nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

change block size back to 64x64

Browse Source
This commit is contained in:
RandNMR73
2025-12-24 23:36:37 +00:00
parent f6080f7cee
commit f2eb855c2a
2 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastvideo/attention/backends/sageattn/api.py
+2 -2
View File
@@ -24,8 +24,8 @@ import fp4quant_cuda
# Centralized block size configuration for sageattn_blackwell kernels
# These should match the values in fastvideo/attention/backends/sageattn/blackwell/block_config.h
BLOCK_M = 128 # Block size for M dimension (query sequence length)
BLOCK_N = 128 # Block size for N dimension (key/value sequence length)
BLOCK_M = 64 # Block size for M dimension (query sequence length)
BLOCK_N = 64 # Block size for N dimension (key/value sequence length)
@triton.jit
fastvideo/attention/backends/sageattn/blackwell/block_config.h
+2 -2
View File
@@ -20,9 +20,9 @@
// Block sizes for M and N dimensions
namespace flash {
// Block size for M dimension (query sequence length)
static constexpr int BLOCK_M = 128;
static constexpr int BLOCK_M = 64;
// Block size for N dimension (key/value sequence length)
static constexpr int BLOCK_N = 128;
static constexpr int BLOCK_N = 64;
}