Merge branch 'ggerganov:master' into vlm

.github/workflows/build.yml

@@ -19,6 +19,11 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  contents: write # for creating release
+
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3

.github/workflows/close-issue.yml

@@ -3,6 +3,11 @@ on:
   schedule:
     - cron: "42 0 * * *"
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  issues: write
+
 jobs:
   close-issues:
     runs-on: ubuntu-latest

.github/workflows/nix-ci-aarch64.yml

@@ -21,6 +21,13 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -12,6 +12,13 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-eval:
     strategy:

CONTRIBUTING.md

@@ -1,24 +1,23 @@
 # Pull requests (for contributors)
 
 - Test your changes:
-  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
+  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the `ggml` library
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
-- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
-  - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
-- Consider allowing write access to your branch for faster review
+- Optionally rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs
+- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)
 
 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
-- Optionally, pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
+- Optionally pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
 
 # Coding guidelines
 
 - Avoid adding third-party dependencies, extra files, extra headers, etc.
 - Always consider cross-compatibility with other operating systems and architectures
-- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
+- Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)

README.md

@@ -169,6 +169,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
 - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
+- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 

ci/run.sh

@@ -1,4 +1,4 @@
-#/bin/bash
+#!/bin/bash
 #
 # sample usage:
 #
@@ -751,7 +751,8 @@ function gg_run_rerank_tiny {
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s><s>hi\nwhat is panda?</s><s>it's a bear\nwhat is panda?</s><s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    # for this model, the SEP token is "</s>"
+    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029
@@ -774,7 +775,7 @@ function gg_run_rerank_tiny {
 
     check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
     check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
-    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.15" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log
 
     set +e
 }

common/arg.cpp

@@ -911,7 +911,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
     ).set_sparam());
     add_opt(llama_arg(
         {"-s", "--seed"}, "SEED",
-        format("RNG seed (default: %u, use random seed for %u)", params.sparams.seed, LLAMA_DEFAULT_SEED),
+        format("RNG seed (default: %d, use random seed for %d)", params.sparams.seed, LLAMA_DEFAULT_SEED),
         [](gpt_params & params, const std::string & value) {
             params.sparams.seed = std::stoul(value);
         }

common/common.cpp

@@ -838,6 +838,31 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         return iparams;
     }
 
+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free_model(model);
+
+            return iparams;
+        }
+    }
+
     auto cparams = llama_context_params_from_gpt_params(params);
 
     llama_context * lctx = llama_new_context_with_model(model, cparams);
@@ -855,6 +880,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
+
             return iparams;
         }
 
@@ -867,6 +893,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
+
             return iparams;
         }
     }
@@ -889,7 +916,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
-    if (params.sparams.ignore_eos && llama_token_eos(model) == -1) {
+    if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
         LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sparams.ignore_eos = false;
     }
@@ -930,6 +957,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
 
     iparams.model   = model;
     iparams.context = lctx;
+
     return iparams;
 }
 

examples/main/README.md

@@ -69,7 +69,7 @@ In this section, we cover the most commonly used options for running the `llama-
 -   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 -   `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\'
 -   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has.
--   -   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 
 ## Input Prompts
 

examples/server/README.md

@@ -100,7 +100,7 @@ The project is under active development, and we are [looking for feedback and co
 | Argument | Explanation |
 | -------- | ----------- |
 | `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
-| `-s, --seed SEED` | RNG seed (default: 4294967295, use random seed for 4294967295) |
+| `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
 | `--penalize-nl` | penalize newline tokens (default: false) |

examples/server/server.cpp

@@ -2027,7 +2027,7 @@ struct server_context {
                                 continue;
                             }
 
-                            // prompt: <s>query</s><s>doc</s>
+                            // prompt: [BOS]query[EOS][SEP]doc[EOS]
                             prompt_tokens.clear();
                             prompt_tokens.push_back(llama_token_bos(model));
                             {
@@ -2035,7 +2035,7 @@ struct server_context {
                                 prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());
                             }
                             prompt_tokens.push_back(llama_token_eos(model));
-                            prompt_tokens.push_back(llama_token_bos(model));
+                            prompt_tokens.push_back(llama_token_sep(model));
                             {
                                 const auto part = tokenize(slot.prompt[1], false);
                                 prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());

ggml/include/ggml-alloc.h

@@ -24,7 +24,7 @@ GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, st
 // Graph allocator
 /*
   Example usage:
-    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
 
     // optional: create a worst-case graph and reserve the buffers to avoid reallocations
     ggml_gallocr_reserve(galloc, build_graph(max_batch));

ggml/src/ggml-cuda.cu

@@ -2448,6 +2448,7 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
     GGML_UNUSED(backend);
 }
 
+#ifdef USE_CUDA_GRAPH
 static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
     graph_node_properties->node_address = node->data;
     graph_node_properties->node_op = node->op;
@@ -2498,6 +2499,7 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
 
     return true;
 }
+#endif
 
 static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;

ggml/src/ggml-metal.m

@@ -3258,7 +3258,7 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t s
 }
 
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
+    struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context));
 
     const size_t size_page = sysconf(_SC_PAGESIZE);
 
@@ -3340,7 +3340,7 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
 // buffer from ptr
 
 ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
-    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
+    struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context));
 
     ctx->all_data = data;
     ctx->all_size = size;

ggml/src/ggml-vulkan.cpp

@@ -1082,10 +1082,25 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
     try {
         buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
     } catch (const vk::SystemError& e) {
-        // Out of Host/Device memory, clean up buffer
-        device->device.destroyBuffer(buf->buffer);
-        buf->size = 0;
-        throw e;
+        if (buf->memory_property_flags != fallback_flags) {
+            // Try again with fallback flags
+            memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
+            buf->memory_property_flags = fallback_flags;
+
+            try {
+                buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
+            }
+            catch (const vk::SystemError& e) {
+                device->device.destroyBuffer(buf->buffer);
+                buf->size = 0;
+                throw e;
+            }
+        } else {
+            // Out of Host/Device memory, clean up buffer
+            device->device.destroyBuffer(buf->buffer);
+            buf->size = 0;
+            throw e;
+        }
     }
     buf->ptr = nullptr;
 

scripts/sync-ggml.last

@@ -1 +1 @@
-e5c233e5edbfcfa1d808b9293de9065035c40751
+564f42082f858f9674b2a2e06e9e779d9ed2c754

src/llama-vocab.h

@@ -40,17 +40,17 @@ struct llama_vocab {
     id special_bos_id  = 1;
     id special_eos_id  = 2;
     id special_unk_id  = 0;
-    id special_sep_id  = -1;
-    id special_pad_id  = -1;
-    id special_cls_id  = -1;
-    id special_mask_id = -1;
+    id special_sep_id  = LLAMA_TOKEN_NULL;
+    id special_pad_id  = LLAMA_TOKEN_NULL;
+    id special_cls_id  = LLAMA_TOKEN_NULL;
+    id special_mask_id = LLAMA_TOKEN_NULL;
 
     id linefeed_id       = 13;
-    id special_prefix_id = -1;
-    id special_suffix_id = -1;
-    id special_middle_id = -1;
-    id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
-    id special_eom_id    = -1;
+    id special_prefix_id = LLAMA_TOKEN_NULL;
+    id special_suffix_id = LLAMA_TOKEN_NULL;
+    id special_middle_id = LLAMA_TOKEN_NULL;
+    id special_eot_id    = LLAMA_TOKEN_NULL; // TODO: move above after "eos_id", and here add "file separator" token
+    id special_eom_id    = LLAMA_TOKEN_NULL;
 
     // set of all tokens that cause "end of generation"
     std::set<id> special_eog_ids;

src/llama.cpp

@@ -2412,7 +2412,7 @@ struct llama_hparams {
 
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
-    llama_token dec_start_token_id = -1;
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;