llama-fit-params: free memory target per device (#18679)

common/arg.cpp

@@ -2255,7 +2255,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             std::vector<std::string> split_arg{ it, {} };
             if (split_arg.size() >= llama_max_devices()) {
                 throw std::invalid_argument(
-                    string_format("got %d input configs, but system only has %d devices", (int)split_arg.size(), (int)llama_max_devices())
+                    string_format("got %zu input configs, but system only has %zu devices", split_arg.size(), llama_max_devices())
                 );
             }
             for (size_t i = 0; i < llama_max_devices(); ++i) {
@@ -2295,10 +2295,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_FIT"));
     add_opt(common_arg(
-        { "-fitt", "--fit-target" }, "MiB",
-        string_format("target margin per device for --fit option, default: %zu", params.fit_params_target/(1024*1024)),
-        [](common_params & params, int value) {
-            params.fit_params_target = value * size_t(1024*1024);
+        { "-fitt", "--fit-target" }, "MiB0,MiB1,MiB2,...",
+        string_format("target margin per device for --fit, comma-separated list of values, "
+            "single value is broadcast across all devices, default: %zu", params.fit_params_target[0]/(1024*1024)),
+        [](common_params & params, const std::string & value) {
+            std::string arg_next = value;
+
+            // split string by , and /
+            const std::regex regex{ R"([,/]+)" };
+            std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
+            std::vector<std::string> split_arg{ it, {} };
+            if (split_arg.size() >= llama_max_devices()) {
+                throw std::invalid_argument(
+                    string_format("got %zu input configs, but system only has %zu devices", split_arg.size(), llama_max_devices())
+                );
+            }
+            if (split_arg.size() == 1) {
+                std::fill(params.fit_params_target.begin(), params.fit_params_target.end(), std::stoul(split_arg[0]) * 1024*1024);
+                return;
+            }
+            for (size_t i = 0; i < split_arg.size(); i++) {
+                params.fit_params_target[i] = std::stoul(split_arg[i]) * 1024*1024;
+            }
         }
     ).set_env("LLAMA_ARG_FIT_TARGET"));
     add_opt(common_arg(

common/common.cpp

@@ -1097,7 +1097,7 @@ common_init_result::common_init_result(common_params & params) :
     if (params.fit_params) {
         LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 

common/common.h

@@ -332,12 +332,14 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM, -1 is auto, <= -2 is all
-    int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
-    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
-    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
-    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
+    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
+    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
+    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
+
+    // margin per device in bytes for fitting parameters to free memory:
+    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 

include/llama.h

@@ -495,7 +495,7 @@ extern "C" {
                     struct llama_context_params * cparams,
                                           float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
         struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
-                                         size_t   margin,                // margin of memory to leave per device in bytes
+                                         size_t * margins,               // margins of memory to leave per device in bytes
                                        uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
                             enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 

src/llama.cpp

@@ -147,9 +147,8 @@ class llama_params_fit_exception : public std::runtime_error {
 static void llama_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     constexpr int64_t MiB = 1024*1024;
-    const int64_t margin = margin_s; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
     typedef std::vector<llama_device_memory_data> dmds_t;
     const llama_model_params default_mparams = llama_model_default_params();
 
@@ -168,6 +167,12 @@ static void llama_params_fit_impl(
         return;
     }
 
+    std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
+    margins.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        margins.push_back(margins_s[id]);
+    }
+
     std::vector<std::string> dev_names;
     {
         dev_names.reserve(nd);
@@ -187,9 +192,10 @@ static void llama_params_fit_impl(
 
     int64_t sum_free            = 0;
     int64_t sum_projected_free  = 0;
-    int64_t min_projected_free  = INT64_MAX;
     int64_t sum_projected_used  = 0;
     int64_t sum_projected_model = 0;
+    std::vector<int64_t> projected_free_per_device;
+    projected_free_per_device.reserve(nd);
 
     if (nd > 1) {
         LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
@@ -199,45 +205,63 @@ static void llama_params_fit_impl(
 
         const int64_t projected_used = dmd.mb.total();
         const int64_t projected_free = dmd.free - projected_used;
+        projected_free_per_device.push_back(projected_free);
 
         sum_free            += dmd.free;
         sum_projected_used  += projected_used;
         sum_projected_free  += projected_free;
-        min_projected_free   = std::min(min_projected_free, projected_free);
         sum_projected_model += dmd.mb.model;
 
         if (nd > 1) {
-            LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
-                __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, std::abs(projected_free)/MiB,
-                projected_free >= 0 ? "surplus" : "deficit");
+            LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " free vs. target of %6" PRId64 "\n",
+                __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, projected_free/MiB, margins[id]/MiB);
         }
     }
     assert(sum_free >= 0 && sum_projected_used >= 0);
     LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
         __func__, sum_projected_used/MiB, sum_free/MiB);
-    if (min_projected_free >= margin) {
-        if (nd == 1) {
+    if (nd == 1) {
+        if (projected_free_per_device[0] >= margins[0]) {
             LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
-                __func__, min_projected_free/MiB, margin/MiB);
+                __func__, projected_free_per_device[0]/MiB, margins[0]/MiB);
+            return;
+        }
+    } else {
+        bool changes_needed = false;
+        for (size_t id = 0; id < nd; id++) {
+            if (projected_free_per_device[id] < margins[id]) {
+                changes_needed = true;
+                break;
+            }
+        }
+        if (!changes_needed) {
+            LLAMA_LOG_INFO("%s: targets for free memory can be met on all devices, no changes needed\n", __func__);
             return;
         }
-        LLAMA_LOG_INFO("%s: will leave at least %" PRId64 " >= %" PRId64 " MiB of free memory on all devices, no changes needed\n",
-            __func__, min_projected_free/MiB, margin/MiB);
-        return;
     }
 
     // step 2: try reducing memory use by reducing the context size
 
     {
-        int64_t global_surplus = sum_projected_free - int64_t(nd)*margin;
+        int64_t global_surplus = sum_projected_free;
+        for (size_t id = 0; id < nd; id++) {
+            global_surplus -= margins[id];
+        }
         if (global_surplus < 0) {
-            LLAMA_LOG_INFO(nd == 1 ?
-                "%s: cannot fulfill margin of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n" :
-                "%s: cannot fulfill margin of %" PRId64 " MiB on all devices, need to use %" PRId64 " MiB less in total\n",
-                __func__, margin/MiB, -global_surplus/MiB);
+            if (nd == 1) {
+                LLAMA_LOG_INFO("%s: cannot meet free memory target of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n",
+                    __func__, margins[0]/MiB, -global_surplus/MiB);
+            } else {
+                LLAMA_LOG_INFO(
+                    "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
+                    __func__, -global_surplus/MiB);
+            }
             if (cparams->n_ctx == 0) {
                 if (hp_nct > n_ctx_min) {
-                    int64_t sum_used_target = sum_free - nd*margin_s;
+                    int64_t sum_used_target = sum_free;
+                    for (size_t id = 0; id < nd; id++) {
+                        sum_used_target -= margins[id];
+                    }
                     if (nd > 1) {
                         // for multiple devices we need to be more conservative in terms of how much context we think can fit:
                         //   - for dense models only whole layers can be assigned to devices
@@ -448,9 +472,9 @@ static void llama_params_fit_impl(
         const dmds_t dmds_cpu_moe = llama_get_device_memory_data(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
 
-        for (const llama_device_memory_data & dmd : dmds_cpu_moe) {
-            global_surplus_cpu_moe += dmd.free;
-            global_surplus_cpu_moe -= int64_t(dmd.mb.total()) + margin;
+        for (size_t id = 0; id < nd; id++) {
+            global_surplus_cpu_moe += dmds_cpu_moe[id].free;
+            global_surplus_cpu_moe -= int64_t(dmds_cpu_moe[id].mb.total()) + margins[id];
         }
 
         if (global_surplus_cpu_moe > 0) {
@@ -469,7 +493,7 @@ static void llama_params_fit_impl(
     std::vector<int64_t> targets; // maximum acceptable memory use per device
     targets.reserve(nd);
     for (size_t id = 0; id < nd; id++) {
-        targets.push_back(dmds_full[id].free - margin);
+        targets.push_back(dmds_full[id].free - margins[id]);
         LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
     }
 
@@ -701,11 +725,11 @@ static void llama_params_fit_impl(
 enum llama_params_fit_status llama_params_fit(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+        size_t * margins, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
     llama_params_fit_status status = LLAMA_PARAMS_FIT_STATUS_SUCCESS;
     try {
-        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
+        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
         LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
     } catch (const llama_params_fit_exception & e) {
         LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());

tools/fit-params/fit-params.cpp

@@ -27,7 +27,7 @@ int main(int argc, char ** argv) {
     auto mparams = common_model_params_to_llama(params);
     auto cparams = common_context_params_to_llama(params);
     const llama_params_fit_status status = llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-        params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+        params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
         params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     if (status != LLAMA_PARAMS_FIT_STATUS_SUCCESS) {
         LOG_ERR("%s: failed to fit CLI arguments to free memory, exiting...\n", __func__);