmaPipelinedFromSharedMemory : public MmaBaseFromSharedMemory<
// For API compatibility with MmaMultistageFromSharedMemory
// but not supported as it worsens perf: older gpus < sm80 don't
// support async tranfers and have to waste registers
// support async transfers and have to waste registers
CUTLASS_DEVICE
void set_prologue_done(bool value) {}
CUTLASS_DEVICE

2
examples/41_fused_multi_head_attention/iterators/default_warp_iterator_from_smem.h
View File

@@ -29,7 +29,7 @@
*
**************************************************************************************************/
/*! \file
\brief Instanciates the right WarpIterator to read from shared memory
\brief Instantiates the right WarpIterator to read from shared memory
The class `DefaultWarpIteratorAFromSharedMemory` is useful when reading
data dumped with `B2bGemm::accumToSmem`.
*/

2
examples/41_fused_multi_head_attention/iterators/predicated_tile_iterator_residual_last.h
View File

@@ -86,7 +86,7 @@ namespace threadblock {
/// To be efficient, this assumes the iterator will be dereferenced and advanced
/// at least once outside any looping structure to minimize integer arithmetic.
///
/// Acceses out of bounds are safe so long as `clear_mask()` is called prior to
/// Accesses out of bounds are safe so long as `clear_mask()` is called prior to
/// dereferencing the iterator.
///
///