diff --git a/llama/patches/0036-backport-kernels-for-gemma4.patch b/llama/patches/0036-backport-kernels-for-gemma4.patch index 86123ccc5..7c2186e08 100644 --- a/llama/patches/0036-backport-kernels-for-gemma4.patch +++ b/llama/patches/0036-backport-kernels-for-gemma4.patch @@ -16,9 +16,11 @@ Subject: [PATCH] backport kernels for gemma4 ...attn-mma-f16-instance-ncols1_4-ncols2_8.cu | 1 + ...attn-mma-f16-instance-ncols1_8-ncols2_4.cu | 1 + ...attn-mma-f16-instance-ncols1_8-ncols2_8.cu | 1 + + .../fattn-tile-instance-dkq512-dv512.cu | 5 +++ ggml/src/ggml-metal/ggml-metal-device.m | 1 + ggml/src/ggml-metal/ggml-metal.metal | 19 ++++++++++ - 14 files changed, 96 insertions(+), 10 deletions(-) + 15 files changed, 101 insertions(+), 10 deletions(-) + create mode 100644 ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh index 3dea2205e..b4282d3cb 100644 @@ -316,6 +318,17 @@ index 680a13ca6..4bc60d62f 100644 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 8); DECL_FATTN_MMA_F16_CASE(256, 256, 8, 8); +DECL_FATTN_MMA_F16_CASE(512, 512, 8, 8); +diff --git a/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu b/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +new file mode 100644 +index 000000000..7c61d8d2e +--- /dev/null ++++ b/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +@@ -0,0 +1,5 @@ ++// This file has been autogenerated by generate_cu_files.py, do not edit manually. ++ ++#include "../fattn-tile.cuh" ++ ++DECL_FATTN_TILE_CASE(512, 512); diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m index 4e5acfbe5..11457f2b1 100644 --- a/ggml/src/ggml-metal/ggml-metal-device.m diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu b/ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu new file mode 100644 index 000000000..7c61d8d2e --- /dev/null +++ b/ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-tile.cuh" + +DECL_FATTN_TILE_CASE(512, 512);