LLM inference in C/C++
C++ ggml
cuda : fix compiler warnings in mmq-load-tiles.cuh (#30443)
Adrien Gallouëtbfe65e1dfa8f
diff --git a/ggml/src/ggml-cuda/mmq-load-tiles.cuh b/ggml/src/ggml-cuda/mmq-load-tiles.cuhindex 0e59c0347b..017b93b546 100644--- a/ggml/src/ggml-cuda/mmq-load-tiles.cuh+++ b/ggml/src/ggml-cuda/mmq-load-tiles.cuh@@ -9,9 +9,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -100,9 +100,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -189,9 +189,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -252,9 +252,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -315,9 +315,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -395,9 +395,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -473,9 +473,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K);#else@@ -539,9 +539,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -600,9 +600,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -713,9 +713,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);#else@@ -780,7 +780,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);#pragma unroll- for (int l = 0; l < sizeof(int); ++l) {+ for (int l = 0; l < int(sizeof(int)); ++l) {x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);}}@@ -824,9 +824,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -948,9 +948,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K);@@ -1038,9 +1038,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1100,9 +1100,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1164,9 +1164,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1229,9 +1229,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1297,9 +1297,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1361,9 +1361,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1430,9 +1430,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1497,9 +1497,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1566,9 +1566,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_constexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, GGML_PREC_Q8);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else@@ -1675,9 +1675,9 @@ template <ggml_type type, int J, bool fallback, ggml_prec prec_src1> static __deconstexpr int warp_size = ggml_cuda_get_physical_warp_size();constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1) / warp_size;constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, prec_src1);- constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, prec_src1);#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback, prec_src1);int * x_qs = (int *) x_tile;float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);#else