mrfakename commited on
Commit
cd5cb58
·
1 Parent(s): 669c9d8

Pin compatible CUTLASS DSL

Browse files
Files changed (1) hide show
  1. requirements.txt +2 -1
requirements.txt CHANGED
@@ -22,7 +22,8 @@ comfy-kitchen==0.2.26
22
  # NVIDIA Sol-Engine's released sparse-attention package, pinned to the H3 release. The native SM120 path uses
23
  # CuTe DSL; the package retains its portable Triton fallback if that optional backend cannot initialize.
24
  sol-attn @ git+https://github.com/NVlabs/Sana.git@46031940ba8af5d18054217e571149579424c0b1#subdirectory=techniques/sparse_backends
25
- nvidia-cutlass-dsl>=4.5.0
 
26
  cuda-python>=13.0
27
  # Sol-Attn compiles its CuTe callable with `--enable-tvm-ffi`, but its package currently leaves this transitive
28
  # runtime undeclared. Without it the SM120 path compiles up to dispatch and then falls back to dense attention.
 
22
  # NVIDIA Sol-Engine's released sparse-attention package, pinned to the H3 release. The native SM120 path uses
23
  # CuTe DSL; the package retains its portable Triton fallback if that optional backend cannot initialize.
24
  sol-attn @ git+https://github.com/NVlabs/Sana.git@46031940ba8af5d18054217e571149579424c0b1#subdirectory=techniques/sparse_backends
25
+ # Sol-Engine's Aug-03 H3 kernel targets the 4.5 `cute.math.fmax` signature; 4.6 changed that API.
26
+ nvidia-cutlass-dsl==4.5.3
27
  cuda-python>=13.0
28
  # Sol-Attn compiles its CuTe callable with `--enable-tvm-ffi`, but its package currently leaves this transitive
29
  # runtime undeclared. Without it the SM120 path compiles up to dispatch and then falls back to dense attention.