Review: fix formatting, remove useless type conversion, fix naming for bools

Author: am17an

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1861,7 +1861,7 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
 
         const auto convert_func = traits::get_nc_converter(src1->type);
         GGML_ASSERT(convert_func != nullptr);
-        convert_func((const void*)((const char*)src1->data), src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+        convert_func(src1->data, src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
         src1_ptr = src1_alloc.get();
         s11 = ne10;
         s12 = ne11*s11;
@@ -1922,7 +1922,7 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
                 ne01, ne11, ne10,
                 alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
                        src1_ptr, cu_data_type_b, s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1954,7 +1954,7 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
                 ne01, ne11, ne10,
                 alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, nb01/nb00,
                        (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type,   ne0,
                 ne23,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -2033,10 +2033,11 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
+    //TODO update for generic tensor parallelism
     const int cc                     = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    bool can_use_batched_cublas_f16  = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
-    bool can_use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
-    bool can_use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
+    bool use_batched_cublas_f16  = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
+    bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
+    bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
 
     if (!split && use_mul_mat_vec) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
@@ -2046,7 +2047,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_q) {
         ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
-    } else if (!split && (can_use_batched_cublas_f16 || can_use_batched_cublas_bf16 || can_use_batched_cublas_f32)
+    } else if (!split && (use_batched_cublas_f16 || use_batched_cublas_bf16 || use_batched_cublas_f32)
         && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);