llama : add ability to cancel model loading (#4462)

* llama : Add ability to cancel model load

Updated llama_progress_callback so that if it returns false, the model
loading is aborted.

* llama : Add test for model load cancellation

* Fix bool return in llama_model_load, remove std::ignore use

* Update llama.cpp

Co-authored-by: Jared Van Bortel <redacted>
* Fail test if model file is missing

* Revert "Fail test if model file is missing"

This reverts commit 32ebd525bf7e5a87ee8a3dbaab3d92ce79fbf23d.

* Add test-model-load-cancel to Makefile

* Revert "Revert "Fail test if model file is missing""

This reverts commit 2796953257ee5383fa7c8fe8fa8fc888c048fb0b.

* Simplify .gitignore for tests, clang-tidy fixes

* Label all ctest tests

* ci : ctest uses -L main

* Attempt at writing ctest_with_model

* ci : get ci/run.sh working with test-model-load-cancel

* ci : restrict .github/workflows/build.yml ctest to -L main

* update requirements.txt

* Disable test-model-load-cancel in make

* Remove venv before creation

* Restructure requirements.txt

Top-level now imports the specific additional requirements for each
python file. Using `pip install -r requirements.txt` will fail if
versions become mismatched in the per-file requirements.

* Make per-python-script requirements work alone

This doesn't break the main requirements.txt.

* Add comment

* Add convert-persimmon-to-gguf.py to new requirements.txt scheme

* Add check-requirements.sh script and GitHub workflow

* Remove shellcheck installation step from workflow

* Add nocleanup special arg

* Fix merge

see: https://github.com/ggerganov/llama.cpp/pull/4462#discussion_r1434593573

* reset to upstream/master

* Redo changes for cancelling model load

---------

Co-authored-by: Georgi Gerganov <redacted>
Co-authored-by: Jared Van Bortel <redacted>

diff --git a/llama.cpp b/llama.cpp
index d6c192441fbf04b5fd7eef22db07feb5ce3324d3..cb0546c952d20f3e6110205c5a5a004bd4b3119e 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -2372,7 +2372,8 @@ struct llama_model_loader {
         }
     }
 
-    void load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) const {
+    // Returns false if cancelled by progress_callback
+    bool load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) const {
         size_t size_data = 0;
 
         for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
@@ -2404,7 +2405,9 @@ struct llama_model_loader {
             GGML_ASSERT(cur); // unused tensors should have been caught by load_data already
 
             if (progress_callback) {
-                progress_callback((float) size_done / size_data, progress_callback_user_data);
+                if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
+                    return false;
+                }
             }
 
             const size_t offs = file_offset(ggml_get_name(cur));
@@ -2466,8 +2469,11 @@ struct llama_model_loader {
         }
 
         if (progress_callback) {
-            progress_callback(1.0f, progress_callback_user_data);
+            // Even though the model is done loading, we still honor
+            // cancellation since we need to free allocations.
+            return progress_callback(1.0f, progress_callback_user_data);
         }
+        return true;
     }
 };
 
@@ -3044,7 +3050,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     if (vocab.linefeed_id    != -1) { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, vocab.linefeed_id,    vocab.id_to_token[vocab.linefeed_id].text.c_str() );    }
 }
 
-static void llm_load_tensors(
+// Returns false if cancelled by progress_callback
+static bool llm_load_tensors(
         llama_model_loader & ml,
         llama_model & model,
         int n_gpu_layers,
@@ -3722,16 +3729,20 @@ static void llm_load_tensors(
         model.tensors_by_name.emplace_back(ggml_get_name(cur), cur);
     }
 
-    ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf_mmap, use_mlock ? &model.mlock_mmap : NULL);
+    if (!ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf_mmap, use_mlock ? &model.mlock_mmap : NULL)) {
+        return false;
+    }
 
     model.mapping = std::move(ml.mapping);
 
     // loading time will be recalculate after the first eval, so
     // we take page faults deferred by mmap() into consideration
     model.t_load_us = ggml_time_us() - model.t_start_us;
+    return true;
 }
 
-static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) {
+// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
+static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) {
     try {
         llama_model_loader ml(fname, params.use_mmap, params.kv_overrides);
 
@@ -3749,19 +3760,21 @@ static bool llama_model_load(const std::string & fname, llama_model & model, con
 
         if (params.vocab_only) {
             LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
-            return true;
+            return 0;
         }
 
-        llm_load_tensors(
+        if (!llm_load_tensors(
             ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock,
             params.progress_callback, params.progress_callback_user_data
-        );
+        )) {
+            return -2;
+        }
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("error loading model: %s\n", err.what());
-        return false;
+        return -1;
     }
 
-    return true;
+    return 0;
 }
 
 //
@@ -9141,11 +9154,18 @@ struct llama_model * llama_load_model_from_file(
                     LLAMA_LOG_INFO("\n");
                 }
             }
+            return true;
         };
     }
 
-    if (!llama_model_load(path_model, *model, params)) {
-        LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
+    int status = llama_model_load(path_model, *model, params);
+    GGML_ASSERT(status <= 0);
+    if (status < 0) {
+        if (status == -1) {
+            LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
+        } else if (status == -2) {
+            LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
+        }
         delete model;
         return nullptr;
     }

diff --git a/llama.h b/llama.h
index 0be4b1337b963e05a029d2ee4d01f98209dd4253..af76bae2d2a156258bee4ca1488e68089a555e85 100644 (file)
--- a/llama.h
+++ b/llama.h
@@ -127,7 +127,7 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef void (*llama_progress_callback)(float progress, void *ctx);
+    typedef bool (*llama_progress_callback)(float progress, void *ctx);
 
     // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
@@ -180,7 +180,9 @@ extern "C" {
         int32_t main_gpu;     // the GPU that is used for scratch and small tensors
         const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
-        // called with a progress value between 0 and 1, pass NULL to disable
+        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
+        // If the provided progress_callback returns true, model loading continues.
+        // If it returns false, model loading is immediately aborted.
         llama_progress_callback progress_callback;
 
         // context pointer passed to the progress callback