llama : allow other bufts when overriding to CPU, add --no-repack option (#14990)

* vulkan: fix debug mode issues

* vulkan: remove broken check_results GGML_OP_SET_ROWS support

common/arg.cpp

@@ -2095,6 +2095,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.no_kv_offload = true;
         }
     ).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
+    add_opt(common_arg(
+        {"-nr", "--no-repack"},
+        "disable weight repacking",
+        [](common_params & params) {
+            params.no_extra_bufts = true;
+        }
+    ).set_env("LLAMA_ARG_NO_REPACK"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
         string_format(

common/common.cpp

@@ -1122,6 +1122,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+    mparams.use_extra_bufts = !params.no_extra_bufts;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;

common/common.h

@@ -359,6 +359,7 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
     bool no_op_offload     = false; // globally disable offload host tensor operations to device
+    bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
 
     bool single_turn       = false; // single turn chat conversation
 

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -5225,9 +5225,9 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
 }
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << ggml_type_name(src0->type) << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << ggml_type_name(dst->type) << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
@@ -11168,7 +11168,7 @@ size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
 static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
     ggml_tensor * tensor = cgraph->nodes[tensor_idx];
-    if (tensor->op == GGML_OP_TRANSPOSE) {
+    if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
 
@@ -11288,7 +11288,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
     } else if (tensor->op == GGML_OP_SCALE) {
         const float * params = (const float *)tensor->op_params;
-        tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]);
+        tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SIN) {
@@ -11399,8 +11399,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
         }
-    } else if (tensor->op == GGML_OP_SET_ROWS) {
-        tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONT) {
         tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
     } else if (tensor->op == GGML_OP_RESHAPE) {
@@ -11508,7 +11506,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
 
 static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
     ggml_tensor * tensor = cgraph->nodes[tensor_idx];
-    if (tensor->op == GGML_OP_TRANSPOSE) {
+    if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
     bool fused_rms_norm_mul = false;
@@ -11568,6 +11566,9 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph *
                         } else if (tensor->type == GGML_TYPE_F16) {
                             correct = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
                             result  = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
+                        } else if (tensor->type == GGML_TYPE_BF16) {
+                            correct = ggml_bf16_to_fp32(*(ggml_bf16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
+                            result  = ggml_bf16_to_fp32(*(ggml_bf16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
                         } else if (tensor->type == GGML_TYPE_I32) {
                             correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
                             result  = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);

include/llama.h

@@ -284,10 +284,11 @@ extern "C" {
         const struct llama_model_kv_override * kv_overrides;
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only;    // only load the vocabulary, no weights
-        bool use_mmap;      // use mmap if possible
-        bool use_mlock;     // force system to keep model in RAM
-        bool check_tensors; // validate model tensor data
+        bool vocab_only;      // only load the vocabulary, no weights
+        bool use_mmap;        // use mmap if possible
+        bool use_mlock;       // force system to keep model in RAM
+        bool check_tensors;   // validate model tensor data
+        bool use_extra_bufts; // use extra buffer types (used for weight repacking)
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations

src/llama-model.cpp

@@ -290,7 +290,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara
 }
 
 // CPU: ACCEL -> GPU host -> CPU extra -> CPU
-static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices) {
+static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts) {
     buft_list_t buft_list;
 
     // add ACCEL buffer types
@@ -319,21 +319,22 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types, only if no GPU device is present
-    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    if (cpu_dev == nullptr) {
-        throw std::runtime_error(format("%s: no CPU backend found", __func__));
-    }
+    // add extra buffer types
+    if (use_extra_bufts) {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (cpu_dev == nullptr) {
+            throw std::runtime_error(format("%s: no CPU backend found", __func__));
+        }
 
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_list.emplace_back(cpu_dev, *extra_bufts);
+                ++extra_bufts;
+            }
         }
     }
 
@@ -1839,7 +1840,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
 
     // build a list of buffer types for the CPU and GPU devices
-    pimpl->cpu_buft_list = make_cpu_buft_list(devices);
+    pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts);
     for (auto * dev : devices) {
         buft_list_t buft_list = make_gpu_buft_list(dev, split_mode, tensor_split);
         // add CPU buffer types as a fallback
@@ -2044,7 +2045,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
                     std::regex pattern(overrides->pattern);
                     if (std::regex_search(tensor_name, pattern)) {
-                        buft = overrides->buft;
+                        if (overrides->buft == ggml_backend_cpu_buffer_type()) {
+                            // when overriding to a CPU buffer, consider the extra buffer types
+                            buft = select_weight_buft(hparams, t_meta, op, pimpl->cpu_buft_list);
+                        } else {
+                            buft = overrides->buft;
+                        }
+
                         LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n",
                                 tensor_name.c_str(),
                                 ggml_nbytes(t_meta) / 1024 / 1024, ggml_type_name(t_meta->type),
@@ -17839,6 +17846,7 @@ llama_model_params llama_model_default_params() {
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
+        /*.use_extra_bufts             =*/ true,
     };
 
 #ifdef GGML_USE_METAL