llama_cpp 0.1.2 → 0.1.3

data/ext/llama_cpp/src/ggml-opencl.h

@@ -1,23 +1,21 @@
 #pragma once
 
+#include "ggml.h"
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
 void ggml_cl_init(void);
 
-enum ggml_blas_order {
-    GGML_BLAS_ORDER_ROW_MAJOR = 101,
-    GGML_BLAS_ORDER_COLUMN_MAJOR = 102,
-};
+bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
 
-enum ggml_blas_op {
-    GGML_BLAS_OP_N = 111,
-    GGML_BLAS_OP_T = 112,
-    GGML_BLAS_OP_C = 113,
-};
+void * ggml_cl_host_malloc(size_t size);
+void   ggml_cl_host_free(void * ptr);
 
-void ggml_cl_sgemm_wrapper(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype);
+void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus
 }

data/ext/llama_cpp/src/ggml.c

@@ -9431,7 +9431,7 @@ static void ggml_compute_forward_rms_norm_back(
 
 // ggml_compute_forward_mul_mat
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
 // helper function to determine if it is better to use BLAS or not
 // for large matrices, BLAS is faster
 static bool ggml_compute_forward_mul_mat_use_blas(
@@ -9472,7 +9472,7 @@ static void ggml_compute_forward_mul_mat_f32(
     const int64_t ne02 = src0->ne[2];
     const int64_t ne03 = src0->ne[3];
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     const int64_t ne10 = src1->ne[0];
 #endif
     const int64_t ne11 = src1->ne[1];
@@ -9536,9 +9536,16 @@ static void ggml_compute_forward_mul_mat_f32(
         }
         return;
     }
+#elif defined(GGML_USE_CLBLAST)
+    if (ggml_cl_can_mul_mat(src0, src1, dst)) {
+        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
+            ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
+        }
+        return;
+    }
 #endif
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         if (params->ith != 0) {
             return;
@@ -9558,21 +9565,11 @@ static void ggml_compute_forward_mul_mat_f32(
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
                 float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
-#if defined(GGML_USE_CLBLAST)
-                // zT = y * xT
-                ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne10,
-                        0.0f,    d, ne01,
-                        GGML_TYPE_F32);
-#else
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
                                  x, ne00,
                         0.0f,    d, ne01);
-#endif
             }
         }
         //printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
@@ -9711,9 +9708,16 @@ static void ggml_compute_forward_mul_mat_f16_f32(
         }
         return;
     }
+#elif defined(GGML_USE_CLBLAST)
+    if (ggml_cl_can_mul_mat(src0, src1, dst)) {
+        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
+            ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
+        }
+        return;
+    }
 #endif
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         GGML_ASSERT(nb10 == sizeof(float));
 
@@ -9743,20 +9747,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                     assert(id*sizeof(float) <= params->wsize);
                 }
 
-#if defined(GGML_USE_CLBLAST)
-                const float * x = wdata;
-                const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
-
-                float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
-
-                // zT = y * xT
-                ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne10,
-                        0.0f,    d, ne01,
-                        GGML_TYPE_F32);
-#else
                 const float * x = wdata;
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
 
@@ -9768,7 +9758,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                         1.0f,    y, ne10,
                                  x, ne00,
                         0.0f,    d, ne01);
-#endif
             }
         }
 
@@ -9931,9 +9920,16 @@ static void ggml_compute_forward_mul_mat_q_f32(
         }
         return;
     }
+#elif defined(GGML_USE_CLBLAST)
+    if (ggml_cl_can_mul_mat(src0, src1, dst)) {
+        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
+            ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
+        }
+        return;
+    }
 #endif
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         if (params->ith != 0) {
             return;
@@ -9956,9 +9952,6 @@ static void ggml_compute_forward_mul_mat_q_f32(
 
                 float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
-#if defined(GGML_USE_CLBLAST)
-                const void* x = (char *) src0->data + i03*nb03 + i02*nb02;
-#else
                 {
                     size_t id = 0;
                     for (int64_t i01 = 0; i01 < ne01; ++i01) {
@@ -9970,23 +9963,12 @@ static void ggml_compute_forward_mul_mat_q_f32(
                 }
 
                 const float * x = wdata;
-#endif
 
-#if defined(GGML_USE_CLBLAST)
-                // zT = y * xT
-                ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne10,
-                        0.0f,    d, ne01,
-                        type);
-#else
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
                                  x, ne00,
                         0.0f,    d, ne01);
-#endif
             }
         }
 
@@ -14165,9 +14147,16 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                             cur = ggml_cuda_mul_mat_get_wsize(node->src0, node->src1, node);
                         }
                         else
+#elif defined(GGML_USE_CLBLAST)
+                        if (ggml_cl_can_mul_mat(node->src0, node->src1, node)) {
+                            node->n_tasks = 1; // TODO: this actually is doing nothing
+                                                //       the threads are still spinning
+                            cur = ggml_cl_mul_mat_get_wsize(node->src0, node->src1, node);
+                        }
+                        else
 #endif
                         if (node->src0->type == GGML_TYPE_F16 && node->src1->type == GGML_TYPE_F32) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1; // TODO: this actually is doing nothing
                                                    //       the threads are still spinning
@@ -14181,13 +14170,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 #endif
                         } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
                             cur = 0;
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
                             }
 #endif
                         } else if (ggml_is_quantized(node->src0->type) && node->src1->type == GGML_TYPE_F32) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);

data/ext/llama_cpp/src/ggml.h

@@ -249,6 +249,7 @@ extern "C" {
     enum ggml_backend {
         GGML_BACKEND_CPU = 0,
         GGML_BACKEND_CUDA = 1,
+        GGML_BACKEND_CL = 2,
     };
 
     // model file types

data/ext/llama_cpp/src/llama.cpp

@@ -12,6 +12,8 @@
 #include "ggml.h"
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
+#elif defined(GGML_USE_CLBLAST)
+#include "ggml-opencl.h"
 #endif
 
 #include <array>
@@ -1092,7 +1094,7 @@ static void llama_model_load_internal(
             fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__);
         }
         fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
-#else
+#elif !defined(GGML_USE_CLBLAST)
         (void) n_gpu_layers;
 #endif
     }
@@ -1125,7 +1127,33 @@ static void llama_model_load_internal(
             done_size += lt.size;
         }
     }
-#endif // GGML_USE_CUBLAS
+#elif defined(GGML_USE_CLBLAST)
+    {
+        const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
+
+        fprintf(stderr, "ggml_opencl: offloading %d layers to GPU\n", n_gpu);
+
+        size_t vram_total = 0;
+
+        for (int i = 0; i < n_gpu; ++i) {
+            const auto & layer = model.layers[i];
+
+            ggml_cl_transform_tensor(layer.wq); vram_total += ggml_nbytes(layer.wq);
+            ggml_cl_transform_tensor(layer.wk); vram_total += ggml_nbytes(layer.wk);
+            ggml_cl_transform_tensor(layer.wv); vram_total += ggml_nbytes(layer.wv);
+            ggml_cl_transform_tensor(layer.wo); vram_total += ggml_nbytes(layer.wo);
+            ggml_cl_transform_tensor(layer.w1); vram_total += ggml_nbytes(layer.w1);
+            ggml_cl_transform_tensor(layer.w2); vram_total += ggml_nbytes(layer.w2);
+            ggml_cl_transform_tensor(layer.w3); vram_total += ggml_nbytes(layer.w3);
+        }
+        if (n_gpu_layers > (int) hparams.n_layer) {
+            fprintf(stderr, "ggml_opencl: offloading output layer to GPU\n");
+            ggml_cl_transform_tensor(model.output); vram_total += ggml_nbytes(model.output);
+        }
+
+        fprintf(stderr, "ggml_opencl: total VRAM used: %zu MB\n", vram_total / 1024 / 1024);
+    }
+#endif
 
     if (progress_callback) {
         progress_callback(1.0f, progress_callback_user_data);

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.1.2'
+  VERSION = '0.1.3'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'master-265db98'
+  LLAMA_CPP_VERSION = 'master-66874d4'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.1.2
+  version: 0.1.3
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-05-22 00:00:00.000000000 Z
+date: 2023-05-26 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email:
@@ -27,7 +27,7 @@ files:
 - ext/llama_cpp/llama_cpp.h
 - ext/llama_cpp/src/LICENSE
 - ext/llama_cpp/src/ggml-cuda.h
-- ext/llama_cpp/src/ggml-opencl.c
+- ext/llama_cpp/src/ggml-opencl.cpp
 - ext/llama_cpp/src/ggml-opencl.h
 - ext/llama_cpp/src/ggml.c
 - ext/llama_cpp/src/ggml.h