update

768ab4df · xuxzh1 · 44f79c7e · 768ab4df · 768ab4df · 768ab4df
Commit 768ab4df authored Nov 13, 2024 by xuxzh1 🎱
20 changed files
--- a/llm/llama.cpp/ggml/src/ggml-cuda/mmq.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/mmq.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/mmq.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/mmq.cuh
@@ -2580,7 +2580,7 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
    __launch_bounds__(WARP_SIZE*nwarps, 2)
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-static __global__ __launch_bounds__(1024) void mul_mat_q(
+static __global__ __launch_bounds__(256) void mul_mat_q(
    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
    const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {

--- a/llm/llama.cpp/ggml/src/ggml-cuda/mmvq.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/mmvq.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/mmvq.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/mmvq.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/norm.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/norm.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/norm.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/norm.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/pad.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/pad.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/pad.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/pad.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/pool2d.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/pool2d.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/pool2d.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/pool2d.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/quantize.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/quantize.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/quantize.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/quantize.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/rope.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/rope.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/rope.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/rope.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/scale.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/scale.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/scale.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/scale.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/softmax.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/softmax.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/softmax.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/softmax.cuh
--- a/llm/llama.cpp/ggml/src/ggml-cuda/sumrows.cu
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/sumrows.cu
--- a/llm/llama.cpp/ggml/src/ggml-cuda/sumrows.cuh
+++ b/llm/llama.cpp/ggml/src/ggml-cuda/sumrows.cuh