add additional hipblas conditions for cublas

AAbushady · Jun 21, 2023 · 222cbbb · 222cbbb
1 parent e1f9581
commit 222cbbb
Show file tree

Hide file tree

Showing 9 changed files with 36 additions and 36 deletions.
diff --git a/examples/common.cpp b/examples/common.cpp
@@ -304,7 +304,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
             params.main_gpu = std::stoi(argv[i]);
 #else
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
@@ -314,7 +314,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
             std::string arg_next = argv[i];
 
             // split string by , and /
@@ -334,7 +334,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--low-vram" || arg == "-lv") {
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
             params.low_vram = true;
 #else
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
@@ -414,7 +414,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         exit(1);
     }
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
     if (!params.lora_adapter.empty() && params.n_gpu_layers > 0) {
         fprintf(stderr, "%s: error: the simultaneous use of LoRAs and GPU acceleration is not supported", __func__);
         exit(1);

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
@@ -560,7 +560,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
             std::string arg_next = argv[i];
 
             // split string by , and /
@@ -583,7 +583,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
         }
         else if (arg == "--low-vram" || arg == "-lv")
         {
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
             params.low_vram = true;
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
@@ -594,7 +594,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
             params.main_gpu = std::stoi(argv[i]);
 #else
             LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {});

diff --git a/ggml.c b/ggml.c
@@ -161,7 +161,7 @@ inline static void* ggml_aligned_malloc(size_t size) {
 #endif
 #elif defined(GGML_USE_OPENBLAS)
 #include <cblas.h>
-#elif defined(GGML_USE_CUBLAS) | defined(GGML_USE_HIPBLAS)
+#elif defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
 #include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
@@ -4116,7 +4116,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
         ggml_init_cublas();
 #elif defined(GGML_USE_CLBLAST)
         ggml_cl_init();
@@ -14875,7 +14875,7 @@ static void ggml_compute_forward_cross_entropy_loss_back(
 static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
     GGML_ASSERT(params);
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
     bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
     if (skip_cpu) {
         return;
@@ -16362,7 +16362,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 
                         size_t cur = 0;
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
                         if (ggml_cuda_can_mul_mat(node->src0, node->src1, node)) {
                             node->n_tasks = 1; // TODO: this actually is doing nothing
                                                 //       the threads are still spinning
@@ -18637,7 +18637,7 @@ int ggml_cpu_has_wasm_simd(void) {
 }
 
 int ggml_cpu_has_blas(void) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS) || defined(GGML_USE_CLBLAST)
     return 1;
 #else
     return 0;

diff --git a/llama-util.h b/llama-util.h
@@ -441,7 +441,7 @@ struct llama_buffer {
     llama_buffer& operator=(llama_buffer&&) = delete;
 };
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
 #include "ggml-cuda.h"
 struct llama_ctx_buffer {
     uint8_t * addr = NULL;

diff --git a/llama.cpp b/llama.cpp
@@ -10,7 +10,7 @@
 #include "llama.h"
 
 #include "ggml.h"
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
 #include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
@@ -175,7 +175,7 @@ struct llama_kv_cache {
             ggml_free(ctx);
         }
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
         ggml_cuda_free_data(k);
         ggml_cuda_free_data(v);
 #endif // GGML_USE_CUBLAS
@@ -220,7 +220,7 @@ struct llama_model {
             ggml_free(ctx);
         }
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cuda_free_data(tensors_by_name[i].second);
         }
@@ -791,7 +791,7 @@ struct llama_model_loader {
                         lmlock->grow_to(lock_size);
                     }
                     break;
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
                 case GGML_BACKEND_GPU:
                 case GGML_BACKEND_GPU_SPLIT:
                     ggml_cuda_transform_tensor(lt.data, lt.ggml_tensor);
@@ -911,7 +911,7 @@ static bool kv_cache_init(
     ggml_set_name(cache.v, "cache_v");
 
     (void) n_gpu_layers;
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
     if (n_gpu_layers > n_layer + 1) {
         ggml_cuda_assign_buffers_no_scratch(cache.v);
     }
@@ -1141,7 +1141,7 @@ static void llama_model_load_internal(
     }
 
     (void) main_gpu;
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
     fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__);
     ggml_cuda_set_main_device(main_gpu);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU
@@ -1252,7 +1252,7 @@ static void llama_model_load_internal(
 
         (void) vram_scratch;
         (void) n_batch;
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
         if (low_vram) {
             fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
             ggml_cuda_set_scratch_size(0); // disable scratch
@@ -1265,7 +1265,7 @@ static void llama_model_load_internal(
             }
         }
 #endif // GGML_USE_CUBLAS
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
         fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
@@ -1305,7 +1305,7 @@ static void llama_model_load_internal(
     }
 
     (void) tensor_split;
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
     {
         ggml_cuda_set_tensor_split(tensor_split);
     }
@@ -1425,7 +1425,7 @@ static bool llama_eval_internal(
     offload_func_t offload_func_kq = llama_nop;
     offload_func_t offload_func_v  = llama_nop;
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
         if (n_gpu_layers > n_layer) {
             offload_func_nr = ggml_cuda_assign_buffers;
         }
@@ -1440,7 +1440,7 @@ static bool llama_eval_internal(
     for (int il = 0; il < n_layer; ++il) {
         offload_func_t offload_func = llama_nop;
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
         if (il >= i_gpu_start) {
             offload_func = ggml_cuda_assign_buffers;
         }

diff --git a/llama.h b/llama.h
@@ -2,7 +2,7 @@
 #define LLAMA_H
 
 #include "ggml.h"
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
 #include "ggml-cuda.h"
 #define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
 #else
@@ -38,7 +38,7 @@
 #define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif

diff --git a/otherarch/ggml_v2.c b/otherarch/ggml_v2.c
@@ -139,7 +139,7 @@ inline static void* ggml_v2_aligned_malloc(size_t size) {
 #include <Accelerate/Accelerate.h>
 #elif defined(GGML_USE_OPENBLAS)
 #include <cblas.h>
-#elif defined(GGML_USE_CUBLAS)
+#elif defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
 #include "ggml_v2-cuda.h"
 #endif
 #if defined(GGML_USE_CLBLAST)
@@ -3894,7 +3894,7 @@ struct ggml_v2_context * ggml_v2_init(struct ggml_v2_init_params params) {
             GGML_V2_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
         ggml_v2_init_cublas();
 #elif defined(GGML_USE_CLBLAST)
         if(quants_unshuffled)
@@ -9448,7 +9448,7 @@ static void ggml_v2_compute_forward_mul_mat_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
     if (ggml_v2_cuda_can_mul_mat(src0, src1, dst)) {
         if (params->ith == 0 && params->type == GGML_V2_TASK_COMPUTE) {
             ggml_v2_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize);
@@ -9642,7 +9642,7 @@ static void ggml_v2_compute_forward_mul_mat_f16_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
     if (ggml_v2_cuda_can_mul_mat(src0, src1, dst)) {
         if (params->ith == 0 && params->type == GGML_V2_TASK_COMPUTE) {
             ggml_v2_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize);
@@ -9881,7 +9881,7 @@ static void ggml_v2_compute_forward_mul_mat_q_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
     if (ggml_v2_cuda_can_mul_mat(src0, src1, dst)) {
         if (params->ith == 0 && params->type == GGML_V2_TASK_COMPUTE) {
             ggml_v2_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize);
@@ -14061,7 +14061,7 @@ void ggml_v2_graph_compute(struct ggml_v2_context * ctx, struct ggml_v2_cgraph *
 
                         size_t cur = 0;
 
-#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS)
                         if (ggml_v2_cuda_can_mul_mat(node->src0, node->src1, node)) {
                             node->n_tasks = 1; // TODO: this actually is doing nothing
                                                 //       the threads are still spinning
@@ -15559,7 +15559,7 @@ int ggml_v2_cpu_has_wasm_simd(void) {
 }
 
 int ggml_v2_cpu_has_blas(void) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_HIPBLAS) || defined(GGML_USE_CLBLAST)
     return 1;
 #else
     return 0;

diff --git a/otherarch/llama_v2-util.h b/otherarch/llama_v2-util.h
@@ -415,7 +415,7 @@ struct llama_v2_buffer {
     llama_v2_buffer& operator=(llama_v2_buffer&&) = delete;
 };
 
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
 #include "ggml_v2-cuda.h"
 struct llama_v2_ctx_buffer {
     uint8_t * addr = NULL;

diff --git a/otherarch/llama_v2.cpp b/otherarch/llama_v2.cpp
@@ -9,7 +9,7 @@
 #include "llama_v2.h"
 
 #include "ggml_v2.h"
-#ifdef GGML_USE_CUBLAS
+#if defined GGML_USE_CUBLAS || defined GGML_USE_HIPBLAS
 #include "ggml_v2-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml_v2-opencl.h"
@@ -3088,4 +3088,4 @@ std::vector<llama_token> llama_v2_tokenize(struct llama_v2_context * ctx, const
     res.resize(n);
 
     return res;
-}
+}