zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-07-25 03:31:17 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

sdpa specialization for head dim 256 (#2007)

Browse Source
This commit is contained in:
Awni Hannun 2025-03-27 19:31:25 -07:00 committed by GitHub
parent a6b5d6e759
commit bc62932984
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
mlx/backend/metal/kernels/scaled_dot_product_attention.metal
View File

@ -32,9 +32,11 @@ using namespace metal;
instantiate_sdpa_vector(type, 64, 64) \
instantiate_sdpa_vector(type, 96, 96) \
instantiate_sdpa_vector(type, 128, 128) \
instantiate_sdpa_vector(type, 256, 256) \
instantiate_sdpa_vector_aggregation(type, 64) \
instantiate_sdpa_vector_aggregation(type, 96) \
instantiate_sdpa_vector_aggregation(type, 128)
instantiate_sdpa_vector_aggregation(type, 128) \
instantiate_sdpa_vector_aggregation(type, 256)
instantiate_sdpa_vector_heads(float)
instantiate_sdpa_vector_heads(bfloat16_t)

3
mlx/fast.cpp
View File

@ -720,7 +720,8 @@ array scaled_dot_product_attention(
const bool sdpa_vector_supported_head_dim =
query_head_dim == value_head_dim &&
(query_head_dim == 64 || query_head_dim == 96 || query_head_dim == 128);
(query_head_dim == 64 || query_head_dim == 96 || query_head_dim == 128 ||
query_head_dim == 256);
const bool sdpa_full_supported_head_dim = query_head_dim == value_head_dim &&
(query_head_dim == 64 || query_head_dim == 80 || query_head_dim == 128);