merge upstream

.dockerignore

@@ -1,7 +1,7 @@
 *.o
 *.a
 .cache/
-.git/
+# Do not ignore .git directory, otherwise the reported build number will always be 0
 .github/
 .gitignore
 .vs/

.github/workflows/docker.yml

@@ -15,11 +15,17 @@ on:
     branches:
       - master
     paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+  workflow_dispatch: # allows manual triggering, useful for debugging
 
 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:
+  packages: write
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
@@ -46,6 +52,8 @@ jobs:
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
@@ -60,6 +68,34 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+
+          # determine tag name postfix (build number, commit hash)
+          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
+            TAG_POSTFIX="b${BUILD_NUMBER}"
+          else
+            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
+            TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
+          fi
+
+          # list all tags possible
+          TAGS=""
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
+
+          echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
+          echo "output_tags=$TAGS"  # print out for debugging
+        env:
+          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
       # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
       - name: Free Disk Space (Ubuntu)
         uses: jlumbroso/free-disk-space@main
@@ -77,31 +113,13 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Downcase github.repository_owner
-        run: |
-          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
       - name: Build and push Docker image (tagged + versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v6
         with:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          # tag list is generated from step above
+          tags: ${{ steps.tag.outputs.output_tags }}
           file: ${{ matrix.config.dockerfile }}

README.md

@@ -112,6 +112,7 @@ Typically finetunes of the base models below are supported as well.
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
+- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)

common/arg.cpp

@@ -691,7 +691,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params) {
             params.ctx_shift = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
     add_opt(llama_arg(
         {"--chunks"}, "N",
         format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
@@ -1102,7 +1102,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; }
             else { throw std::invalid_argument("invalid value"); }
         }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING"));
     add_opt(llama_arg(
         {"--attention"}, "{causal,non,causal}",
         "attention type for embeddings, use model default if unspecified",
@@ -1121,77 +1121,77 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
             else { throw std::invalid_argument("invalid value"); }
         }
-    ));
+    ).set_env("LLAMA_ARG_ROPE_SCALING_TYPE"));
     add_opt(llama_arg(
         {"--rope-scale"}, "N",
         "RoPE context scaling factor, expands context by a factor of N",
         [](gpt_params & params, const std::string & value) {
             params.rope_freq_scale = 1.0f / std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_ROPE_SCALE"));
     add_opt(llama_arg(
         {"--rope-freq-base"}, "N",
         "RoPE base frequency, used by NTK-aware scaling (default: loaded from model)",
         [](gpt_params & params, const std::string & value) {
             params.rope_freq_base = std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_ROPE_FREQ_BASE"));
     add_opt(llama_arg(
         {"--rope-freq-scale"}, "N",
         "RoPE frequency scaling factor, expands context by a factor of 1/N",
         [](gpt_params & params, const std::string & value) {
             params.rope_freq_scale = std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_ROPE_FREQ_SCALE"));
     add_opt(llama_arg(
         {"--yarn-orig-ctx"}, "N",
         format("YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx),
         [](gpt_params & params, int value) {
             params.yarn_orig_ctx = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_YARN_ORIG_CTX"));
     add_opt(llama_arg(
         {"--yarn-ext-factor"}, "N",
         format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor),
         [](gpt_params & params, const std::string & value) {
             params.yarn_ext_factor = std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_YARN_EXT_FACTOR"));
     add_opt(llama_arg(
         {"--yarn-attn-factor"}, "N",
         format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor),
         [](gpt_params & params, const std::string & value) {
             params.yarn_attn_factor = std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_YARN_ATTN_FACTOR"));
     add_opt(llama_arg(
         {"--yarn-beta-slow"}, "N",
         format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow),
         [](gpt_params & params, const std::string & value) {
             params.yarn_beta_slow = std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_YARN_BETA_SLOW"));
     add_opt(llama_arg(
         {"--yarn-beta-fast"}, "N",
         format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast),
         [](gpt_params & params, const std::string & value) {
             params.yarn_beta_fast = std::stof(value);
         }
-    ));
+    ).set_env("LLAMA_ARG_YARN_BETA_FAST"));
     add_opt(llama_arg(
         {"-gan", "--grp-attn-n"}, "N",
         format("group-attention factor (default: %d)", params.grp_attn_n),
         [](gpt_params & params, int value) {
             params.grp_attn_n = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_GRP_ATTN_N"));
     add_opt(llama_arg(
         {"-gaw", "--grp-attn-w"}, "N",
         format("group-attention width (default: %.1f)", (double)params.grp_attn_w),
         [](gpt_params & params, int value) {
             params.grp_attn_w = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_GRP_ATTN_W"));
     add_opt(llama_arg(
         {"-dkvc", "--dump-kv-cache"},
         "verbose print of the KV cache",
@@ -1205,23 +1205,23 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params) {
             params.no_kv_offload = true;
         }
-    ));
+    ).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
     add_opt(llama_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
         format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()),
         [](gpt_params & params, const std::string & value) {
             // TODO: get the type right here
             params.cache_type_k = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_CACHE_TYPE_K"));
     add_opt(llama_arg(
         {"-ctv", "--cache-type-v"}, "TYPE",
         format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()),
         [](gpt_params & params, const std::string & value) {
             // TODO: get the type right here
             params.cache_type_v = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_CACHE_TYPE_V"));
     add_opt(llama_arg(
         {"--perplexity", "--all-logits"},
         format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"),
@@ -1355,22 +1355,22 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params, const std::string & value) {
             params.rpc_servers = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_RPC"));
 #endif
     add_opt(llama_arg(
         {"--mlock"},
         "force system to keep model in RAM rather than swapping or compressing",
         [](gpt_params & params) {
             params.use_mlock = true;
         }
-    ));
+    ).set_env("LLAMA_ARG_MLOCK"));
     add_opt(llama_arg(
         {"--no-mmap"},
         "do not memory-map model (slower load but may reduce pageouts if not using mlock)",
         [](gpt_params & params) {
             params.use_mmap = false;
         }
-    ));
+    ).set_env("LLAMA_ARG_NO_MMAP"));
     add_opt(llama_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"
@@ -1385,7 +1385,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
             else { throw std::invalid_argument("invalid value"); }
         }
-    ));
+    ).set_env("LLAMA_ARG_NUMA"));
     add_opt(llama_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1433,7 +1433,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
                 fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting the split mode has no effect.\n");
             }
         }
-    ));
+    ).set_env("LLAMA_ARG_SPLIT_MODE"));
     add_opt(llama_arg(
         {"-ts", "--tensor-split"}, "N0,N1,N2,...",
         "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1",
@@ -1460,7 +1460,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
                 fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting a tensor split has no effect.\n");
             }
         }
-    ));
+    ).set_env("LLAMA_ARG_TENSOR_SPLIT"));
     add_opt(llama_arg(
         {"-mg", "--main-gpu"}, "INDEX",
         format("the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu),
@@ -1470,7 +1470,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
                 fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting the main GPU has no effect.\n");
             }
         }
-    ));
+    ).set_env("LLAMA_ARG_MAIN_GPU"));
     add_opt(llama_arg(
         {"--check-tensors"},
         format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"),
@@ -1533,7 +1533,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params, const std::string & value) {
             params.model_alias = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ALIAS"));
     add_opt(llama_arg(
         {"-m", "--model"}, "FNAME",
         ex == LLAMA_EXAMPLE_EXPORT_LORA
@@ -1741,7 +1741,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params, const std::string & value) {
             params.public_path = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
     add_opt(llama_arg(
         {"--embedding", "--embeddings"},
         format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"),
@@ -1779,22 +1779,22 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params, const std::string & value) {
             params.ssl_file_key = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_KEY_FILE"));
     add_opt(llama_arg(
         {"--ssl-cert-file"}, "FNAME",
         "path to file a PEM-encoded SSL certificate",
         [](gpt_params & params, const std::string & value) {
             params.ssl_file_cert = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
     add_opt(llama_arg(
         {"-to", "--timeout"}, "N",
         format("server read/write timeout in seconds (default: %d)", params.timeout_read),
         [](gpt_params & params, int value) {
             params.timeout_read  = value;
             params.timeout_write = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TIMEOUT"));
     add_opt(llama_arg(
         {"--threads-http"}, "N",
         format("number of threads used to process HTTP requests (default: %d)", params.n_threads_http),

common/log.cpp

@@ -82,7 +82,7 @@ struct gpt_log_entry {
             }
         }
 
-        if (level != GGML_LOG_LEVEL_NONE && prefix) {
+        if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
             if (timestamp) {
                 // [M.s.ms.us]
                 fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",

common/log.h

@@ -83,8 +83,10 @@ void gpt_log_set_timestamps(struct gpt_log * log,       bool   timestamps); // w
 #define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
 #define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
 #define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
 
 #define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)
 #define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
 #define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
+#define LOG_CNTV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  verbosity, __VA_ARGS__)

common/sampling.cpp

@@ -209,7 +209,15 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
             GGML_ASSERT(false && "unknown mirostat version");
         }
     } else {
-        llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+        if (params.n_probs > 0) {
+            // some use cases require to sample greedily, but still obtain the probabilities of the top tokens
+            // ref: https://github.com/ggerganov/llama.cpp/pull/9605
+            //
+            // the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
+            // it is much faster, since we avoid sorting all tokens and should give a good approximation
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+        }
         llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
     }