app : allow --version, --licenses & --help (#25054 )

Signed-off-by: Adrien Gallouët <angt@huggingface.co>
sched : reintroduce less synchronizations during split compute (#20793 )
2026-06-27 08:07:40 +02:00 · 2026-06-26 23:18:11 +02:00 · 2026-06-26 17:18:30 +03:00 · 2026-06-26 15:15:48 +02:00 · 2026-06-26 14:36:03 +02:00
15 changed files with 102 additions and 38 deletions
@@ -145,7 +145,7 @@ ENTRYPOINT ["/app/tools.sh"]
 # ==============================================================================
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 ENTRYPOINT [ "/app/llama-cli" ]

@@ -156,7 +156,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]

@@ -104,7 +104,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -115,7 +115,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -113,7 +113,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -124,7 +124,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -141,7 +141,7 @@ ENTRYPOINT ["/app/tools.sh"]
 FROM base AS light

 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -153,7 +153,7 @@ FROM base AS server
 ENV LLAMA_ARG_HOST=0.0.0.0

 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -115,7 +115,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -126,7 +126,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -214,7 +214,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app/
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app/

 WORKDIR /app

@@ -225,7 +225,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app/
+COPY --from=build /app/full/llama /app/full/llama-server /app/

 WORKDIR /app

@@ -127,7 +127,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -138,7 +138,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -124,7 +124,7 @@ WORKDIR /llama.cpp/bin

 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin

 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]

@@ -138,7 +138,7 @@ WORKDIR /llama.cpp/bin

 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-server /llama.cpp/bin

 EXPOSE 8080

@@ -107,7 +107,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -118,7 +118,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -97,7 +97,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -108,7 +108,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -50,6 +50,7 @@ struct command {
    std::vector<std::string> aliases;
    bool hidden;
    int (*func)(int, char **);
+    bool flags = false; // allow --name
 };

 #ifdef LLAMA_INSTALL_BUILD
@@ -69,9 +70,9 @@ static const command cmds[] = {
    {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,          llama_fit_params   },
    {"quantize",      "Quantize a model",                                   {},           true,          llama_quantize     },
    {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,          llama_perplexity   },
-    {"version",       "Show version",                                       {},           false,         version            },
-    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses           },
-    {"help",          "Show available commands",                            {},           false,         help               },
+    {"version",       "Show version",                                       {},           false,         version,           true },
+    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses,          true },
+    {"help",          "Show available commands",                            {},           false,         help,              true },
 };

 #undef UPDATE_HIDDEN
@@ -108,7 +109,10 @@ static int help(int argc, char ** argv) {
    return 0;
 }

-static bool matches(const std::string & arg, const command & cmd) {
+static bool matches(std::string arg, const command & cmd) {
+    if (cmd.flags && arg.size() > 2 && arg[0] == '-' && arg[1] == '-') {
+        arg.erase(0, 2);
+    }
    if (arg == cmd.name) {
        return true;
    }
@@ -352,6 +352,8 @@ static std::string get_default_local_path(const std::string & url) {

 common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex) {
    common_download_hf_plan plan;
+    common_download_hf_plan plan_spec;
+    common_download_hf_plan plan_voc;
    common_download_opts opts;

    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
@@ -377,7 +379,15 @@ common_models_handler common_models_handler_init(const common_params & params, l
        plan = common_download_get_hf_plan(params.model, opts);
    }

-    return common_models_handler{plan, opts};
+    if (!params.speculative.draft.mparams.hf_repo.empty()) {
+        plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts);
+    }
+
+    if (!params.vocoder.model.hf_repo.empty()) {
+        plan_voc = common_download_get_hf_plan(params.vocoder.model, opts);
+    }
+
+    return common_models_handler{plan, plan_spec, plan_voc, opts};
 }

 bool common_models_handler_is_preset_repo(const common_models_handler & handler) {
@@ -425,7 +435,9 @@ static std::vector<common_download_task> build_url_tasks(const common_params_mod
 void common_models_handler_apply(common_models_handler & handler, common_params & params, common_download_callback * callback) {
    std::vector<common_download_task> tasks;

-    auto & plan = handler.plan;
+    auto & plan      = handler.plan;
+    auto & plan_spec = handler.plan_spec;
+    auto & plan_voc  = handler.plan_voc;

    auto opts = handler.opts; // copy
    opts.callback = callback;
@@ -484,19 +496,22 @@ void common_models_handler_apply(common_models_handler & handler, common_params
    }

    // handle hf_plan tasks
-    if (!plan.model_files.empty()) {
-        for (size_t i = 0; i < plan.model_files.size(); ++i) {
-            auto & model_file = plan.model_files[i];
+    auto add_tasks = [&opts, &tasks](const hf_cache::hf_files & model_files, common_params_model & model) {
+        for (size_t i = 0; i < model_files.size(); ++i) {
+            auto & model_file = model_files[i];
            bool is_first = (i == 0);
            tasks.emplace_back(model_file, opts, [&, is_first]() {
                if (is_first) {
                    // only use first part as model path
-                    params.model.path = hf_cache::finalize_file(model_file);
+                    model.path = hf_cache::finalize_file(model_file);
                } else {
                    hf_cache::finalize_file(model_file);
                }
            });
        }
+    };
+    if (!plan.model_files.empty()) {
+        add_tasks(plan.model_files, params.model);
    }
    if (!plan.mmproj.local_path.empty()) {
        tasks.emplace_back(plan.mmproj, opts, [&]() {
@@ -522,9 +537,31 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        });
    }

+    // handle plan_spec (e.g. --spec-draft-hf)
+    if (!plan_spec.model_files.empty()) {
+        add_tasks(plan_spec.model_files, params.speculative.draft.mparams);
+    }
+
+    // handle vocoder plan (e.g. --hf-repo-v)
+    if (!plan_voc.model_files.empty()) {
+        add_tasks(plan_voc.model_files, params.vocoder.model);
+    }
+
    // run all tasks in parallel
    if (!params.offline) {
-        common_download_run_tasks(tasks);
+        // if duplicated files are found, only download once (but still call on_done for each task)
+        std::unordered_map<std::string, common_download_task *> unique_tasks;
+        for (auto & task : tasks) {
+            auto it = unique_tasks.find(task.local_path);
+            if (it == unique_tasks.end()) {
+                unique_tasks[task.local_path] = &task;
+            }
+        }
+        std::vector<common_download_task> unique_tasks_vec;
+        for (auto & pair : unique_tasks) {
+            unique_tasks_vec.push_back(*pair.second);
+        }
+        common_download_run_tasks(unique_tasks_vec);
    }

    // download successful, update params with the downloaded paths
@@ -3711,6 +3748,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        "draft model for speculative decoding (default: unused)",
        [](common_params & params, const std::string & value) {
            params.speculative.draft.mparams.path = value;
+            params.speculative.draft.mparams.hf_file = value; // will be used if --spec-draft-hf is set
        }
    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_MODEL"));
    add_opt(common_arg(
@@ -133,6 +133,8 @@ void common_params_add_preset_options(std::vector<common_arg> & args);

 struct common_models_handler {
    common_download_hf_plan plan;
+    common_download_hf_plan plan_spec;
+    common_download_hf_plan plan_voc;
    common_download_opts opts;
 };

@@ -1551,6 +1551,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
        int split_backend_id = split->backend_id;
        ggml_backend_t split_backend = sched->backends[split_backend_id];

+        ggml_backend_synchronize(split_backend);
+
        // copy the input tensors to the split backend
        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1561,15 +1563,15 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                } else {
+                } else if (!split_backend->iface.cpy_tensor_async) {
                    ggml_backend_synchronize(split_backend);
                }
-                ggml_backend_tensor_copy(input, input_cpy);
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
            } else {
                // wait for the split backend to finish using the input before overwriting it
                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
-                } else {
+                } else if (!split_backend->iface.cpy_tensor_async) {
                    ggml_backend_synchronize(split_backend);
                }

@@ -1674,6 +1676,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
            }
        }

+        ggml_backend_synchronize(split_backend);
+
        if (!sched->callback_eval) {
            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
            if (ec != GGML_STATUS_SUCCESS) {
@@ -3192,11 +3192,24 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
    ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
    ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;

-    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
+    // Enables async copies from CPU to CUDA, instead of only CUDA-to-CUDA
+    // Excluding this path for HIP and MUSA as a precaution.
+    // According to the summary in https://github.com/ggml-org/llama.cpp/pull/20793#issuecomment-4275794315, this change is not beneficial for hip anyways.
+    // Additionally, there is a lot of anectodal evidence that hip/musa stream behavior might not always 1:1 match CUDA behavior.
+    // e.g. https://github.com/ROCm/rocm-systems/issues/5109
+    // It thus makes sense to exclude this path for HIP and MUSA. This PR was not aimed these backends, the majority of testing happened on CUDA.
+    // This can be revisited in the future if enabling copy_from_host benefits hip/MUSA, and if the PR author can extensively test on these backends.
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+    const bool copy_from_host = false;
+#else
+    const bool copy_from_host = ggml_backend_buffer_is_host(buf_src) && ggml_backend_dev_type(backend_src->device) == GGML_BACKEND_DEVICE_TYPE_CPU;
+#endif
+
+    if (!(copy_from_host || ggml_backend_is_cuda(backend_src)) || !ggml_backend_is_cuda(backend_dst)) {
        return false;
    }

-    if (!ggml_backend_buffer_is_cuda(buf_src) || !ggml_backend_buffer_is_cuda(buf_dst)) {
+    if (!(copy_from_host || ggml_backend_buffer_is_cuda(buf_src)) || !ggml_backend_buffer_is_cuda(buf_dst)) {
        return false;
    }

@@ -3207,14 +3220,17 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *) buf_src->context;
    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *) buf_dst->context;

-    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+    if ((copy_from_host && cuda_ctx_dst->device != buf_ctx_dst->device) ||
+        !copy_from_host && (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device)) {
 #ifndef NDEBUG
        GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
 #endif // NDEBUG
        return false;
    }

-    if (backend_src != backend_dst) {
+    if (copy_from_host) {
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyHostToDevice, cuda_ctx_dst->stream()));
+    } else if (backend_src != backend_dst) {
        // copy on src stream
        if (cuda_ctx_src->device == cuda_ctx_dst->device) {
            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
Author	SHA1	Message	Date
Adrien Gallouët	050ee92d04	app : allow --version, --licenses & --help (#25054 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2026-06-26 23:18:11 +02:00
Andreas Kieslinger	3fc4e10527	sched : reintroduce less synchronizations during split compute (#20793 ) * CUDA: Improve performance via less synchronizations between token (#17795) * Adds CPU-to-CUDA copy capability to ggml_backend_cuda_cpy_tensor_async() * Adds function to relax sync requirements between input copies on supported backends (CUDA for now) * Exchanges synchronous copy with async copy function. * Adds macro guards to allow compilation in non-CUDA builds * Reworked backend detection in ggml-backend.cpp to avoid linking conflicts * Relax requirement of checks in async CUDA copies from backend and buffer type to just buffer type, to avoid linking issues * Minor cleanup * Makes opt-in to relax use of explicit syncs more general. Backends like vulkan which require a synchronization between HtoD copies and graph execution could also adopt this change now. * Reintroduces stricter check for CPU->CUDA backend async copy via GGML_DEVICE_TYPE_CPU. * Corrects initialization of ggml_backend_sync_mode in ggml_backend_sched_split initialization * Simplifies synchronizations to adhere to `saaasg` pattern. * Apply suggestion from @ggerganov (src->buffer to buf_src) Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Apply suggestion from @ggerganov (src->buffer to buf_src) v2 Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Apply suggestions from @johannesgaessler code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * Adds single-GPU synchronizations to multi-GPU settings to fix hip backend pipeline parallel bugs. * Scheduler Hardening: Exclude hip/MUSA from copy_from_host CPU split -> GPU split optimization * Scheduler Hardening: Re-adding original additional synchronizations for non-async backends * Adds disclaimer to hip/musa exclusion of copy_from_host. Highlights that it is out of precaution, but that no perf-impact is visible, and that it can be revisited separately anytime. --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2026-06-26 17:18:30 +03:00
Adrien Gallouët	5d8ccdf9d1	devops : add llama in all docker images (#25035 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2026-06-26 15:15:48 +02:00
Xuan-Son Nguyen	024930c6ad	arg: fix handling --spec-draft-hf and --hf-repo-v (#25043 ) * arg: fix handling --spec-draft-hf and --hf-repo-v * fix missing mparams.hf_file	2026-06-26 14:36:03 +02:00