llama : add --no-host to disable host buffers (#16310)

* implement --no-host to disable host buffer

* fix equal_mparams

* move no-host enumeration order together with other model params

---------

Co-authored-by: slaren <redacted>

diff --git a/common/arg.cpp b/common/arg.cpp
index a020ac44132907ba808cc1ae6b0abaf11c75c484..5d09edd14da5ff217c1d3416231bcc5e881b84af 100644 (file)
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -2584,6 +2584,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.no_extra_bufts = true;
         }
     ).set_env("LLAMA_ARG_NO_REPACK"));
+    add_opt(common_arg(
+        {"--no-host"},
+        "bypass host buffer allowing extra buffers to be used",
+        [](common_params & params) {
+            params.no_host = true;
+        }
+    ).set_env("LLAMA_ARG_NO_HOST"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
         string_format(

diff --git a/common/common.cpp b/common/common.cpp
index c1e736c44cf56abcaa60de141fc030c76e6e3be6..b0591e84b066880c4fc1aa9bf383ac7bc4d35014 100644 (file)
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -1133,6 +1133,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;
+    mparams.no_host         = params.no_host;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;

diff --git a/common/common.h b/common/common.h
index d33788bd100b236695130c94c88cb479fa7fbfc0..8a8ecd667f2cc7375e9febee5c761cf19cd8a1c0 100644 (file)
--- a/common/common.h
+++ b/common/common.h
@@ -392,6 +392,7 @@ struct common_params {
     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 no_host           = false; // bypass host buffer allowing extra buffers to be used
 
     bool single_turn       = false; // single turn chat conversation
 

diff --git a/include/llama.h b/include/llama.h
index 8fc3d7db5a917424b8765061771d9adcec3431d7..a0a660bff88dae73307bdd3e88d44980233b7ae2 100644 (file)
--- a/include/llama.h
+++ b/include/llama.h
@@ -296,6 +296,7 @@ extern "C" {
         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)
+        bool no_host;         // bypass host buffer allowing extra buffers to be used
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations

diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 4c2d481a41d42247e3b8e2bcef112144cfec538c..ba4e9bf3f4f5c131d63ccc7977de1e8f2afc1f1c 100644 (file)
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -310,7 +310,7 @@ 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, bool use_extra_bufts) {
+static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts, bool no_host) {
     buft_list_t buft_list;
 
     // add ACCEL buffer types
@@ -331,11 +331,13 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
     // generally, this will be done using the first device in the list
     // a better approach would be to handle this on a weight-by-weight basis using the offload_op
     // function of the device to determine if it would benefit from being stored in a host buffer
-    for (auto * dev : devices) {
-        ggml_backend_buffer_type_t buft = ggml_backend_dev_host_buffer_type(dev);
-        if (buft) {
-            buft_list.emplace_back(dev, buft);
-            break;
+    if (!no_host) {
+        for (auto * dev : devices) {
+            ggml_backend_buffer_type_t buft = ggml_backend_dev_host_buffer_type(dev);
+            if (buft) {
+                buft_list.emplace_back(dev, buft);
+                break;
+            }
         }
     }
 
@@ -2083,7 +2085,7 @@ 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, params.use_extra_bufts);
+    pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts, params.no_host);
     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
@@ -19865,6 +19867,7 @@ llama_model_params llama_model_default_params() {
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
+        /*.no_host                     =*/ false,
     };
 
     return result;

diff --git a/tools/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp
index 89bc01b48546c220163185192154e8350e1a2f31..0de07b98112681f571ca4056d90da818a69b8026 100644 (file)
--- a/tools/llama-bench/llama-bench.cpp
+++ b/tools/llama-bench/llama-bench.cpp
@@ -336,6 +336,7 @@ struct cmd_params {
     std::vector<bool>                use_mmap;
     std::vector<bool>                embeddings;
     std::vector<bool>                no_op_offload;
+    std::vector<bool>                no_host;
     ggml_numa_strategy               numa;
     int                              reps;
     ggml_sched_priority              prio;
@@ -373,6 +374,7 @@ static const cmd_params cmd_params_defaults = {
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
     /* no_op_offload        */ { false },
+    /* no_host              */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
     /* prio                 */ GGML_SCHED_PRIO_NORMAL,
@@ -453,6 +455,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -ot --override-tensor <tensor name pattern>=<buffer type>;...\n");
     printf("                                            (default: disabled)\n");
     printf("  -nopo, --no-op-offload <0|1>              (default: 0)\n");
+    printf("  --no-host <0|1>                           (default: %s)\n",
+           join(cmd_params_defaults.no_host, ",").c_str());
     printf("\n");
     printf(
         "Multiple values can be given for each parameter by separating them with ','\n"
@@ -782,6 +786,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 auto p = string_split<bool>(argv[i], split_delim);
                 params.no_op_offload.insert(params.no_op_offload.end(), p.begin(), p.end());
+            } else if (arg == "--no-host") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.no_host.insert(params.no_host.end(), p.begin(), p.end());
             } else if (arg == "-ts" || arg == "--tensor-split") {
                 if (++i >= argc) {
                     invalid_param = true;
@@ -1003,6 +1014,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.no_op_offload.empty()) {
         params.no_op_offload = cmd_params_defaults.no_op_offload;
     }
+    if (params.no_host.empty()) {
+        params.no_host = cmd_params_defaults.no_host;
+    }
     if (params.n_threads.empty()) {
         params.n_threads = cmd_params_defaults.n_threads;
     }
@@ -1044,6 +1058,7 @@ struct cmd_params_instance {
     bool               use_mmap;
     bool               embeddings;
     bool               no_op_offload;
+    bool               no_host;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -1056,6 +1071,7 @@ struct cmd_params_instance {
         mparams.main_gpu     = main_gpu;
         mparams.tensor_split = tensor_split.data();
         mparams.use_mmap     = use_mmap;
+        mparams.no_host      = no_host;
 
         if (n_cpu_moe <= 0) {
             if (tensor_buft_overrides.empty()) {
@@ -1101,6 +1117,7 @@ struct cmd_params_instance {
                split_mode == other.split_mode &&
                main_gpu == other.main_gpu && use_mmap == other.use_mmap && tensor_split == other.tensor_split &&
                devices == other.devices &&
+               no_host == other.no_host &&
                vec_tensor_buft_override_equal(tensor_buft_overrides, other.tensor_buft_overrides);
     }
 
@@ -1136,6 +1153,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & ts : params.tensor_split)
     for (const auto & ot : params.tensor_buft_overrides)
     for (const auto & mmp : params.use_mmap)
+    for (const auto & noh : params.no_host)
     for (const auto & embd : params.embeddings)
     for (const auto & nopo : params.no_op_offload)
     for (const auto & nb : params.n_batch)
@@ -1178,6 +1196,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
                 /* .no_op_offload= */ nopo,
+                /* .no_host      = */ noh,
             };
             instances.push_back(instance);
         }
@@ -1211,6 +1230,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
                 /* .no_op_offload= */ nopo,
+                /* .no_host      = */ noh,
             };
             instances.push_back(instance);
         }
@@ -1244,6 +1264,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
                 /* .no_op_offload= */ nopo,
+                /* .no_host      = */ noh,
             };
             instances.push_back(instance);
         }
@@ -1282,6 +1303,7 @@ struct test {
     bool                     use_mmap;
     bool                     embeddings;
     bool                     no_op_offload;
+    bool                     no_host;
     int                      n_prompt;
     int                      n_gen;
     int                      n_depth;
@@ -1318,6 +1340,7 @@ struct test {
         use_mmap       = inst.use_mmap;
         embeddings     = inst.embeddings;
         no_op_offload  = inst.no_op_offload;
+        no_host        = inst.no_host;
         n_prompt       = inst.n_prompt;
         n_gen          = inst.n_gen;
         n_depth        = inst.n_depth;
@@ -1375,8 +1398,8 @@ struct test {
             "type_k",         "type_v",         "n_gpu_layers",  "n_cpu_moe",      "split_mode",
             "main_gpu",       "no_kv_offload",  "flash_attn",    "devices",        "tensor_split",
             "tensor_buft_overrides",            "use_mmap",      "embeddings",     "no_op_offload",
-            "n_prompt",       "n_gen",          "n_depth",       "test_time",      "avg_ns",
-            "stddev_ns",      "avg_ts",         "stddev_ts"
+            "no_host",        "n_prompt",       "n_gen",          "n_depth",       "test_time",
+            "avg_ns",         "stddev_ns",      "avg_ts",         "stddev_ts"
         };
         return fields;
     }
@@ -1391,7 +1414,7 @@ struct test {
             return INT;
         }
         if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
-            field == "use_mmap" || field == "embeddings") {
+            field == "use_mmap" || field == "embeddings" || field == "no_host") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -1466,6 +1489,7 @@ struct test {
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
                                             std::to_string(no_op_offload),
+                                            std::to_string(no_host),
                                             std::to_string(n_prompt),
                                             std::to_string(n_gen),
                                             std::to_string(n_depth),
@@ -1654,6 +1678,9 @@ struct markdown_printer : public printer {
         if (field == "no_op_offload") {
             return 4;
         }
+        if (field == "no_host") {
+            return 4;
+        }
 
         int width = std::max((int) field.length(), 10);
 
@@ -1688,6 +1715,9 @@ struct markdown_printer : public printer {
         if (field == "no_op_offload") {
             return "nopo";
         }
+        if (field == "no_host") {
+            return "noh";
+        }
         if (field == "devices") {
             return "dev";
         }
@@ -1768,6 +1798,9 @@ struct markdown_printer : public printer {
         if (params.no_op_offload.size() > 1 || params.no_op_offload != cmd_params_defaults.no_op_offload) {
             fields.emplace_back("no_op_offload");
         }
+        if (params.no_host.size() > 1 || params.no_host != cmd_params_defaults.no_host) {
+            fields.emplace_back("no_host");
+        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");