llama: automatically fit args to free memory

Author: JohannesGaessler

common/arg.cpp

@@ -20,6 +20,7 @@
 #include <nlohmann/json.hpp>
 
 #include <algorithm>
+#include <cinttypes>
 #include <climits>
 #include <cstdarg>
 #include <filesystem>
@@ -1434,7 +1435,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.kv_overrides.back().key[0] = 0;
     }
 
-    if (!params.tensor_buft_overrides.empty()) {
+    // pad tensor_buft_overrides for llama_params_fit:
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+    while (params.tensor_buft_overrides.size() < ntbo) {
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
@@ -2961,6 +2964,34 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_MAIN_GPU"));
+    add_opt(common_arg(
+        { "-fit", "--fit" }, "[on|off]",
+        string_format("whether to adjust unset arguments to fit in device memory ('on' or 'off', default: '%s')", params.fit_params ? "on" : "off"),
+        [](common_params & params, const std::string & value) {
+            if (is_truthy(value)) {
+                params.fit_params = true;
+            } else if (is_falsey(value)) {
+                params.fit_params = false;
+            } else {
+                throw std::runtime_error(
+                    string_format("error: unkown value for --fit: '%s'\n", value.c_str()));
+            }
+        }
+    ).set_env("LLAMA_ARG_FIT"));
+    add_opt(common_arg(
+        { "-fitm", "--fit-margin" }, "MiB",
+        string_format("target margin per device for --fit option, default: %zu", params.fit_params_margin/(1024*1024)),
+        [](common_params & params, int value) {
+            params.fit_params_margin = value * size_t(1024*1024);
+        }
+    ).set_env("LLAMA_ARG_FIT_MARGIN"));
+    add_opt(common_arg(
+        { "-fitc", "--fit-ctx" }, "N",
+        string_format("minimum ctx size that can be set by --fit option, default: %" PRIu32, params.fit_params_min_ctx),
+        [](common_params & params, int value) {
+            params.fit_params_min_ctx = value;
+        }
+    ).set_env("LLAMA_ARG_FIT_CTX"));
     add_opt(common_arg(
         {"--check-tensors"},
         string_format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"),

common/common.cpp

@@ -916,6 +916,19 @@ std::string fs_get_cache_file(const std::string & filename) {
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (params.fit_params) {
+        const bool fit_successful = llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_margin, params.fit_params_min_ctx,
+            params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+
+        if (fit_successful) {
+            LOG_INF("%s: successfully fit parameters to device memory\n", __func__);
+        } else {
+            LOG_WRN("%s: failed to fit parameters to device memory, may crash during allocation\n", __func__);
+        }
+    }
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
@@ -926,8 +939,6 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    auto cparams = common_context_params_to_llama(params);
-
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
         LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",

common/common.h

@@ -274,8 +274,8 @@ struct lr_opt {
 struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
 
 struct common_params {
-    int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =  4096; // context size
+    int32_t n_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
+    int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -296,9 +296,12 @@ 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 - use default)
-    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
+    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
+    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_margin  = 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
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 

ggml/include/ggml-alloc.h

@@ -53,7 +53,14 @@ GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 // call with a worst-case graph to avoid buffer reallocations
 // not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
+// ggml_gallocr_resrve_n_size writes the buffer sizes per galloc buffer that would be allocated by ggml_gallocr_reserve_n to sizes
 GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API void ggml_gallocr_reserve_n_size(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids,
+    size_t * sizes);
 GGML_API bool ggml_gallocr_reserve_n(
     ggml_gallocr_t galloc,
     struct ggml_cgraph * graph,
@@ -68,6 +75,8 @@ GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_i
 
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
+// ggml_backend_alloc_ctx_tensors_from_buft_size returns the size of the buffer that would be allocated by ggml_backend_alloc_ctx_tensors_from_buft
+GGML_API size_t                       ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
 

ggml/include/ggml-backend.h

@@ -307,6 +307,7 @@ extern "C" {
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph
+    GGML_API void                 ggml_backend_sched_reserve_size(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph, size_t * sizes);
     GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); // returns success
 
     GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);

ggml/include/ggml.h

@@ -2509,7 +2509,8 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
+    GGML_API void ggml_log_get(ggml_log_callback * log_callback, void ** user_data);
+    GGML_API void ggml_log_set(ggml_log_callback   log_callback, void *  user_data);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 

ggml/src/ggml-alloc.c

@@ -595,7 +595,9 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
 }
 
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
+    return t->data != NULL // tensor data already set externally
+        || t->buffer // tensor on external buffer (but not yet allocated)
+        || ggml_gallocr_is_own(galloc, t); // tensor will be allocated by galloc
 }
 
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
@@ -791,7 +793,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     }
 }
 
-bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+static bool ggml_gallocr_reserve_n_impl(
+        ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids, bool no_alloc) {
     size_t min_hash_size = graph->n_nodes + graph->n_leafs;
     // add 25% margin to avoid hash collisions
     min_hash_size += min_hash_size / 4;
@@ -893,21 +896,41 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         if (realloc) {
 #ifndef NDEBUG
             size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
-            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n",
+                __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
 
             ggml_vbuffer_free(galloc->buffers[i]);
-            galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
-            if (galloc->buffers[i] == NULL) {
-                GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
-                return false;
+            if (no_alloc) {
+                galloc->buffers[i] = NULL;
+            } else {
+                galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+                if (galloc->buffers[i] == NULL) {
+                    GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                    return false;
+                }
             }
         }
     }
 
     return true;
 }
 
+void ggml_gallocr_reserve_n_size(
+        ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids, size_t * sizes) {
+    GGML_ASSERT(ggml_gallocr_reserve_n_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids, /*no_alloc =*/ true));
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        sizes[i] = 0;
+        for (int c = 0; c < galloc->buf_tallocs[i]->n_chunks; c++) {
+            sizes[i] += galloc->buf_tallocs[i]->chunks[c]->max_size;
+        }
+    }
+}
+
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    return ggml_gallocr_reserve_n_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids, /*no_alloc =*/ false);
+}
+
 bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
     return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
 }
@@ -1110,14 +1133,16 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
     return true;
 }
 
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+static ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft_impl(
+        struct ggml_context * ctx, ggml_backend_buffer_type_t buft, size_t * nbytes_total, bool no_alloc) {
     GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
 
     size_t alignment = ggml_backend_buft_get_alignment(buft);
     size_t max_size = ggml_backend_buft_get_max_size(buft);
 
     ggml_backend_buffer_t * buffers = NULL;
     size_t n_buffers = 0;
+    *nbytes_total = 0;
 
     size_t cur_buf_size = 0;
     struct ggml_tensor * first = ggml_get_first_tensor(ctx);
@@ -1129,10 +1154,11 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
         if (cur_buf_size > 0 && (cur_buf_size + this_size) > max_size) {
             // allocate tensors in the current buffer
-            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+            if (!no_alloc && !alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
                 return NULL;
             }
             first = t;
+            *nbytes_total += cur_buf_size;
             cur_buf_size = this_size;
         } else {
             cur_buf_size += this_size;
@@ -1141,15 +1167,21 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     // allocate remaining tensors
     if (cur_buf_size > 0) {
-        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+        *nbytes_total += cur_buf_size;
+        if (!no_alloc && !alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
             return NULL;
         }
     }
 
+    if (no_alloc) {
+        return NULL;
+    }
+
     if (n_buffers == 0) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
 #endif
+        GGML_ASSERT(!buffers);
         return NULL;
     }
 
@@ -1159,10 +1191,24 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
     } else {
         buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
     }
-    free(buffers);
+    if (buffers) {
+        free(buffers); // can be NULL if context is empty or no_alloc
+    }
     return buffer;
 }
 
+size_t ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    size_t nbytes_total = 0;
+    ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft_impl(ctx, buft, &nbytes_total, /*no_alloc=*/ true);
+    GGML_ASSERT(!buf);
+    return nbytes_total;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    size_t nbytes_total = 0;
+    return ggml_backend_alloc_ctx_tensors_from_buft_impl(ctx, buft, &nbytes_total, /*no_alloc =*/ false);
+}
+
 ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
     return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
 }

ggml/src/ggml-backend.cpp

@@ -36,12 +36,11 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
 }
 
 ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    GGML_ASSERT(buft);
     if (size == 0) {
         // return a dummy buffer for zero-sized allocations
         return ggml_backend_buffer_init(buft, {}, NULL, 0);
     }
-
-    GGML_ASSERT(buft);
     return buft->iface.alloc_buffer(buft, size);
 }
 
@@ -1694,6 +1693,20 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
     sched->is_alloc = false;
 }
 
+void ggml_backend_sched_reserve_size(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph, size_t * sizes) {
+    GGML_ASSERT(sched);
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
+    GGML_ASSERT(sizes);
+
+    ggml_backend_sched_reset(sched);
+
+    ggml_backend_sched_synchronize(sched);
+
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
+    ggml_gallocr_reserve_n_size(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids, sizes);
+}
+
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
     GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);

ggml/src/ggml.c

@@ -7273,6 +7273,11 @@ size_t ggml_quantize_chunk(
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void ggml_log_get(ggml_log_callback * log_callback, void ** user_data) {
+    *log_callback = g_logger_state.log_callback;
+    *user_data    = g_logger_state.log_callback_user_data;
+}
+
 void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
     g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
     g_logger_state.log_callback_user_data = user_data;

include/llama.h

@@ -297,6 +297,7 @@ extern "C" {
         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
+        bool no_alloc;        // only load metadata and simulate memory allocations
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
@@ -450,10 +451,23 @@ extern "C" {
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
+    // fits mparams and cparams to free device memory (assumes system memory is unlimited)
+    // returns true if the parameters could be successfully modified to fit device memory
+    LLAMA_API bool llama_params_fit(
+                                   const char   * path_model,
+                    struct llama_model_params   * mparams,
+                    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
+                                       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
+
     LLAMA_API int64_t llama_time_us(void);
 
     LLAMA_API size_t llama_max_devices(void);
     LLAMA_API size_t llama_max_parallel_sequences(void);
+    LLAMA_API size_t llama_max_tensor_buft_overrides(void);
 
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
@@ -1332,7 +1346,8 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
+    LLAMA_API void llama_log_get(ggml_log_callback * log_callback, void ** user_data);
+    LLAMA_API void llama_log_set(ggml_log_callback   log_callback, void *  user_data);
 
     //
     // Performance utils