Revert "CUDA: fuse ffn_up*unary_op(ffn_gate) for MMVQ (V2) (ikawrakow#864)"

This reverts commit f76e985.

Author: Nexesenex

ggml/src/CMakeLists.txt

@@ -357,8 +357,6 @@ if (GGML_CUDA)
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
         file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB   SRCS "ggml-cuda/template-instances/mmvq-instance*.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
 
         if (GGML_CUDA_FA_ALL_QUANTS)
             file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")

ggml/src/ggml-cuda.cu

@@ -6,7 +6,127 @@
 
 #include "common.cuh"
 
-struct mmvq_args;
+void mul_mat_vec_iq2_k_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
 
-void iqk_mul_mat_vec_q(ggml_type type, const mmvq_args & args, cudaStream_t stream);
+void mul_mat_vec_iq3_k_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
 
+void mul_mat_vec_iq2_kl_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq3_ks_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq4_k_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq5_k_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq5_ks_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq6_k_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq4_ks_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq4_kss_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq2_ks_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq1_bn_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq2_bn_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq2_k_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq3_k_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq4_k_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq5_k_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq4_ks_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq5_ks_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq1_s_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq1_m_r4_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq1_kt_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq2_kt_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq3_kt_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);
+
+void mul_mat_vec_iq4_kt_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const char * ids_data,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst,
+    const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream);