llama : consistently catch and throw only exceptions deriving from std::exception (#1599)

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

ggml-metal.m

@@ -195,14 +195,25 @@ bool ggml_metal_add_buffer(
             }
         }
 
+        size_t page_size = getpagesize();
+        size_t aligned_size = size;
+        if ((aligned_size % page_size) != 0) {
+            aligned_size += (page_size - (aligned_size % page_size));
+        }
+
         ctx->buffers[ctx->n_buffers].name = name;
         ctx->buffers[ctx->n_buffers].data = data;
         ctx->buffers[ctx->n_buffers].size = size;
-        ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytes:data length:size options:MTLResourceStorageModeShared];
+        ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:aligned_size options:MTLResourceStorageModeShared deallocator:nil];
+
+        if (ctx->buffers[ctx->n_buffers].metal == nil) {
+            fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
+            return false;
+        } else {
+            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
+        }
 
         ++ctx->n_buffers;
-
-        fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, size / 1024.0 / 1024.0);
     }
 
     return true;

ggml.c

@@ -22,6 +22,10 @@
 #include <float.h>
 #include <limits.h>
 
+#ifdef GGML_USE_METAL
+#include <unistd.h>
+#endif
+
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
 #ifndef static_assert
@@ -122,7 +126,11 @@ typedef void* thread_ret_t;
 #else
 inline static void* ggml_aligned_malloc(size_t size) {
     void* aligned_memory = NULL;
+#ifdef GGML_USE_METAL
+    int result = posix_memalign(&aligned_memory, getpagesize(), size);
+#else
     int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
+#endif
     if (result != 0) {
         // Handle allocation failure
         return NULL;

llama-util.h

@@ -405,13 +405,29 @@ struct llama_buffer {
     llama_buffer() = default;
 
     void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
         delete[] addr;
         addr = new uint8_t[len];
+#endif
         size = len;
     }
 
     ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
         delete[] addr;
+#endif
+        addr = NULL;
     }
 
     // disable copy and move

llama.cpp

@@ -53,7 +53,6 @@ enum e_model {
     MODEL_65B,
 };
 
-
 static const size_t MB = 1024*1024;
 
 // computed for n_ctx == 2048
@@ -290,15 +289,15 @@ template <typename T>
 static T checked_mul(T a, T b) {
     T ret = a * b;
     if (a != 0 && ret / a != b) {
-        throw format("overflow multiplying %llu * %llu",
-                     (unsigned long long) a, (unsigned long long) b);
+        throw std::runtime_error(format("overflow multiplying %llu * %llu",
+                     (unsigned long long) a, (unsigned long long) b));
     }
     return ret;
 }
 
 static size_t checked_div(size_t a, size_t b) {
     if (b == 0 || a % b != 0) {
-        throw format("error dividing %zu / %zu", a, b);
+        throw std::runtime_error(format("error dividing %zu / %zu", a, b));
     }
     return a / b;
 }
@@ -362,7 +361,7 @@ struct llama_load_tensor {
         const auto & first_shard = shards.at(0);
         for (const auto & shard : shards) {
             if (shard.type != first_shard.type) {
-                throw format("inconsistent tensor shard type in '%s'", name.c_str());
+                throw std::runtime_error(format("inconsistent tensor shard type in '%s'", name.c_str()));
             }
         }
         type = first_shard.type;
@@ -385,8 +384,8 @@ struct llama_load_tensor {
         const auto & first_shard = shards.at(0);
         for (const auto & shard : shards) {
             if (shard.ne != first_shard.ne) {
-                throw format("inconsistent tensor shard shape in '%s': first was %s, other was %s",
-                             name.c_str(), llama_format_tensor_shape(first_shard.ne).c_str(), llama_format_tensor_shape(shard.ne).c_str());
+                throw std::runtime_error(format("inconsistent tensor shard shape in '%s': first was %s, other was %s",
+                             name.c_str(), llama_format_tensor_shape(first_shard.ne).c_str(), llama_format_tensor_shape(shard.ne).c_str()));
             }
         }
         ne = first_shard.ne;
@@ -464,8 +463,8 @@ struct llama_file_loader {
                 }
         }
 
-        throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?",
-                     magic, version);
+        throw std::runtime_error(format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?",
+                     magic, version));
     }
     void read_hparams() {
         hparams.n_vocab = file.read_u32();
@@ -505,7 +504,7 @@ struct llama_file_loader {
             file.read_raw(shard.ne.data(), sizeof(shard.ne[0]) * n_dims);
             std::string name = file.read_string(name_len);
             if (n_dims < 1 || n_dims > 2) {
-                throw format("llama.cpp: tensor '%s' should not be %u-dimensional", name.c_str(), n_dims);
+                throw std::runtime_error(format("llama.cpp: tensor '%s' should not be %u-dimensional", name.c_str(), n_dims));
             }
             switch (shard.type) {
                 case GGML_TYPE_F32:
@@ -522,7 +521,7 @@ struct llama_file_loader {
                 case GGML_TYPE_Q6_K:
                     break;
                 default: {
-                    throw format("unrecognized tensor type %u\n", shard.type);
+                    throw std::runtime_error(format("unrecognized tensor type %u\n", shard.type));
                 }
             }
 
@@ -631,7 +630,7 @@ struct llama_model_loader {
             auto * ith_file = new llama_file_loader(fname.c_str(), i, tensors_map);
             file_loaders.emplace_back(ith_file);
             if (ith_file->hparams != first_file->hparams) {
-                throw format("llama.cpp: hparams inconsistent between files");
+                throw std::runtime_error(format("llama.cpp: hparams inconsistent between files"));
             }
         }
         if (!llama_mmap::SUPPORTED) {
@@ -661,7 +660,7 @@ struct llama_model_loader {
     uint32_t guess_n_parts() const {
         auto it = tensors_map.name_to_idx.find("tok_embeddings.weight");
         if (it == tensors_map.name_to_idx.end()) {
-            throw std::string("missing tok_embeddings.weight");
+            throw std::runtime_error(std::string("missing tok_embeddings.weight"));
         }
         const llama_load_tensor & lt = tensors_map.tensors.at(it->second);
         return file_loaders.at(0)->hparams.n_embd / lt.shards.at(0).ne.at(0);
@@ -678,12 +677,12 @@ struct llama_model_loader {
     struct ggml_tensor * get_tensor(const std::string & name, const std::vector<uint32_t> & ne, ggml_backend backend) {
         auto it = tensors_map.name_to_idx.find(name);
         if (it == tensors_map.name_to_idx.end()) {
-            throw format("llama.cpp: tensor '%s' is missing from model", name.c_str());
+            throw std::runtime_error(std::runtime_error(format("llama.cpp: tensor '%s' is missing from model", name.c_str())));
         }
         llama_load_tensor & lt = tensors_map.tensors.at(it->second);
         if (lt.ne != ne) {
-            throw format("llama.cpp: tensor '%s' has wrong shape; expected %s, got %s",
-                         name.c_str(), llama_format_tensor_shape(ne).c_str(), llama_format_tensor_shape(lt.ne).c_str());
+            throw std::runtime_error(format("llama.cpp: tensor '%s' has wrong shape; expected %s, got %s",
+                         name.c_str(), llama_format_tensor_shape(ne).c_str(), llama_format_tensor_shape(lt.ne).c_str()));
         }
 
         return get_tensor_for(lt, backend);
@@ -707,7 +706,7 @@ struct llama_model_loader {
 
     void done_getting_tensors() const {
         if (num_ggml_tensors_created != tensors_map.tensors.size()) {
-            throw std::string("llama.cpp: file contained more tensors than expected");
+            throw std::runtime_error(std::string("llama.cpp: file contained more tensors than expected"));
         }
     }
 
@@ -995,7 +994,7 @@ static void llama_model_load_internal(
         if (hparams.ftype != LLAMA_FTYPE_ALL_F32     &&
             hparams.ftype != LLAMA_FTYPE_MOSTLY_F16  &&
             hparams.ftype != LLAMA_FTYPE_MOSTLY_Q8_0) {
-            throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1405)");
+            throw std::runtime_error(format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1405)"));
         }
     }
 
@@ -1003,7 +1002,7 @@ static void llama_model_load_internal(
         if (hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
             hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_1 ||
             hparams.ftype == LLAMA_FTYPE_MOSTLY_Q8_0) {
-            throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1508)");
+            throw std::runtime_error(format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1508)"));
         }
     }
 
@@ -1034,7 +1033,7 @@ static void llama_model_load_internal(
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
-            throw format("ggml_init() failed");
+            throw std::runtime_error(format("ggml_init() failed"));
         }
     }
 
@@ -1215,8 +1214,8 @@ static bool llama_model_load(
         llama_model_load_internal(fname, lctx, n_ctx, n_gpu_layers, memory_type, use_mmap, use_mlock,
                                   vocab_only, progress_callback, progress_callback_user_data);
         return true;
-    } catch (const std::string & err) {
-        fprintf(stderr, "error loading model: %s\n", err.c_str());
+    } catch (const std::exception & err) {
+        fprintf(stderr, "error loading model: %s\n", err.what());
         return false;
     }
 }
@@ -1281,12 +1280,6 @@ static bool llama_eval_internal(
     ggml_set_name(embd, "embd");
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
 
-#ifdef GGML_USE_METAL
-    if (lctx.ctx_metal && N == 1) {
-        ggml_metal_set_tensor(lctx.ctx_metal, embd);
-    }
-#endif
-
     struct ggml_tensor * cur;
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
 
@@ -1484,12 +1477,6 @@ static bool llama_eval_internal(
         }
 
         ggml_graph_compute(ctx0, &gf);
-
-        if (lctx.ctx_metal) {
-            // We need to sync the CPU KV cache with the GPU KV cache
-            ggml_metal_set_tensor(lctx.ctx_metal, kv_self.k);
-            ggml_metal_set_tensor(lctx.ctx_metal, kv_self.v);
-        }
     }
 #else
     ggml_graph_compute(ctx0, &gf);
@@ -2133,8 +2120,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break;
         case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break;
         case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break;
+
         // K-quants
-        case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
         case LLAMA_FTYPE_MOSTLY_Q3_K_M:
         case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@@ -2142,8 +2130,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q4_K_M: quantized_type = GGML_TYPE_Q4_K; break;
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
-        case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break;
-        default: throw format("invalid output file type %d\n", ftype);
+        case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
+        default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
 
     if (nthread <= 0) {
@@ -2244,7 +2232,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                     f32_data[i] = ggml_fp16_to_fp32(f16_data[i]);
                 }
             } else {
-                throw format("type %s unsupported for integer quantization", ggml_type_name(tensor.type));
+                throw std::runtime_error(format("type %s unsupported for integer quantization", ggml_type_name(tensor.type)));
             }
 
             printf("quantizing .. ");
@@ -2446,8 +2434,8 @@ int llama_model_quantize(
     try {
         llama_model_quantize_internal(fname_inp, fname_out, ftype, nthread);
         return 0;
-    } catch (const std::string & err) {
-        fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.c_str());
+    } catch (const std::exception & err) {
+        fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.what());
         return 1;
     }
 }
@@ -2700,8 +2688,8 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
     try {
         return llama_apply_lora_from_file_internal(ctx, path_lora, path_base_model, n_threads);
-    } catch (const std::string & err) {
-        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.c_str());
+    } catch (const std::exception & err) {
+        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what());
         return 1;
     }
 }