diff --git a/common/arg.cpp b/common/arg.cpp index aad8266ed7..f1e2bf6908 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -2644,6 +2644,27 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.mtmd_batch_max_tokens = value; } ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS")); + add_opt(common_arg( + {"--video-fps"}, "N", + string_format("target video frame rate (default: %.1f)", params.video_fps), + [](common_params & params, const std::string & value) { + params.video_fps = std::stof(value); + } + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FPS")); + add_opt(common_arg( + {"--video-timestamp-interval"}, "N", + string_format("interval in milliseconds between text timestamps (default: %" PRId64 ")", params.video_timestamp_interval_ms), + [](common_params & params, int value) { + params.video_timestamp_interval_ms = value; + } + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_TIMESTAMP_INTERVAL")); + add_opt(common_arg( + {"--video-ffmpeg-dir"}, "DIR", + "path to the directory containing ffmpeg and ffprobe (default: search in PATH)", + [](common_params & params, const std::string & value) { + params.video_ffmpeg_bin_dir = value; + } + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FFMPEG_DIR")); if (params.is_gen_docs || llama_supports_rpc()) { add_opt(common_arg( {"--rpc"}, "SERVERS", @@ -2699,6 +2720,19 @@ common_params_context common_params_parser_init(common_params & params, llama_ex else { throw std::invalid_argument("invalid value"); } } ).set_env("LLAMA_ARG_LOAD_MODE")); + add_opt(common_arg( + {"--tensor-read-lazy"}, "MODE", + "on-demand reading of certain tensors, for example per-layer embeddings (default: auto)\n" + "- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)\n" + "- auto: on, but only for tensors larger than 4 GiB\n" + "- off: always keep them resident", + [](common_params & params, const std::string & value) { + /**/ if (value == "on") { params.tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_ON; } + else if (value == "auto") { params.tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_AUTO; } + else if (value == "off") { params.tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_OFF; } + else { throw std::invalid_argument("invalid value"); } + } + ).set_env("LLAMA_ARG_TENSOR_READ_LAZY")); add_opt(common_arg( {"--numa"}, "TYPE", "attempt optimizations that help on some NUMA systems\n" @@ -4111,6 +4145,38 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.speculative.draft.n_min = value; } ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_MIN")); + add_opt(common_arg( + {"--spec-synth-len"}, "L", + "target mean synthetic acceptance length, including the target token (benchmarking only)", + [](common_params & params, const std::string & value) { + const std::string text = string_strip(value); + size_t pos = 0; + const double length = std::stod(text, &pos); + if (pos != text.size() || length == -1.0) { + throw std::invalid_argument("invalid value"); + } + params.speculative.synth_len = length; + } + ).set_spec().set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_SYNTH_LEN")); + add_opt(common_arg( + {"--spec-synth-rates"}, "P0,P1,...", + "comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)", + [](common_params & params, const std::string & value) { + const auto values = string_split(value, ','); + std::vector rates; + rates.reserve(values.size()); + for (const auto & raw : values) { + const std::string text = string_strip(raw); + size_t pos = 0; + const double rate = std::stod(text, &pos); + if (pos != text.size()) { + throw std::invalid_argument("invalid value"); + } + rates.push_back(rate); + } + params.speculative.synth_rates = std::move(rates); + } + ).set_spec().set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_SYNTH_RATES")); add_opt(common_arg( {"--spec-draft-p-split", "--draft-p-split"}, "P", diff --git a/common/common.cpp b/common/common.cpp index 3d54bd6002..347e8e9fc4 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1688,6 +1688,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) { mparams.main_gpu = params.main_gpu; mparams.split_mode = params.split_mode; mparams.load_mode = params.load_mode; + mparams.tensor_read_lazy = params.tensor_read_lazy; mparams.tensor_split = params.tensor_split; mparams.check_tensors = params.check_tensors; mparams.use_extra_bufts = !params.no_extra_bufts; diff --git a/common/common.h b/common/common.h index 51518f343a..82fed22092 100644 --- a/common/common.h +++ b/common/common.h @@ -370,6 +370,9 @@ struct common_params_speculative_ngram_cache { struct common_params_speculative { std::vector types = { COMMON_SPECULATIVE_TYPE_NONE }; + double synth_len = -1.0; + std::vector synth_rates; + // used by Simple, MTP, Eagle3, etc. - all methods that require some kind of draft model common_params_speculative_draft draft; @@ -384,6 +387,10 @@ struct common_params_speculative { return !draft.mparams.empty(); } + bool has_synth() const { + return synth_len != -1.0 || !synth_rates.empty(); + } + uint32_t need_n_rs_seq() const { bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) { return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK; @@ -476,6 +483,8 @@ struct common_params { enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs enum llama_load_mode load_mode = LLAMA_LOAD_MODE_AUTO; // how to load the model + enum llama_tensor_read_lazy tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_AUTO; // on-demand reading of tensors marked by the arch + common_cpu_params cpuparams; common_cpu_params cpuparams_batch; @@ -590,6 +599,11 @@ struct common_params { int image_max_tokens = -1; int mtmd_batch_max_tokens = 1024; + // for video input + float video_fps = 4.0f; + int64_t video_timestamp_interval_ms = 5000; + std::string video_ffmpeg_bin_dir = ""; + // finetune struct lr_opt lr; enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW; diff --git a/common/speculative.cpp b/common/speculative.cpp index 4eef2212e7..393a73bc39 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -14,6 +14,7 @@ #include #include +#include #include #include #include @@ -138,6 +139,7 @@ struct common_speculative_impl { const common_speculative_type type; uint32_t n_seq; + int32_t n_max; // maximum draft length after implementation-specific limits size_t n_call_begin = 0; // number of times this implementation was called for refresh. size_t n_call_draft = 0; // number of times this implementation was called for generation. @@ -157,7 +159,7 @@ struct common_speculative_impl { int64_t t_draft_us = 0; // total time spent in generating drafts in this implementation in microseconds. int64_t t_accept_us = 0; // total time spent in accumulation of this implementation in microseconds. - common_speculative_impl(common_speculative_type type, uint32_t n_seq) : type(type), n_seq(n_seq) {} + common_speculative_impl(common_speculative_type type, uint32_t n_seq, int32_t n_max) : type(type), n_seq(n_seq), n_max(n_max) {} virtual ~common_speculative_impl() = default; @@ -182,7 +184,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl { std::vector smpls; common_speculative_impl_draft_simple(const common_params_speculative & params, uint32_t n_seq) - : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, n_seq, params.draft.n_max) , params(params.draft) { auto * ctx_dft = this->params.ctx_dft; @@ -452,7 +454,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl { std::vector g_embd_buf; common_speculative_impl_draft_eagle3(const common_params_speculative & params, uint32_t n_seq) - : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq, params.draft.n_max) , params(params.draft) { SPC_TRC("%s", "adding speculative implementation 'draft-eagle3'\n"); @@ -937,7 +939,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl { common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq, common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH) - : common_speculative_impl(type, n_seq) + : common_speculative_impl(type, n_seq, params.draft.n_max) , params(params.draft) , is_dspark(type == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK) { @@ -983,6 +985,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl { this->params.n_max = std::min(this->params.n_max, n_draft_max); this->params.n_min = std::min(this->params.n_min, n_draft_max); } + this->n_max = this->params.n_max; batch = llama_batch_init(llama_n_batch(ctx_dft), 0, n_seq); batch_inject = llama_batch_init(llama_n_batch(ctx_dft), n_embd_dec, n_seq); @@ -1315,7 +1318,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { std::vector> chain_h; common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq) - : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq, params.draft.n_max) , params(params.draft) { auto * ctx_tgt = this->params.ctx_tgt; @@ -1382,6 +1385,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl { c.reserve((size_t) (this->params.n_max + 1) * n_embd); } } + this->n_max = this->params.n_max; pending_h.assign(n_seq, std::vector(n_embd, 0.0f)); @@ -1726,7 +1730,7 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl { common_speculative_impl_ngram_simple( const common_params_speculative & params, uint32_t n_seq, common_ngram_simple_config config) - : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq, params.ngram_simple.size_m) , params(params.ngram_simple) , config(config) { @@ -1770,7 +1774,7 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl { const common_ngram_map & config, uint32_t n_seq) : common_speculative_impl(config.key_only ? COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K - : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq) + : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq, config.size_value) { for (uint32_t i = 0; i < n_seq; i++) { this->config.push_back(config); @@ -1841,7 +1845,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl { common_speculative_impl_ngram_mod( const common_params_speculative & params, uint32_t n_seq) - : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, n_seq, params.ngram_mod.n_max) , params(params.ngram_mod) , mod(params.ngram_mod.n_match, 4*1024*1024) , verbose(std::getenv("LLAMA_TRACE") != nullptr) { @@ -2017,7 +2021,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl { const std::string & path_dynamic, bool save_dynamic, bool save_static) - : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq) + : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq, n_draft) , params(params.ngram_cache) , n_draft(n_draft) , save_dynamic(save_dynamic) @@ -2138,6 +2142,8 @@ struct common_speculative { // which implementaion was used for a given seq_id std::vector impl_last; + + std::vector synth_probs; }; static common_ngram_map get_common_ngram_map( @@ -2316,6 +2322,101 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) { return n_max; } +int32_t common_speculative_n_max(const common_speculative * spec) { + int32_t n_max = 0; + + if (spec == nullptr) { + return n_max; + } + + for (const auto & impl : spec->impls) { + n_max = std::max(n_max, std::max(0, impl->n_max)); + } + + return n_max; +} + +std::vector common_speculative_synth_rates_resolve(const common_params_speculative * spec, int32_t n_max) { + const bool has_length = spec->synth_len != -1.0; + const bool has_rates = !spec->synth_rates.empty(); + + if (!has_length && !has_rates) { + return {}; + } + if (has_length && has_rates) { + throw std::invalid_argument("synthetic acceptance length and rates are mutually exclusive"); + } + + if (n_max <= 0) { + throw std::invalid_argument("synthetic acceptance requires at least one speculative token"); + } + + if (has_rates) { + const auto & rates = spec->synth_rates; + if (rates.size() != (size_t) n_max) { + throw std::invalid_argument(string_format( + "synthetic acceptance rates must contain %d values, got %zu", n_max, rates.size())); + } + + for (size_t i = 0; i < rates.size(); ++i) { + if (!std::isfinite(rates[i]) || rates[i] < 0.0 || rates[i] > 1.0) { + throw std::invalid_argument("synthetic acceptance rates must be finite and within [0, 1]"); + } + if (i > 0 && rates[i] > rates[i - 1]) { + throw std::invalid_argument("synthetic acceptance rates must be monotonically non-increasing"); + } + } + + return rates; + } + + const double length = spec->synth_len; + const double length_max = (double) n_max + 1.0; + if (!std::isfinite(length) || length < 1.0 || length > length_max) { + throw std::invalid_argument(string_format( + "synthetic acceptance length must be finite and within [1, %.0f]", length_max)); + } + + double p = 0.0; + if (length == length_max) { + p = 1.0; + } else if (length > 1.0) { + double p_min = 0.0; + double p_max = 1.0; + for (int i = 0; i < 32; ++i) { + const double p_mid = 0.5 * (p_min + p_max); + double sum = 0.0; + double term = p_mid; + for (int32_t j = 0; j < n_max; ++j) { + sum += term; + term *= p_mid; + } + + if (sum < length - 1.0) { + p_min = p_mid; + } else { + p_max = p_mid; + } + } + p = 0.5 * (p_min + p_max); + } + + std::vector rates; + rates.reserve(n_max); + double rate = p; + for (int32_t i = 0; i < n_max; ++i) { + rates.push_back(rate); + rate *= p; + } + + return rates; +} + +const std::vector & common_speculative_get_synth_probs(const common_speculative * spec) { + GGML_ASSERT(spec); + return spec->synth_probs; +} + common_params common_base_params_to_speculative(const common_params & params) { const bool has_draft = params.speculative.has_dft(); @@ -2568,13 +2669,39 @@ common_speculative * common_speculative_init(common_params_speculative & params, return nullptr; } - auto * result = new common_speculative { - /* .dparams = */ common_speculative_draft_params_vec(n_seq), - /* .impls = */ std::move(impls), - /* .impl_last = */ std::vector(n_seq, nullptr) - }; + common_speculative_ptr result(new common_speculative { + /* .dparams = */ common_speculative_draft_params_vec(n_seq), + /* .impls = */ std::move(impls), + /* .impl_last = */ std::vector(n_seq, nullptr), + /* .synth_probs = */ {}, + }); - return result; + const int32_t n_max_configured = common_speculative_n_max(¶ms); + const int32_t n_max_effective = common_speculative_n_max(result.get()); + const auto rates = common_speculative_synth_rates_resolve(¶ms, n_max_effective); + + std::vector rates_str; + rates_str.reserve(rates.size()); + result->synth_probs.reserve(rates.size()); + double rate_prev = 1.0; + double acceptance_length = 1.0; + for (const double rate : rates) { + result->synth_probs.push_back(rate_prev > 0.0 ? rate / rate_prev : 0.0); + rates_str.push_back(string_format("%.6g", rate)); + rate_prev = rate; + acceptance_length += rate; + } + if (!result->synth_probs.empty()) { + SPC_WRN("%s", "synthetic speculative acceptance is enabled for benchmarking; generated output is not valid\n"); + if (n_max_effective != n_max_configured) { + SPC_WRN("synthetic acceptance draft limit was reduced from %d to %d by the initialized speculative implementations\n", + n_max_configured, n_max_effective); + } + SPC_INF("synthetic acceptance: n_max = %zu, mean length = %.6f, rates = [%s]\n", + rates.size(), acceptance_length, string_join(rates_str, ", ").c_str()); + } + + return result.release(); } void common_speculative_free(common_speculative * spec) { diff --git a/common/speculative.h b/common/speculative.h index 12ae31b7de..22505891f7 100644 --- a/common/speculative.h +++ b/common/speculative.h @@ -26,6 +26,15 @@ std::string common_speculative_type_to_str(enum common_speculative_type type); // return the max number of draft tokens based on the speculative parameters int32_t common_speculative_n_max(const common_params_speculative * spec); +// return the max number of draft tokens from the initialized implementations +int32_t common_speculative_n_max(const common_speculative * spec); + +// validate and resolve the unconditional synthetic acceptance rates +std::vector common_speculative_synth_rates_resolve(const common_params_speculative * spec, int32_t n_max); + +// return the conditional synthetic acceptance probabilities +const std::vector & common_speculative_get_synth_probs(const common_speculative * spec); + common_params common_base_params_to_speculative(const common_params & params); struct common_speculative_output_limits { diff --git a/conversion/nemotron.py b/conversion/nemotron.py index cd0d48c8f0..07fbc65314 100644 --- a/conversion/nemotron.py +++ b/conversion/nemotron.py @@ -302,6 +302,10 @@ class NemotronHModel(GraniteHybridModel): ) if not keep: return None + # PEFT names adapter tensors using model.layers.*, while Nemotron-H checkpoints + # and the GGUF tensor map use backbone.layers.* + if name.startswith("model.layers.") and ".mixer." in name: + name = name.replace("model.layers.", "backbone.layers.", 1) return super().filter_tensors((name, gen)) def prepare_metadata(self, vocab_only: bool): diff --git a/docs/speculative.md b/docs/speculative.md index 0f9f8a3d97..ffb1e34c7f 100644 --- a/docs/speculative.md +++ b/docs/speculative.md @@ -212,6 +212,15 @@ Use `--backend-sampling` to run supported target-model samplers on the model bac Unsupported samplers and device layouts fall back to CPU sampling. Tensor split mode does not support backend sampling. A fixed seed produces repeatable random draws, but stochastic CPU and backend sampling can still select different tokens because floating-point operations can differ between implementations and devices. Use greedy sampling when exact output matching is required. +### Synthetic Acceptance + +`llama-server` and `llama-cli` can replace normal speculative verification with synthetic decisions for benchmarking. The generated output is not valid model output because accepted draft tokens do not have to match the target model. + +Use exactly one of these options: + +- `--spec-synth-rates P0,P1,...` sets unconditional per-position acceptance probabilities. Entry `i` is the probability that the first `i+1` draft tokens are all accepted. The number of entries must match the effective maximum draft length. Values must be finite, within `[0, 1]`, and monotonically non-increasing. +- `--spec-synth-len L` sets the target mean acceptance length, including the target token. For `K` maximum draft tokens, `L` must be within `[1, K+1]`. The server finds a constant conditional probability `p` such that `p + p^2 + ... + p^K = L - 1`, then uses unconditional rates `[p, p^2, ..., p^K]`. + ### General Speculative Parameters ``` diff --git a/ggml/src/ggml-metal/ggml-metal-context.m b/ggml/src/ggml-metal/ggml-metal-context.m index 32d97cd5d0..1227ed39a0 100644 --- a/ggml/src/ggml-metal/ggml-metal-context.m +++ b/ggml/src/ggml-metal/ggml-metal-context.m @@ -84,106 +84,108 @@ struct ggml_metal { ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) { GGML_LOG_INFO("%s: allocating\n", __func__); + @autoreleasepool { #if TARGET_OS_OSX && !GGML_METAL_NDEBUG - // Show all the Metal device instances in the system - NSArray * devices = MTLCopyAllDevices(); - for (id device in devices) { - GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]); - } - [devices release]; // since it was created by a *Copy* C method + // Show all the Metal device instances in the system + NSArray * devices = MTLCopyAllDevices(); + for (id device in devices) { + GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]); + } + [devices release]; // since it was created by a *Copy* C method #endif - // init context - ggml_metal_t res = calloc(1, sizeof(struct ggml_metal)); + // init context + ggml_metal_t res = calloc(1, sizeof(struct ggml_metal)); - id device = ggml_metal_device_get_obj(dev); + id device = ggml_metal_device_get_obj(dev); - GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); - - // TODO: would it be better to have one queue for the backend and one queue for the device? - // the graph encoders and async ops would use the backend queue while the sync ops would use the device queue? - //res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND] - id queue = ggml_metal_device_get_queue(dev); - if (queue == nil) { - GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__); - return NULL; - } - - res->dev = dev; - res->lib = ggml_metal_device_get_library(dev); - if (res->lib == NULL) { - GGML_LOG_WARN("%s: the device does not have a precompiled Metal library - this is unexpected\n", __func__); - GGML_LOG_WARN("%s: will try to compile it on the fly\n", __func__); - - res->lib = ggml_metal_library_init(dev); - if (res->lib == NULL) { - GGML_LOG_ERROR("%s: error: failed to initialize the Metal library\n", __func__); - - free(res); + GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); + // TODO: would it be better to have one queue for the backend and one queue for the device? + // the graph encoders and async ops would use the backend queue while the sync ops would use the device queue? + //res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND] + id queue = ggml_metal_device_get_queue(dev); + if (queue == nil) { + GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__); return NULL; } - } - res->ev_cpy = ggml_metal_device_event_init(dev); + res->dev = dev; + res->lib = ggml_metal_device_get_library(dev); + if (res->lib == NULL) { + GGML_LOG_WARN("%s: the device does not have a precompiled Metal library - this is unexpected\n", __func__); + GGML_LOG_WARN("%s: will try to compile it on the fly\n", __func__); - const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev); + res->lib = ggml_metal_library_init(dev); + if (res->lib == NULL) { + GGML_LOG_ERROR("%s: error: failed to initialize the Metal library\n", __func__); - snprintf(res->name, sizeof(res->name), "%s", props_dev->name); + free(res); - res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); - - res->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil; - res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil; - - { - const char * val = getenv("GGML_METAL_GRAPH_DEBUG"); - res->debug_graph = val ? atoi(val) : 0; - } - - { - const char * val = getenv("GGML_METAL_FUSION_DEBUG"); - res->debug_fusion = val ? atoi(val) : 0; - } - - res->use_graph_optimize = true; - - if (getenv("GGML_METAL_GRAPH_OPTIMIZE_DISABLE") != NULL) { - res->use_graph_optimize = false; - } - - memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt)); - - GGML_LOG_INFO("%s: use fusion = %s\n", __func__, res->use_fusion ? "true" : "false"); - GGML_LOG_INFO("%s: use concurrency = %s\n", __func__, res->use_concurrency ? "true" : "false"); - GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false"); - - res->capture_compute = 0; - res->capture_started = false; - res->capture_scope = nil; - - { - const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE"); - if (val) { - res->capture_compute = atoi(val); + return NULL; + } } + + res->ev_cpy = ggml_metal_device_event_init(dev); + + const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev); + + snprintf(res->name, sizeof(res->name), "%s", props_dev->name); + + res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); + + res->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil; + res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil; + + { + const char * val = getenv("GGML_METAL_GRAPH_DEBUG"); + res->debug_graph = val ? atoi(val) : 0; + } + + { + const char * val = getenv("GGML_METAL_FUSION_DEBUG"); + res->debug_fusion = val ? atoi(val) : 0; + } + + res->use_graph_optimize = true; + + if (getenv("GGML_METAL_GRAPH_OPTIMIZE_DISABLE") != NULL) { + res->use_graph_optimize = false; + } + + memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt)); + + GGML_LOG_INFO("%s: use fusion = %s\n", __func__, res->use_fusion ? "true" : "false"); + GGML_LOG_INFO("%s: use concurrency = %s\n", __func__, res->use_concurrency ? "true" : "false"); + GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false"); + + res->capture_compute = 0; + res->capture_started = false; + res->capture_scope = nil; + + { + const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE"); + if (val) { + res->capture_compute = atoi(val); + } + } + + res->has_error = false; + + res->gf = nil; + res->encode_async = nil; + for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) { + res->cmd_bufs[i].obj = nil; + } + + res->cmd_bufs_ext = [[NSMutableArray alloc] init]; + + res->cmd_buf_last = nil; + + res->pipelines_ext = ggml_metal_pipelines_init(); + + return res; } - - res->has_error = false; - - res->gf = nil; - res->encode_async = nil; - for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) { - res->cmd_bufs[i].obj = nil; - } - - res->cmd_bufs_ext = [[NSMutableArray alloc] init]; - - res->cmd_buf_last = nil; - - res->pipelines_ext = ggml_metal_pipelines_init(); - - return res; } void ggml_metal_free(ggml_metal_t ctx) { diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m index 19c57820e8..41ce90dc8a 100644 --- a/ggml/src/ggml-metal/ggml-metal-device.m +++ b/ggml/src/ggml-metal/ggml-metal-device.m @@ -778,7 +778,9 @@ void ggml_metal_encoder_free(ggml_metal_encoder_t encoder) { } void ggml_metal_encoder_debug_group_push(ggml_metal_encoder_t encoder, const char * name) { - [encoder->obj pushDebugGroup:[NSString stringWithCString:name encoding:NSUTF8StringEncoding]]; + @autoreleasepool { + [encoder->obj pushDebugGroup:[NSString stringWithCString:name encoding:NSUTF8StringEncoding]]; + } } void ggml_metal_encoder_debug_group_pop (ggml_metal_encoder_t encoder) { @@ -1023,249 +1025,251 @@ ggml_metal_device_t ggml_metal_device_init(int device, int n_devices) { assert(dev != NULL); - if (dev->mtl_device == nil) { - dev->mtl_device = MTLCreateSystemDefaultDevice(); + @autoreleasepool { + if (dev->mtl_device == nil) { + dev->mtl_device = MTLCreateSystemDefaultDevice(); - if (dev->mtl_device) { - dev->mtl_queue = [dev->mtl_device newCommandQueue]; - if (dev->mtl_queue == nil) { - GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__); - } + if (dev->mtl_device) { + dev->mtl_queue = [dev->mtl_device newCommandQueue]; + if (dev->mtl_queue == nil) { + GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__); + } - dev->addr_virt = 0x000000400ULL; + dev->addr_virt = 0x000000400ULL; - dev->props.device = device; + dev->props.device = device; - // the Metal backend uses the system default device as the single physical device; - // additional (virtual) devices are emulated on top of it via GGML_METAL_DEVICES - dev->props.device_phys = 0; - dev->props.device_virt = device; + // the Metal backend uses the system default device as the single physical device; + // additional (virtual) devices are emulated on top of it via GGML_METAL_DEVICES + dev->props.device_phys = 0; + dev->props.device_virt = device; - dev->props.has_simdgroup_reduction = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7]; - dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; + dev->props.has_simdgroup_reduction = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7]; + dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; - dev->props.has_simdgroup_mm = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7]; - dev->props.has_unified_memory = dev->mtl_device.hasUnifiedMemory; + dev->props.has_simdgroup_mm = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7]; + dev->props.has_unified_memory = dev->mtl_device.hasUnifiedMemory; - dev->props.has_bfloat = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; - dev->props.has_bfloat |= [dev->mtl_device supportsFamily:MTLGPUFamilyApple6]; - if (getenv("GGML_METAL_BF16_DISABLE") != NULL) { - dev->props.has_bfloat = false; - } + dev->props.has_bfloat = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; + dev->props.has_bfloat |= [dev->mtl_device supportsFamily:MTLGPUFamilyApple6]; + if (getenv("GGML_METAL_BF16_DISABLE") != NULL) { + dev->props.has_bfloat = false; + } - dev->props.has_tensor = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal4_GGML]; - if (getenv("GGML_METAL_TENSOR_DISABLE") != NULL) { - dev->props.has_tensor = false; - } - - // note: disable the tensor API by default for old chips because with the current implementation it is not useful - // - M2 Ultra: ~5% slower - // - M4, M4 Max: no significant difference - // - // TODO: try to update the tensor API kernels to at least match the simdgroup performance - if (getenv("GGML_METAL_TENSOR_ENABLE") == NULL && - ![[dev->mtl_device name] containsString:@"M5"] && - ![[dev->mtl_device name] containsString:@"M6"] && - ![[dev->mtl_device name] containsString:@"A19"] && - ![[dev->mtl_device name] containsString:@"A20"]) { - GGML_LOG_INFO("%s: tensor API disabled for pre-M5 and pre-A19 devices\n", __func__); - dev->props.has_tensor = false; - } - - // double-check that the tensor API compiles - if (dev->props.has_tensor) { - const char * src_tensor_f16 = "\n" - "#include \n" - "#include \n" - "#include \n" - " \n" - "using namespace metal; \n" - "using namespace mpp::tensor_ops; \n" - " \n" - "kernel void dummy_kernel( \n" - " tensor> A [[buffer(0)]], \n" - " tensor> B [[buffer(1)]], \n" - " device float * C [[buffer(2)]], \n" - " uint2 tgid [[threadgroup_position_in_grid]]) \n" - "{ \n" - " auto tA = A.slice(0, (int)tgid.y); \n" - " auto tB = B.slice((int)tgid.x, 0); \n" - " \n" - " matmul2d< \n" - " matmul2d_descriptor(16, 16, dynamic_extent), \n" - " execution_simdgroups<4>> mm; \n" - " \n" - " auto cT = mm.get_destination_cooperative_tensor(); \n" - " \n" - " auto sA = tA.slice(0, 0); \n" - " auto sB = tB.slice(0, 0); \n" - " mm.run(sB, sA, cT); \n" - " \n" - " auto tC = tensor, tensor_inline>(C, dextents(16, 16)); \n" - " \n" - " cT.store(tC); \n" - "}"; - - GGML_LOG_INFO("%s: testing tensor API for f16 support\n", __func__); - ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_f16, false); - if (lib == NULL) { - GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__); + dev->props.has_tensor = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal4_GGML]; + if (getenv("GGML_METAL_TENSOR_DISABLE") != NULL) { dev->props.has_tensor = false; - } else { - struct ggml_metal_pipeline_with_params ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil); - if (!ppl.pipeline) { + } + + // note: disable the tensor API by default for old chips because with the current implementation it is not useful + // - M2 Ultra: ~5% slower + // - M4, M4 Max: no significant difference + // + // TODO: try to update the tensor API kernels to at least match the simdgroup performance + if (getenv("GGML_METAL_TENSOR_ENABLE") == NULL && + ![[dev->mtl_device name] containsString:@"M5"] && + ![[dev->mtl_device name] containsString:@"M6"] && + ![[dev->mtl_device name] containsString:@"A19"] && + ![[dev->mtl_device name] containsString:@"A20"]) { + GGML_LOG_INFO("%s: tensor API disabled for pre-M5 and pre-A19 devices\n", __func__); + dev->props.has_tensor = false; + } + + // double-check that the tensor API compiles + if (dev->props.has_tensor) { + const char * src_tensor_f16 = "\n" + "#include \n" + "#include \n" + "#include \n" + " \n" + "using namespace metal; \n" + "using namespace mpp::tensor_ops; \n" + " \n" + "kernel void dummy_kernel( \n" + " tensor> A [[buffer(0)]], \n" + " tensor> B [[buffer(1)]], \n" + " device float * C [[buffer(2)]], \n" + " uint2 tgid [[threadgroup_position_in_grid]]) \n" + "{ \n" + " auto tA = A.slice(0, (int)tgid.y); \n" + " auto tB = B.slice((int)tgid.x, 0); \n" + " \n" + " matmul2d< \n" + " matmul2d_descriptor(16, 16, dynamic_extent), \n" + " execution_simdgroups<4>> mm; \n" + " \n" + " auto cT = mm.get_destination_cooperative_tensor(); \n" + " \n" + " auto sA = tA.slice(0, 0); \n" + " auto sB = tB.slice(0, 0); \n" + " mm.run(sB, sA, cT); \n" + " \n" + " auto tC = tensor, tensor_inline>(C, dextents(16, 16)); \n" + " \n" + " cT.store(tC); \n" + "}"; + + GGML_LOG_INFO("%s: testing tensor API for f16 support\n", __func__); + ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_f16, false); + if (lib == NULL) { GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__); dev->props.has_tensor = false; + } else { + struct ggml_metal_pipeline_with_params ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil); + if (!ppl.pipeline) { + GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__); + dev->props.has_tensor = false; + } + + ggml_metal_library_free(lib); } - - ggml_metal_library_free(lib); } - } - // try to compile a dummy kernel to determine if the tensor API is supported for bfloat - if (dev->props.has_tensor && dev->props.has_bfloat) { - const char * src_tensor_bf16 = "\n" - "#include \n" - "#include \n" - "#include \n" - " \n" - "using namespace metal; \n" - "using namespace mpp::tensor_ops; \n" - " \n" - "kernel void dummy_kernel( \n" - " tensor> A [[buffer(0)]], \n" - " tensor> B [[buffer(1)]], \n" - " device float * C [[buffer(2)]], \n" - " uint2 tgid [[threadgroup_position_in_grid]]) \n" - "{ \n" - " auto tA = A.slice(0, (int)tgid.y); \n" - " auto tB = B.slice((int)tgid.x, 0); \n" - " \n" - " matmul2d< \n" - " matmul2d_descriptor(16, 16, dynamic_extent), \n" - " execution_simdgroups<4>> mm; \n" - " \n" - " auto cT = mm.get_destination_cooperative_tensor(); \n" - " \n" - " auto sA = tA.slice(0, 0); \n" - " auto sB = tB.slice(0, 0); \n" - " mm.run(sB, sA, cT); \n" - " \n" - " auto tC = tensor, tensor_inline>(C, dextents(16, 16)); \n" - " \n" - " cT.store(tC); \n" - "}"; + // try to compile a dummy kernel to determine if the tensor API is supported for bfloat + if (dev->props.has_tensor && dev->props.has_bfloat) { + const char * src_tensor_bf16 = "\n" + "#include \n" + "#include \n" + "#include \n" + " \n" + "using namespace metal; \n" + "using namespace mpp::tensor_ops; \n" + " \n" + "kernel void dummy_kernel( \n" + " tensor> A [[buffer(0)]], \n" + " tensor> B [[buffer(1)]], \n" + " device float * C [[buffer(2)]], \n" + " uint2 tgid [[threadgroup_position_in_grid]]) \n" + "{ \n" + " auto tA = A.slice(0, (int)tgid.y); \n" + " auto tB = B.slice((int)tgid.x, 0); \n" + " \n" + " matmul2d< \n" + " matmul2d_descriptor(16, 16, dynamic_extent), \n" + " execution_simdgroups<4>> mm; \n" + " \n" + " auto cT = mm.get_destination_cooperative_tensor(); \n" + " \n" + " auto sA = tA.slice(0, 0); \n" + " auto sB = tB.slice(0, 0); \n" + " mm.run(sB, sA, cT); \n" + " \n" + " auto tC = tensor, tensor_inline>(C, dextents(16, 16)); \n" + " \n" + " cT.store(tC); \n" + "}"; - GGML_LOG_INFO("%s: testing tensor API for bfloat support\n", __func__); - ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_bf16, false); - if (lib == NULL) { - GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__); - dev->props.has_bfloat = false; - } else { - struct ggml_metal_pipeline_with_params ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil); - if (!ppl.pipeline) { + GGML_LOG_INFO("%s: testing tensor API for bfloat support\n", __func__); + ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_bf16, false); + if (lib == NULL) { GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__); dev->props.has_bfloat = false; + } else { + struct ggml_metal_pipeline_with_params ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil); + if (!ppl.pipeline) { + GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__); + dev->props.has_bfloat = false; + } + + ggml_metal_library_free(lib); } - - ggml_metal_library_free(lib); } - } - dev->props.use_residency_sets = true; + dev->props.use_residency_sets = true; #if defined(GGML_METAL_HAS_RESIDENCY_SETS) - dev->props.use_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil; + dev->props.use_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil; #endif - dev->props.use_shared_buffers = dev->props.has_unified_memory; + dev->props.use_shared_buffers = dev->props.has_unified_memory; #if TARGET_OS_OSX - // In case of eGPU, shared memory may be preferable. - dev->props.use_shared_buffers |= [dev->mtl_device location] == MTLDeviceLocationExternal; + // In case of eGPU, shared memory may be preferable. + dev->props.use_shared_buffers |= [dev->mtl_device location] == MTLDeviceLocationExternal; #endif - if (getenv("GGML_METAL_SHARED_BUFFERS_DISABLE") != NULL) { - dev->props.use_shared_buffers = false; - } - if (getenv("GGML_METAL_SHARED_BUFFERS_ENABLE") != NULL) { - dev->props.use_shared_buffers = true; - } + if (getenv("GGML_METAL_SHARED_BUFFERS_DISABLE") != NULL) { + dev->props.use_shared_buffers = false; + } + if (getenv("GGML_METAL_SHARED_BUFFERS_ENABLE") != NULL) { + dev->props.use_shared_buffers = true; + } - dev->props.supports_gpu_family_apple7 = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7]; + dev->props.supports_gpu_family_apple7 = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7]; - dev->props.device_id = ggml_metal_device_id_parse([[dev->mtl_device name] UTF8String]); + dev->props.device_id = ggml_metal_device_id_parse([[dev->mtl_device name] UTF8String]); - dev->props.op_offload_min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32; + dev->props.op_offload_min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32; - dev->props.max_buffer_size = dev->mtl_device.maxBufferLength; - dev->props.max_theadgroup_memory_size = dev->mtl_device.maxThreadgroupMemoryLength; - if (@available(macOS 10.12, iOS 16.0, *)) { - dev->props.max_working_set_size = dev->mtl_device.recommendedMaxWorkingSetSize; - } else { - dev->props.max_working_set_size = dev->mtl_device.maxBufferLength; - } + dev->props.max_buffer_size = dev->mtl_device.maxBufferLength; + dev->props.max_theadgroup_memory_size = dev->mtl_device.maxThreadgroupMemoryLength; + if (@available(macOS 10.12, iOS 16.0, *)) { + dev->props.max_working_set_size = dev->mtl_device.recommendedMaxWorkingSetSize; + } else { + dev->props.max_working_set_size = dev->mtl_device.maxBufferLength; + } - snprintf(dev->props.name, sizeof(dev->props.name), "%s%d", "MTL", device); - const char * gpu_name = [[dev->mtl_device name] UTF8String]; - if (n_devices > 1) { - snprintf(dev->props.desc, sizeof(dev->props.desc), "%s (dev p%d/v%d)", - gpu_name, dev->props.device_phys, dev->props.device_virt); - } else { - snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", gpu_name); - } + snprintf(dev->props.name, sizeof(dev->props.name), "%s%d", "MTL", device); + const char * gpu_name = [[dev->mtl_device name] UTF8String]; + if (n_devices > 1) { + snprintf(dev->props.desc, sizeof(dev->props.desc), "%s (dev p%d/v%d)", + gpu_name, dev->props.device_phys, dev->props.device_virt); + } else { + snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", gpu_name); + } - dev->library = ggml_metal_library_init(dev); - if (!dev->library) { - GGML_LOG_ERROR("%s: error: failed to create library\n", __func__); - } + dev->library = ggml_metal_library_init(dev); + if (!dev->library) { + GGML_LOG_ERROR("%s: error: failed to create library\n", __func__); + } - if (dev->props.use_residency_sets) { - dev->rsets = ggml_metal_rsets_init(dev); - } else { - dev->rsets = nil; - } + if (dev->props.use_residency_sets) { + dev->rsets = ggml_metal_rsets_init(dev); + } else { + dev->rsets = nil; + } - // print MTL GPU family: - GGML_LOG_INFO("%s: GPU name: %s (%s)\n", __func__, dev->props.name, dev->props.desc); + // print MTL GPU family: + GGML_LOG_INFO("%s: GPU name: %s (%s)\n", __func__, dev->props.name, dev->props.desc); - // determine max supported GPU family - // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf - // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf - { - for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) { - if ([dev->mtl_device supportsFamily:i]) { - dev->props.gpu_family = i - (int) MTLGPUFamilyApple1 + 1; - GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, dev->props.gpu_family, i); - break; + // determine max supported GPU family + // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf + // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf + { + for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) { + if ([dev->mtl_device supportsFamily:i]) { + dev->props.gpu_family = i - (int) MTLGPUFamilyApple1 + 1; + GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, dev->props.gpu_family, i); + break; + } + } + + for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) { + if ([dev->mtl_device supportsFamily:i]) { + GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i); + break; + } + } + + for (int i = MTLGPUFamilyMetal3_GGML + 5; i >= MTLGPUFamilyMetal3_GGML; --i) { + if ([dev->mtl_device supportsFamily:i]) { + GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3_GGML + 3, i); + break; + } } } - for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) { - if ([dev->mtl_device supportsFamily:i]) { - GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i); - break; - } - } - - for (int i = MTLGPUFamilyMetal3_GGML + 5; i >= MTLGPUFamilyMetal3_GGML; --i) { - if ([dev->mtl_device supportsFamily:i]) { - GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3_GGML + 3, i); - break; - } - } - } - - GGML_LOG_INFO("%s: simdgroup reduction = %s\n", __func__, dev->props.has_simdgroup_reduction ? "true" : "false"); - GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, dev->props.has_simdgroup_mm ? "true" : "false"); - GGML_LOG_INFO("%s: has unified memory = %s\n", __func__, dev->props.has_unified_memory ? "true" : "false"); - GGML_LOG_INFO("%s: has bfloat = %s\n", __func__, dev->props.has_bfloat ? "true" : "false"); - GGML_LOG_INFO("%s: has tensor = %s\n", __func__, dev->props.has_tensor ? "true" : "false"); - GGML_LOG_INFO("%s: use residency sets = %s\n", __func__, dev->props.use_residency_sets ? "true" : "false"); - GGML_LOG_INFO("%s: use shared buffers = %s\n", __func__, dev->props.use_shared_buffers ? "true" : "false"); + GGML_LOG_INFO("%s: simdgroup reduction = %s\n", __func__, dev->props.has_simdgroup_reduction ? "true" : "false"); + GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, dev->props.has_simdgroup_mm ? "true" : "false"); + GGML_LOG_INFO("%s: has unified memory = %s\n", __func__, dev->props.has_unified_memory ? "true" : "false"); + GGML_LOG_INFO("%s: has bfloat = %s\n", __func__, dev->props.has_bfloat ? "true" : "false"); + GGML_LOG_INFO("%s: has tensor = %s\n", __func__, dev->props.has_tensor ? "true" : "false"); + GGML_LOG_INFO("%s: use residency sets = %s\n", __func__, dev->props.use_residency_sets ? "true" : "false"); + GGML_LOG_INFO("%s: use shared buffers = %s\n", __func__, dev->props.use_shared_buffers ? "true" : "false"); #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15) - if (@available(macOS 10.12, iOS 16.0, *)) { - GGML_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, dev->props.max_working_set_size / 1e6); - } + if (@available(macOS 10.12, iOS 16.0, *)) { + GGML_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, dev->props.max_working_set_size / 1e6); + } #endif + } } } diff --git a/include/llama.h b/include/llama.h index 84313cfd74..b8020a7fce 100644 --- a/include/llama.h +++ b/include/llama.h @@ -214,6 +214,12 @@ extern "C" { LLAMA_API const char * llama_load_mode_name(enum llama_load_mode load_mode); LLAMA_API enum llama_load_mode llama_load_mode_from_str(const char * str); + enum llama_tensor_read_lazy { + LLAMA_TENSOR_READ_LAZY_OFF = 0, // always read the whole tensor up front + LLAMA_TENSOR_READ_LAZY_AUTO = 1, // lazy only for marked tensors larger than 4 GiB (requires mmap) + LLAMA_TENSOR_READ_LAZY_ON = 2, // read the rows of tensors marked by the arch on demand (requires mmap) + }; + enum llama_context_type { LLAMA_CONTEXT_TYPE_DEFAULT = 0, LLAMA_CONTEXT_TYPE_MTP = 1, @@ -315,6 +321,8 @@ extern "C" { enum llama_split_mode split_mode; // how to split the model across multiple GPUs enum llama_load_mode load_mode; // how to load the model + enum llama_tensor_read_lazy tensor_read_lazy; // on-demand reading of tensors marked by the arch + // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE int32_t main_gpu; diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp index ed572da7fb..4d183cbc9c 100644 --- a/src/llama-mmap.cpp +++ b/src/llama-mmap.cpp @@ -438,11 +438,34 @@ void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); } // llama_mmap +#if defined(_POSIX_MAPPED_FILES) || defined(_WIN32) +// merge `ranges` and return their complement within [0, limit) +static llama_mmap::ranges ranges_complement(llama_mmap::ranges ranges, size_t limit) { + llama_mmap::ranges res; + std::sort(ranges.begin(), ranges.end()); + + size_t pos = 0; + for (const auto & range : ranges) { + const size_t beg = std::min(range.first, limit); + const size_t end = std::min(range.second, limit); + if (beg > pos) { + res.emplace_back(pos, beg); + } + pos = std::max(pos, end); + } + if (pos < limit) { + res.emplace_back(pos, limit); + } + + return res; +} +#endif + struct llama_mmap::impl { #ifdef _POSIX_MAPPED_FILES std::vector> mapped_fragments; - impl(struct llama_file * file, size_t prefetch, bool numa) { + impl(struct llama_file * file, size_t prefetch, bool numa, const llama_mmap::ranges & lazy_ranges) { size = file->size(); int fd = file->file_id(); int flags = MAP_SHARED; @@ -452,18 +475,34 @@ struct llama_mmap::impl { LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n", strerror(errno)); } - if (prefetch) { flags |= MAP_POPULATE; } + // MAP_POPULATE would fault in the lazy ranges too + if (prefetch && lazy_ranges.empty()) { flags |= MAP_POPULATE; } #endif addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0); if (addr == MAP_FAILED) { throw std::runtime_error(format("mmap failed: %s", strerror(errno))); } - if (prefetch > 0) { - if (posix_madvise(addr, std::min(file->size(), prefetch), POSIX_MADV_WILLNEED)) { - LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n", - strerror(errno)); + // page-aligned madvise over [beg, end), clamped to the file + auto advise = [&](size_t beg, size_t end, int advice, const char * name) { + const size_t page_size = sysconf(_SC_PAGESIZE); + beg = beg & ~(page_size - 1); + end = std::min((end + page_size - 1) & ~(page_size - 1), file->size()); + if (beg >= end) { + return; } + if (posix_madvise((char *) addr + beg, end - beg, advice)) { + LLAMA_LOG_WARN("warning: posix_madvise(.., %s) failed: %s\n", name, strerror(errno)); + } + }; + + if (prefetch > 0) { + for (const auto & range : ranges_complement(lazy_ranges, std::min(file->size(), prefetch))) { + advise(range.first, range.second, POSIX_MADV_WILLNEED, "POSIX_MADV_WILLNEED"); + } + } + for (const auto & range : lazy_ranges) { + advise(range.first, range.second, POSIX_MADV_RANDOM, "POSIX_MADV_RANDOM"); } if (numa) { if (posix_madvise(addr, file->size(), POSIX_MADV_RANDOM)) { @@ -533,7 +572,7 @@ struct llama_mmap::impl { #elif defined(_WIN32) HANDLE hMapping = nullptr; - impl(struct llama_file * file, size_t prefetch, bool numa) { + impl(struct llama_file * file, size_t prefetch, bool numa, const llama_mmap::ranges & lazy_ranges) { GGML_UNUSED(numa); size = file->size(); @@ -563,10 +602,15 @@ struct llama_mmap::impl { pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory"); if (pPrefetchVirtualMemory) { - WIN32_MEMORY_RANGE_ENTRY range; - range.VirtualAddress = addr; - range.NumberOfBytes = (SIZE_T) std::min(size, prefetch); - if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) { + std::vector entries; + for (const auto & range : ranges_complement(lazy_ranges, std::min(size, prefetch))) { + WIN32_MEMORY_RANGE_ENTRY entry; + entry.VirtualAddress = (char *) addr + range.first; + entry.NumberOfBytes = (SIZE_T) (range.second - range.first); + entries.push_back(entry); + } + if (!entries.empty() && + !pPrefetchVirtualMemory(GetCurrentProcess(), (ULONG_PTR) entries.size(), entries.data(), 0)) { LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n", llama_format_win_err(GetLastError()).c_str()); } @@ -597,10 +641,11 @@ struct llama_mmap::impl { } } #else - impl(struct llama_file * file, size_t prefetch, bool numa) { + impl(struct llama_file * file, size_t prefetch, bool numa, const llama_mmap::ranges & lazy_ranges) { GGML_UNUSED(file); GGML_UNUSED(prefetch); GGML_UNUSED(numa); + GGML_UNUSED(lazy_ranges); throw std::runtime_error("mmap not supported"); } @@ -617,7 +662,8 @@ struct llama_mmap::impl { size_t size; }; -llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa) : pimpl(std::make_unique(file, prefetch, numa)) {} +llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa, + const ranges & lazy_ranges) : pimpl(std::make_unique(file, prefetch, numa, lazy_ranges)) {} llama_mmap::~llama_mmap() = default; size_t llama_mmap::size() const { return pimpl->size; } diff --git a/src/llama-mmap.h b/src/llama-mmap.h index b7d5c61e95..cc28c8a73f 100644 --- a/src/llama-mmap.h +++ b/src/llama-mmap.h @@ -2,6 +2,7 @@ #include #include +#include #include #include @@ -41,8 +42,12 @@ private: }; struct llama_mmap { + // list of [first, last) byte ranges within a file + using ranges = std::vector>; + llama_mmap(const llama_mmap &) = delete; - llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false); + llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false, + const ranges & lazy_ranges = {}); ~llama_mmap(); size_t size() const; diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp index 5071556a1f..201176edbf 100644 --- a/src/llama-model-loader.cpp +++ b/src/llama-model-loader.cpp @@ -1287,6 +1287,18 @@ struct ggml_tensor * llama_model_loader::create_tensor( return NULL; } + if ((flags & TENSOR_READ_LAZY) && use_mmap && tensor_read_lazy != LLAMA_TENSOR_READ_LAZY_OFF) { + // in auto mode, small tensors are cheap enough to keep resident + constexpr size_t auto_lazy_min_size = 4ull * 1024 * 1024 * 1024; + if (tensor_read_lazy == LLAMA_TENSOR_READ_LAZY_ON || ggml_nbytes(cur) > auto_lazy_min_size) { + const auto & w = require_weight(tn.str().c_str()); + lazy_tensor_ranges[w.idx].emplace_back(w.offs, w.offs + ggml_nbytes(cur)); + + LLAMA_LOG_INFO("%s: tensor %s (size = %zu MiB) lazy read enabled\n", + __func__, tn.str().c_str(), ggml_nbytes(cur)/1024/1024); + } + } + ggml_tensor t_meta = *cur; if (flags & TENSOR_ALLOW_RESHAPE) { for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) { @@ -1354,7 +1366,9 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps if (use_mmap) { mappings.reserve(files.size()); mmaps_used.reserve(files.size()); - for (const auto & file : files) { + for (uint32_t idx = 0; idx < files.size(); idx++) { + const auto & file = files[idx]; + bool is_numa = false; auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); @@ -1366,7 +1380,11 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps } } - std::unique_ptr mapping = std::make_unique(file.get(), prefetch ? -1 : 0, is_numa); + const auto it_lazy = lazy_tensor_ranges.find(idx); + static const llama_mmap::ranges no_lazy_ranges; + + std::unique_ptr mapping = std::make_unique(file.get(), prefetch ? -1 : 0, is_numa, + it_lazy != lazy_tensor_ranges.end() ? it_lazy->second : no_lazy_ranges); mmaps_used.emplace_back(mapping->size(), 0); if (mlock_mmaps) { std::unique_ptr mlock_mmap(new llama_mlock()); diff --git a/src/llama-model-loader.h b/src/llama-model-loader.h index e9fe3592d4..ec5b692463 100644 --- a/src/llama-model-loader.h +++ b/src/llama-model-loader.h @@ -68,6 +68,7 @@ struct llama_model_loader { static const int TENSOR_SKIP = 1 << 2; static const int TENSOR_SKIP_IF_VIRTUAL = 1 << 3; static const int TENSOR_ALLOW_RESHAPE = 1 << 4; + static const int TENSOR_READ_LAZY = 1 << 5; // read rows on demand instead of loading whole tensor; requires mmap for now int n_kv = 0; int n_tensors = 0; @@ -82,12 +83,18 @@ struct llama_model_loader { bool no_alloc; bool load_mtp; + // set by the caller before the create_tensor() calls + enum llama_tensor_read_lazy tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_OFF; + llama_files files; llama_ftype ftype; llama_fver fver; llama_mmaps mappings; + // byte ranges of TENSOR_READ_LAZY tensors, per file index + std::map lazy_tensor_ranges; + std::map weights_map; std::unordered_map kv_overrides; const llama_model_tensor_buft_override * tensor_buft_overrides; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index e4f9f7bea4..f1d84e5168 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2681,6 +2681,7 @@ llama_model_params llama_model_default_params() { /*.n_gpu_layers =*/ -1, /*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER, /*.load_mode =*/ LLAMA_LOAD_MODE_AUTO, + /*.tensor_read_lazy =*/ LLAMA_TENSOR_READ_LAZY_AUTO, /*.main_gpu =*/ 0, /*.tensor_split =*/ nullptr, /*.progress_callback =*/ nullptr, @@ -3118,7 +3119,8 @@ llama_model_base::llama_model_base(const struct llama_model_params & params) : l TENSOR_NOT_REQUIRED (llama_model_loader::TENSOR_NOT_REQUIRED), TENSOR_SKIP (llama_model_loader::TENSOR_SKIP), TENSOR_SKIP_IF_VIRTUAL(llama_model_loader::TENSOR_SKIP_IF_VIRTUAL), - TENSOR_ALLOW_RESHAPE (llama_model_loader::TENSOR_ALLOW_RESHAPE) {} + TENSOR_ALLOW_RESHAPE (llama_model_loader::TENSOR_ALLOW_RESHAPE), + TENSOR_READ_LAZY (llama_model_loader::TENSOR_READ_LAZY) {} ggml_tensor * llama_model_base::create_tensor(const LLM_TN_IMPL & tn, const std::initializer_list & ne, int flags) { GGML_ASSERT(ml != nullptr); diff --git a/src/llama-model.h b/src/llama-model.h index a2c25c6381..8bb2d7c0bf 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -777,6 +777,7 @@ struct llama_model_base : public llama_model { const int TENSOR_SKIP; const int TENSOR_SKIP_IF_VIRTUAL; const int TENSOR_ALLOW_RESHAPE; + const int TENSOR_READ_LAZY; explicit llama_model_base(const llama_model_params & params); virtual ~llama_model_base() = default; diff --git a/src/llama.cpp b/src/llama.cpp index 1609fec88d..9c841ee352 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -318,6 +318,8 @@ static std::pair llama_model_load(struct gguf_context * meta llama_model_loader ml(metadata, set_tensor_data, set_tensor_data_ud, fname, splits, file, params.load_mode, params.check_tensors, params.no_alloc, params.load_mtp, params.kv_overrides, params.tensor_buft_overrides); + ml.tensor_read_lazy = params.tensor_read_lazy; + ml.print_info(); std::unique_ptr model_ptr(llama_model_create(ml, params)); diff --git a/src/models/gemma4.cpp b/src/models/gemma4.cpp index e44f423bdb..aa518c6df5 100644 --- a/src/models/gemma4.cpp +++ b/src/models/gemma4.cpp @@ -50,7 +50,7 @@ void llama_model_gemma4::load_arch_tensors(llama_model_loader &) { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); if (n_embd_per_layer > 0) { - per_layer_tok_embd = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_per_layer * n_layer, n_vocab}, 0); + per_layer_tok_embd = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_per_layer * n_layer, n_vocab}, TENSOR_READ_LAZY); per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0); per_layer_proj_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM, "weight", 0), {n_embd_per_layer}, 0); } diff --git a/src/models/minimax-01.cpp b/src/models/minimax-01.cpp index a6ccee1917..f14626b2c7 100644 --- a/src/models/minimax-01.cpp +++ b/src/models/minimax-01.cpp @@ -174,11 +174,9 @@ public: bool can_reuse(const llm_graph_params & params) override { bool res = true; - if (params.ubatch.n_seq_tokens > 1) { - res &= ( inp_q_decay && inp_q_decay->ne[2] == params.ubatch.n_seq_tokens); - res &= ( inp_k_decay && inp_k_decay->ne[2] == params.ubatch.n_seq_tokens); - res &= (inp_diag_decay && inp_diag_decay->ne[1] == params.ubatch.n_seq_tokens); - } + res &= ( inp_q_decay && inp_q_decay->ne[2] == params.ubatch.n_seq_tokens); + res &= ( inp_k_decay && inp_k_decay->ne[2] == params.ubatch.n_seq_tokens); + res &= (inp_diag_decay && inp_diag_decay->ne[1] == params.ubatch.n_seq_tokens); return res; } @@ -223,19 +221,17 @@ llama_model_minimax_01::graph::graph(const llama_model & model, const llm_graph_ ggml_set_input(inp->inp_slopes); cb(inp->inp_slopes, "slopes", -1); - if (n_seq_tokens != 1) { - inp->inp_q_decay = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, n_head, n_seq_tokens, n_seqs); - ggml_set_input(inp->inp_q_decay); - cb(inp->inp_q_decay, "q_decay_exp", -1); + inp->inp_q_decay = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, n_head, n_seq_tokens, n_seqs); + ggml_set_input(inp->inp_q_decay); + cb(inp->inp_q_decay, "q_decay_exp", -1); - inp->inp_k_decay = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, n_head, n_seq_tokens, n_seqs); - ggml_set_input(inp->inp_k_decay); - cb(inp->inp_k_decay, "k_decay_exp", -1); + inp->inp_k_decay = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, n_head, n_seq_tokens, n_seqs); + ggml_set_input(inp->inp_k_decay); + cb(inp->inp_k_decay, "k_decay_exp", -1); - inp->inp_diag_decay = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_seq_tokens, n_seq_tokens, n_head, n_seqs); - ggml_set_input(inp->inp_diag_decay); - cb(inp->inp_diag_decay, "diag_decay_exp", -1); - } + inp->inp_diag_decay = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_seq_tokens, n_seq_tokens, n_head, n_seqs); + ggml_set_input(inp->inp_diag_decay); + cb(inp->inp_diag_decay, "diag_decay_exp", -1); la = (llm_graph_input_la *) res->add_input(std::move(inp)); @@ -319,41 +315,8 @@ llama_model_minimax_01::graph::graph(const llama_model & model, const llm_graph_ ggml_tensor * qkv = nullptr; ggml_tensor * kv_new = nullptr; - - if (n_seq_tokens == 1) { - // lightning attention - optimized single token case for TG - - ggml_tensor * slopes_neg = ggml_scale(ctx0, slope_rate, -1.0); - cb(slopes_neg, "slopes_neg", il); - - ggml_tensor * ratio = ggml_exp(ctx0, slopes_neg); - cb(ratio, "ratio", il); - - ggml_tensor * ratio_3d = ggml_reshape_3d(ctx0, ratio, 1, 1, n_head); - cb(ratio_3d, "ratio3d", il); - - ggml_tensor * v_trans = ggml_cont(ctx0, ggml_permute(ctx0, Vcur, 1, 2, 0, 3)); - cb(v_trans, "v_trans", il); - - ggml_tensor * k_trans = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 1, 2, 0, 3)); - cb(k_trans, "k_trans", il); - - ggml_tensor * kv_cur = ggml_mul_mat(ctx0, k_trans, v_trans); - cb(kv_cur, "kv_cur", il); - - ggml_tensor * kv_old_s = ggml_mul(ctx0, kv_old, ratio_3d); - cb(kv_old_s, "kv_old_s", il); - - kv_new = ggml_add(ctx0, kv_old_s, kv_cur); - cb(kv_new, "kv_new", il); - - ggml_tensor * q_trans = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - cb(q_trans, "q_trans", il); - - qkv = ggml_mul_mat(ctx0, kv_new, q_trans); - cb(qkv, "qkv", il); - } else if(n_seq_tokens > 1) { - // lightning attention - general multi token case for PP + { + // lightning attention ggml_tensor * q_decay_exp = la->inp_q_decay; ggml_tensor * k_decay_exp = la->inp_k_decay; diff --git a/tests/test-arg-parser.cpp b/tests/test-arg-parser.cpp index ba58f852eb..e0907631ab 100644 --- a/tests/test-arg-parser.cpp +++ b/tests/test-arg-parser.cpp @@ -4,6 +4,7 @@ #include "llama.h" #include "speculative.h" +#include #include #include #include @@ -34,6 +35,62 @@ static void test(void) { std::numeric_limits::max(), std::numeric_limits::max()); + { + common_params_speculative spec; + spec.synth_len = 3.4; + + auto assert_invalid = [](const common_params_speculative & value, int32_t n_max) { + try { + common_speculative_synth_rates_resolve(&value, n_max); + assert(false); + } catch (const std::invalid_argument &) { + } + }; + + const auto rates = common_speculative_synth_rates_resolve(&spec, 4); + assert(rates.size() == 4); + assert(std::abs(rates[0] - 0.80581) < 1e-5); + assert(std::abs(rates[1] - 0.64933) < 1e-5); + assert(std::abs(rates[2] - 0.52323) < 1e-5); + assert(std::abs(rates[3] - 0.42163) < 1e-5); + assert(std::abs(1.0 + rates[0] + rates[1] + rates[2] + rates[3] - 3.4) < 1e-8); + + spec.synth_len = 1.0; + assert(common_speculative_synth_rates_resolve(&spec, 4) == std::vector({0.0, 0.0, 0.0, 0.0})); + + spec.synth_len = 5.0; + assert(common_speculative_synth_rates_resolve(&spec, 4) == std::vector({1.0, 1.0, 1.0, 1.0})); + + spec.synth_len = 5.1; + assert_invalid(spec, 4); + + spec.synth_len = std::numeric_limits::quiet_NaN(); + assert_invalid(spec, 4); + + spec.synth_len = 0.0; + assert_invalid(spec, 4); + + spec.synth_len = -1.0; + spec.synth_rates = {0.8, 0.6, 0.4}; + assert_invalid(spec, 4); + + spec.synth_rates = {0.8, 0.6, 0.4, 0.2}; + assert(common_speculative_synth_rates_resolve(&spec, 4) == spec.synth_rates); + + spec.synth_rates = {0.8, 0.9, 0.4, 0.2}; + assert_invalid(spec, 4); + + spec.synth_rates = {0.8, std::numeric_limits::quiet_NaN(), 0.4, 0.2}; + assert_invalid(spec, 4); + + spec.synth_rates = {0.8, 0.6, 0.4, -0.2}; + assert_invalid(spec, 4); + + spec.synth_rates = {0.8, 0.6, 0.4, 0.2}; + spec.synth_len = 3.0; + assert_invalid(spec, 4); + } + { common_params base; base.n_parallel = 4; @@ -197,6 +254,26 @@ static void test(void) { assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SPECULATIVE)); assert(params.speculative.draft.n_max == 123); + { + common_params synth_params; + argv = {"binary_name", "--spec-synth-len", "3.4"}; + assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), synth_params, LLAMA_EXAMPLE_SERVER)); + assert(synth_params.speculative.synth_len == 3.4); + } + + { + common_params synth_params; + argv = {"binary_name", "--spec-synth-rates", "0.8,0.6,0.2"}; + assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), synth_params, LLAMA_EXAMPLE_SERVER)); + assert(synth_params.speculative.synth_rates == std::vector({0.8, 0.6, 0.2})); + } + + { + common_params synth_params; + argv = {"binary_name", "--spec-synth-len", "3.4x"}; + assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), synth_params, LLAMA_EXAMPLE_SERVER)); + } + argv = {"binary_name", "-lm", "none"}; assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.load_mode == LLAMA_LOAD_MODE_NONE); diff --git a/tools/cli/README.md b/tools/cli/README.md index c9cbacafcd..163ee4fbaf 100644 --- a/tools/cli/README.md +++ b/tools/cli/README.md @@ -59,12 +59,14 @@ | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)
(env: LLAMA_ARG_MMAP) | | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available
(env: LLAMA_ARG_DIO) | | `-lm, --load-mode MODE` | model loading mode (default: auto)
- auto: mmap, unless a device does not support it
- none: no special loading mode
- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)
- mlock: force system to keep model in RAM rather than swapping or compressing
- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing
- dio: use DirectIO if available

(env: LLAMA_ARG_LOAD_MODE) | +| `--tensor-read-lazy MODE` | on-demand reading of certain tensors, for example per-layer embeddings (default: auto)
- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)
- auto: on, but only for tensors larger than 4 GiB
- off: always keep them resident
(env: LLAMA_ARG_TENSOR_READ_LAZY) | | `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggml-org/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | | `-dev, --device ` | comma-separated list of devices to use for offloading (none = don't offload)
use --list-devices to see a list of available devices
(env: LLAMA_ARG_DEVICE) | | `--list-devices` | print list of available devices and exit | | `-ot, --override-tensor =,...` | override tensor buffer type
(env: LLAMA_ARG_OVERRIDE_TENSOR) | | `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU
(env: LLAMA_ARG_CPU_MOE) | | `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU
(env: LLAMA_ARG_N_CPU_MOE) | +| `-ncffn, --n-cpu-ffn N` | keep the dense FFN weights of the first N layers in the CPU
(dense models; for MoE expert weights use --n-cpu-moe)
(env: LLAMA_ARG_N_CPU_FFN) | | `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
(env: LLAMA_ARG_N_GPU_LAYERS) | | `-sm, --split-mode {none,layer,row,tensor}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs (pipelined)
- row: split weight across GPUs by rows (parallelized)
- tensor: split weights and KV across GPUs (parallelized, EXPERIMENTAL)
(env: LLAMA_ARG_SPLIT_MODE) | | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1
(env: LLAMA_ARG_TENSOR_SPLIT) | @@ -154,7 +156,6 @@ | `-sysf, --system-prompt-file FNAME` | a file containing the system prompt (default: none) | | `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode | | `-sp, --special` | special tokens output enabled (default: false) | -| `-cnv, --conversation, -no-cnv, --no-conversation` | whether to run in conversation mode:
- does not print special tokens and suffix/prefix
- interactive mode is also enabled
(default: auto enabled if chat template is available) | | `-st, --single-turn` | run conversation for a single turn only, then exit when done
will not be interactive if first turn is predefined with --prompt
(default: false) | | `-mli, --multiline-input` | allows you to write or paste multiple lines without ending each in '\' | | `--warmup, --no-warmup` | whether to perform warmup with an empty run (default: enabled) | @@ -166,6 +167,9 @@ | `--image, --audio, --video FILE` | path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files | | `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)
(env: LLAMA_ARG_IMAGE_MIN_TOKENS) | | `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)
(env: LLAMA_ARG_IMAGE_MAX_TOKENS) | +| `--video-fps N` | target video frame rate (default: 4.0)
(env: LLAMA_ARG_VIDEO_FPS) | +| `--video-timestamp-interval N` | interval in milliseconds between text timestamps (default: 5000)
(env: LLAMA_ARG_VIDEO_TIMESTAMP_INTERVAL) | +| `--video-ffmpeg-dir DIR` | path to the directory containing ffmpeg and ffprobe (default: search in PATH)
(env: LLAMA_ARG_VIDEO_FFMPEG_DIR) | | `-o, --output, --output-file FNAME` | output file (default: '') | | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'
(env: LLAMA_ARG_CHAT_TEMPLATE_KWARGS) | | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)
(env: LLAMA_ARG_JINJA) | @@ -197,6 +201,8 @@ | `--spec-draft-n-cpu-moe, --spec-draft-ncmoe, -ncmoed, --n-cpu-moe-draft N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model
(env: LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE) | | `--spec-draft-n-max N` | number of tokens to draft for speculative decoding (default: 3)
(env: LLAMA_ARG_SPEC_DRAFT_N_MAX) | | `--spec-draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)
(env: LLAMA_ARG_SPEC_DRAFT_N_MIN) | +| `--spec-synth-len L` | target mean synthetic acceptance length, including the target token (benchmarking only)
(env: LLAMA_ARG_SPEC_SYNTH_LEN) | +| `--spec-synth-rates P0,P1,...` | comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)
(env: LLAMA_ARG_SPEC_SYNTH_RATES) | | `--spec-draft-p-split, --draft-p-split P` | speculative decoding split probability (default: 0.10)
(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT) | | `--spec-draft-p-min, --draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.00)
(env: LLAMA_ARG_SPEC_DRAFT_P_MIN) | | `--spec-draft-backend-sampling, --no-spec-draft-backend-sampling` | offload draft sampling to the backend (default: enabled)
(env: LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING) | diff --git a/tools/completion/README.md b/tools/completion/README.md index 833687dcad..0cd86bac70 100644 --- a/tools/completion/README.md +++ b/tools/completion/README.md @@ -142,12 +142,14 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)
(env: LLAMA_ARG_MMAP) | | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available
(env: LLAMA_ARG_DIO) | | `-lm, --load-mode MODE` | model loading mode (default: auto)
- auto: mmap, unless a device does not support it
- none: no special loading mode
- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)
- mlock: force system to keep model in RAM rather than swapping or compressing
- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing
- dio: use DirectIO if available

(env: LLAMA_ARG_LOAD_MODE) | +| `--tensor-read-lazy MODE` | on-demand reading of certain tensors, for example per-layer embeddings (default: auto)
- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)
- auto: on, but only for tensors larger than 4 GiB
- off: always keep them resident
(env: LLAMA_ARG_TENSOR_READ_LAZY) | | `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggml-org/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | | `-dev, --device ` | comma-separated list of devices to use for offloading (none = don't offload)
use --list-devices to see a list of available devices
(env: LLAMA_ARG_DEVICE) | | `--list-devices` | print list of available devices and exit | | `-ot, --override-tensor =,...` | override tensor buffer type
(env: LLAMA_ARG_OVERRIDE_TENSOR) | | `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU
(env: LLAMA_ARG_CPU_MOE) | | `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU
(env: LLAMA_ARG_N_CPU_MOE) | +| `-ncffn, --n-cpu-ffn N` | keep the dense FFN weights of the first N layers in the CPU
(dense models; for MoE expert weights use --n-cpu-moe)
(env: LLAMA_ARG_N_CPU_FFN) | | `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
(env: LLAMA_ARG_N_GPU_LAYERS) | | `-sm, --split-mode {none,layer,row,tensor}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs (pipelined)
- row: split weight across GPUs by rows (parallelized)
- tensor: split weights and KV across GPUs (parallelized, EXPERIMENTAL)
(env: LLAMA_ARG_SPLIT_MODE) | | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1
(env: LLAMA_ARG_TENSOR_SPLIT) | diff --git a/tools/mtmd/mtmd-cli.cpp b/tools/mtmd/mtmd-cli.cpp index f6c787fdb6..97678c6b2e 100644 --- a/tools/mtmd/mtmd-cli.cpp +++ b/tools/mtmd/mtmd-cli.cpp @@ -87,6 +87,9 @@ struct mtmd_cli_context { mtmd::bitmaps bitmaps; std::vector videos; + mtmd_helper_init_opt init_opt = mtmd_helper_init_opt_default(); + std::string video_ffmpeg_bin_dir; + mtmd::batch_ptr mbatch; // chat template @@ -170,6 +173,12 @@ struct mtmd_cli_context { LOG_ERR("Failed to load vision model from %s\n", clip_path); exit(1); } + + video_ffmpeg_bin_dir = params.video_ffmpeg_bin_dir; + init_opt.video_params.fps_target = params.video_fps; + init_opt.video_params.timestamp_interval_ms = params.video_timestamp_interval_ms; + init_opt.video_params.ffmpeg_bin_dir = video_ffmpeg_bin_dir.empty() + ? nullptr : video_ffmpeg_bin_dir.c_str(); } bool check_antiprompt(const llama_tokens & generated_tokens) { @@ -184,7 +193,7 @@ struct mtmd_cli_context { } bool load_media(const std::string & fname) { - auto res = mtmd_helper_bitmap_init_from_file(ctx_vision.get(), fname.c_str(), false); + auto res = mtmd_helper_bitmap_init_from_file(ctx_vision.get(), fname.c_str(), false, init_opt); if (!res.bitmap) { return false; } diff --git a/tools/mtmd/mtmd-helper.cpp b/tools/mtmd/mtmd-helper.cpp index f1defb6477..77f9d58fe2 100644 --- a/tools/mtmd/mtmd-helper.cpp +++ b/tools/mtmd/mtmd-helper.cpp @@ -369,14 +369,18 @@ static bool is_webp_file(const unsigned char * buf, size_t len) { } #ifdef MTMD_VIDEO -static mtmd_bitmap * decode_webp_with_ffmpeg(mtmd_context * mctx, const unsigned char * buf, size_t len, bool placeholder); +static mtmd_bitmap * decode_webp_with_ffmpeg(mtmd_context * mctx, const unsigned char * buf, size_t len, bool placeholder, + const mtmd_helper_video_init_params & params); #endif -mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len, bool placeholder) { +mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len, bool placeholder, + mtmd_helper_init_opt opt) { // calculate the hash if needed std::string id; mtmd_bitmap * result = nullptr; + GGML_UNUSED(opt); // only used by video code paths + if (!placeholder) { // use sha256 to prevent cache poisoning id = hash_sha256_hex(buf, len); @@ -414,7 +418,7 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, #ifdef MTMD_VIDEO // stb_image does not support webp; decode it with ffmpeg as a single frame if (!result && is_webp_file(buf, len)) { - result = decode_webp_with_ffmpeg(ctx, buf, len, placeholder); + result = decode_webp_with_ffmpeg(ctx, buf, len, placeholder, opt.video_params); if (!result) { LOG_ERR("%s: failed to decode webp buffer\n", __func__); return {nullptr, nullptr}; @@ -427,8 +431,7 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, // last try: load as video #ifdef MTMD_VIDEO if (!result) { - auto params = mtmd_helper_video_init_params_default(); - auto video_ctx = mtmd_helper_video_init_from_buf(ctx, buf, len, params); + auto video_ctx = mtmd_helper_video_init_from_buf(ctx, buf, len, opt.video_params); if (!video_ctx) { LOG_ERR("%s: failed to decode buffer as either image/audio/video\n", __func__); return {nullptr, nullptr}; @@ -456,7 +459,8 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, return {nullptr, nullptr}; } -mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname, bool placeholder) { +mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname, bool placeholder, + mtmd_helper_init_opt opt) { #ifdef _WIN32 int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0); if (!wlen) { @@ -497,7 +501,7 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, return {nullptr, nullptr}; } - return mtmd_helper_bitmap_init_from_buf(ctx, buf.data(), buf.size(), placeholder); + return mtmd_helper_bitmap_init_from_buf(ctx, buf.data(), buf.size(), placeholder, opt); } bool mtmd_helper_support_video(mtmd_context * ctx) { @@ -855,6 +859,12 @@ mtmd_helper_video_init_params mtmd_helper_video_init_params_default() { }; } +mtmd_helper_init_opt mtmd_helper_init_opt_default() { + return { + /* video_params */ mtmd_helper_video_init_params_default(), + }; +} + static std::string video_resolve_bin(const char * bin_dir, const char * name) { if (!bin_dir || bin_dir[0] == '\0') { return name; // rely on PATH @@ -876,8 +886,8 @@ static std::string video_resolve_bin(const char * bin_dir, const char * name) { } #ifdef MTMD_VIDEO -static mtmd_bitmap * decode_webp_with_ffmpeg(mtmd_context * mctx, const unsigned char * buf, size_t len, bool placeholder) { - auto params = mtmd_helper_video_init_params_default(); +static mtmd_bitmap * decode_webp_with_ffmpeg(mtmd_context * mctx, const unsigned char * buf, size_t len, bool placeholder, + const mtmd_helper_video_init_params & params) { mtmd_helper_video vctx; vctx.mctx = mctx; vctx.input_buf.assign(buf, buf + len); diff --git a/tools/mtmd/mtmd-helper.h b/tools/mtmd/mtmd-helper.h index 58dfb15250..772e0f091b 100644 --- a/tools/mtmd/mtmd-helper.h +++ b/tools/mtmd/mtmd-helper.h @@ -23,6 +23,23 @@ extern "C" { struct mtmd_helper_video; typedef struct mtmd_helper_video mtmd_helper_video; +struct mtmd_helper_video_init_params { + float fps_target; // desired output fps; <= 0 means use the video's native fps, defaulted to 4.0f + const char * ffmpeg_bin_dir; // directory containing ffmpeg/ffprobe binaries; NULL means search PATH + int64_t timestamp_interval_ms; // interval for adding timestamp as text chunk (example: "[10m50.5s]"); <= 0 means no timestamp, defaulted to 5000ms + // TODO @ngxson : allow "placeholder" bitmap output for counting tokens +}; + +MTMD_API struct mtmd_helper_video_init_params mtmd_helper_video_init_params_default(void); + +// opt for mtmd_helper_bitmap_init_from_*() +struct mtmd_helper_init_opt { + struct mtmd_helper_video_init_params video_params; +}; +typedef struct mtmd_helper_init_opt mtmd_helper_init_opt; + +MTMD_API struct mtmd_helper_init_opt mtmd_helper_init_opt_default(void); + // Set callback for all future logging events. // If this is not called, or NULL is supplied, everything is output on stderr. // Note: this also call mtmd_log_set() internally @@ -40,7 +57,11 @@ struct mtmd_helper_bitmap_wrapper { // it calls mtmd_helper_bitmap_init_from_buf() internally // returns nullptr on failure // this function is thread-safe -MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname, bool placeholder); +MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file( + mtmd_context * ctx, + const char * fname, + bool placeholder, + struct mtmd_helper_init_opt opt); // helper function to construct a mtmd_bitmap from a buffer containing a file // supported formats: @@ -53,7 +74,11 @@ MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtm // - output bitmap will have SHA-256 hash (hex string) as the ID // returns nullptr on failure // this function is thread-safe -MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len, bool placeholder); +MTMD_API struct mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf( + mtmd_context * ctx, + const unsigned char * buf, size_t len, + bool placeholder, + struct mtmd_helper_init_opt opt); // helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache MTMD_API size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks); @@ -124,14 +149,7 @@ struct mtmd_helper_video_info { int32_t n_frames; // estimated total frames at effective fps (-1 if unknown) }; -struct mtmd_helper_video_init_params { - float fps_target; // desired output fps; <= 0 means use the video's native fps, defaulted to 4.0f - const char * ffmpeg_bin_dir; // directory containing ffmpeg/ffprobe binaries; NULL means search PATH - int64_t timestamp_interval_ms; // interval for adding timestamp as text chunk (example: "[10m50.5s]"); <= 0 means no timestamp, defaulted to 5000ms - // TODO @ngxson : allow "placeholder" bitmap output for counting tokens -}; - -MTMD_API struct mtmd_helper_video_init_params mtmd_helper_video_init_params_default(void); +// note: mtmd_helper_video_init_params is defined at the top, as it is part of mtmd_helper_init_opt // returns NULL on failure (ffprobe not found, file unreadable, etc.) MTMD_API mtmd_helper_video * mtmd_helper_video_init( diff --git a/tools/server/README.md b/tools/server/README.md index 93736c3edf..07e58fe916 100644 --- a/tools/server/README.md +++ b/tools/server/README.md @@ -76,12 +76,14 @@ For the full list of features, please refer to [server's changelog](https://gith | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)
(env: LLAMA_ARG_MMAP) | | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available
(env: LLAMA_ARG_DIO) | | `-lm, --load-mode MODE` | model loading mode (default: auto)
- auto: mmap, unless a device does not support it
- none: no special loading mode
- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)
- mlock: force system to keep model in RAM rather than swapping or compressing
- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing
- dio: use DirectIO if available

(env: LLAMA_ARG_LOAD_MODE) | +| `--tensor-read-lazy MODE` | on-demand reading of certain tensors, for example per-layer embeddings (default: auto)
- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)
- auto: on, but only for tensors larger than 4 GiB
- off: always keep them resident
(env: LLAMA_ARG_TENSOR_READ_LAZY) | | `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggml-org/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | | `-dev, --device ` | comma-separated list of devices to use for offloading (none = don't offload)
use --list-devices to see a list of available devices
(env: LLAMA_ARG_DEVICE) | | `--list-devices` | print list of available devices and exit | | `-ot, --override-tensor =,...` | override tensor buffer type
(env: LLAMA_ARG_OVERRIDE_TENSOR) | | `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU
(env: LLAMA_ARG_CPU_MOE) | | `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU
(env: LLAMA_ARG_N_CPU_MOE) | +| `-ncffn, --n-cpu-ffn N` | keep the dense FFN weights of the first N layers in the CPU
(dense models; for MoE expert weights use --n-cpu-moe)
(env: LLAMA_ARG_N_CPU_FFN) | | `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
(env: LLAMA_ARG_N_GPU_LAYERS) | | `-sm, --split-mode {none,layer,row,tensor}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs (pipelined)
- row: split weight across GPUs by rows (parallelized)
- tensor: split weights and KV across GPUs (parallelized, EXPERIMENTAL)
(env: LLAMA_ARG_SPLIT_MODE) | | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1
(env: LLAMA_ARG_TENSOR_SPLIT) | @@ -182,6 +184,9 @@ For the full list of features, please refer to [server's changelog](https://gith | `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)
(env: LLAMA_ARG_IMAGE_MIN_TOKENS) | | `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)
(env: LLAMA_ARG_IMAGE_MAX_TOKENS) | | `--mtmd-batch-max-tokens N` | maximum number of image tokens per batch when encoding images (default: 1024)
(env: LLAMA_ARG_MTMD_BATCH_MAX_TOKENS) | +| `--video-fps N` | target video frame rate (default: 4.0)
(env: LLAMA_ARG_VIDEO_FPS) | +| `--video-timestamp-interval N` | interval in milliseconds between text timestamps (default: 5000)
(env: LLAMA_ARG_VIDEO_TIMESTAMP_INTERVAL) | +| `--video-ffmpeg-dir DIR` | path to the directory containing ffmpeg and ffprobe (default: search in PATH)
(env: LLAMA_ARG_VIDEO_FFMPEG_DIR) | | `-a, --alias STRING` | set model name aliases, comma-separated (to be used by API)
(env: LLAMA_ARG_ALIAS) | | `--tags STRING` | set model tags, comma-separated (informational, not used for routing)
(env: LLAMA_ARG_TAGS) | | `--embd-normalize N` | normalisation for embeddings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm) | @@ -256,6 +261,8 @@ For the full list of features, please refer to [server's changelog](https://gith | `--spec-draft-n-cpu-moe, --spec-draft-ncmoe, -ncmoed, --n-cpu-moe-draft N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model
(env: LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE) | | `--spec-draft-n-max N` | number of tokens to draft for speculative decoding (default: 3)
(env: LLAMA_ARG_SPEC_DRAFT_N_MAX) | | `--spec-draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)
(env: LLAMA_ARG_SPEC_DRAFT_N_MIN) | +| `--spec-synth-len L` | target mean synthetic acceptance length, including the target token (benchmarking only)
(env: LLAMA_ARG_SPEC_SYNTH_LEN) | +| `--spec-synth-rates P0,P1,...` | comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)
(env: LLAMA_ARG_SPEC_SYNTH_RATES) | | `--spec-draft-p-split, --draft-p-split P` | speculative decoding split probability (default: 0.10)
(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT) | | `--spec-draft-p-min, --draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.00)
(env: LLAMA_ARG_SPEC_DRAFT_P_MIN) | | `--spec-draft-backend-sampling, --no-spec-draft-backend-sampling` | offload draft sampling to the backend (default: enabled)
(env: LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING) | diff --git a/tools/server/server-common.cpp b/tools/server/server-common.cpp index 7997d4016a..c30955e89f 100644 --- a/tools/server/server-common.cpp +++ b/tools/server/server-common.cpp @@ -910,12 +910,17 @@ size_t validate_utf8(const std::string& text) { return len; } -server_tokens process_mtmd_prompt(mtmd_context * mctx, const std::string & prompt, const std::vector & files, bool is_placeholder) { +server_tokens process_mtmd_prompt( + mtmd_context * mctx, + const std::string & prompt, + const std::vector & files, + const mtmd_helper_init_opt & init_opt, + bool is_placeholder) { // these will be freed upon going out of scope mtmd::bitmaps bitmaps; std::vector videos; for (auto & file : files) { - auto out = mtmd_helper_bitmap_init_from_buf(mctx, file.data(), file.size(), is_placeholder); + auto out = mtmd_helper_bitmap_init_from_buf(mctx, file.data(), file.size(), is_placeholder, init_opt); if (!out.bitmap) { throw std::runtime_error("Failed to load image or audio file"); } @@ -956,7 +961,7 @@ server_tokens process_mtmd_prompt(mtmd_context * mctx, const std::string & promp * - "prompt": [12, 34, "string", 56, 78] * - "prompt": { "prompt_string": "string", "multimodal_data": [ "base64" ] } */ -static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special) { +static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special, const mtmd_helper_init_opt & init_opt) { constexpr char JSON_STRING_PROMPT_KEY[] = "prompt_string"; constexpr char JSON_MTMD_DATA_KEY[] = "multimodal_data"; const bool has_mtmd = mctx != nullptr; @@ -979,7 +984,7 @@ static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_co for (const auto & entry : json_prompt.at(JSON_MTMD_DATA_KEY)) { files.push_back(base64_decode(entry)); } - return process_mtmd_prompt(mctx, json_prompt.at(JSON_STRING_PROMPT_KEY), files); + return process_mtmd_prompt(mctx, json_prompt.at(JSON_STRING_PROMPT_KEY), files, init_opt); } else { // Not multimodal, but contains a subobject. llama_tokens tmp = tokenize_mixed(vocab, json_prompt.at(JSON_STRING_PROMPT_KEY), add_special, parse_special); @@ -990,15 +995,15 @@ static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_co } } -std::vector tokenize_input_prompts(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special) { +std::vector tokenize_input_prompts(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special, const mtmd_helper_init_opt & init_opt) { std::vector result; if (json_prompt.is_array() && !json_is_array_and_contains_numbers(json_prompt)) { result.reserve(json_prompt.size()); for (const auto & p : json_prompt) { - result.push_back(tokenize_input_subprompt(vocab, mctx, p,add_special, parse_special)); + result.push_back(tokenize_input_subprompt(vocab, mctx, p, add_special, parse_special, init_opt)); } } else { - result.push_back(tokenize_input_subprompt(vocab, mctx, json_prompt, add_special, parse_special)); + result.push_back(tokenize_input_subprompt(vocab, mctx, json_prompt, add_special, parse_special, init_opt)); } if (result.empty()) { throw std::runtime_error("\"prompt\" must not be empty"); @@ -1787,7 +1792,8 @@ server_tokens format_prompt_rerank( const struct llama_vocab * vocab, mtmd_context * mctx, const std::string & query, - const std::string & doc) { + const std::string & doc, + const mtmd_helper_init_opt & init_opt) { server_tokens result = {}; const char * rerank_prompt = llama_model_chat_template(model, "rerank"); @@ -1796,12 +1802,12 @@ server_tokens format_prompt_rerank( std::string prompt = rerank_prompt; string_replace_all(prompt, "{query}" , query); string_replace_all(prompt, "{document}", doc ); - server_tokens tokens = tokenize_input_subprompt(vocab, mctx, prompt, false, true); + server_tokens tokens = tokenize_input_subprompt(vocab, mctx, prompt, false, true, init_opt); result.push_back(tokens); } else { // Get EOS token - use SEP token as fallback if EOS is not available - server_tokens query_tokens = tokenize_input_subprompt(vocab, mctx, query, false, false); - server_tokens doc_tokens = tokenize_input_subprompt(vocab, mctx, doc, false, false); + server_tokens query_tokens = tokenize_input_subprompt(vocab, mctx, query, false, false, init_opt); + server_tokens doc_tokens = tokenize_input_subprompt(vocab, mctx, doc, false, false, init_opt); llama_token eos_token = llama_vocab_eos(vocab); if (eos_token == LLAMA_TOKEN_NULL) { eos_token = llama_vocab_sep(vocab); diff --git a/tools/server/server-common.h b/tools/server/server-common.h index f8ea82ef4c..6c681a2cf5 100644 --- a/tools/server/server-common.h +++ b/tools/server/server-common.h @@ -5,6 +5,7 @@ #include "llama.h" #include "chat.h" #include "mtmd.h" +#include "mtmd-helper.h" #include "json.h" @@ -269,7 +270,12 @@ size_t validate_utf8(const std::string& text); // process mtmd prompt, return the server_tokens containing both text tokens and media chunks // if is_placeholder is true, the media chunk will be treated as placeholder for counting tokens; the output tokens are not usable for actual inference (e.g. for submitting a task to server_queue) -server_tokens process_mtmd_prompt(mtmd_context * mctx, const std::string & prompt, const std::vector & files, bool is_placeholder = false); +server_tokens process_mtmd_prompt( + mtmd_context * mctx, + const std::string & prompt, + const std::vector & files, + const mtmd_helper_init_opt & init_opt, + bool is_placeholder = false); /** * break the input "prompt" object into multiple prompt if needed, then tokenize them @@ -289,7 +295,8 @@ std::vector tokenize_input_prompts( mtmd_context * mctx, const json & json_prompt, bool add_special, - bool parse_special); + bool parse_special, + const mtmd_helper_init_opt & init_opt); // // OAI utils @@ -538,7 +545,8 @@ server_tokens format_prompt_rerank( const struct llama_vocab * vocab, mtmd_context * mctx, const std::string & query, - const std::string & doc); + const std::string & doc, + const mtmd_helper_init_opt & init_opt); // simple implementation of a pipe // used for streaming data between threads diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp index a9edbd7be8..e6c991f7cf 100644 --- a/tools/server/server-context.cpp +++ b/tools/server/server-context.cpp @@ -23,6 +23,7 @@ #include #include #include +#include #include #include @@ -51,6 +52,50 @@ static common_speculative_output_limits server_output_limits(const common_params return result; } +// synthetic draft verification for benchmarking - accept draft tokens at random instead of by match with the target +// on replay the draft was already accepted before a context checkpoint restore, so repeat the same decisions +static std::vector server_sample_and_accept_synth( + common_sampler * smpl, + llama_context * ctx, + const std::vector & idxs, + const llama_tokens & draft, + const std::vector & synth_probs, + std::mt19937 & rng, + bool is_replay) { + GGML_ASSERT(idxs.size() == draft.size() + 1); + GGML_ASSERT(synth_probs.size() >= draft.size()); + + std::vector result; + result.reserve(idxs.size()); + + const llama_vocab * vocab = llama_model_get_vocab(llama_get_model(ctx)); + std::uniform_real_distribution dist(0.0, 1.0); + for (size_t i = 0; i < draft.size(); ++i) { + const llama_token id = common_sampler_sample(smpl, ctx, idxs[i]); + const bool accept = is_replay || dist(rng) < synth_probs[i]; + // do not accept a drafted EOG token - it would end the generation early + // on replay the last token is from the target and can be EOG, so skip this check + if (accept && (is_replay || !llama_vocab_is_eog(vocab, draft[i]))) { + // synthetic draft tokens do not advance grammar or reasoning state + // the last replay token is from the target and must advance both + const bool is_replay_target = is_replay && i + 1 == draft.size(); + common_sampler_accept(smpl, draft[i], is_replay_target); + result.push_back(draft[i]); + continue; + } + + common_sampler_accept(smpl, id, true); + result.push_back(id); + return result; + } + + const llama_token id = common_sampler_sample(smpl, ctx, idxs[draft.size()]); + common_sampler_accept(smpl, id, true); + result.push_back(id); + + return result; +} + // state diagram: https://github.com/ggml-org/llama.cpp/pull/9283 enum slot_state { SLOT_STATE_IDLE, @@ -211,6 +256,7 @@ struct server_slot { std::vector spec_i_batch; common_prompt_checkpoint spec_ckpt; bool spec_is_replay = false; + std::mt19937 spec_synth_rng; // TODO: move members that belong to the task (such as `generated_text`, `has_new_line`) to task_results_state // see https://github.com/ggml-org/llama.cpp/pull/18283#issuecomment-3710175837 @@ -794,6 +840,8 @@ public: llama_model * model_tgt = nullptr; mtmd_context * mctx = nullptr; + // note: video_params.ffmpeg_bin_dir points into params_base, which outlives this struct + mtmd_helper_init_opt init_opt = mtmd_helper_init_opt_default(); const llama_vocab * vocab = nullptr; server_queue queue_tasks; @@ -1118,6 +1166,11 @@ private: } SRV_INF("loaded multimodal model, '%s'\n", mmproj_path.c_str()); + init_opt.video_params.fps_target = params_base.video_fps; + init_opt.video_params.timestamp_interval_ms = params_base.video_timestamp_interval_ms; + init_opt.video_params.ffmpeg_bin_dir = params_base.video_ffmpeg_bin_dir.empty() + ? nullptr : params_base.video_ffmpeg_bin_dir.c_str(); + if (params_base.ctx_shift) { params_base.ctx_shift = false; SRV_WRN("%s\n", "ctx_shift is not supported by multimodal, it will be disabled"); @@ -1187,6 +1240,9 @@ private: spec.reset(common_speculative_init(params_base.speculative, params_base.n_parallel)); } catch (const std::exception & e) { SRV_ERR("failed to initialize speculative decoding context: %s\n", e.what()); + if (params_base.speculative.has_synth()) { + return false; + } } } @@ -1202,6 +1258,11 @@ private: model_dft = nullptr; } + if (!spec && params_base.speculative.has_synth()) { + SRV_ERR("%s", "synthetic acceptance requires an initialized speculative decoding context\n"); + return false; + } + for (int i = 0; i < params_base.n_parallel; i++) { server_slot & slot = slots[i]; @@ -1710,6 +1771,13 @@ private: SLT_TRC(slot, "sampler chain: %s\n", common_sampler_print(slot.smpl.get()).c_str()); SLT_TRC(slot, "sampler params: \n%s\n", task.params.sampling.print().c_str()); + + if (spec && !common_speculative_get_synth_probs(spec.get()).empty()) { + const uint32_t seed = task.params.sampling.seed == LLAMA_DEFAULT_SEED + ? std::random_device{}() + : task.params.sampling.seed; + slot.spec_synth_rng.seed(seed); + } } else { slot.smpl.reset(); } @@ -2134,9 +2202,9 @@ private: try { auto & prompt = task.cli_prompt; if (mctx != nullptr) { - task.tokens = process_mtmd_prompt(mctx, prompt, task.cli_files); + task.tokens = process_mtmd_prompt(mctx, prompt, task.cli_files, init_opt); } else { - task.tokens = std::move(tokenize_input_prompts(vocab, mctx, prompt, true, true)[0]); + task.tokens = std::move(tokenize_input_prompts(vocab, mctx, prompt, true, true, init_opt)[0]); } task.cli_prompt.clear(); task.cli_files.clear(); @@ -3795,7 +3863,12 @@ private: common_sampler_ptr smpl_save(common_sampler_clone(slot.smpl.get())); GGML_ASSERT(slot.spec_i_batch.size() == n_draft + 1); - auto accepted = common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft); + const auto & synth_probs = common_speculative_get_synth_probs(spec.get()); + auto accepted = synth_probs.empty() + ? common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft) + : server_sample_and_accept_synth( + slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft, + synth_probs, slot.spec_synth_rng, slot.spec_is_replay); slot.spec_i_batch.clear(); GGML_ASSERT(accepted.size() >= 1); @@ -3861,7 +3934,7 @@ private: auto & n_accepted_per_pos = slot.n_accepted_per_pos; if (n_accepted_per_pos.empty()) { - n_accepted_per_pos.resize(common_speculative_n_max(¶ms_base.speculative), 0); + n_accepted_per_pos.resize(common_speculative_n_max(spec.get()), 0); } for (size_t i = 0; i < n_accepted && i < n_accepted_per_pos.size(); ++i) { n_accepted_per_pos[i]++; @@ -4165,10 +4238,10 @@ std::unique_ptr server_routes::handle_completions_impl( if (res_type != TASK_RESPONSE_TYPE_NONE && ctx_server.mctx != nullptr) { // This is the case used by OAI compatible chat path with MTMD. TODO It can be moved to the path below. - inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get(), files)); + inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get(), files, ctx_server.init_opt)); } else { // Everything else, including multimodal completions. - inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true); + inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true, ctx_server.init_opt); } // tasks.reserve(inputs.size()); // TODO: this is inaccurate due to child tasks @@ -4752,7 +4825,7 @@ void server_routes::init_routes() { data["input_extra"] = input_extra; // default to empty array if it's not exist std::string prompt = json_value(data, "prompt", std::string()); - std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, false, true); + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, false, true, ctx_server.init_opt); SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); data["prompt"] = format_prompt_infill( ctx_server.vocab, @@ -4816,7 +4889,7 @@ void server_routes::init_routes() { }; this->post_chat_completions_tok = [this](const server_http_req & req) { - return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, req, TASK_RESPONSE_TYPE_OAI_CHAT); + return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, ctx_server.init_opt, req, TASK_RESPONSE_TYPE_OAI_CHAT); }; this->post_control = [this](const server_http_req & req) { @@ -4875,7 +4948,7 @@ void server_routes::init_routes() { }; this->post_responses_tok_oai = [this](const server_http_req & req) { - return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, req, TASK_RESPONSE_TYPE_OAI_RESP); + return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, ctx_server.init_opt, req, TASK_RESPONSE_TYPE_OAI_RESP); }; this->post_transcriptions_oai = [this](const server_http_req & req) { @@ -4925,7 +4998,7 @@ void server_routes::init_routes() { }; this->post_anthropic_count_tokens = [this](const server_http_req & req) { - return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, req, TASK_RESPONSE_TYPE_ANTHROPIC); + return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, ctx_server.init_opt, req, TASK_RESPONSE_TYPE_ANTHROPIC); }; // same with handle_chat_completions, but without inference part @@ -5058,7 +5131,7 @@ void server_routes::init_routes() { std::vector tasks; tasks.reserve(documents.size()); for (size_t i = 0; i < documents.size(); i++) { - auto tmp = format_prompt_rerank(ctx_server.model_tgt, ctx_server.vocab, ctx_server.mctx, query, documents[i]); + auto tmp = format_prompt_rerank(ctx_server.model_tgt, ctx_server.vocab, ctx_server.mctx, query, documents[i], ctx_server.init_opt); server_task task = server_task(SERVER_TASK_TYPE_RERANK); task.id = rd.get_new_id(); task.tokens = std::move(tmp); @@ -5296,7 +5369,7 @@ std::unique_ptr server_routes::handle_embeddings_impl(cons } } - auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true); + auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true, ctx_server.init_opt); for (const auto & tokens : tokenized_prompts) { // this check is necessary for models that do not add BOS token to the input if (tokens.empty()) { @@ -5357,7 +5430,7 @@ std::unique_ptr server_routes::handle_embeddings_impl(cons return res; } -std::unique_ptr server_routes::handle_count_tokens(const llama_vocab * vocab, mtmd_context * mctx, const server_http_req & req, task_response_type res_type) { +std::unique_ptr server_routes::handle_count_tokens(const llama_vocab * vocab, mtmd_context * mctx, const mtmd_helper_init_opt & init_opt, const server_http_req & req, task_response_type res_type) { auto res = create_response(); std::vector files; json body = json::parse(req.body); @@ -5395,7 +5468,7 @@ std::unique_ptr server_routes::handle_count_tokens(const l if (!prompt.is_string()) { throw std::runtime_error("for mtmd, input prompt must be a string."); } - n_tokens = process_mtmd_prompt(mctx, prompt.get(), files, true).size(); + n_tokens = process_mtmd_prompt(mctx, prompt.get(), files, init_opt, true).size(); } else { n_tokens = tokenize_mixed(vocab, prompt, true, true).size(); } diff --git a/tools/server/server-context.h b/tools/server/server-context.h index 5d464b8e8c..0acbbffa9e 100644 --- a/tools/server/server-context.h +++ b/tools/server/server-context.h @@ -169,7 +169,7 @@ private: std::unique_ptr handle_slots_restore(const server_http_req & req, int id_slot); std::unique_ptr handle_slots_erase(const server_http_req &, int id_slot); std::unique_ptr handle_embeddings_impl(const server_http_req & req, task_response_type res_type); - std::unique_ptr handle_count_tokens(const llama_vocab * vocab, mtmd_context * mctx, const server_http_req & req, task_response_type res_type); + std::unique_ptr handle_count_tokens(const llama_vocab * vocab, mtmd_context * mctx, const mtmd_helper_init_opt & init_opt, const server_http_req & req, task_response_type res_type); // using unique_ptr to allow late initialization of const std::unique_ptr meta; diff --git a/tools/server/tests/unit/test_speculative.py b/tools/server/tests/unit/test_speculative.py index 5837195006..22b523954e 100644 --- a/tools/server/tests/unit/test_speculative.py +++ b/tools/server/tests/unit/test_speculative.py @@ -52,6 +52,18 @@ def test_with_and_without_draft(): assert tokens_no_draft == tokens_draft + server.stop() + create_server() + assert server.spec_draft_n_max is not None + server.spec_synth_rates = [0.0] * server.spec_draft_n_max + server.start() + res = server.make_request("POST", "/completion", data=request) + + assert res.status_code == 200 + assert res.body["timings"]["draft_n"] > 0 + assert res.body["timings"]["draft_n_accepted"] == 0 + assert res.body["tokens"] == tokens_no_draft + def test_different_draft_min_draft_max(): global server @@ -80,6 +92,66 @@ def test_different_draft_min_draft_max(): last_content = res.body["content"] +def test_synth_is_deterministic(): + global server + assert server.spec_draft_n_max is not None + server.spec_synth_rates = [0.75 ** (i + 1) for i in range(server.spec_draft_n_max)] + server.start() + + request = { + "prompt": "I believe the meaning of life is", + "temperature": 0.2, + "top_k": 5, + "seed": 4242, + "n_predict": 32, + } + responses = [server.make_request("POST", "/completion", data=request) for _ in range(2)] + + for res in responses: + assert res.status_code == 200 + assert res.body["timings"]["draft_n"] > 0 + assert responses[0].body["timings"]["draft_n"] == responses[1].body["timings"]["draft_n"] + assert responses[0].body["timings"]["draft_n_accepted"] == responses[1].body["timings"]["draft_n_accepted"] + + +def test_synth_ignores_target_tokens(): + global server + assert server.spec_draft_n_max is not None + server.spec_synth_rates = [1.0] * server.spec_draft_n_max + server.start() + + res = server.make_request("POST", "/completion", data={ + "prompt": "I believe the meaning of life is", + "temperature": 0.0, + "seed": 4242, + "n_predict": 32, + }) + + assert res.status_code == 200 + assert res.body["timings"]["draft_n"] > 0 + assert res.body["timings"]["draft_n_accepted"] == res.body["timings"]["draft_n"] + + res = server.make_request("POST", "/completion", data={ + "prompt": "I believe the meaning of life is", + "temperature": 0.0, + "seed": 4242, + "n_predict": 6, + "grammar": 'root ::= "a"{5,5}', + }) + assert res.status_code == 200, res.body + + res = server.make_request("POST", "/completion", data={ + "prompt": "Respond with only: OK", + "temperature": 0.0, + "seed": 4242, + "n_predict": 64, + "ignore_eos": True, + }) + assert res.status_code == 200, res.body + assert res.body["tokens_predicted"] == 64 + assert res.body["stop_type"] == "limit" + + def test_slot_ctx_not_exceeded(): global server server.n_ctx = 256 diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py index a0d2dfa3c5..5a4f31a53e 100644 --- a/tools/server/tests/utils.py +++ b/tools/server/tests/utils.py @@ -99,6 +99,8 @@ class ServerProcess: spec_type: str | None = None spec_draft_n_min: int | None = None spec_draft_n_max: int | None = None + spec_synth_len: float | None = None + spec_synth_rates: List[float] | None = None no_ui: bool | None = None jinja: bool | None = None reasoning_format: Literal['deepseek', 'none', 'nothink'] | None = None @@ -245,6 +247,11 @@ class ServerProcess: server_args.extend(["--spec-draft-n-max", self.spec_draft_n_max]) if self.spec_draft_n_min: server_args.extend(["--spec-draft-n-min", self.spec_draft_n_min]) + if self.spec_synth_len is not None: + server_args.extend(["--spec-synth-len", self.spec_synth_len]) + if self.spec_synth_rates is not None: + rates = ",".join(str(rate) for rate in self.spec_synth_rates) + server_args.extend(["--spec-synth-rates", rates]) if self.no_ui: server_args.append("--no-ui") if self.no_models_autoload: diff --git a/tools/tts/tts.cpp b/tools/tts/tts.cpp index 368123baf5..6fd1936324 100644 --- a/tools/tts/tts.cpp +++ b/tools/tts/tts.cpp @@ -103,7 +103,7 @@ int main(int argc, char ** argv) { mtmd::bitmap_ptr speaker_bitmap; if (!params.tts_speaker_file.empty()) { - auto wrapper = mtmd_helper_bitmap_init_from_file(mctx.get(), params.tts_speaker_file.c_str(), false); + auto wrapper = mtmd_helper_bitmap_init_from_file(mctx.get(), params.tts_speaker_file.c_str(), false, mtmd_helper_init_opt_default()); if (!wrapper.bitmap) { LOG_ERR("failed to load speaker file %s\n", params.tts_speaker_file.c_str()); return 1; diff --git a/tools/ui/src/lib/components/app/badges/BadgesModality.svelte b/tools/ui/src/lib/components/app/badges/BadgesModality.svelte index 4eb3e7838d..83b1b46aff 100644 --- a/tools/ui/src/lib/components/app/badges/BadgesModality.svelte +++ b/tools/ui/src/lib/components/app/badges/BadgesModality.svelte @@ -1,5 +1,5 @@ -{#each modalities as modality (modality)} - {#if modality === ModelModality.VISION || modality === ModelModality.AUDIO || modality === ModelModality.VIDEO} - - {#if modality === ModelModality.VISION} - +{#each visible as modality (modality)} + {@const ModalityIcon = MODALITY_ICONS[modality]} + + - Vision (Image) - {:else if modality === ModelModality.VIDEO} - - {/if} + {MODALITY_LABELS[modality]} + {/each} diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte index 40819d6f1a..893f8077dd 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte @@ -23,7 +23,8 @@ FileExtensionText, KeyboardKey, MimeTypeText, - SpecialFileType + SpecialFileType, + ToolSource } from '$lib/enums'; import { useChatFormPickers } from '$lib/hooks/use-chat-form-pickers.svelte'; import { @@ -73,7 +74,6 @@ disabled?: boolean; isLoading?: boolean; placeholder?: string; - showMcpPromptButton?: boolean; showAddButton?: boolean; showModelSelector?: boolean; @@ -103,7 +103,6 @@ onValueChange, placeholder = 'Type a message...', showAddButton = true, - showMcpPromptButton = false, showModelSelector = true, uploadedFiles = $bindable([]), value = $bindable('') @@ -152,9 +151,18 @@ getServerHome: () => toolsStore.serverHome ?? null, getShowModelSelector: () => showModelSelector, getValue: () => value, - hasCwdTools: () => toolsStore.hasEnabledCwdTools, - hasPrompts: () => - mcpStore.hasPromptsCapability(conversationsStore.preferences.getAllMcpServerOverrides()), + hasCwdTools: () => conversationsStore.preferences.hasEnabledCwdTools(), + // policy-aware, same rule as the agentic flow: MCP category on and at + // least one globally-enabled server whose group key is not disabled + hasPrompts: () => { + const prefs = conversationsStore.preferences; + + if (!prefs.isCategoryEnabled(ToolSource.MCP)) return false; + + return mcpStore + .getServers() + .some((s) => s.enabled && prefs.isServerToolsEnabled(s.id) && s.url.trim()); + }, openModelSelector: () => chatFormActionsRef?.openModelSelector(), setCaretOffset: (offset) => inputRef?.setCaretOffset(offset), setValue: (v) => { @@ -620,8 +628,6 @@ isReasoning={chatStore.isReasoning} {isRecording} onFileUpload={handleFileUpload} - onMcpPromptClick={showMcpPromptButton ? () => pickers.openPromptPicker() : undefined} - onMcpResourcesClick={() => (isResourceDialogOpen = true)} onMcpSettingsClick={() => (isMcpServersDialogOpen = true)} onMicClick={handleMicClick} {onStop} @@ -635,7 +641,7 @@ - {#if toolsStore.hasEnabledCwdTools} + {#if conversationsStore.preferences.hasEnabledCwdTools()} - import { File, MessageSquare, Plus } from '@lucide/svelte'; + import { File, Image, MessageSquare, Mic, Plus, Video } from '@lucide/svelte'; import { ChatFormActionAddToolsSubmenu, McpLogo } from '$lib/components/app'; import { buttonVariants } from '$lib/components/ui/button'; import * as DropdownMenu from '$lib/components/ui/dropdown-menu'; @@ -8,10 +8,10 @@ import { ATTACHMENT_FILE_ITEMS, ATTACHMENT_TOOLTIP_TEXT, - ICON_CLASS_DEFAULT, - TOOLTIP_DELAY_DURATION + ICON_CLASS_DEFAULT } from '$lib/constants'; import { getChatFormActionsContext } from '$lib/contexts'; + import { AttachmentAction, AttachmentItemEnabledWhen } from '$lib/enums'; import { useAttachmentMenu } from '$lib/hooks/use-attachment-menu.svelte'; interface Props { @@ -30,21 +30,29 @@ const attachmentMenu = useAttachmentMenu( () => ({ hasAudioModality: chatFormActions.hasAudioModality, - hasMcpPromptsSupport: chatFormActions.hasMcpPromptsSupport, - hasMcpResourcesSupport: chatFormActions.hasMcpResourcesSupport, hasVideoModality: chatFormActions.hasVideoModality, hasVisionModality: chatFormActions.hasVisionModality }), () => ({ onFileUpload: chatFormActions.onFileUpload, - onMcpPromptClick: chatFormActions.onMcpPromptClick, - onMcpResourcesClick: chatFormActions.onMcpResourcesClick, onSystemPromptClick: chatFormActions.onSystemPromptClick }), () => { dropdownOpen = false; } ); + + const FILE_MODALITY_ICONS: Record = { + [AttachmentItemEnabledWhen.HAS_AUDIO_MODALITY]: { icon: Mic, label: 'Audio' }, + [AttachmentItemEnabledWhen.HAS_VIDEO_MODALITY]: { icon: Video, label: 'Video' }, + [AttachmentItemEnabledWhen.HAS_VISION_MODALITY]: { icon: Image, label: 'Vision' } + }; + + const supportedModalities = $derived.by(() => + ATTACHMENT_FILE_ITEMS.filter((item) => attachmentMenu.isItemEnabled(item.enabledWhen)) + .map((item) => FILE_MODALITY_ICONS[item.enabledWhen ?? '']) + .filter((modality) => modality !== undefined) + );
@@ -84,50 +92,32 @@ } }} > - - - + attachmentMenu.callbacks[AttachmentAction.FILE_UPLOAD]()} + > + + Add files - - - {#each ATTACHMENT_FILE_ITEMS as item (item.id)} - {@const enabled = attachmentMenu.isItemEnabled(item.enabledWhen)} - {#if enabled} - attachmentMenu.callbacks[item.action]()} - > - + {#if supportedModalities.length > 0} + + {#each supportedModalities as modality (modality.label)} + + + + - {item.label} - - {:else if item.disabledTooltip} - - - {#snippet child({ props })} -
- - - - {item.label} - -
- {/snippet} - - - -

{item.disabledTooltip}

- - - {/if} - {/each} - - + +

{modality.label}

+ + + {/each} + + {/if} + + - import { FolderOpen, Server, Zap } from '@lucide/svelte'; - import { McpLogo } from '$lib/components/app'; - import * as DropdownMenu from '$lib/components/ui/dropdown-menu'; - import { ICON_CLASS_DEFAULT } from '$lib/constants'; - import { getChatFormActionsContext } from '$lib/contexts'; - - const chatFormActions = getChatFormActionsContext(); - - function handleServersClick() { - chatFormActions.onMcpSettingsClick?.(); - } - - - - - - - MCP - - - - - - - Servers - - - {#if chatFormActions.hasMcpPromptsSupport} - - - - Prompts - - {/if} - - {#if chatFormActions.hasMcpResourcesSupport} - - - - Resources - - {/if} - - diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte index 1b6fc4b020..197d6c2e59 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddReasoningSubmenu.svelte @@ -8,70 +8,68 @@ const reasoning = useReasoningMenu(); -{#if reasoning.modelSupportsThinking} - - - {#if reasoning.thinkingEnabled} - - {:else if reasoning.isOff} - - {:else} - - {/if} + + + {#if reasoning.isReasoningActive} + + {:else if reasoning.isOff} + + {:else} + + {/if} - - Reasoning - - - {reasoning.currentEffort} - - - - - - {#each reasoning.levels as level (level.value)} - {@const tokenLabel = reasoning.tokenLabel(level)} - reasoning.select(level)} - > - {#if reasoning.isSelected(level)} - - {:else} -
- {/if} + Reasoning - {level.label} + + {reasoning.currentEffort} + + + - {#if tokenLabel} - - {tokenLabel} - - {/if} + + {#each reasoning.levels as level (level.value)} + {@const tokenLabel = reasoning.tokenLabel(level)} + reasoning.select(level)} + > + {#if reasoning.isSelected(level)} + + {:else} +
+ {/if} - {#if level.hasInfo} - - - - + {level.label} - -

Maximum reasoning effort with extended context usage

- - - {/if} - - {/each} - - -{/if} + {#if tokenLabel} + + {tokenLabel} + + {/if} + + {#if level.hasInfo} + + + + + + +

Maximum reasoning effort with extended context usage

+ + + {/if} + + {/each} + + diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte index 2f69dc96de..acd0f4d211 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte @@ -1,18 +1,18 @@
@@ -194,80 +186,15 @@ - (mcpExpanded = open)} open={mcpExpanded}> - - {#if mcpExpanded} - - {:else} - - {/if} + - {/each} - - {#if mcpServers.length === 0} -
- No MCP servers configured -
- {/if} -
- - + System Message + {#if toolsPanel.totalToolCount > 0} (toolsExpanded = open)} open={toolsExpanded}> @@ -289,40 +216,12 @@
- {#each toolsPanel.activeGroups as group (group.key)} - {@const checked = toolsPanel.isGroupChecked(group)} - {@const enabledCount = toolsPanel.getEnabledToolCount(group)} - {@const favicon = toolsPanel.getFavicon(group)} + {#each toolsPanel.categoryGroups as group (group.key)} + {@render sheetGroupRow(group)} + {/each} - + {#each toolsPanel.mcpGroups as group (group.key)} + {@render sheetGroupRow(group)} {/each}
@@ -331,38 +230,55 @@ - - {#if chatFormActions.hasMcpPromptsSupport} - - {/if} - - {#if chatFormActions.hasMcpResourcesSupport} - - {/if}
+ +{#snippet sheetGroupRow(group: ToolGroup)} + {@const checkState = toolsPanel.getGroupCheckState(group)} + {@const enabledCount = toolsPanel.getEnabledToolCount(group)} + {@const favicon = toolsPanel.getFavicon(group)} + {@const groupDisabled = toolsPanel.isGroupDisabled(group)} + + +{/snippet} diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte index 40fed27c70..f495441714 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte @@ -7,6 +7,7 @@ import { CLI_FLAGS, ICON_CLASS_DEFAULT } from '$lib/constants'; import { useToolsPanel } from '$lib/hooks/use-tools-panel.svelte'; import { mcpStore, toolsStore } from '$lib/stores'; + import type { ToolGroup } from '$lib/types'; const toolsPanel = useToolsPanel(); const hasMcpServersAvailable = $derived(mcpStore.getServers().length > 0); @@ -62,95 +63,108 @@ {/if} {:else}
- {#each toolsPanel.activeGroups as group (group.key)} - {@const isExpanded = toolsPanel.expandedGroups.has(group.key)} - {@const checked = toolsPanel.isGroupChecked(group)} - {@const favicon = toolsPanel.getFavicon(group)} + {#each toolsPanel.categoryGroups as group (group.key)} + {@render groupRow(group)} + {/each} - toolsPanel.toggleGroupExpanded(group.key)} - open={isExpanded} - > -
- - {#if isExpanded} - - {:else} - - {/if} - - - {#if favicon} - { - (e.currentTarget as HTMLImageElement).style.display = 'none'; - }} - src={favicon} - /> - {/if} - - {group.label} - - - - {toolsPanel.getEnabledToolCount(group)}/{group.tools.length} - - - - - - {#snippet child({ props })} - toolsPanel.toggleGroupByKey(group.key)} - /> - {/snippet} - - - -

- {checked ? 'Disable' : 'Enable'} - {group.tools.length} tool{group.tools.length !== 1 ? 's' : ''} -

- - -
- - -
- {#each group.tools as entry (entry.key)} - {@const enabled = toolsStore.isToolEnabled(entry.key)} - - {/each} -
- - + {#each toolsPanel.mcpGroups as group (group.key)} + {@render groupRow(group)} {/each}
{/if} + +{#snippet groupRow(group: ToolGroup)} + {@const isExpanded = toolsPanel.expandedGroups.has(group.key)} + {@const checkState = toolsPanel.getGroupCheckState(group)} + {@const favicon = toolsPanel.getFavicon(group)} + {@const groupDisabled = toolsPanel.isGroupDisabled(group)} + + toolsPanel.toggleGroupExpanded(group.key)} + open={isExpanded} + > +
+ + {#if isExpanded} + + {:else} + + {/if} + + + {#if favicon} + { + (e.currentTarget as HTMLImageElement).style.display = 'none'; + }} + src={favicon} + /> + {/if} + + {group.label} + + + + {toolsPanel.getEnabledToolCount(group)}/{group.tools.length} + + + + + + {#snippet child({ props })} + toolsPanel.toggleGroupByKey(group.key)} + /> + {/snippet} + + + +

+ {checkState.checked ? 'Disable' : 'Enable'} + {group.tools.length} tool{group.tools.length !== 1 ? 's' : ''} +

+ + +
+ + +
+ {#each group.tools as entry (entry.key)} + {@const enabled = toolsPanel.isToolEnabled(entry)} + {@const parentDisabled = toolsPanel.isToolParentDisabled(entry)} + + {/each} +
+ + +{/snippet} diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte index f1aa743693..395f2cfbe1 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte @@ -13,7 +13,7 @@ import { setChatFormActionsContext } from '$lib/contexts'; import { FileTypeCategory, MessageRole } from '$lib/enums'; import { ChatService } from '$lib/services'; - import { chatStore, conversationsStore, mcpStore, settingsStore } from '$lib/stores'; + import { chatStore, conversationsStore, settingsStore } from '$lib/stores'; import { getFileTypeCategory } from '$lib/utils'; interface Props { @@ -31,8 +31,6 @@ onMicClick?: () => void; onStop?: () => void; onSystemPromptClick?: () => void; - onMcpPromptClick?: () => void; - onMcpResourcesClick?: () => void; onMcpSettingsClick?: () => void; } @@ -45,8 +43,6 @@ isReasoning = false, isRecording = false, onFileUpload, - onMcpPromptClick, - onMcpResourcesClick, onMcpSettingsClick, onMicClick, onStop, @@ -58,18 +54,6 @@ let currentConfig = $derived(settingsStore.config); - let hasMcpPromptsSupport = $derived.by(() => { - const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides(); - - return mcpStore.hasPromptsCapability(perChatOverrides); - }); - - let hasMcpResourcesSupport = $derived.by(() => { - const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides(); - - return mcpStore.hasResourcesCapability(perChatOverrides); - }); - let hasAudioModality = $state(false); let hasVideoModality = $state(false); let hasVisionModality = $state(false); @@ -142,12 +126,6 @@ get hasAudioModality() { return hasAudioModality; }, - get hasMcpPromptsSupport() { - return hasMcpPromptsSupport; - }, - get hasMcpResourcesSupport() { - return hasMcpResourcesSupport; - }, get hasVideoModality() { return hasVideoModality; }, @@ -157,12 +135,6 @@ get onFileUpload() { return onFileUpload; }, - get onMcpPromptClick() { - return onMcpPromptClick; - }, - get onMcpResourcesClick() { - return onMcpResourcesClick; - }, get onMcpSettingsClick() { return onMcpSettingsClick; }, diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormCurrentWorkingDirectory/ChatFormCurrentWorkingDirectory.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormCurrentWorkingDirectory/ChatFormCurrentWorkingDirectory.svelte index 307d0e702e..ecd1698467 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormCurrentWorkingDirectory/ChatFormCurrentWorkingDirectory.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormCurrentWorkingDirectory/ChatFormCurrentWorkingDirectory.svelte @@ -5,12 +5,12 @@ import SearchInput from '$lib/components/app/forms/SearchInput.svelte'; import * as Popover from '$lib/components/ui/popover'; import { DEFAULT_MOBILE_BREAKPOINT, HOME_TILDE, SEARCH, UI_DATA_ATTRS } from '$lib/constants'; - import { BuiltInTool, GlobSearchType, KeyboardKey } from '$lib/enums'; + import { BuiltInTool, GlobSearchType, KeyboardKey, ToolSource } from '$lib/enums'; import { useDebouncedSearch } from '$lib/hooks/use-debounced-search.svelte'; import { usePickerNavigation } from '$lib/hooks/use-picker-navigation.svelte'; import { useScrollActiveRow } from '$lib/hooks/use-scroll-active-row.svelte'; import { ToolsService } from '$lib/services/tools.service'; - import { toolsStore } from '$lib/stores'; + import { conversationsStore, toolsStore } from '$lib/stores'; import type { GlobEntry } from '$lib/types'; import { abbreviateHome, @@ -63,8 +63,11 @@ // unavailable instead of firing searches that would only fail. Browse is // hidden too: it resolves the picked folder name through the same tool. const fileSearchKey = $derived(toolsStore.getPermissionKey(BuiltInTool.SERVER_FILE_GLOB_SEARCH)); + // effective policy: the active conversation's tool policy, or global defaults const fileSearchEnabled = $derived( - fileSearchKey !== null && toolsStore.isToolEnabled(fileSearchKey) + fileSearchKey !== null && + conversationsStore.preferences.isToolEnabled(fileSearchKey) && + conversationsStore.preferences.isCategoryEnabled(ToolSource.SERVER) ); const searchUnavailableMessage = $derived( fileSearchKey === null diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMcpPrompts/ChatFormPickerMcpPrompts.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMcpPrompts/ChatFormPickerMcpPrompts.svelte index 353d6e7baf..f6a3ee1347 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMcpPrompts/ChatFormPickerMcpPrompts.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMcpPrompts/ChatFormPickerMcpPrompts.svelte @@ -9,7 +9,7 @@ } from '$lib/components/app/chat'; import Badge from '$lib/components/ui/badge/badge.svelte'; import { KeyboardKey } from '$lib/enums'; - import { conversationsStore, mcpStore } from '$lib/stores'; + import { mcpStore } from '$lib/stores'; import type { GetPromptResult, MCPPromptInfo, MCPServerSettingsEntry } from '$lib/types'; import { debounce, uuid } from '$lib/utils'; import { SvelteMap } from 'svelte/reactivity'; @@ -87,8 +87,7 @@ isLoading = true; try { - const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides(); - const initialized = await mcpStore.ensureInitialized(perChatOverrides); + const initialized = await mcpStore.ensureInitialized(); if (!initialized) { prompts = []; diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMention.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMention.svelte index 5a5c8320f9..d09a3d3cb8 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMention.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormPickers/ChatFormPickerMention.svelte @@ -5,10 +5,16 @@ import * as Popover from '$lib/components/ui/popover'; import * as Tooltip from '$lib/components/ui/tooltip'; import { FILE_GLOB_SEARCH_PICKERS, HOME_TILDE, SEARCH } from '$lib/constants'; - import { BuiltInTool, FileMentionEntryType, GlobSearchType, KeyboardKey } from '$lib/enums'; + import { + BuiltInTool, + FileMentionEntryType, + GlobSearchType, + KeyboardKey, + ToolSource + } from '$lib/enums'; import { useDebouncedSearch } from '$lib/hooks/use-debounced-search.svelte'; import { usePickerNavigation } from '$lib/hooks/use-picker-navigation.svelte'; - import { deviceStore, settingsStore, toolsStore } from '$lib/stores'; + import { conversationsStore, deviceStore, settingsStore, toolsStore } from '$lib/stores'; import type { FileMentionEntry, GlobEntryResult } from '$lib/types'; import { abbreviateHome, runGlobSearchWithChildren } from '$lib/utils'; @@ -52,8 +58,11 @@ // --tools) or the user disabled it, the picker still opens but explains // why instead of firing searches that would only fail. const fileSearchKey = $derived(toolsStore.getPermissionKey(BuiltInTool.SERVER_FILE_GLOB_SEARCH)); + // effective policy: the active conversation's tool policy, or global defaults const fileSearchEnabled = $derived( - fileSearchKey !== null && toolsStore.isToolEnabled(fileSearchKey) + fileSearchKey !== null && + conversationsStore.preferences.isToolEnabled(fileSearchKey) && + conversationsStore.preferences.isCategoryEnabled(ToolSource.SERVER) ); let searchResults = $state([]); diff --git a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte index 41d79387b6..6e30cebec6 100644 --- a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte @@ -111,7 +111,6 @@ onValueChange={editCtx.setContent} placeholder="Edit your message..." showAddButton={editCtx.messageRole === MessageRole.USER} - showMcpPromptButton showModelSelector={editCtx.messageRole === MessageRole.USER} value={editCtx.editedContent} /> diff --git a/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte b/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte index 9825b4b90b..962b6774fc 100644 --- a/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte @@ -160,6 +160,5 @@ onSubmit={handleSubmit} onSystemPromptClick={handleSystemPromptClick} onUploadedFileRemove={handleUploadedFileRemove} - showMcpPromptButton /> diff --git a/tools/ui/src/lib/components/app/chat/index.ts b/tools/ui/src/lib/components/app/chat/index.ts index 61ec242e90..d7d7745df1 100644 --- a/tools/ui/src/lib/components/app/chat/index.ts +++ b/tools/ui/src/lib/components/app/chat/index.ts @@ -220,19 +220,6 @@ export { default as ChatFormActionModels } from './ChatForm/ChatFormActions/Chat */ export { default as ChatFormActionAddToolsSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte'; -/** - * Dropdown submenu for MCP prompts and resources in the chat form. - * - * Shows an "MCP" sub-menu item with entries for MCP Prompts and MCP - * Resources. Only visible when the server supports them. - * - * @example - * ```svelte - * - * ``` - */ -export { default as ChatFormActionAddMcpSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpSubmenu.svelte'; - /** * Dropdown submenu for selecting reasoning effort level. * diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMcpResourcesBrowser.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMcpResourcesBrowser.svelte index 6dfcdb856c..82a07477d1 100644 --- a/tools/ui/src/lib/components/app/dialogs/DialogMcpResourcesBrowser.svelte +++ b/tools/ui/src/lib/components/app/dialogs/DialogMcpResourcesBrowser.svelte @@ -8,7 +8,7 @@ import { Button } from '$lib/components/ui/button'; import * as Dialog from '$lib/components/ui/dialog'; import { ICON_CLASS_DEFAULT } from '$lib/constants'; - import { conversationsStore, mcpStore } from '$lib/stores'; + import { mcpStore } from '$lib/stores'; import type { MCPResourceContent, MCPResourceInfo, MCPResourceTemplateInfo } from '$lib/types'; import { getResourceDisplayName } from '$lib/utils'; import { SvelteSet } from 'svelte/reactivity'; @@ -48,8 +48,7 @@ }); async function loadResources() { - const perChatOverrides = conversationsStore.preferences.getAllMcpServerOverrides(); - const initialized = await mcpStore.ensureInitialized(perChatOverrides); + const initialized = await mcpStore.ensureInitialized(); if (initialized) { await mcpStore.fetchAllResources(); diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte index fab45aa970..bc28a754c6 100644 --- a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte +++ b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte @@ -10,7 +10,7 @@ RECOMMENDED_MCP_SERVERS } from '$lib/constants'; import { BooleanString, HealthCheckStatus } from '$lib/enums'; - import { conversationsStore, mcpStore } from '$lib/stores'; + import { mcpStore } from '$lib/stores'; import { canonicalizeServerUrl, parseHeadersToArray, uuid } from '$lib/utils'; interface Props { @@ -234,8 +234,6 @@ useProxy: newServerUseProxy }); - conversationsStore.preferences.setMcpServerOverride(newServerId, true); - handleOpenChange(false); } diff --git a/tools/ui/src/lib/components/app/dialogs/DialogModelInformation.svelte b/tools/ui/src/lib/components/app/dialogs/DialogModelInformation.svelte index 811c24d6b7..e200c004e5 100644 --- a/tools/ui/src/lib/components/app/dialogs/DialogModelInformation.svelte +++ b/tools/ui/src/lib/components/app/dialogs/DialogModelInformation.svelte @@ -76,22 +76,19 @@ - - + + + - - Model Information +
+
+ Model Information - Current model details and capabilities - + Current model details and capabilities +
-
{#if isLoadingModels || isLoadingRouterProps}
Loading model information...
@@ -100,17 +97,15 @@ {@const modelMeta = firstModel.meta} {#if serverProps} - + +