llama: automatically set parameters not set by the user in such a way that maximizes GPU utilization (#16653)

* llama: automatically fit args to free memory

llama-fit-params tool

* fix CI

* hints for bug reports, ensure no reallocation

* fix segfault with Vulkan

* add llama-fit-params to CI

* fix CI

* fix CI

* fix CI

* minor adjustments

* fix assignment of 1 dense layer

* fix logger not being reset on model load failure

* remove --n-gpu-layer hint on model load failure

* fix llama-fit-params verbosity

* fix edge case

* fix typo [no ci]

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -11,7 +11,7 @@ body:
         (i.e. the generated text) are incorrect or llama.cpp crashes during model evaluation.
         If you encountered the issue while using an external UI (e.g. ollama),
         please reproduce your issue using one of the examples/binaries in this repository.
-        The `llama-cli` binary can be used for simple and reproducible model inference.
+        The `llama-completion` binary can be used for simple and reproducible model inference.
   - type: textarea
     id: version
     attributes:
@@ -74,9 +74,12 @@ body:
         Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
         that information would be very much appreciated by us.
+
+        If possible, please try to reproduce the issue using `llama-completion` with `-fit off`.
+        If you can only reproduce the issue with `-fit on`, please provide logs both with and without `--verbose`.
       placeholder: >
-        e.g. when I run llama-cli with -ngl 99 I get garbled outputs.
-        When I use -ngl 0 it works correctly.
+        e.g. when I run llama-completion with `-fa on` I get garbled outputs for very long prompts.
+        With short prompts or `-fa off` it works correctly.
         Here are the exact commands that I used: ...
     validations:
       required: true

ci/run.sh

@@ -398,6 +398,8 @@ function gg_run_qwen3_0_6b {
     ./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
     ./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
 
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
     (time ./bin/llama-completion -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
     (time ./bin/llama-completion -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -523,6 +525,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
     (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
@@ -563,6 +567,8 @@ function gg_run_rerank_tiny {
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
     # for this model, the SEP token is "</s>"
     (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 

common/arg.cpp

@@ -20,6 +20,7 @@
 #include <nlohmann/json.hpp>
 
 #include <algorithm>
+#include <cinttypes>
 #include <climits>
 #include <cstdarg>
 #include <fstream>
@@ -529,7 +530,9 @@ 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});
     }
 
@@ -2153,6 +2156,34 @@ 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(
+        { "-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);
+        }
+    ).set_env("LLAMA_ARG_FIT_TARGET"));
+    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

@@ -1088,7 +1088,15 @@ struct common_init_result::impl {
 
 common_init_result::common_init_result(common_params & params) :
     pimpl(new impl{}) {
-    const auto mparams = common_model_params_to_llama(params);
+    auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (params.fit_params) {
+        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\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.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+    }
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
@@ -1103,8 +1111,6 @@ common_init_result::common_init_result(common_params & params) :
     // TODO: fix naming
     common_init_sampler_from_model(model, params.sampling);
 
-    auto cparams = common_context_params_to_llama(params);
-
     if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
         LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sampling.ignore_eos = false;
@@ -1143,8 +1149,7 @@ common_init_result::common_init_result(common_params & 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",
-                __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         return;
     }
 
@@ -1176,15 +1181,13 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
     llama_model * model = res->model();
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
-            __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return res;
     }
 
     llama_context * lctx = res->context();
     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",
-            __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         return res;
     }
 

common/common.h

@@ -99,6 +99,7 @@ enum llama_example {
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
+    LLAMA_EXAMPLE_FIT_PARAMS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -306,8 +307,8 @@ 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
@@ -328,9 +329,12 @@ 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_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
 
     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

@@ -2615,7 +2615,8 @@ 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

@@ -594,7 +594,9 @@ 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
 }
 
 // free the extra space at the end if the new tensor is smaller
@@ -823,7 +825,8 @@ 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;
@@ -928,16 +931,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
                 size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
                 if (cur_size > 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);
+                        __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;
+                }
             }
         }
     }
@@ -945,6 +951,21 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     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);
 }
@@ -1147,7 +1168,8 @@ 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);
@@ -1155,6 +1177,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     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);
@@ -1166,10 +1189,11 @@ 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;
@@ -1178,15 +1202,21 @@ 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;
     }
 
@@ -1196,10 +1226,24 @@ 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);
 }
 
@@ -128,6 +127,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
         return NULL;
     }
 
+    // FIXME JG: a multi_buffer has a non-zero size, according to the above comment get_base is not optional,
+    //     I don't know whether the above comment is correct
+    if (!buffer->iface.get_base) {
+        return NULL;
+    }
+
     void * base = buffer->iface.get_base(buffer);
 
     GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
@@ -1727,6 +1732,20 @@ 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

@@ -7566,6 +7566,11 @@ 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

@@ -313,6 +313,7 @@ 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
@@ -466,10 +467,24 @@ 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
+    // this function is NOT thread safe because it modifies the global llama logger state
+    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);
@@ -1354,7 +1369,9 @@ 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);
+    // The logger state is global so these functions are NOT thread safe.
+    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

src/llama-context.cpp

@@ -258,6 +258,7 @@ llama_context::llama_context(
 
         backend_buft.clear();
         backend_ptrs.clear();
+        backend_buf_exp_size.clear();
 
         for (auto & backend : backends) {
             auto * buft = ggml_backend_get_default_buffer_type(backend.get());
@@ -274,6 +275,7 @@ llama_context::llama_context(
 
             backend_buft.push_back(buft);
             backend_ptrs.push_back(backend.get());
+            backend_buf_exp_size.push_back(0);
         }
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
@@ -389,7 +391,8 @@ llama_context::llama_context(
 
         // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(),
+                model.hparams.no_alloc, model.hparams.no_alloc ? backend_buf_exp_size.data() : nullptr);
             if (!gf) {
                 if (pipeline_parallel) {
                     LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
@@ -407,7 +410,7 @@ llama_context::llama_context(
 
         // reserve with tg (token generation) graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -422,7 +425,7 @@ llama_context::llama_context(
             //
             // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
             //
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -431,11 +434,13 @@ llama_context::llama_context(
         for (size_t i = 0; i < backend_ptrs.size(); ++i) {
             ggml_backend_t             backend = backend_ptrs[i];
             ggml_backend_buffer_type_t buft    = backend_buft[i];
-            size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend);
-            if (size > 1) {
+            if (!model.hparams.no_alloc) {
+                backend_buf_exp_size[i] = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            }
+            if (backend_buf_exp_size[i] > 1) {
                 LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                         ggml_backend_buft_name(buft),
-                        size / 1024.0 / 1024.0);
+                        backend_buf_exp_size[i] / 1024.0 / 1024.0);
             }
         }
 
@@ -454,6 +459,23 @@ llama_context::llama_context(
 }
 
 llama_context::~llama_context() {
+    // FIXME this currently results in a use-after-free bug if the model is freed before the context
+    // if (!model.hparams.no_alloc) {
+    //     for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+    //         ggml_backend_t             backend = backend_ptrs[i];
+    //         ggml_backend_buffer_type_t buft    = backend_buft[i];
+
+    //         const size_t size_exp = backend_buf_exp_size[i];
+    //         const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    //         if (size_exp == size_act) {
+    //             LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         } else {
+    //             LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         }
+    //     }
+    // }
     ggml_opt_free(opt_ctx);
 }
 
@@ -1428,7 +1450,8 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
+ggml_cgraph * llama_context::graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only, size_t * sizes) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
     GGML_ASSERT(n_outputs >= 1);
 
@@ -1465,8 +1488,13 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     // initialize scheduler with the specified graph
     if (split_only) {
-        ggml_backend_sched_split_graph(sched.get(), gf);
+        if (sizes) {
+            ggml_backend_sched_reserve_size(sched.get(), gf, sizes);
+        } else {
+            ggml_backend_sched_split_graph(sched.get(), gf);
+        }
     } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+        GGML_ASSERT(!sizes);
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
@@ -2088,15 +2116,26 @@ void llama_context::perf_reset() {
 
 std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> llama_context::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> ret;
-    for (const auto & buft_size : model.memory_breakdown()) {
-        ret[buft_size.first].model += buft_size.second;
+    for (const auto & [buft, size] : model.memory_breakdown()) {
+        ret[buft].model += size;
     }
-    for (const auto & buft_size : memory->memory_breakdown()) {
-        ret[buft_size.first].context += buft_size.second;
+    if (memory) {
+        for (const auto & [buft, size] : memory->memory_breakdown()) {
+            ret[buft].context += size;
+        }
     }
-    for (const auto & backend_ptr : backends) {
-        ggml_backend_t backend = backend_ptr.get();
-        ret[ggml_backend_sched_get_buffer_type(sched.get(), backend)].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    if (model.hparams.no_alloc) {
+        for (size_t i = 0; i < backends.size(); ++i) {
+            ggml_backend_t             backend = backends[i].get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += backend_buf_exp_size[i];
+        }
+    } else {
+        for (const auto & backend_ptr : backends) {
+            ggml_backend_t             backend = backend_ptr.get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+        }
     }
     return ret;
 }

src/llama-context.h

@@ -26,6 +26,10 @@ struct llama_memory_breakdown_data {
     size_t model   = 0; // memory allocated for the model
     size_t context = 0; // memory allocated for the context
     size_t compute = 0; // memory allocated for temporary compute buffers
+
+    size_t total() const {
+        return model + context + compute;
+    }
 };
 
 struct llama_context {
@@ -206,7 +210,8 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
+    ggml_cgraph * graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false, size_t * sizes = nullptr);
 
 private:
     llm_graph_params graph_params(
@@ -281,9 +286,10 @@ private:
 
     std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
 
-    // buffer types used for the compute buffer of each backend
+    // pointers and buffer types used for the compute buffer of each backend
     std::vector<ggml_backend_t>             backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
+    std::vector<size_t>                     backend_buf_exp_size; // expected buffer sizes
 
     llm_graph_result_ptr gf_res_prev;
     llm_graph_result_ptr gf_res_reserve;

src/llama-hparams.h

@@ -34,6 +34,7 @@ struct llama_hparams_convnext {
 
 struct llama_hparams {
     bool vocab_only;
+    bool no_alloc;
     bool rope_finetuned;
     bool use_par_res;
     bool swin_norm;

src/llama-impl.cpp

@@ -25,6 +25,10 @@ time_meas::~time_meas() {
     }
 }
 
+void llama_log_get(ggml_log_callback * log_callback, void ** user_data) {
+    ggml_log_get(log_callback, user_data);
+}
+
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);
     g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;

src/llama-kv-cache.cpp

@@ -175,7 +175,15 @@ llama_kv_cache::llama_kv_cache(
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
     for (auto & [buft, ctx] : ctx_map) {
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
+        ggml_backend_buffer_t buf;
+        if (model.hparams.no_alloc) {
+            buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
+                t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
+            }
+        } else {
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
+        }
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -482,9 +490,18 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [ctx, buf] : ctxs_bufs) {
+        ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
+
+        if (hparams.no_alloc) {
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
+            ret[buft] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
+
     return ret;
 }
 

src/llama-model-loader.cpp

@@ -473,6 +473,7 @@ llama_model_loader::llama_model_loader(
         std::vector<std::string> & splits,
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
     int trace = 0;
@@ -716,6 +717,7 @@ llama_model_loader::llama_model_loader(
 
     this->use_mmap = use_mmap;
     this->check_tensors = check_tensors;
+    this->no_alloc = no_alloc;
 }
 
 std::string llama_model_loader::get_arch_name() const {

src/llama-model-loader.h

@@ -71,6 +71,7 @@ struct llama_model_loader {
 
     bool use_mmap = false;
     bool check_tensors;
+    bool no_alloc;
 
     llama_files files;
     llama_ftype ftype;
@@ -97,6 +98,7 @@ struct llama_model_loader {
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
 

src/llama-model.cpp

@@ -6606,9 +6606,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
         std::vector<ggml_backend_buffer_ptr> bufs;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
+            GGML_ASSERT(!ml.no_alloc);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
-                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer,
+                //     then we could just use metal for all layers
                 // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
                 void * addr = nullptr;
                 size_t first, last; // NOLINT
@@ -6624,9 +6626,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
-        }
-        else {
-            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        } else {
+            ggml_backend_buffer_t buf;
+            if (ml.no_alloc) {
+                buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+                for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+                    t->buffer = buf; // set dummy buffer for weights so that the backend scheduler won't try to allocate them
+                }
+            } else {
+                buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); // real buffer
+            }
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
@@ -6681,6 +6690,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         }
     }
 
+    if (ml.no_alloc) {
+        return true;
+    }
+
     // load tensor data
     for (auto & [ctx, buf_map] : ctx_buf_maps) {
         if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
@@ -6723,9 +6736,18 @@ size_t llama_model::n_devices() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
-        for (const auto & buf : bufs) {
-            ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [ctx, bufs] : pimpl->ctxs_bufs) {
+        if (hparams.no_alloc) {
+            GGML_ASSERT(bufs.size() == 1);
+            ggml_backend_buffer_t buf = bufs[0].get();
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf) == nullptr);
+            ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            for (const auto & buf : bufs) {
+                // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
+                ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+            }
         }
     }
     return ret;
@@ -6770,6 +6792,7 @@ void llama_model::print_info() const {
     // hparams
     LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, arch_name().c_str());
     LLAMA_LOG_INFO("%s: vocab_only       = %d\n",     __func__, hparams.vocab_only);
+    LLAMA_LOG_INFO("%s: no_alloc         = %d\n",     __func__, hparams.no_alloc);
 
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
@@ -7618,6 +7641,7 @@ llama_model_params llama_model_default_params() {
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
         /*.no_host                     =*/ false,
+        /*.no_alloc                    =*/ false,
     };
 
     return result;

src/llama-quant.cpp

@@ -596,7 +596,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());

src/llama.cpp

@@ -1,6 +1,9 @@
+#include "llama.h"
+
 #include "llama-impl.h"
 
 #include "llama-chat.h"
+#include "llama-context.h"
 #include "llama-mmap.h"
 #include "llama-vocab.h"
 #include "llama-model-loader.h"
@@ -11,11 +14,14 @@
 #include "ggml-backend.h"
 
 #include <algorithm>
+#include <cassert>
+#include <cinttypes>
 #include <cstddef>
 #include <cstdint>
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <stdexcept>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -37,6 +43,643 @@ const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_ty
     GGML_ABORT("fatal error");
 }
 
+struct llama_device_memory_data {
+    int64_t total;
+    int64_t free;
+    llama_memory_breakdown_data mb;
+};
+
+static std::vector<llama_device_memory_data> llama_get_device_memory_data(
+        const char * path_model, const llama_model_params * mparams, const llama_context_params * cparams,
+        std::vector<ggml_backend_dev_t> & devs, uint32_t & hp_ngl, uint32_t & hp_n_ctx_train, uint32_t & hp_n_expert,
+        const ggml_log_level log_level) {
+    struct user_data_t {
+        struct {
+            ggml_log_callback callback;
+            void * user_data;
+        } original_logger;
+        ggml_log_level min_level; // prints below this log level go to debug log
+    };
+    user_data_t ud;
+    llama_log_get(&ud.original_logger.callback, &ud.original_logger.user_data);
+    ud.min_level = log_level;
+
+    llama_log_set([](ggml_log_level level, const char * text, void * user_data) {
+        const user_data_t * ud = (const user_data_t *) user_data;
+        const ggml_log_level level_eff = level >= ud->min_level ? level : GGML_LOG_LEVEL_DEBUG;
+        ud->original_logger.callback(level_eff, text, ud->original_logger.user_data);
+    }, &ud);
+
+    llama_model_params mparams_copy = *mparams;
+    mparams_copy.no_alloc = true;
+    mparams_copy.use_mmap = false;
+
+    llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
+    if (model == nullptr) {
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to load model");
+    }
+
+    llama_context * ctx = llama_init_from_model(model, *cparams);
+    if (ctx == nullptr) {
+        llama_model_free(model);
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to create llama_context from model");
+    }
+
+    std::vector<llama_device_memory_data> ret(model->devices.size());
+
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown = ctx->memory_breakdown();
+
+    for (const auto & [buft, mb] : memory_breakdown) {
+        if (ggml_backend_buft_is_host(buft)) {
+            continue;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            continue;
+        }
+        for (size_t i = 0; i < ret.size(); i++) {
+            if (model->devices[i] == dev) {
+                ret[i].mb.model   += mb.model;
+                ret[i].mb.context += mb.context;
+                ret[i].mb.compute += mb.compute;
+                break;
+            }
+        }
+    }
+    for (size_t i = 0; i < ret.size(); i++) {
+        size_t free, total;
+        ggml_backend_dev_memory(model->devices[i], &free, &total);
+        ret[i].free  = free;
+        ret[i].total = total;
+    }
+
+    devs           = model->devices;
+    hp_ngl         = model->hparams.n_layer;
+    hp_n_ctx_train = model->hparams.n_ctx_train;
+    hp_n_expert    = model->hparams.n_expert;
+
+    llama_memory_breakdown_print(ctx); // goes to debug log
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+    return ret;
+}
+
+// enum to identify part of a layer for distributing its tensors:
+enum layer_fraction_t {
+    LAYER_FRACTION_NONE = 0, // nothing
+    LAYER_FRACTION_ATTN = 1, // attention
+    LAYER_FRACTION_UP   = 2, // attention + up
+    LAYER_FRACTION_GATE = 3, // attention + up + gate
+    LAYER_FRACTION_MOE  = 4, // everything but sparse MoE weights
+};
+// this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
+
+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) {
+    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();
+
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    // step 1: get data for default parameters and check whether any changes are necessary in the first place
+
+    LLAMA_LOG_DEBUG("%s: getting device memory data for initial parameters:\n", __func__);
+    const dmds_t dmds_full = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    const size_t nd = devs.size(); // number of devices
+    if (nd == 0) {
+        LLAMA_LOG_INFO("%s: no devices with dedicated memory found\n", __func__);
+        return;
+    }
+
+    std::vector<std::string> dev_names;
+    {
+        dev_names.reserve(nd);
+        size_t max_length = 0;
+        for (ggml_backend_dev_t dev : devs) {
+            std::string name = ggml_backend_dev_name(dev);
+            name += " (";
+            name += ggml_backend_dev_description(dev);
+            name += ")";
+            dev_names.push_back(name);
+            max_length = std::max(max_length, name.length());
+        }
+        for (std::string & dn : dev_names) {
+            dn.insert(dn.end(), max_length - dn.length(), ' ');
+        }
+    }
+
+    int64_t sum_total          = 0;
+    int64_t sum_projected_free = 0;
+    int64_t min_projected_free = INT64_MAX;
+    int64_t sum_projected_used = 0;
+    int64_t sum_projected_ctx  = 0;
+
+    if (nd > 1) {
+        LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
+    }
+    for (size_t id = 0; id < nd; id++) {
+        const llama_device_memory_data & dmd = dmds_full[id];
+
+        const int64_t projected_used = dmd.mb.total();
+        const int64_t projected_free = dmd.free - projected_used;
+
+        sum_total          += dmd.total;
+        sum_projected_used += projected_used;
+        sum_projected_free += projected_free;
+        min_projected_free  = std::min(min_projected_free, projected_free);
+        sum_projected_ctx  += dmd.mb.context;
+
+        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");
+        }
+    }
+    assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
+    assert(sum_projected_used >= sum_projected_ctx);
+    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_total/MiB);
+    if (min_projected_free >= margin) {
+        if (nd == 1) {
+            LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
+                __func__, min_projected_free/MiB, margin/MiB);
+            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;
+        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 (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
+                    const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
+                    const uint32_t ctx_reduction = std::min(
+                        uint32_t((-global_surplus + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
+                    cparams->n_ctx = hp_nct - ctx_reduction;
+                    const int64_t memory_reduction = ctx_reduction * bytes_per_ctx;
+                    global_surplus += memory_reduction;
+                    LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                        __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                } else {
+                    LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                        __func__, hp_nct, n_ctx_min);
+                }
+            } else {
+                LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
+            }
+        }
+        if (global_surplus >= 0) {
+            LLAMA_LOG_INFO("%s: entire model can be fit across devices by reducing context\n", __func__);
+            return;
+        }
+    }
+
+    if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
+        throw std::runtime_error("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+    }
+    if (nd > 1) {
+        if (!tensor_split) {
+            throw std::runtime_error("did not provide a buffer to write the tensor_split to, abort");
+        }
+        if (mparams->tensor_split) {
+            for (size_t id = 0; id < nd; id++) {
+                if (mparams->tensor_split[id] != 0.0f) {
+                    throw std::runtime_error("model_params::tensor_split already set by user, abort");
+                }
+            }
+        }
+        if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            throw std::runtime_error("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
+        }
+        if (hp_ngl < 2*nd) {
+            throw std::runtime_error("model has only " + std::to_string(hp_ngl) + " layers but need at least "
+                + std::to_string(2*nd) + " to fit memory for " + std::to_string(nd) + " devices, abort");
+        }
+    }
+    if (!tensor_buft_overrides) {
+        throw std::runtime_error("did not provide buffer to set tensor_buft_overrides, abort");
+    }
+    if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
+        throw std::runtime_error("model_params::tensor_buft_overrides already set by user, abort");
+    }
+
+    // step 3: iteratively fill the back to front with "dense" layers
+    //   - for a dense model simply fill full layers, giving each device a contiguous slice of the model
+    //   - for a MoE model, same as dense model but with all MoE tensors in system memory
+
+    // utility function that returns a static C string matching the tensors for a specific layer index and layer fraction:
+    auto get_overflow_pattern = [&](const size_t il, const layer_fraction_t lf) -> const char * {
+        constexpr size_t n_strings = 1000;
+        if (il >= n_strings) {
+            throw std::runtime_error("at most " + std::to_string(n_strings) + " model layers are supported");
+        }
+        switch (lf) {
+            case LAYER_FRACTION_ATTN: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|gate|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_UP: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_GATE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_down.*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_MOE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate)_(ch|)exps";
+                }
+                return patterns[il].c_str();
+            }
+            default:
+                GGML_ABORT("fatal error");
+        }
+    };
+
+    struct ngl_t {
+        uint32_t n_layer = 0; // number of total layers
+        uint32_t n_part  = 0; // number of partial layers, <= n_layer
+
+        // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
+        layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+    };
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+
+    // utility function to set n_gpu_layers and tensor_split
+    auto set_ngl_tensor_split_tbo = [&](
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            llama_model_params & mparams,
+            const bool add_nonrepeating) {
+        mparams.n_gpu_layers = 0;
+        for (size_t id = 0; id < nd; id++) {
+            mparams.n_gpu_layers += ngl_per_device[id].n_layer;
+            if (nd > 1) {
+                tensor_split[id] = ngl_per_device[id].n_layer;
+            }
+        }
+        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl);
+        uint32_t il0 = hp_ngl - mparams.n_gpu_layers; // start index for tensor buft overrides
+
+        if (add_nonrepeating) {
+            mparams.n_gpu_layers += 1;
+            tensor_split[nd - 1] += 1;
+        }
+        mparams.tensor_split = tensor_split;
+
+        size_t itbo = 0;
+        for (size_t id = 0; id < nd; id++) {
+            il0 += ngl_per_device[id].n_layer - ngl_per_device[id].n_part;
+            for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
+                if (itbo + 1 >= ntbo) {
+                    tensor_buft_overrides[itbo].pattern = nullptr;
+                    tensor_buft_overrides[itbo].buft    = nullptr;
+                    itbo++;
+                    mparams.tensor_buft_overrides = tensor_buft_overrides;
+                    throw std::runtime_error("llama_params_fit_n_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model\n");
+                }
+                tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
+                tensor_buft_overrides[itbo].buft = overflow_bufts[id];
+                itbo++;
+            }
+            il0 += ngl_per_device[id].n_part;
+        }
+        tensor_buft_overrides[itbo].pattern = nullptr;
+        tensor_buft_overrides[itbo].buft    = nullptr;
+        itbo++;
+        mparams.tensor_buft_overrides = tensor_buft_overrides;
+    };
+
+    // utility function that returns the memory use per device for given numbers of layers per device
+    auto get_memory_for_layers = [&](
+            const char * func_name,
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            const bool add_nonrepeating) -> std::vector<int64_t> {
+        llama_model_params mparams_copy = *mparams;
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy, add_nonrepeating);
+
+        const dmds_t dmd_nl = llama_get_device_memory_data(
+            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        LLAMA_LOG_DEBUG("%s: memory for test allocation by device:\n", func_name);
+        for (size_t id = 0; id < nd; id++) {
+            const ngl_t & n = ngl_per_device[id];
+            LLAMA_LOG_DEBUG(
+                "%s: id=%zu, n_layer=%2" PRIu32 ", n_part=%2" PRIu32 ", overflow_type=%d, mem=%6" PRId64 " MiB\n",
+                func_name, id, n.n_layer, n.n_part, int(n.overflow_type), dmd_nl[id].mb.total()/MiB);
+        }
+
+        std::vector<int64_t> ret;
+        ret.reserve(nd);
+        for (const llama_device_memory_data & dmd : dmd_nl) {
+            ret.push_back(dmd.mb.total());
+        }
+        return ret;
+    };
+
+    int64_t global_surplus_cpu_moe = 0;
+    if (hp_nex > 0) {
+        const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate)_(ch|)exps"; // matches all MoE tensors
+        ggml_backend_buffer_type_t cpu_buft = ggml_backend_cpu_buffer_type();
+        tensor_buft_overrides[0] = {pattern_moe_all.c_str(), cpu_buft};
+        tensor_buft_overrides[1] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+
+        LLAMA_LOG_DEBUG("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
+        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;
+        }
+
+        if (global_surplus_cpu_moe > 0) {
+            LLAMA_LOG_INFO("%s: with only dense weights in device memory there is a total surplus of %" PRId64 " MiB\n",
+                __func__, global_surplus_cpu_moe/MiB);
+        } else {
+            LLAMA_LOG_INFO("%s: with only dense weights in device memory there is still a total deficit of %" PRId64 " MiB\n",
+                __func__, -global_surplus_cpu_moe/MiB);
+        }
+
+        // reset
+        tensor_buft_overrides[0] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+    }
+
+    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);
+        LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
+    }
+
+    // whether for the optimal memory use we expect to load at least some MoE tensors:
+    const bool partial_moe = hp_nex > 0 && global_surplus_cpu_moe > 0;
+
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the partial layers of a device overflow to:
+    overflow_bufts.reserve(nd);
+    for (size_t id = 0; id < nd - 1; ++id) {
+        overflow_bufts.push_back(ggml_backend_dev_buffer_type(devs[id + 1]));
+    }
+    overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
+
+    std::vector<ngl_t> ngl_per_device(nd);
+    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts, partial_moe);
+    if (hp_nex > 0) {
+        for (size_t id = 0; id < nd; id++) {
+            ngl_per_device[id].overflow_type = LAYER_FRACTION_MOE;
+        }
+    }
+
+    // optimize the number of layers per device using the method of false position:
+    //   - ngl_per_device has 0 layers for each device, lower bound
+    //   - try a "high" configuration where a device is given all unassigned layers
+    //   - interpolate the memory use / layer between low and high linearly to get a guess where it meets our target
+    //   - check memory use of our guess, replace either the low or high bound
+    //   - once we only have a difference of a single layer, stop and return the lower bound that just barely still fits
+    if (hp_nex == 0) {
+        LLAMA_LOG_INFO("%s: filling dense layers back-to-front:\n", __func__);
+    } else {
+        LLAMA_LOG_INFO("%s: filling dense-only layers back-to-front:\n", __func__);
+    }
+    uint32_t n_unassigned = hp_ngl;
+    for (int id = nd - 1; id >= 0; id--) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        ngl_per_device_high[id].n_layer = n_unassigned;
+        if (hp_nex > 0) {
+            ngl_per_device_high[id].n_part = ngl_per_device_high[id].n_layer;
+        }
+        if (ngl_per_device_high[id].n_layer > 0) {
+            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+            if (mem_high[id] > targets[id]) {
+                uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                while (delta > 1) {
+                    uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                    step_size = std::max(step_size, uint32_t(1));
+                    step_size = std::min(step_size, delta - 1);
+
+                    std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                    ngl_per_device_test[id].n_layer += step_size;
+                    if (hp_nex) {
+                        ngl_per_device_test[id].n_part += step_size;
+                    }
+                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+
+                    if (mem_test[id] <= targets[id]) {
+                        ngl_per_device  = ngl_per_device_test;
+                        mem             = mem_test;
+                        n_unassigned   -= ngl_per_device[id].n_layer;
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    } else {
+                        ngl_per_device_high = ngl_per_device_test;
+                        mem_high            = mem_test;
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    }
+                    delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                }
+            } else {
+                ngl_per_device  = ngl_per_device_high;
+                n_unassigned   -= ngl_per_device[id].n_layer;
+                LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers, %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, mem[id]/MiB, projected_margin/MiB);
+    }
+    if (hp_nex == 0 || global_surplus_cpu_moe <= 0) {
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+        return;
+    }
+
+    // step 4: for a MoE model where all dense tensors fit,
+    //     convert the dense-only layers in the back to full layers in the front until all devices are full
+    // essentially the same procedure as for the dense-only layers except front-to-back
+    // also, try fitting at least part of one more layer to reduce waste for "small" GPUs with e.g. 24 GiB VRAM
+
+    size_t id_dense_start = nd;
+    for (int id = nd - 1; id >= 0; id--) {
+        if (ngl_per_device[id].n_layer > 0) {
+            id_dense_start = id;
+            continue;
+        }
+        break;
+    }
+    assert(id_dense_start < nd);
+
+    LLAMA_LOG_INFO("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
+    for (size_t id = 0; id <= id_dense_start; id++) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        for (size_t jd = id_dense_start; jd < nd; jd++) {
+            const uint32_t n_layer_move = ngl_per_device_high[jd].n_layer;
+            ngl_per_device_high[id].n_layer += n_layer_move;
+            ngl_per_device_high[jd].n_layer -= n_layer_move;
+            ngl_per_device_high[jd].n_part = 0;
+        }
+        size_t id_dense_start_high = nd - 1;
+        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+
+        if (mem_high[id] > targets[id]) {
+            assert(ngl_per_device_high[id].n_layer >= ngl_per_device_high[id].n_part);
+            assert(ngl_per_device[id].n_layer >= ngl_per_device[id].n_part);
+            assert((ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                   >= ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            uint32_t delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            while (delta > 1) {
+                uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                step_size = std::max(step_size, uint32_t(1));
+                step_size = std::min(step_size, delta - 1);
+
+                std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                size_t id_dense_start_test = id_dense_start;
+                uint32_t n_converted_test = 0;
+                for (;id_dense_start_test < nd; id_dense_start_test++) {
+                    const uint32_t n_convert_jd = std::min(step_size - n_converted_test, ngl_per_device_test[id_dense_start_test].n_part);
+                    ngl_per_device_test[id_dense_start_test].n_layer -= n_convert_jd;
+                    ngl_per_device_test[id_dense_start_test].n_part -= n_convert_jd;
+                    ngl_per_device_test[id].n_layer += n_convert_jd;
+                    n_converted_test += n_convert_jd;
+
+                    if (ngl_per_device_test[id_dense_start_test].n_layer > 0) {
+                        break;
+                    }
+                }
+                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+
+                if (mem_test[id] <= targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                } else {
+                    ngl_per_device_high = ngl_per_device_test;
+                    mem_high            = mem_test;
+                    id_dense_start_high = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
+                        __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
+                }
+                delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                    - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            }
+        } else {
+            ngl_per_device = ngl_per_device_high;
+            id_dense_start = id_dense_start_high;
+            LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+        }
+
+        // try to fit at least part of one more layer
+        if (ngl_per_device[id_dense_start].n_layer > 0) {
+            std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+            size_t id_dense_start_test = id_dense_start;
+            ngl_per_device_test[id_dense_start_test].n_layer--;
+            ngl_per_device_test[id_dense_start_test].n_part--;
+            ngl_per_device_test[id].n_layer++;
+            ngl_per_device_test[id].n_part++;
+            if (ngl_per_device_test[id_dense_start_test].n_layer == 0) {
+                id_dense_start_test++;
+            }
+            ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+            if (mem_test[id] < targets[id]) {
+                ngl_per_device = ngl_per_device_test;
+                mem            = mem_test;
+                id_dense_start = id_dense_start_test;
+                LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
+                    __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
+                LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                if (mem_test[id] < targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            } else {
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
+                LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                if (mem_test[id] < targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+}
+
+bool 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) {
+    const int64_t t0_us = llama_time_us();
+    bool ok = true;
+    try {
+        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
+        LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
+    } catch (const std::runtime_error & e) {
+        LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
+        ok = false;
+    }
+    const int64_t t1_us = llama_time_us();
+    LLAMA_LOG_INFO("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
+    return ok;
+}
+
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
         /*.no_perf                     =*/ true,
@@ -49,6 +692,10 @@ size_t llama_max_devices(void) {
     return 16;
 }
 
+size_t llama_max_tensor_buft_overrides() {
+    return 4096;
+}
+
 bool llama_supports_mmap(void) {
     return llama_mmap::SUPPORTED;
 }
@@ -108,11 +755,12 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
     model.t_start_us = tm.t_start_us;
 
     try {
-        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides, params.tensor_buft_overrides);
+        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
 
         ml.print_info();
 
         model.hparams.vocab_only = params.vocab_only;
+        model.hparams.no_alloc   = params.no_alloc;
 
         try {
             model.load_arch(ml);

tools/CMakeLists.txt

@@ -37,4 +37,5 @@ else()
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
     endif()
+    add_subdirectory(fit-params)
 endif()

tools/fit-params/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-fit-params)
+add_executable(${TARGET} fit-params.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/fit-params/README.md

@@ -0,0 +1,55 @@
+# fit-params
+
+llama.cpp binaries can automatically fit the projected memory use of a model to the free device memory available at runtime,
+this is controlled using the CLI arguments starting with `-fit`/`--fit`.
+Internally the code is calling `llama_params_fit` to adjust the `llama_model_params` and `llama_context_params` structs.
+`llama-fit-params` is a simple utility that prints the CLI arguments corresponding to these adjustments to stdout.
+Example usage:
+
+``` bash
+# First, run llama-fit-params and store the results in a file:
+> ./build/bin/llama-fit-params --model /opt/models/qwen_3-30b3a-f16.gguf | tee args.txt
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
+build: 6895 (4341dc8bc) with cc (GCC) 15.2.1 20250813 for x86_64-pc-linux-gnu
+llama_params_fit_impl: projected to use 61807 MiB of device memory vs. 24077 MiB of free device memory
+llama_params_fit_impl: cannot fulfill margin of 1024 MiB, need to reduce device memory by 42444 MiB
+llama_params_fit_impl: context size reduced from 40960 to 4096 -> need 3456 MiB less memory in total
+llama_params_fit_impl: with only dense weights in device memory there is a total surplus of 16164 MiB
+llama_params_fit_impl: distributing layers across devices with overflow to next device/system memory:
+llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 4090): 48 layers (34 overflowing),  19187 MiB used,   1199 MiB free
+llama_params_fit: successfully fit params to free device memory
+llama_params_fit: fitting params to free memory took 1.15 seconds
+Printing fitted CLI arguments to stdout...
+-c 4096 -ngl 48 -ot blk\.14\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.15\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.16\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.17\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.18\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.19\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.20\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.21\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.22\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.23\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.24\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.25\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.26\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.27\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.28\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.29\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.30\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.31\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.32\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.33\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.34\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.35\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.36\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.37\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.38\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.39\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.40\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.41\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.42\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.43\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.44\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.45\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.46\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.47\.ffn_(up|down|gate)_(ch|)exps=CPU
+
+# Next, use those results for a llama.cpp binary:
+> cat args.txt | xargs ./build/bin/llama-server --model /opt/models/qwen_3-30b3a-f16.gguf
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
+build: 6895 (4341dc8bc) with cc (GCC) 15.2.1 20250813 for x86_64-pc-linux-gnu
+system info: n_threads = 16, n_threads_batch = 16, total_threads = 32
+
+system_info: n_threads = 16 (n_threads_batch = 16) / 32 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
+
+main: binding port with default address family
+main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 31
+main: loading model
+srv    load_model: loading model '/opt/models/qwen_3-30b3a-f16.gguf'
+llama_params_fit_impl: projected to use 19187 MiB of device memory vs. 24077 MiB of free device memory
+llama_params_fit_impl: will leave 1199 >= 1024 MiB of free device memory, no changes needed
+llama_params_fit: successfully fit params to free device memory
+llama_params_fit: fitting params to free memory took 0.28 seconds
+[...]
+main: server is listening on http://127.0.0.1:8080 - starting the main loop
+srv  update_slots: all slots are idle
+^Csrv    operator(): operator(): cleaning up before exit...
+
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free     self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - CUDA0 (RTX 4090)   | 24077 =  945 + (19187 = 17904 +     384 +     898) +        3945 |
+llama_memory_breakdown_print: |   - Host               |                 58271 = 58259 +       0 +      12                |
+```

tools/fit-params/fit-params.cpp

@@ -0,0 +1,62 @@
+#include "llama.h"
+
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+
+#include <iostream>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+        return 1;
+    }
+
+    common_init();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+    auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+    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.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+
+    LOG_INF("Printing fitted CLI arguments to stdout...\n");
+    std::cout << "-c "    << cparams.n_ctx;
+    std::cout << " -ngl " << mparams.n_gpu_layers;
+
+    size_t nd = llama_max_devices();
+    while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) {
+        nd--;
+    }
+    if (nd > 1) {
+        for (size_t id = 0; id < nd; id++) {
+            if (id == 0) {
+                std::cout << " -ts ";
+            }
+            if (id > 0) {
+                std::cout << ",";
+            }
+            std::cout << mparams.tensor_split[id];
+        }
+    }
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+    for (size_t itbo = 0; itbo < ntbo && mparams.tensor_buft_overrides[itbo].pattern != nullptr; itbo++) {
+        if (itbo == 0) {
+            std::cout << " -ot ";
+        }
+        if (itbo > 0) {
+            std::cout << ",";
+        }
+        std::cout << mparams.tensor_buft_overrides[itbo].pattern << "=" << ggml_backend_buft_name(mparams.tensor_buft_overrides[itbo].buft);
+    }
+    std::cout << "\n";
+
+    return 0;
+}