llama-swap SIGTERMs a model when the owner switches models (only one large model fits in RAM), and switching back re-prefills the whole conversation: about 150 s for a 30k-token chat on the 35B. With --cache-disk-path DIR, each slot with a prompt is saved to DIR on shutdown. After the restart, a request whose prompt starts with a saved prompt is restored from the file, so only the suffix is processed. Off by default. Save, in server_context::start_loop() after the loop returns: - disk_cache_collect() copies the state of each slot with >= 256 text tokens to host memory (llama_state_seq_get_data_ext, flags NONE) while the device buffers still exist. A cold server reads its released state through llama_context_restore_memory_host(), with no device use and no extra copy. - vram_share_exit() then frees the VRAM and releases the GPU token as before, so only the D2H copy delays the token release, never the disk IO. - disk_cache_flush() writes the files. A slot is skipped if its KV does not cover its token list (seq_pos_max + 1 != n_tokens: the last batch was not decoded) or if the same token list is already on disk. File (server_prompt_disk_cache, server-task.cpp): - A 48-byte header (magic, version, key hash, sizes, payload hash), then a payload of 8-byte padded sections: tokens, checkpoint table, state, the checkpoint states. The hash is FNV-1a over the 64-bit words of the payload. - Plain fwrite to <name>.tmp, fdatasync, POSIX_FADV_DONTNEED, rename. The read is a plain fread plus DONTNEED. There is no mmap: upstream #28092 livelocked the kernel on AMD + Vulkan by DMA-ing into a MAP_SHARED file. - The index is built at startup from the headers. A bad magic, a size that does not match the header, or an inconsistent table means the file is removed. A hash mismatch at load also removes it, and nothing is restored. Other keys and other versions are kept. A .tmp older than 1 h is removed. - --cache-disk-size N MiB (default 4096, 0 = no limit): before a write, the least recently used *.lpcd files of every model in DIR are removed until the new file fits; a restore updates the file mtime. Entries the new prompt extends are dropped after the write and do not count against the limit. A write also needs its size + 1 GiB free on the file system. A write error or a full disk is logged; the file is not saved and the exit stays clean. Key: the build commit, the model path + size + mtime, LoRA and control vectors, n_ctx, n_ctx_seq, n_seq_max, n_ubatch, kv_unified, swa_full, the K/V types, the flash-attn type, and the RoPE/YaRN parameters. Restore, in get_available_slot() after the RAM cache: - Only while warm (holding the GPU token), with cache_prompt on and no media tokens. The file must reuse at least 256 more tokens than the slot holds. - Reusable means one of two things. Either the saved token list is an exact prefix of the new prompt, or (for a recurrent/SWA model) a saved checkpoint at or below the common prefix can be used. A plain-KV model can be cut back to any prefix. Exact prefix alone is not enough for OpenWebUI on Qwen: the template rewrites the previous assistant turn (the reasoning is dropped), so the saved list is almost never a prefix of the next turn. The checkpoints are exact snapshots, and update_slots() rolls back to one exactly as it does for the RAM cache. They are restored with id_task -1, so one that fails to load means a re-prefill, not an abort. - The slot's current prompt goes to the RAM cache first. A set_data failure clears the slot and removes the file. Not supported with speculative decoding (warning, disabled): the draft state is not saved. Only the slots are saved, not the RAM cache entries. Measured on the pod, CPU only, Qwen3.5-0.8B (hybrid, 6 of 24 layers attention): - 573-token chat: 64.5 MiB file with 2 checkpoints, copy 18 ms, write 171 ms, exit 314 ms (125 ms without the cache). After the restart: read 166 ms, restore 3 ms, turn 2 prompt_n 58 of 596 in 462 ms against a 2913 ms full prefill. The answer is identical to a server that never restarted. - 12k tokens: 200.7 MiB, copy 110 ms, write 350 ms, exit 641 ms, peak RSS +162 MiB during the exit save. After the restart: read 307 ms, restore 18 ms, turn 2 prompt_n 35 of 12188 in 1.2 s against a 72 s prefill. - The /root PVC writes at about 0.66 GB/s with fdatasync. /tmp (the node overlay) stalled once: 5.8 s for 64.5 MiB. llama-swap v239 waits for the old process to exit before it starts the next model, and SIGKILLs a process that is still there after 30 s. So the exit save adds to the model switch time: about 1 s per GB on the PVC. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2385 lines
83 KiB
C++
2385 lines
83 KiB
C++
#include "server-task.h"
|
|
|
|
#include "build-info.h"
|
|
#include "server-chat.h"
|
|
#include "chat.h"
|
|
#include "common.h"
|
|
#include "json-schema-to-grammar.h"
|
|
#include "llama.h"
|
|
#include "sampling.h"
|
|
#include "speculative.h"
|
|
#include "server-common.h"
|
|
|
|
#include <algorithm>
|
|
#include <cerrno>
|
|
#include <chrono>
|
|
#include <cstdio>
|
|
#include <cstring>
|
|
#include <filesystem>
|
|
#include <sstream>
|
|
|
|
#if defined(__linux__)
|
|
#include <fcntl.h>
|
|
#include <unistd.h>
|
|
#endif
|
|
|
|
//
|
|
// task_params
|
|
//
|
|
|
|
json task_params::format_logit_bias(const std::vector<llama_logit_bias> & logit_bias) const {
|
|
json data = json::array();
|
|
for (const auto & lb : logit_bias) {
|
|
data.push_back(json{
|
|
{"bias", lb.bias},
|
|
{"token", lb.token},
|
|
});
|
|
}
|
|
return data;
|
|
}
|
|
|
|
json task_params::to_json(bool only_metrics) const {
|
|
std::vector<std::string> samplers;
|
|
samplers.reserve(sampling.samplers.size());
|
|
for (const auto & sampler : sampling.samplers) {
|
|
samplers.emplace_back(common_sampler_type_to_str(sampler));
|
|
}
|
|
|
|
json lora = json::array();
|
|
for (auto & it : this->lora) {
|
|
lora.push_back({{"id", it.first}, {"scale", it.second}});
|
|
}
|
|
|
|
if (only_metrics) {
|
|
return json {
|
|
{"seed", sampling.seed},
|
|
{"temperature", sampling.temp},
|
|
{"dynatemp_range", sampling.dynatemp_range},
|
|
{"dynatemp_exponent", sampling.dynatemp_exponent},
|
|
{"top_k", sampling.top_k},
|
|
{"top_p", sampling.top_p},
|
|
{"min_p", sampling.min_p},
|
|
{"top_n_sigma", sampling.top_n_sigma},
|
|
{"xtc_probability", sampling.xtc_probability},
|
|
{"xtc_threshold", sampling.xtc_threshold},
|
|
{"typical_p", sampling.typ_p},
|
|
{"repeat_last_n", sampling.penalty_last_n},
|
|
{"repeat_penalty", sampling.penalty_repeat},
|
|
{"presence_penalty", sampling.penalty_present},
|
|
{"frequency_penalty", sampling.penalty_freq},
|
|
{"dry_multiplier", sampling.dry_multiplier},
|
|
{"dry_base", sampling.dry_base},
|
|
{"dry_allowed_length", sampling.dry_allowed_length},
|
|
{"dry_penalty_last_n", sampling.dry_penalty_last_n},
|
|
{"mirostat", sampling.mirostat},
|
|
{"mirostat_tau", sampling.mirostat_tau},
|
|
{"mirostat_eta", sampling.mirostat_eta},
|
|
{"adaptive_target", sampling.adaptive_target},
|
|
{"adaptive_decay", sampling.adaptive_decay},
|
|
{"max_tokens", n_predict},
|
|
{"n_predict", n_predict}, // TODO: deduplicate?
|
|
{"n_keep", n_keep},
|
|
{"n_discard", n_discard},
|
|
{"ignore_eos", sampling.ignore_eos},
|
|
{"stream", stream},
|
|
{"n_probs", sampling.n_probs},
|
|
{"min_keep", sampling.min_keep},
|
|
{"chat_format", common_chat_format_name(chat_format)},
|
|
{"reasoning_format", common_reasoning_format_name(reasoning_format)},
|
|
{"reasoning_in_content", stream && reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY},
|
|
{"generation_prompt", sampling.generation_prompt},
|
|
{"samplers", samplers},
|
|
{"speculative.types", common_speculative_type_name_str(speculative.types)},
|
|
{"timings_per_token", timings_per_token},
|
|
{"post_sampling_probs", post_sampling_probs},
|
|
{"backend_sampling", sampling.backend_sampling},
|
|
{"lora", lora},
|
|
};
|
|
}
|
|
|
|
auto grammar_triggers = json::array();
|
|
for (const auto & trigger : sampling.grammar_triggers) {
|
|
server_grammar_trigger ct(trigger);
|
|
grammar_triggers.push_back(ct.to_json());
|
|
}
|
|
|
|
return json {
|
|
{"seed", sampling.seed},
|
|
{"temperature", sampling.temp},
|
|
{"dynatemp_range", sampling.dynatemp_range},
|
|
{"dynatemp_exponent", sampling.dynatemp_exponent},
|
|
{"top_k", sampling.top_k},
|
|
{"top_p", sampling.top_p},
|
|
{"min_p", sampling.min_p},
|
|
{"top_n_sigma", sampling.top_n_sigma},
|
|
{"xtc_probability", sampling.xtc_probability},
|
|
{"xtc_threshold", sampling.xtc_threshold},
|
|
{"typical_p", sampling.typ_p},
|
|
{"repeat_last_n", sampling.penalty_last_n},
|
|
{"repeat_penalty", sampling.penalty_repeat},
|
|
{"presence_penalty", sampling.penalty_present},
|
|
{"frequency_penalty", sampling.penalty_freq},
|
|
{"dry_multiplier", sampling.dry_multiplier},
|
|
{"dry_base", sampling.dry_base},
|
|
{"dry_allowed_length", sampling.dry_allowed_length},
|
|
{"dry_penalty_last_n", sampling.dry_penalty_last_n},
|
|
{"dry_sequence_breakers", sampling.dry_sequence_breakers},
|
|
{"mirostat", sampling.mirostat},
|
|
{"mirostat_tau", sampling.mirostat_tau},
|
|
{"mirostat_eta", sampling.mirostat_eta},
|
|
{"adaptive_target", sampling.adaptive_target},
|
|
{"adaptive_decay", sampling.adaptive_decay},
|
|
{"stop", antiprompt},
|
|
{"max_tokens", n_predict},
|
|
{"n_predict", n_predict}, // TODO: deduplicate?
|
|
{"n_keep", n_keep},
|
|
{"n_discard", n_discard},
|
|
{"ignore_eos", sampling.ignore_eos},
|
|
{"stream", stream},
|
|
{"logit_bias", format_logit_bias(sampling.logit_bias)},
|
|
{"n_probs", sampling.n_probs},
|
|
{"min_keep", sampling.min_keep},
|
|
{"grammar", common_grammar_value(sampling.grammar)},
|
|
{"grammar_lazy", sampling.grammar_lazy},
|
|
{"grammar_triggers", grammar_triggers},
|
|
{"preserved_tokens", sampling.preserved_tokens},
|
|
{"chat_format", common_chat_format_name(chat_format)},
|
|
{"reasoning_format", common_reasoning_format_name(reasoning_format)},
|
|
{"reasoning_in_content", stream && reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY},
|
|
{"generation_prompt", sampling.generation_prompt},
|
|
{"samplers", samplers},
|
|
{"speculative.types", common_speculative_type_name_str(speculative.types)},
|
|
{"timings_per_token", timings_per_token},
|
|
{"post_sampling_probs", post_sampling_probs},
|
|
{"backend_sampling", sampling.backend_sampling},
|
|
{"lora", lora},
|
|
};
|
|
}
|
|
|
|
//
|
|
// task_result_state
|
|
//
|
|
task_result_state::task_result_state(common_chat_session session)
|
|
: chat_session(std::move(session))
|
|
, chat_msg(chat_session.msg())
|
|
, oai_resp_id("resp_" + random_string())
|
|
, oai_resp_reasoning_id("rs_" + random_string())
|
|
, oai_resp_message_id("msg_" + random_string()) {
|
|
}
|
|
|
|
common_chat_msg task_result_state::update_chat_msg(
|
|
const common_chat_input & added,
|
|
bool is_partial,
|
|
std::vector<common_chat_msg_diff> & diffs,
|
|
bool filter_tool_calls) {
|
|
auto msg_prv_copy = chat_msg;
|
|
auto new_msg = is_partial ? chat_session.feed(added) : chat_session.finish(added);
|
|
if (!new_msg.empty()) {
|
|
new_msg.set_tool_call_ids(generated_tool_call_ids, gen_tool_call_id);
|
|
chat_msg = new_msg;
|
|
auto all_diffs = common_chat_msg_diff::compute_diffs(msg_prv_copy, chat_msg);
|
|
|
|
if (!filter_tool_calls) {
|
|
diffs = std::move(all_diffs);
|
|
} else {
|
|
for (auto & d : all_diffs) {
|
|
// If this is a new type of delta, flush all currently pending tool call names
|
|
for (size_t i = 0; i < chat_msg.tool_calls.size(); ++i) {
|
|
if (sent_tool_call_names.count(i) || chat_msg.tool_calls[i].name.empty()) {
|
|
continue;
|
|
}
|
|
if (d.tool_call_index != i || !d.tool_call_delta.arguments.empty()) {
|
|
common_chat_msg_diff header;
|
|
header.tool_call_index = i;
|
|
header.tool_call_delta.id = chat_msg.tool_calls[i].id;
|
|
header.tool_call_delta.name = chat_msg.tool_calls[i].name;
|
|
diffs.push_back(std::move(header));
|
|
sent_tool_call_names.insert(i);
|
|
}
|
|
}
|
|
|
|
if (d.tool_call_index == std::string::npos) {
|
|
diffs.push_back(std::move(d));
|
|
} else {
|
|
size_t i = d.tool_call_index;
|
|
if (sent_tool_call_names.count(i)) {
|
|
if (!d.tool_call_delta.arguments.empty()) {
|
|
d.tool_call_delta.name = "";
|
|
d.tool_call_delta.id = "";
|
|
diffs.push_back(std::move(d));
|
|
}
|
|
} else {
|
|
// Not sent yet.
|
|
if (!d.tool_call_delta.arguments.empty() || !is_partial) {
|
|
d.tool_call_delta.name = chat_msg.tool_calls[i].name;
|
|
d.tool_call_delta.id = chat_msg.tool_calls[i].id;
|
|
diffs.push_back(std::move(d));
|
|
sent_tool_call_names.insert(i);
|
|
} else {
|
|
// Suppress
|
|
}
|
|
}
|
|
}
|
|
}
|
|
// Final check at EOF
|
|
if (!is_partial) {
|
|
for (size_t i = 0; i < chat_msg.tool_calls.size(); ++i) {
|
|
if (!sent_tool_call_names.count(i) && !chat_msg.tool_calls[i].name.empty()) {
|
|
common_chat_msg_diff header;
|
|
header.tool_call_index = i;
|
|
header.tool_call_delta.id = chat_msg.tool_calls[i].id;
|
|
header.tool_call_delta.name = chat_msg.tool_calls[i].name;
|
|
diffs.push_back(std::move(header));
|
|
sent_tool_call_names.insert(i);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
return chat_msg;
|
|
}
|
|
|
|
//
|
|
// result_prompt_progress
|
|
//
|
|
json result_prompt_progress::to_json() const {
|
|
return json {
|
|
{"total", total},
|
|
{"cache", cache},
|
|
{"processed", processed},
|
|
{"time_ms", time_ms},
|
|
};
|
|
}
|
|
|
|
static inline std::string stop_type_to_str(stop_type type) {
|
|
switch (type) {
|
|
case STOP_TYPE_EOS: return "eos";
|
|
case STOP_TYPE_WORD: return "word";
|
|
case STOP_TYPE_LIMIT: return "limit";
|
|
default: return "none";
|
|
}
|
|
}
|
|
|
|
//
|
|
// completion_token_output
|
|
//
|
|
|
|
json completion_token_output::to_json(bool post_sampling_probs) const {
|
|
json probs_for_token = json::array();
|
|
for (const auto & p : probs) {
|
|
std::string txt(p.txt);
|
|
txt.resize(validate_utf8(txt));
|
|
probs_for_token.push_back(json {
|
|
{"id", p.tok},
|
|
{"token", txt},
|
|
{"bytes", str_to_bytes(p.txt)},
|
|
{
|
|
post_sampling_probs ? "prob" : "logprob",
|
|
post_sampling_probs ? p.prob : logarithm(p.prob)
|
|
},
|
|
});
|
|
}
|
|
return probs_for_token;
|
|
}
|
|
|
|
json completion_token_output::probs_vector_to_json(const std::vector<completion_token_output> & probs, bool post_sampling_probs) {
|
|
json out = json::array();
|
|
for (const auto & p : probs) {
|
|
std::string txt(p.text_to_send);
|
|
txt.resize(validate_utf8(txt));
|
|
out.push_back(json {
|
|
{"id", p.tok},
|
|
{"token", txt},
|
|
{"bytes", str_to_bytes(p.text_to_send)},
|
|
{
|
|
post_sampling_probs ? "prob" : "logprob",
|
|
post_sampling_probs ? p.prob : logarithm(p.prob)
|
|
},
|
|
{
|
|
post_sampling_probs ? "top_probs" : "top_logprobs",
|
|
p.to_json(post_sampling_probs)
|
|
},
|
|
});
|
|
}
|
|
return out;
|
|
}
|
|
|
|
float completion_token_output::logarithm(float x) {
|
|
// the JSON library converts -inf to null, so we need to prevent that
|
|
return x == 0.0f ? std::numeric_limits<float>::lowest() : std::log(x);
|
|
}
|
|
|
|
std::vector<unsigned char> completion_token_output::str_to_bytes(const std::string & str) {
|
|
std::vector<unsigned char> bytes;
|
|
for (unsigned char c : str) {
|
|
bytes.push_back(c);
|
|
}
|
|
return bytes;
|
|
}
|
|
|
|
//
|
|
// server_task_result_cmpl_final
|
|
//
|
|
json server_task_result_cmpl_final::to_json() {
|
|
GGML_ASSERT(is_updated && "update() must be called before to_json()");
|
|
switch (res_type) {
|
|
case TASK_RESPONSE_TYPE_NONE:
|
|
return to_json_non_oaicompat();
|
|
case TASK_RESPONSE_TYPE_OAI_CMPL:
|
|
return to_json_oaicompat();
|
|
case TASK_RESPONSE_TYPE_OAI_CHAT:
|
|
return stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat();
|
|
case TASK_RESPONSE_TYPE_OAI_RESP:
|
|
return stream ? to_json_oaicompat_resp_stream() : to_json_oaicompat_resp();
|
|
case TASK_RESPONSE_TYPE_OAI_ASR:
|
|
return to_json_oaicompat_asr();
|
|
case TASK_RESPONSE_TYPE_ANTHROPIC:
|
|
return stream ? to_json_anthropic_stream() : to_json_anthropic();
|
|
default:
|
|
GGML_ASSERT(false && "Invalid task_response_type");
|
|
}
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_non_oaicompat() {
|
|
json res = json {
|
|
{"index", index},
|
|
{"content", content.text},
|
|
{"tokens", tokens},
|
|
{"id_slot", id_slot},
|
|
{"stop", true},
|
|
{"model", oaicompat_model},
|
|
{"tokens_predicted", n_decoded},
|
|
{"tokens_evaluated", n_prompt_tokens},
|
|
{"generation_settings", generation_params.to_json()},
|
|
{"prompt", prompt},
|
|
{"has_new_line", has_new_line},
|
|
{"truncated", truncated},
|
|
{"stop_type", stop_type_to_str(stop)},
|
|
{"stopping_word", stopping_word},
|
|
{"tokens_cached", n_tokens_cached},
|
|
{"timings", stats.to_json()},
|
|
};
|
|
if (!stream && !probs_output.empty()) {
|
|
res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs);
|
|
}
|
|
return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
|
|
}
|
|
|
|
json server_task_result_cmpl_final::usage_json_oaicompat() {
|
|
return json {
|
|
{"completion_tokens", n_decoded},
|
|
{"prompt_tokens", n_prompt_tokens},
|
|
{"total_tokens", n_decoded + n_prompt_tokens},
|
|
{"prompt_tokens_details", json { {"cached_tokens", n_prompt_tokens_cache} }},
|
|
};
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_oaicompat() {
|
|
std::time_t t = std::time(0);
|
|
json logprobs = json(nullptr); // OAI default to null
|
|
if (!stream && probs_output.size() > 0) {
|
|
logprobs = json{
|
|
{"content", completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs)},
|
|
};
|
|
}
|
|
json finish_reason = "length";
|
|
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
|
|
finish_reason = "stop";
|
|
}
|
|
json res = json {
|
|
{"choices", json::array({
|
|
json{
|
|
{"text", content.text},
|
|
{"index", index},
|
|
{"logprobs", logprobs},
|
|
{"finish_reason", finish_reason},
|
|
}
|
|
})},
|
|
{"created", t},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "text_completion"},
|
|
{"usage", usage_json_oaicompat()},
|
|
{"id", oaicompat_cmpl_id}
|
|
};
|
|
|
|
// extra fields for debugging purposes
|
|
if (verbose) {
|
|
res["__verbose"] = to_json_non_oaicompat();
|
|
}
|
|
if (stats.is_set()) {
|
|
res["timings"] = stats.to_json();
|
|
}
|
|
|
|
return res;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_oaicompat_chat() {
|
|
std::string finish_reason = "length";
|
|
common_chat_msg msg;
|
|
if (!oaicompat_msg.empty()) {
|
|
msg = oaicompat_msg;
|
|
} else {
|
|
msg.role = "assistant";
|
|
msg.content = content.text;
|
|
}
|
|
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
|
|
finish_reason = msg.tool_calls.empty() ? "stop" : "tool_calls";
|
|
}
|
|
|
|
json choice {
|
|
{"finish_reason", finish_reason},
|
|
{"index", index},
|
|
{"message", msg.to_json_oaicompat()},
|
|
};
|
|
|
|
if (!stream && probs_output.size() > 0) {
|
|
choice["logprobs"] = json{
|
|
{"content", completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs)},
|
|
};
|
|
}
|
|
|
|
std::time_t t = std::time(0);
|
|
|
|
json res = json {
|
|
{"choices", json::array({choice})},
|
|
{"created", t},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "chat.completion"},
|
|
{"usage", usage_json_oaicompat()},
|
|
{"id", oaicompat_cmpl_id}
|
|
};
|
|
|
|
// extra fields for debugging purposes
|
|
if (verbose) {
|
|
res["__verbose"] = to_json_non_oaicompat();
|
|
}
|
|
if (stats.is_set()) {
|
|
res["timings"] = stats.to_json();
|
|
}
|
|
|
|
return res;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
|
|
std::time_t t = std::time(0);
|
|
std::string finish_reason = "length";
|
|
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
|
|
finish_reason = oaicompat_msg.tool_calls.empty() ? "stop" : "tool_calls";
|
|
}
|
|
|
|
json deltas = json::array();
|
|
for (const auto & diff : oaicompat_msg_diffs) {
|
|
deltas.push_back({
|
|
{"choices", json::array({
|
|
json {
|
|
{"finish_reason", nullptr},
|
|
{"index", index},
|
|
{"delta", server_chat_msg_diff_to_json_oaicompat(diff)},
|
|
},
|
|
})},
|
|
{"created", t},
|
|
{"id", oaicompat_cmpl_id},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "chat.completion.chunk"},
|
|
});
|
|
}
|
|
|
|
deltas.push_back({
|
|
{"choices", json::array({
|
|
json {
|
|
{"finish_reason", finish_reason},
|
|
{"index", index},
|
|
{"delta", json::object()},
|
|
},
|
|
})},
|
|
{"created", t},
|
|
{"id", oaicompat_cmpl_id},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "chat.completion.chunk"},
|
|
});
|
|
|
|
if (include_usage) {
|
|
// OpenAI API spec for chat.completion.chunks specifies an empty `choices` array for the last chunk when including usage
|
|
// https://platform.openai.com/docs/api-reference/chat_streaming/streaming#chat_streaming/streaming-choices
|
|
deltas.push_back({
|
|
{"choices", json::array()},
|
|
{"created", t},
|
|
{"id", oaicompat_cmpl_id},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "chat.completion.chunk"},
|
|
{"usage", usage_json_oaicompat()},
|
|
});
|
|
}
|
|
|
|
if (stats.is_set()) {
|
|
deltas.back()["timings"] = stats.to_json();
|
|
}
|
|
|
|
// extra fields for debugging purposes
|
|
if (verbose && !deltas.empty()) {
|
|
deltas.front()["__verbose"] = to_json_non_oaicompat();
|
|
}
|
|
|
|
return deltas;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_oaicompat_resp() {
|
|
common_chat_msg msg;
|
|
if (!oaicompat_msg.empty()) {
|
|
msg = oaicompat_msg;
|
|
} else {
|
|
msg.role = "assistant";
|
|
msg.content = content.text;
|
|
}
|
|
|
|
std::vector<json> output;
|
|
|
|
if (msg.reasoning_content != "") {
|
|
output.push_back(json {
|
|
{"id", "rs_" + random_string()},
|
|
{"summary", json::array()},
|
|
{"type", "reasoning"},
|
|
{"content", json::array({ json {
|
|
{"text", msg.reasoning_content},
|
|
{"type", "reasoning_text"},
|
|
}})},
|
|
{"encrypted_content", ""},
|
|
{"status", "completed"},
|
|
});
|
|
}
|
|
|
|
if (msg.content != "") {
|
|
output.push_back(json {
|
|
{"content", json::array({ json {
|
|
{"type", "output_text"},
|
|
{"annotations", json::array()},
|
|
{"logprobs", json::array()},
|
|
{"text", msg.content},
|
|
}})},
|
|
{"id", "msg_" + random_string()},
|
|
{"role", msg.role},
|
|
{"status", "completed"},
|
|
{"type", "message"},
|
|
});
|
|
}
|
|
|
|
for (const common_chat_tool_call & tool_call : oaicompat_msg.tool_calls) {
|
|
output.push_back(json {
|
|
{"id", "fc_" + tool_call.id},
|
|
{"type", "function_call"},
|
|
{"status", "completed"},
|
|
{"arguments", tool_call.arguments},
|
|
{"call_id", "call_" + tool_call.id},
|
|
{"name", tool_call.name},
|
|
});
|
|
}
|
|
|
|
std::time_t t = std::time(0);
|
|
json res = {
|
|
{"completed_at", t},
|
|
{"created_at", t},
|
|
{"id", oai_resp_id},
|
|
{"model", oaicompat_model},
|
|
{"object", "response"},
|
|
{"output", output},
|
|
{"status", "completed"},
|
|
{"usage", json {
|
|
{"input_tokens", n_prompt_tokens},
|
|
{"output_tokens", n_decoded},
|
|
{"total_tokens", n_decoded + n_prompt_tokens},
|
|
{"input_tokens_details", json { {"cached_tokens", n_prompt_tokens_cache} }},
|
|
}},
|
|
};
|
|
|
|
return res;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
|
|
std::vector<json> server_sent_events;
|
|
std::vector<json> output;
|
|
size_t output_index = 0;
|
|
|
|
if (oaicompat_msg.reasoning_content != "") {
|
|
const size_t item_output_index = output_index++;
|
|
const json output_item = json {
|
|
{"id", oai_resp_reasoning_id},
|
|
{"summary", json::array()},
|
|
{"type", "reasoning"},
|
|
{"content", json::array({ json {
|
|
{"text", oaicompat_msg.reasoning_content},
|
|
{"type", "reasoning_text"},
|
|
}})},
|
|
{"encrypted_content", ""},
|
|
};
|
|
|
|
server_sent_events.push_back(json {
|
|
{"event", "response.output_item.done"},
|
|
{"data", json {
|
|
{"type", "response.output_item.done"},
|
|
{"output_index", item_output_index},
|
|
{"item", output_item}
|
|
}}
|
|
});
|
|
output.push_back(output_item);
|
|
}
|
|
|
|
if (oaicompat_msg.content != "") {
|
|
const size_t item_output_index = output_index++;
|
|
server_sent_events.push_back(json {
|
|
{"event", "response.output_text.done"},
|
|
{"data", json {
|
|
{"type", "response.output_text.done"},
|
|
{"item_id", oai_resp_message_id},
|
|
{"output_index", item_output_index},
|
|
{"content_index", 0},
|
|
{"text", oaicompat_msg.content}
|
|
}}
|
|
});
|
|
|
|
const json content_part = {
|
|
{"type", "output_text"},
|
|
{"annotations", json::array()},
|
|
{"logprobs", json::array()},
|
|
{"text", oaicompat_msg.content}
|
|
};
|
|
|
|
server_sent_events.push_back(json {
|
|
{"event", "response.content_part.done"},
|
|
{"data", json {
|
|
{"type", "response.content_part.done"},
|
|
{"item_id", oai_resp_message_id},
|
|
{"output_index", item_output_index},
|
|
{"content_index", 0},
|
|
{"part", content_part}
|
|
}}
|
|
});
|
|
const json output_item = {
|
|
{"type", "message"},
|
|
{"status", "completed"},
|
|
{"id", oai_resp_message_id},
|
|
{"content", json::array({content_part})},
|
|
{"role", "assistant"}
|
|
};
|
|
|
|
server_sent_events.push_back(json {
|
|
{"event", "response.output_item.done"},
|
|
{"data", json {
|
|
{"type", "response.output_item.done"},
|
|
{"output_index", item_output_index},
|
|
{"item", output_item}
|
|
}}
|
|
});
|
|
output.push_back(output_item);
|
|
}
|
|
|
|
for (const common_chat_tool_call & tool_call : oaicompat_msg.tool_calls) {
|
|
const size_t item_output_index = output_index++;
|
|
const json output_item = {
|
|
{"id", "fc_" + tool_call.id},
|
|
{"type", "function_call"},
|
|
{"status", "completed"},
|
|
{"arguments", tool_call.arguments},
|
|
{"call_id", "call_" + tool_call.id},
|
|
{"name", tool_call.name}
|
|
};
|
|
server_sent_events.push_back(json {
|
|
{"event", "response.output_item.done"},
|
|
{"data", json {
|
|
{"type", "response.output_item.done"},
|
|
{"output_index", item_output_index},
|
|
{"item", output_item}
|
|
}}
|
|
});
|
|
output.push_back(output_item);
|
|
}
|
|
|
|
std::time_t t = std::time(0);
|
|
server_sent_events.push_back(json {
|
|
{"event", "response.completed"},
|
|
{"data", json {
|
|
{"type", "response.completed"},
|
|
{"response", json {
|
|
{"id", oai_resp_id},
|
|
{"object", "response"},
|
|
{"created_at", t},
|
|
{"status", "completed"},
|
|
{"model", oaicompat_model},
|
|
{"output", output},
|
|
{"usage", json {
|
|
{"input_tokens", n_prompt_tokens},
|
|
{"output_tokens", n_decoded},
|
|
{"total_tokens", n_decoded + n_prompt_tokens},
|
|
{"input_tokens_details", json { {"cached_tokens", n_prompt_tokens_cache} }},
|
|
}}
|
|
}},
|
|
}}
|
|
});
|
|
|
|
if (stats.is_set()) {
|
|
server_sent_events.back().at("data")["timings"] = stats.to_json();
|
|
}
|
|
|
|
return server_sent_events;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_oaicompat_asr() {
|
|
json event = json {
|
|
{"type", "transcript.text.done"},
|
|
{"text", oaicompat_msg.content},
|
|
{"usage", json {
|
|
{"type", "tokens"},
|
|
{"input_tokens", n_prompt_tokens},
|
|
{"output_tokens", n_decoded},
|
|
{"total_tokens", n_decoded + n_prompt_tokens},
|
|
{"input_tokens_details", json { {"cached_tokens", n_prompt_tokens_cache} }},
|
|
}},
|
|
};
|
|
return event;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_anthropic() {
|
|
std::string stop_reason = "max_tokens";
|
|
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
|
|
stop_reason = oaicompat_msg.tool_calls.empty() ? "end_turn" : "tool_use";
|
|
}
|
|
|
|
json content_blocks = json::array();
|
|
|
|
common_chat_msg msg;
|
|
if (!oaicompat_msg.empty()) {
|
|
msg = oaicompat_msg;
|
|
} else {
|
|
msg.role = "assistant";
|
|
msg.content = content.text;
|
|
}
|
|
|
|
// thinking block comes first (Anthropic extended thinking format)
|
|
if (!msg.reasoning_content.empty()) {
|
|
content_blocks.push_back({
|
|
{"type", "thinking"},
|
|
{"thinking", msg.reasoning_content},
|
|
{"signature", ""} // empty signature for local models (no cryptographic verification)
|
|
});
|
|
}
|
|
|
|
if (!msg.content.empty()) {
|
|
content_blocks.push_back({
|
|
{"type", "text"},
|
|
{"text", msg.content}
|
|
});
|
|
}
|
|
|
|
for (const auto & tool_call : msg.tool_calls) {
|
|
json tool_use_block = {
|
|
{"type", "tool_use"},
|
|
{"id", tool_call.id},
|
|
{"name", tool_call.name}
|
|
};
|
|
|
|
try {
|
|
tool_use_block["input"] = json::parse(tool_call.arguments);
|
|
} catch (const std::exception &) {
|
|
tool_use_block["input"] = json::object();
|
|
}
|
|
|
|
content_blocks.push_back(tool_use_block);
|
|
}
|
|
|
|
json res = {
|
|
{"id", oaicompat_cmpl_id},
|
|
{"type", "message"},
|
|
{"role", "assistant"},
|
|
{"content", content_blocks},
|
|
{"model", oaicompat_model},
|
|
{"stop_reason", stop_reason},
|
|
{"stop_sequence", stopping_word.empty() ? nullptr : json(stopping_word)},
|
|
{"usage", {
|
|
{"cache_read_input_tokens", n_prompt_tokens_cache},
|
|
{"input_tokens", n_prompt_tokens - n_prompt_tokens_cache},
|
|
{"output_tokens", n_decoded}
|
|
}}
|
|
};
|
|
|
|
return res;
|
|
}
|
|
|
|
json server_task_result_cmpl_final::to_json_anthropic_stream() {
|
|
json events = json::array();
|
|
|
|
std::string stop_reason = "max_tokens";
|
|
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
|
|
stop_reason = oaicompat_msg.tool_calls.empty() ? "end_turn" : "tool_use";
|
|
}
|
|
|
|
bool has_thinking = !oaicompat_msg.reasoning_content.empty();
|
|
bool has_text = !oaicompat_msg.content.empty();
|
|
size_t num_tool_calls = oaicompat_msg.tool_calls.size();
|
|
|
|
// content block indices: thinking (0) -> text (0 or 1) -> tool_use (n+)
|
|
size_t thinking_block_index = 0;
|
|
size_t text_block_index = has_thinking ? 1 : 0;
|
|
|
|
bool thinking_block_started = false;
|
|
bool text_block_started = false;
|
|
std::unordered_set<size_t> tool_calls_started;
|
|
|
|
for (const auto & diff : oaicompat_msg_diffs) {
|
|
// handle thinking/reasoning content
|
|
if (!diff.reasoning_content_delta.empty()) {
|
|
if (!thinking_block_started) {
|
|
events.push_back({
|
|
{"event", "content_block_start"},
|
|
{"data", {
|
|
{"type", "content_block_start"},
|
|
{"index", thinking_block_index},
|
|
{"content_block", {
|
|
{"type", "thinking"},
|
|
{"thinking", ""}
|
|
}}
|
|
}}
|
|
});
|
|
thinking_block_started = true;
|
|
}
|
|
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", thinking_block_index},
|
|
{"delta", {
|
|
{"type", "thinking_delta"},
|
|
{"thinking", diff.reasoning_content_delta}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
|
|
// handle regular text content
|
|
if (!diff.content_delta.empty()) {
|
|
if (!text_block_started) {
|
|
events.push_back({
|
|
{"event", "content_block_start"},
|
|
{"data", {
|
|
{"type", "content_block_start"},
|
|
{"index", text_block_index},
|
|
{"content_block", {
|
|
{"type", "text"},
|
|
{"text", ""}
|
|
}}
|
|
}}
|
|
});
|
|
text_block_started = true;
|
|
}
|
|
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", text_block_index},
|
|
{"delta", {
|
|
{"type", "text_delta"},
|
|
{"text", diff.content_delta}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
|
|
// handle tool calls
|
|
if (diff.tool_call_index != std::string::npos) {
|
|
size_t content_block_index = (has_thinking ? 1 : 0) + (has_text ? 1 : 0) + diff.tool_call_index;
|
|
|
|
if (tool_calls_started.find(diff.tool_call_index) == tool_calls_started.end()) {
|
|
const auto & full_tool_call = oaicompat_msg.tool_calls[diff.tool_call_index];
|
|
|
|
events.push_back({
|
|
{"event", "content_block_start"},
|
|
{"data", {
|
|
{"type", "content_block_start"},
|
|
{"index", content_block_index},
|
|
{"content_block", {
|
|
{"type", "tool_use"},
|
|
{"id", full_tool_call.id},
|
|
{"name", full_tool_call.name}
|
|
}}
|
|
}}
|
|
});
|
|
tool_calls_started.insert(diff.tool_call_index);
|
|
}
|
|
|
|
if (!diff.tool_call_delta.arguments.empty()) {
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", content_block_index},
|
|
{"delta", {
|
|
{"type", "input_json_delta"},
|
|
{"partial_json", diff.tool_call_delta.arguments}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
}
|
|
}
|
|
|
|
// close content blocks in order
|
|
if (has_thinking) {
|
|
// Anthropic API requires a signature_delta before closing thinking blocks
|
|
// We use an empty signature since we can't generate a cryptographic signature for local models
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", thinking_block_index},
|
|
{"delta", {
|
|
{"type", "signature_delta"},
|
|
{"signature", ""}
|
|
}}
|
|
}}
|
|
});
|
|
events.push_back({
|
|
{"event", "content_block_stop"},
|
|
{"data", {
|
|
{"type", "content_block_stop"},
|
|
{"index", thinking_block_index}
|
|
}}
|
|
});
|
|
}
|
|
|
|
if (has_text) {
|
|
events.push_back({
|
|
{"event", "content_block_stop"},
|
|
{"data", {
|
|
{"type", "content_block_stop"},
|
|
{"index", text_block_index}
|
|
}}
|
|
});
|
|
}
|
|
|
|
for (size_t i = 0; i < num_tool_calls; i++) {
|
|
size_t content_block_index = (has_thinking ? 1 : 0) + (has_text ? 1 : 0) + i;
|
|
events.push_back({
|
|
{"event", "content_block_stop"},
|
|
{"data", {
|
|
{"type", "content_block_stop"},
|
|
{"index", content_block_index}
|
|
}}
|
|
});
|
|
}
|
|
|
|
events.push_back({
|
|
{"event", "message_delta"},
|
|
{"data", {
|
|
{"type", "message_delta"},
|
|
{"delta", {
|
|
{"stop_reason", stop_reason},
|
|
{"stop_sequence", stopping_word.empty() ? nullptr : json(stopping_word)}
|
|
}},
|
|
{"usage", {
|
|
{"output_tokens", n_decoded}
|
|
}}
|
|
}}
|
|
});
|
|
|
|
events.push_back({
|
|
{"event", "message_stop"},
|
|
{"data", {
|
|
{"type", "message_stop"}
|
|
}}
|
|
});
|
|
|
|
return events;
|
|
}
|
|
|
|
//
|
|
// server_task_result_cmpl_partial
|
|
//
|
|
void server_task_result_cmpl_partial::update(task_result_state & state) {
|
|
is_updated = true;
|
|
if (is_begin) {
|
|
return; // begin marker only flushes headers, skip parsing
|
|
}
|
|
state.update_chat_msg(content, true, oaicompat_msg_diffs);
|
|
|
|
// Copy current state for use in to_json_*() (reflects state BEFORE this chunk)
|
|
thinking_block_started = state.thinking_block_started;
|
|
text_block_started = state.text_block_started;
|
|
|
|
oai_resp_created = state.oai_resp_created;
|
|
oai_resp_id = state.oai_resp_id;
|
|
oai_resp_reasoning_id = state.oai_resp_reasoning_id;
|
|
oai_resp_message_id = state.oai_resp_message_id;
|
|
oai_resp_fc_id = state.oai_resp_fc_id;
|
|
|
|
// track if the accumulated message has any reasoning content
|
|
anthropic_has_reasoning = !state.chat_msg.reasoning_content.empty();
|
|
|
|
if (res_type == TASK_RESPONSE_TYPE_OAI_RESP && !state.oai_resp_created && (is_progress || n_decoded == 1)) {
|
|
state.oai_resp_created = true;
|
|
}
|
|
|
|
// Pre-compute state updates based on diffs (for next chunk)
|
|
for (const common_chat_msg_diff & diff : oaicompat_msg_diffs) {
|
|
if (!diff.reasoning_content_delta.empty() && !state.thinking_block_started) {
|
|
state.thinking_block_started = true;
|
|
}
|
|
if (!diff.content_delta.empty() && !state.text_block_started) {
|
|
state.text_block_started = true;
|
|
}
|
|
if (!diff.tool_call_delta.name.empty()) {
|
|
state.oai_resp_fc_id = diff.tool_call_delta.id;
|
|
}
|
|
}
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json() {
|
|
GGML_ASSERT(is_updated && "update() must be called before to_json()");
|
|
if (is_begin) {
|
|
return nullptr; // simply signal to HTTP handler to send the headers and status code
|
|
}
|
|
switch (res_type) {
|
|
case TASK_RESPONSE_TYPE_NONE:
|
|
return to_json_non_oaicompat();
|
|
case TASK_RESPONSE_TYPE_OAI_CMPL:
|
|
return to_json_oaicompat();
|
|
case TASK_RESPONSE_TYPE_OAI_CHAT:
|
|
return to_json_oaicompat_chat();
|
|
case TASK_RESPONSE_TYPE_OAI_RESP:
|
|
return to_json_oaicompat_resp();
|
|
case TASK_RESPONSE_TYPE_OAI_ASR:
|
|
return to_json_oaicompat_asr();
|
|
case TASK_RESPONSE_TYPE_ANTHROPIC:
|
|
return to_json_anthropic();
|
|
default:
|
|
GGML_ASSERT(false && "Invalid task_response_type");
|
|
}
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json_non_oaicompat() {
|
|
// non-OAI-compat JSON
|
|
json res = json {
|
|
{"index", index},
|
|
{"content", content.text},
|
|
{"tokens", tokens},
|
|
{"stop", false},
|
|
{"id_slot", id_slot},
|
|
{"tokens_predicted", n_decoded},
|
|
{"tokens_evaluated", n_prompt_tokens},
|
|
};
|
|
// populate the timings object when needed (usually for the last response or with timings_per_token enabled)
|
|
if (stats.is_set()) {
|
|
res["timings"] = stats.to_json();
|
|
}
|
|
if (is_progress) {
|
|
res["prompt_progress"] = progress.to_json();
|
|
}
|
|
if (!prob_output.probs.empty()) {
|
|
res["completion_probabilities"] = completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs);
|
|
}
|
|
return res;
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json_oaicompat() {
|
|
std::time_t t = std::time(0);
|
|
json logprobs = json(nullptr); // OAI default to null
|
|
if (prob_output.probs.size() > 0) {
|
|
logprobs = json{
|
|
{"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
|
|
};
|
|
}
|
|
json res = json {
|
|
{"choices", json::array({
|
|
json{
|
|
{"text", content.text},
|
|
{"index", index},
|
|
{"logprobs", logprobs},
|
|
{"finish_reason", nullptr},
|
|
}
|
|
})},
|
|
{"created", t},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "text_completion"},
|
|
{"id", oaicompat_cmpl_id}
|
|
};
|
|
|
|
// extra fields for debugging purposes
|
|
if (verbose) {
|
|
res["__verbose"] = to_json_non_oaicompat();
|
|
}
|
|
if (stats.is_set()) {
|
|
res["timings"] = stats.to_json();
|
|
}
|
|
if (is_progress) {
|
|
res["prompt_progress"] = progress.to_json();
|
|
}
|
|
|
|
return res;
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
|
|
bool first = n_decoded == 1;
|
|
std::time_t t = std::time(0);
|
|
json choices;
|
|
|
|
std::vector<json> deltas;
|
|
auto add_delta = [&](const json & delta) {
|
|
deltas.push_back({
|
|
{"choices", json::array({
|
|
json {
|
|
{"finish_reason", nullptr},
|
|
{"index", index},
|
|
{"delta", delta},
|
|
},
|
|
})},
|
|
{"created", t},
|
|
{"id", oaicompat_cmpl_id},
|
|
{"model", oaicompat_model},
|
|
{"system_fingerprint", std::string(llama_build_info())},
|
|
{"object", "chat.completion.chunk"},
|
|
});
|
|
};
|
|
// We have to send an initial update to conform to openai behavior
|
|
if (first || is_progress) {
|
|
add_delta({
|
|
{"role", "assistant"},
|
|
{"content", nullptr},
|
|
});
|
|
}
|
|
|
|
for (const auto & diff : oaicompat_msg_diffs) {
|
|
add_delta(server_chat_msg_diff_to_json_oaicompat(diff));
|
|
}
|
|
|
|
if (!deltas.empty()) {
|
|
auto & last_json = deltas[deltas.size() - 1];
|
|
GGML_ASSERT(last_json.at("choices").size() >= 1);
|
|
|
|
if (prob_output.probs.size() > 0) {
|
|
last_json.at("choices").at(0)["logprobs"] = json {
|
|
{"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
|
|
};
|
|
}
|
|
|
|
if (stats.is_set()) {
|
|
last_json["timings"] = stats.to_json();
|
|
}
|
|
if (is_progress) {
|
|
last_json["prompt_progress"] = progress.to_json();
|
|
}
|
|
}
|
|
|
|
return deltas;
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
|
|
std::vector<json> events;
|
|
|
|
if (!oai_resp_created) {
|
|
events.push_back(json {
|
|
{"event", "response.created"},
|
|
{"data", json {
|
|
{"type", "response.created"},
|
|
{"response", json {
|
|
{"id", oai_resp_id},
|
|
{"object", "response"},
|
|
{"status", "in_progress"},
|
|
}},
|
|
}},
|
|
});
|
|
events.push_back(json {
|
|
{"event", "response.in_progress"},
|
|
{"data", json {
|
|
{"type", "response.in_progress"},
|
|
{"response", json {
|
|
{"id", oai_resp_id},
|
|
{"object", "response"},
|
|
{"status", "in_progress"},
|
|
}},
|
|
}},
|
|
});
|
|
} else if (is_progress) {
|
|
events.push_back(json {
|
|
{"event", "response.in_progress"},
|
|
{"data", json {
|
|
{"type", "response.in_progress"},
|
|
{"response", json {
|
|
{"id", oai_resp_id},
|
|
{"object", "response"},
|
|
{"status", "in_progress"},
|
|
}},
|
|
}},
|
|
});
|
|
}
|
|
|
|
for (const common_chat_msg_diff & diff : oaicompat_msg_diffs) {
|
|
if (!diff.reasoning_content_delta.empty()) {
|
|
if (!thinking_block_started) {
|
|
events.push_back(json {
|
|
{"event", "response.output_item.added"},
|
|
{"data", json {
|
|
{"type", "response.output_item.added"},
|
|
{"output_index", 0},
|
|
{"item", json {
|
|
{"id", oai_resp_reasoning_id},
|
|
{"summary", json::array()},
|
|
{"type", "reasoning"},
|
|
{"content", json::array()},
|
|
{"encrypted_content", ""},
|
|
{"status", "in_progress"},
|
|
}},
|
|
}},
|
|
});
|
|
thinking_block_started = true;
|
|
}
|
|
events.push_back(json {
|
|
{"event", "response.reasoning_text.delta"},
|
|
{"data", json {
|
|
{"type", "response.reasoning_text.delta"},
|
|
{"output_index", 0},
|
|
{"content_index", 0},
|
|
{"delta", diff.reasoning_content_delta},
|
|
{"item_id", oai_resp_reasoning_id},
|
|
}},
|
|
});
|
|
}
|
|
|
|
if (!diff.content_delta.empty()) {
|
|
if (!text_block_started) {
|
|
events.push_back(json {
|
|
{"event", "response.output_item.added"},
|
|
{"data", json {
|
|
{"type", "response.output_item.added"},
|
|
{"output_index", anthropic_has_reasoning ? 1 : 0},
|
|
{"item", json {
|
|
{"content", json::array()},
|
|
{"id", oai_resp_message_id},
|
|
{"role", "assistant"},
|
|
{"status", "in_progress"},
|
|
{"type", "message"},
|
|
}},
|
|
}},
|
|
});
|
|
events.push_back(json {
|
|
{"event", "response.content_part.added"},
|
|
{"data", json {
|
|
{"type", "response.content_part.added"},
|
|
{"item_id", oai_resp_message_id},
|
|
{"output_index", anthropic_has_reasoning ? 1 : 0},
|
|
{"content_index", 0},
|
|
{"part", json {
|
|
{"type", "output_text"},
|
|
{"text", ""},
|
|
}},
|
|
}},
|
|
});
|
|
text_block_started = true;
|
|
}
|
|
events.push_back(json {
|
|
{"event", "response.output_text.delta"},
|
|
{"data", json {
|
|
{"type", "response.output_text.delta"},
|
|
{"item_id", oai_resp_message_id},
|
|
{"output_index", anthropic_has_reasoning ? 1 : 0},
|
|
{"content_index", 0},
|
|
{"delta", diff.content_delta},
|
|
}},
|
|
});
|
|
}
|
|
|
|
if (!diff.tool_call_delta.name.empty()) {
|
|
events.push_back(json {
|
|
{"event", "response.output_item.added"},
|
|
{"data", json {
|
|
{"type", "response.output_item.added"},
|
|
{"output_index", (anthropic_has_reasoning ? 1 : 0) + (text_block_started ? 1 : 0) + diff.tool_call_index},
|
|
{"item", json {
|
|
{"id", "fc_" + diff.tool_call_delta.id},
|
|
{"arguments", ""},
|
|
{"call_id", "call_" + diff.tool_call_delta.id},
|
|
{"name", diff.tool_call_delta.name},
|
|
{"type", "function_call"},
|
|
{"status", "in_progress"},
|
|
}},
|
|
}},
|
|
});
|
|
oai_resp_fc_id = diff.tool_call_delta.id;
|
|
}
|
|
|
|
if (!diff.tool_call_delta.arguments.empty()) {
|
|
events.push_back(json {
|
|
{"event", "response.function_call_arguments.delta"},
|
|
{"data", json {
|
|
{"type", "response.function_call_arguments.delta"},
|
|
{"delta", diff.tool_call_delta.arguments},
|
|
{"item_id", "fc_" + oai_resp_fc_id},
|
|
{"output_index", (anthropic_has_reasoning ? 1 : 0) + (text_block_started ? 1 : 0) + diff.tool_call_index},
|
|
}},
|
|
});
|
|
}
|
|
}
|
|
|
|
if (!events.empty()) {
|
|
json & data = events.back().at("data");
|
|
if (stats.is_set()) {
|
|
data["timings"] = stats.to_json();
|
|
}
|
|
if (is_progress) {
|
|
data["prompt_progress"] = progress.to_json();
|
|
}
|
|
}
|
|
|
|
return events;
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json_oaicompat_asr() {
|
|
json event = json {
|
|
{"type", "transcript.text.delta"},
|
|
{"delta", content.text},
|
|
};
|
|
return event;
|
|
}
|
|
|
|
json server_task_result_cmpl_partial::to_json_anthropic() {
|
|
json events = json::array();
|
|
bool first = (n_decoded == 1);
|
|
// use member variables to track block state across streaming calls
|
|
// (anthropic_thinking_block_started, anthropic_text_block_started)
|
|
|
|
if (first) {
|
|
events.push_back({
|
|
{"event", "message_start"},
|
|
{"data", {
|
|
{"type", "message_start"},
|
|
{"message", {
|
|
{"id", oaicompat_cmpl_id},
|
|
{"type", "message"},
|
|
{"role", "assistant"},
|
|
{"content", json::array()},
|
|
{"model", oaicompat_model},
|
|
{"stop_reason", nullptr},
|
|
{"stop_sequence", nullptr},
|
|
{"usage", {
|
|
{"cache_read_input_tokens", n_prompt_tokens_cache},
|
|
{"input_tokens", n_prompt_tokens - n_prompt_tokens_cache},
|
|
{"output_tokens", 0}
|
|
}}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
|
|
// content block indices: thinking (0) -> text (0 or 1) -> tool_use (n+)
|
|
size_t thinking_block_index = 0;
|
|
// use anthropic_has_reasoning (set in update()) to know if ANY reasoning was generated
|
|
size_t text_block_index = anthropic_has_reasoning ? 1 : 0;
|
|
|
|
// use local copies of streaming state (copied from task_result_state in update())
|
|
// these reflect the state BEFORE this chunk was processed
|
|
bool thinking_started = thinking_block_started;
|
|
bool text_started = text_block_started;
|
|
|
|
for (const auto & diff : oaicompat_msg_diffs) {
|
|
// handle thinking/reasoning content
|
|
if (!diff.reasoning_content_delta.empty()) {
|
|
if (!thinking_started) {
|
|
events.push_back({
|
|
{"event", "content_block_start"},
|
|
{"data", {
|
|
{"type", "content_block_start"},
|
|
{"index", thinking_block_index},
|
|
{"content_block", {
|
|
{"type", "thinking"},
|
|
{"thinking", ""}
|
|
}}
|
|
}}
|
|
});
|
|
thinking_started = true;
|
|
}
|
|
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", thinking_block_index},
|
|
{"delta", {
|
|
{"type", "thinking_delta"},
|
|
{"thinking", diff.reasoning_content_delta}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
|
|
// handle regular text content
|
|
if (!diff.content_delta.empty()) {
|
|
if (!text_started) {
|
|
events.push_back({
|
|
{"event", "content_block_start"},
|
|
{"data", {
|
|
{"type", "content_block_start"},
|
|
{"index", text_block_index},
|
|
{"content_block", {
|
|
{"type", "text"},
|
|
{"text", ""}
|
|
}}
|
|
}}
|
|
});
|
|
text_started = true;
|
|
}
|
|
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", text_block_index},
|
|
{"delta", {
|
|
{"type", "text_delta"},
|
|
{"text", diff.content_delta}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
|
|
// handle tool calls
|
|
if (diff.tool_call_index != std::string::npos) {
|
|
// use anthropic_has_reasoning for thinking block count (persists across calls)
|
|
size_t content_block_index = (anthropic_has_reasoning ? 1 : 0) + (text_started ? 1 : 0) + diff.tool_call_index;
|
|
|
|
if (!diff.tool_call_delta.name.empty()) {
|
|
events.push_back({
|
|
{"event", "content_block_start"},
|
|
{"data", {
|
|
{"type", "content_block_start"},
|
|
{"index", content_block_index},
|
|
{"content_block", {
|
|
{"type", "tool_use"},
|
|
{"id", diff.tool_call_delta.id},
|
|
{"name", diff.tool_call_delta.name}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
|
|
if (!diff.tool_call_delta.arguments.empty()) {
|
|
events.push_back({
|
|
{"event", "content_block_delta"},
|
|
{"data", {
|
|
{"type", "content_block_delta"},
|
|
{"index", content_block_index},
|
|
{"delta", {
|
|
{"type", "input_json_delta"},
|
|
{"partial_json", diff.tool_call_delta.arguments}
|
|
}}
|
|
}}
|
|
});
|
|
}
|
|
}
|
|
}
|
|
|
|
return events;
|
|
}
|
|
|
|
//
|
|
// server_task_result_embd
|
|
//
|
|
json server_task_result_embd::to_json() {
|
|
return res_type == TASK_RESPONSE_TYPE_OAI_EMBD
|
|
? to_json_oaicompat()
|
|
: to_json_non_oaicompat();
|
|
}
|
|
|
|
json server_task_result_embd::to_json_non_oaicompat() {
|
|
return json {
|
|
{"index", index},
|
|
{"embedding", embedding},
|
|
};
|
|
}
|
|
|
|
json server_task_result_embd::to_json_oaicompat() {
|
|
return json {
|
|
{"index", index},
|
|
{"embedding", embedding[0]},
|
|
{"tokens_evaluated", n_tokens},
|
|
};
|
|
}
|
|
|
|
//
|
|
// server_task_result_rerank
|
|
//
|
|
json server_task_result_rerank::to_json() {
|
|
return json {
|
|
{"index", index},
|
|
{"score", score},
|
|
{"tokens_evaluated", n_tokens},
|
|
};
|
|
}
|
|
|
|
//
|
|
// server_task_result_tts
|
|
//
|
|
json server_task_result_tts::to_json() {
|
|
return json {
|
|
{"sample_rate", sample_rate},
|
|
{"n_bytes", audio.size()},
|
|
{"final", final},
|
|
};
|
|
}
|
|
|
|
//
|
|
// server_task_result_decision
|
|
//
|
|
json server_task_result_decision::to_json() {
|
|
return json {
|
|
{"index", index},
|
|
{"scores", scores},
|
|
{"tokens_evaluated", n_tokens},
|
|
};
|
|
}
|
|
|
|
//
|
|
// server_task_result_error
|
|
//
|
|
json server_task_result_error::to_json() {
|
|
json res = format_error_response(err_msg, err_type);
|
|
if (err_type == ERROR_TYPE_EXCEED_CONTEXT_SIZE) {
|
|
res["n_prompt_tokens"] = n_prompt_tokens;
|
|
res["n_ctx"] = n_ctx;
|
|
}
|
|
return res;
|
|
}
|
|
|
|
//
|
|
// server_task_result_metrics
|
|
//
|
|
json server_task_result_slots::to_json() {
|
|
return slots_data;
|
|
}
|
|
|
|
json server_task_result_metrics::to_json() {
|
|
// not used, /metrics renders prometheus text via to_metrics()
|
|
return json{};
|
|
}
|
|
|
|
// metrics definition: https://prometheus.io/docs/practices/naming/#metric-names
|
|
std::string server_task_result_metrics::to_metrics() {
|
|
const std::vector<metric_item> counters = {
|
|
{
|
|
"prompt_tokens_total",
|
|
"Number of prompt tokens processed, excluding cached tokens",
|
|
(double) metrics.prompt.count
|
|
}, {
|
|
"prompt_tokens_cached_total",
|
|
"Number of prompt tokens reused from the cache",
|
|
(double) metrics.n_prompt_cached
|
|
}, {
|
|
"prompt_seconds_total",
|
|
"Total time spent processing prompts",
|
|
metrics.prompt.time / 1.e6
|
|
}, {
|
|
"tokens_predicted_total",
|
|
"Number of generation tokens processed",
|
|
(double) metrics.predict.count
|
|
}, {
|
|
"tokens_predicted_seconds_total",
|
|
"Total time spent generating tokens",
|
|
metrics.predict.time / 1.e6
|
|
}, {
|
|
"n_decode_total",
|
|
"Total number of llama_decode() calls, excluding speculative decoding and multimodal decoding",
|
|
(double) metrics.n_decode
|
|
}, {
|
|
"n_tokens_max",
|
|
"Largest observed sequence length (prompt + generation)",
|
|
(double) metrics.n_tokens_max
|
|
}, {
|
|
"spec_decode_num_draft_tokens_total",
|
|
"Speculative: Total draft tokens generated",
|
|
(double) metrics.n_draft_tokens
|
|
}, {
|
|
"spec_decode_num_accepted_tokens_total",
|
|
"Speculative: Total draft tokens accepted by the target model",
|
|
(double) metrics.n_draft_accepted
|
|
}, {
|
|
"spec_decode_num_drafts_total",
|
|
"Speculative: Total speculative decoding verification steps",
|
|
(double) metrics.n_draft_verif_steps
|
|
},
|
|
};
|
|
|
|
const std::vector<metric_item> gauges = {
|
|
{
|
|
"prompt_tokens_seconds",
|
|
"Average prompt throughput in tokens/s",
|
|
metrics.prompt_bucket.n_per_second()
|
|
}, {
|
|
"predicted_tokens_seconds",
|
|
"Average generation throughput in tokens/s",
|
|
metrics.predict_bucket.n_per_second()
|
|
}, {
|
|
"requests_processing",
|
|
"Number of requests processing",
|
|
(double) n_processing_slots
|
|
}, {
|
|
"requests_deferred",
|
|
"Number of requests deferred",
|
|
(double) n_tasks_deferred
|
|
}, {
|
|
"n_busy_slots_per_decode",
|
|
"Average number of busy slots per llama_decode() call",
|
|
(double) metrics.n_busy_slots / std::max((double) metrics.n_decode, 1.0)
|
|
},
|
|
};
|
|
|
|
std::stringstream prometheus;
|
|
|
|
auto add_items = [&prometheus](const char * type, const std::vector<metric_item> & items) {
|
|
for (const auto & item : items) {
|
|
prometheus << "# HELP llamacpp:" << item.name << " " << item.description << "\n"
|
|
<< "# TYPE llamacpp:" << item.name << " " << type << "\n"
|
|
<< "llamacpp:" << item.name << " " << item.value << "\n";
|
|
}
|
|
};
|
|
|
|
add_items("counter", counters);
|
|
add_items("gauge", gauges);
|
|
|
|
// labeled counter: one time series per draft position
|
|
if (!metrics.n_accepted_per_pos.empty()) {
|
|
prometheus << "# HELP llamacpp:spec_decode_num_accepted_tokens_per_pos_total"
|
|
" Accepted tokens per draft position\n"
|
|
<< "# TYPE llamacpp:spec_decode_num_accepted_tokens_per_pos_total counter\n";
|
|
for (size_t i = 0; i < metrics.n_accepted_per_pos.size(); i++) {
|
|
prometheus << "llamacpp:spec_decode_num_accepted_tokens_per_pos_total{position=\""
|
|
<< i << "\"} " << metrics.n_accepted_per_pos[i] << "\n";
|
|
}
|
|
}
|
|
|
|
return prometheus.str();
|
|
}
|
|
|
|
//
|
|
// server_task_result_slot_save_load
|
|
//
|
|
json server_task_result_slot_save_load::to_json() {
|
|
if (is_save) {
|
|
return json {
|
|
{ "id_slot", id_slot },
|
|
{ "filename", filename },
|
|
{ "n_saved", n_tokens },
|
|
{ "n_written", n_bytes },
|
|
{ "timings", {
|
|
{ "save_ms", t_ms }
|
|
}},
|
|
};
|
|
}
|
|
|
|
return json {
|
|
{ "id_slot", id_slot },
|
|
{ "filename", filename },
|
|
{ "n_restored", n_tokens },
|
|
{ "n_read", n_bytes },
|
|
{ "timings", {
|
|
{ "restore_ms", t_ms }
|
|
}},
|
|
};
|
|
}
|
|
|
|
//
|
|
// server_task_result_slot_erase
|
|
//
|
|
json server_task_result_slot_erase::to_json() {
|
|
return json {
|
|
{ "id_slot", id_slot },
|
|
{ "n_erased", n_erased },
|
|
};
|
|
}
|
|
|
|
//
|
|
// server_task_result_get_lora
|
|
//
|
|
|
|
json server_task_result_get_lora::to_json() {
|
|
json result = json::array();
|
|
for (size_t i = 0; i < loras.size(); ++i) {
|
|
auto & lora = loras[i];
|
|
json entry = {
|
|
{"id", i},
|
|
{"path", lora.info.path},
|
|
{"scale", lora.info.scale},
|
|
{"task_name", lora.info.task_name},
|
|
{"prompt_prefix", lora.info.prompt_prefix},
|
|
};
|
|
if (!lora.alora_invocation_tokens.empty()) {
|
|
entry["alora_invocation_string"] = lora.alora_invocation_string;
|
|
entry["alora_invocation_tokens"] = lora.alora_invocation_tokens;
|
|
}
|
|
result.push_back(std::move(entry));
|
|
}
|
|
return result;
|
|
}
|
|
|
|
//
|
|
// server_task_result_apply_lora
|
|
//
|
|
|
|
json server_task_result_apply_lora::to_json() {
|
|
return json {{ "success", true }};
|
|
}
|
|
|
|
//
|
|
// server_prompt_cache
|
|
//
|
|
size_t server_prompt_cache::size() const {
|
|
size_t res = 0;
|
|
|
|
for (const auto & state : states) {
|
|
res += state.size();
|
|
}
|
|
|
|
return res;
|
|
}
|
|
|
|
size_t server_prompt_cache::n_tokens() const {
|
|
size_t res = 0;
|
|
|
|
for (const auto & state : states) {
|
|
res += state.prompt.n_tokens();
|
|
}
|
|
|
|
return res;
|
|
}
|
|
|
|
server_prompt_cache_state * server_prompt_cache::alloc(const server_prompt & prompt, size_t state_size_tgt, size_t state_size_dft) {
|
|
// first check if the current state is contained fully in the cache
|
|
for (auto it = states.begin(); it != states.end(); ++it) {
|
|
const int cur_lcp_len = it->prompt.tokens.get_common_prefix(prompt.tokens);
|
|
|
|
if (cur_lcp_len == (int) prompt.tokens.size()) {
|
|
SRV_TRC("%s", " - prompt is already in the cache, skipping\n");
|
|
return nullptr;
|
|
}
|
|
}
|
|
|
|
// calculate checkpoints size to see if it will fit with the prompt
|
|
size_t checkpoints_size = 0;
|
|
for (const auto & ckpt : prompt.checkpoints) {
|
|
checkpoints_size += ckpt.size();
|
|
}
|
|
|
|
const size_t state_size_new = state_size_tgt + state_size_dft + checkpoints_size;
|
|
|
|
// skip over-limit entries to avoid disturbing the cache
|
|
if (limit_size > 0 && state_size_new > limit_size) {
|
|
SRV_WRN(" - prompt state size %.3f MiB exceeds cache size limit %.3f MiB, skipping\n",
|
|
state_size_new / (1024.0 * 1024.0), limit_size / (1024.0 * 1024.0));
|
|
return nullptr;
|
|
}
|
|
|
|
// remove any cached prompts that are fully contained in the current prompt
|
|
for (auto it = states.begin(); it != states.end();) {
|
|
const int len = it->prompt.tokens.get_common_prefix(prompt.tokens);
|
|
|
|
if (len == (int) it->prompt.tokens.size()) {
|
|
SRV_TRC(" - removing obsolete cached prompt with length %d\n", len);
|
|
|
|
it = states.erase(it);
|
|
} else {
|
|
++it;
|
|
}
|
|
}
|
|
|
|
if (limit_size > 0) {
|
|
// make room before allocating the new vectors to avoid breaching the limit
|
|
while (!states.empty() && size() + state_size_new > limit_size) {
|
|
SRV_WRN(" - making room for prompt cache entry, removing oldest entry (size = %.3f MiB)\n",
|
|
states.front().size() / (1024.0 * 1024.0));
|
|
|
|
states.pop_front();
|
|
}
|
|
}
|
|
|
|
std::vector<uint8_t> state_data_tgt;
|
|
std::vector<uint8_t> state_data_dft;
|
|
|
|
// check if we can allocate enough memory for the new state
|
|
try {
|
|
state_data_tgt.resize(state_size_tgt);
|
|
state_data_dft.resize(state_size_dft);
|
|
} catch (const std::bad_alloc & e) {
|
|
SRV_ERR("failed to allocate memory for prompt cache state: %s\n", e.what());
|
|
|
|
limit_size = std::max<size_t>(1, 0.4*size());
|
|
|
|
SRV_WRN(" - cache size limit reduced to %.3f MiB\n", limit_size / (1024.0 * 1024.0));
|
|
|
|
update();
|
|
|
|
return nullptr;
|
|
}
|
|
|
|
states.push_back({
|
|
/*.prompt =*/ {
|
|
/*.tokens =*/ prompt.tokens.clone(),
|
|
/*.checkpoints =*/ prompt.checkpoints,
|
|
},
|
|
/*.data =*/ {
|
|
/*.main =*/ std::move(state_data_tgt),
|
|
/*.drft =*/ std::move(state_data_dft),
|
|
},
|
|
});
|
|
|
|
return &states.back();
|
|
}
|
|
|
|
bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tokens_new, llama_context * ctx_tgt, llama_context * ctx_dft, int32_t id_slot) {
|
|
const int lcp_best = prompt.tokens.get_common_prefix(tokens_new);
|
|
|
|
float f_keep_best = prompt.tokens.size() > 0 ? float(lcp_best) / prompt.tokens.size() : -1.0f; // empty slot: any cache entry wins
|
|
float f_sim_best = float(lcp_best) / tokens_new.size();
|
|
|
|
SRV_TRC(" - looking for better prompt, base f_keep = %.3f, f_sim = %.3f\n", f_keep_best, f_sim_best);
|
|
|
|
auto it_best = states.end();
|
|
|
|
// find the most similar cached prompt, that would also preserve the most context
|
|
for (auto it = states.begin(); it != states.end(); ++it) {
|
|
const int lcp_cur = it->prompt.tokens.get_common_prefix(tokens_new);
|
|
|
|
const float f_keep_cur = float(lcp_cur) / it->prompt.tokens.size();
|
|
const float f_sim_cur = float(lcp_cur) / tokens_new.size();
|
|
|
|
SRV_TRC(" - prompt with length %7zu, lcp = %7d, f_keep = %.3f, f_sim = %.3f\n", it->prompt.tokens.size(), lcp_cur, f_keep_cur, f_sim_cur);
|
|
|
|
// don't trash large prompts
|
|
if (f_keep_cur < 0.25f) {
|
|
continue;
|
|
}
|
|
|
|
if (f_keep_best < f_keep_cur && f_sim_best < f_sim_cur) {
|
|
f_keep_best = f_keep_cur;
|
|
f_sim_best = f_sim_cur;
|
|
|
|
it_best = it;
|
|
}
|
|
}
|
|
|
|
if (it_best != states.end()) {
|
|
SRV_TRC(" - found better prompt with f_keep = %.3f, f_sim = %.3f\n", f_keep_best, f_sim_best);
|
|
|
|
{
|
|
auto & data = it_best->data.main;
|
|
|
|
const size_t size = data.size();
|
|
const size_t n = llama_state_seq_set_data_ext(ctx_tgt, data.data(), size, id_slot, 0);
|
|
if (n != size) {
|
|
SRV_ERR("failed to restore state with size %zu\n", size);
|
|
|
|
return false;
|
|
}
|
|
|
|
data.clear();
|
|
data.shrink_to_fit();
|
|
}
|
|
|
|
{
|
|
auto & data = it_best->data.drft;
|
|
|
|
if (!data.empty()) {
|
|
GGML_ASSERT(ctx_dft);
|
|
|
|
const size_t size = data.size();
|
|
const size_t n = llama_state_seq_set_data_ext(ctx_dft, data.data(), size, id_slot, 0);
|
|
if (n != size) {
|
|
SRV_WRN("failed to restore state with size %zu\n", size);
|
|
|
|
return false;
|
|
}
|
|
|
|
data.clear();
|
|
data.shrink_to_fit();
|
|
}
|
|
}
|
|
|
|
prompt = std::move(it_best->prompt);
|
|
|
|
states.erase(it_best);
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
void server_prompt_cache::update() {
|
|
if (limit_size > 0) {
|
|
while (!states.empty() && size() > limit_size) {
|
|
SRV_WRN(" - cache size limit reached, removing oldest entry (size = %.3f MiB)\n", states.front().size() / (1024.0 * 1024.0));
|
|
|
|
states.pop_front();
|
|
}
|
|
}
|
|
|
|
// average size per token
|
|
const float size_per_token = std::max<float>(1.0f, float(size()) / (std::max<size_t>(1, n_tokens())));
|
|
|
|
// dynamically increase the token limit if it can fit in the memory limit
|
|
const size_t limit_tokens_cur = limit_size > 0 ? std::max<size_t>(limit_tokens, limit_size/size_per_token) : limit_tokens;
|
|
|
|
if (limit_tokens > 0) {
|
|
while (!states.empty() && n_tokens() > limit_tokens_cur) {
|
|
SRV_WRN(" - cache token limit (%zu, est: %zu) reached, removing oldest entry (size = %.3f MiB)\n",
|
|
limit_tokens, limit_tokens_cur, states.front().size() / (1024.0 * 1024.0));
|
|
|
|
states.pop_front();
|
|
}
|
|
}
|
|
|
|
SRV_TRC(" - cache state: %zu prompts, %.3f MiB (limits: %.3f MiB, %zu tokens, %zu est)\n",
|
|
states.size(), size() / (1024.0 * 1024.0), limit_size / (1024.0 * 1024.0), limit_tokens, limit_tokens_cur);
|
|
|
|
for (const auto & state : states) {
|
|
SRV_TRC(" - prompt %p: %7d tokens, checkpoints: %2zu, %9.3f MiB\n",
|
|
(const void *)&state, state.prompt.n_tokens(), state.prompt.checkpoints.size(), state.size() / (1024.0 * 1024.0));
|
|
}
|
|
}
|
|
|
|
//
|
|
// server_prompt_disk_cache
|
|
//
|
|
|
|
// file = header + payload; the payload is a list of sections, each padded to 8 bytes:
|
|
// tokens, checkpoint table, state size, state, the state of each checkpoint
|
|
// hash_payload is FNV-1a over the 64-bit words of the payload, so a truncated or corrupt file is never used
|
|
static constexpr uint32_t DISK_CACHE_MAGIC = 0x4443504c; // "LPCD"
|
|
static constexpr uint32_t DISK_CACHE_VERSION = 1;
|
|
static constexpr const char * DISK_CACHE_EXT = ".lpcd";
|
|
|
|
// keep this much space free on the file system
|
|
static constexpr uint64_t DISK_CACHE_MIN_FREE = 1024ull*1024ull*1024ull;
|
|
|
|
struct disk_cache_header {
|
|
uint32_t magic;
|
|
uint32_t version;
|
|
uint64_t key_hash;
|
|
uint64_t n_tokens;
|
|
uint64_t n_ckpt;
|
|
uint64_t size_payload;
|
|
uint64_t hash_payload;
|
|
};
|
|
|
|
struct disk_cache_ckpt {
|
|
int64_t n_tokens;
|
|
llama_pos pos_min;
|
|
llama_pos pos_max;
|
|
uint64_t size;
|
|
};
|
|
|
|
static_assert(sizeof(disk_cache_header) == 48 && sizeof(disk_cache_ckpt) == 24, "disk cache layout");
|
|
|
|
static uint64_t disk_cache_pad(uint64_t n) {
|
|
return (8 - n % 8) % 8;
|
|
}
|
|
|
|
static uint64_t disk_cache_hash_bytes(const void * data, size_t n) {
|
|
uint64_t h = 0xcbf29ce484222325ull;
|
|
for (size_t i = 0; i < n; i++) {
|
|
h = (h ^ ((const uint8_t *) data)[i]) * 0x100000001b3ull;
|
|
}
|
|
return h;
|
|
}
|
|
|
|
static uint64_t disk_cache_payload_size(uint64_t n_tokens, const std::vector<disk_cache_ckpt> & ckpts, uint64_t n_state) {
|
|
const auto padded = [](uint64_t n) { return n + disk_cache_pad(n); };
|
|
uint64_t res = padded(n_tokens*sizeof(llama_token)) + padded(ckpts.size()*sizeof(disk_cache_ckpt)) + sizeof(uint64_t) + padded(n_state);
|
|
for (const auto & c : ckpts) {
|
|
res += padded(c.size);
|
|
}
|
|
return res;
|
|
}
|
|
|
|
// sequential section IO with the running payload hash
|
|
struct disk_cache_io {
|
|
FILE * f = nullptr;
|
|
uint64_t hash = 0xcbf29ce484222325ull;
|
|
bool ok = true;
|
|
|
|
~disk_cache_io() {
|
|
if (f) {
|
|
fclose(f);
|
|
}
|
|
}
|
|
|
|
void add_hash(const uint8_t * p, size_t size, const uint8_t * pad) {
|
|
const size_t n_full = size - size % 8;
|
|
for (size_t i = 0; i < n_full; i += 8) {
|
|
uint64_t w;
|
|
memcpy(&w, p + i, 8);
|
|
hash = (hash ^ w) * 0x100000001b3ull;
|
|
}
|
|
if (size % 8) {
|
|
uint64_t w;
|
|
memcpy(&w, p + n_full, size % 8);
|
|
memcpy((uint8_t *) &w + size % 8, pad, disk_cache_pad(size));
|
|
hash = (hash ^ w) * 0x100000001b3ull;
|
|
}
|
|
}
|
|
|
|
void write(const void * src, size_t size) {
|
|
static const uint8_t zero[8] = {};
|
|
const size_t n_pad = disk_cache_pad(size);
|
|
ok = ok && fwrite(src, 1, size, f) == size && fwrite(zero, 1, n_pad, f) == n_pad;
|
|
add_hash((const uint8_t *) src, size, zero);
|
|
}
|
|
|
|
void read(void * dst, size_t size) {
|
|
uint8_t pad[8];
|
|
const size_t n_pad = disk_cache_pad(size);
|
|
ok = ok && fread(dst, 1, size, f) == size && fread(pad, 1, n_pad, f) == n_pad;
|
|
if (ok) {
|
|
add_hash((const uint8_t *) dst, size, pad);
|
|
}
|
|
}
|
|
|
|
// the data is in our buffers, the cached pages are waste (see llama_file::drop_cache)
|
|
void drop_cache() {
|
|
#if defined(__linux__)
|
|
posix_fadvise(fileno(f), 0, 0, POSIX_FADV_DONTNEED);
|
|
#endif
|
|
}
|
|
};
|
|
|
|
// open a file and read everything up to the state data: 1 = entry of this key, 0 = other key or version, -1 = bad file
|
|
static int disk_cache_open(const std::string & path, uint64_t key_hash, disk_cache_io & io, disk_cache_header & hdr,
|
|
llama_tokens & tokens, std::vector<disk_cache_ckpt> & ckpts, uint64_t & n_state) {
|
|
std::error_code ec;
|
|
const uint64_t size = std::filesystem::file_size(path, ec);
|
|
io.f = ec ? nullptr : fopen(path.c_str(), "rb");
|
|
if (io.f == nullptr || fread(&hdr, 1, sizeof(hdr), io.f) != sizeof(hdr) || hdr.magic != DISK_CACHE_MAGIC) {
|
|
return -1;
|
|
}
|
|
if (hdr.version != DISK_CACHE_VERSION) {
|
|
return 0;
|
|
}
|
|
if (size != sizeof(hdr) + hdr.size_payload || hdr.n_tokens > size/sizeof(llama_token) || hdr.n_ckpt > size/sizeof(disk_cache_ckpt)) {
|
|
return -1;
|
|
}
|
|
if (hdr.key_hash != key_hash) {
|
|
return 0;
|
|
}
|
|
tokens.resize(hdr.n_tokens);
|
|
ckpts.resize(hdr.n_ckpt);
|
|
io.read(tokens.data(), tokens.size()*sizeof(llama_token));
|
|
io.read(ckpts.data(), ckpts.size()*sizeof(disk_cache_ckpt));
|
|
io.read(&n_state, sizeof(n_state));
|
|
if (!io.ok || n_state > size) {
|
|
return -1;
|
|
}
|
|
for (const auto & c : ckpts) {
|
|
if (c.size > size) {
|
|
return -1;
|
|
}
|
|
}
|
|
return disk_cache_payload_size(hdr.n_tokens, ckpts, n_state) == hdr.size_payload ? 1 : -1;
|
|
}
|
|
|
|
static void disk_cache_touch(const std::string & path) {
|
|
std::error_code ec;
|
|
std::filesystem::last_write_time(path, std::filesystem::file_time_type::clock::now(), ec);
|
|
}
|
|
|
|
server_prompt_disk_cache::server_prompt_disk_cache(const std::string & dir, int32_t limit_size_mib, const std::string & key) : dir(dir) {
|
|
key_hash = disk_cache_hash_bytes(key.data(), key.size());
|
|
limit_size = 1024ull*1024ull*std::max(0, limit_size_mib);
|
|
|
|
size_t n_other = 0;
|
|
size_t n_bad = 0;
|
|
try {
|
|
std::filesystem::create_directories(dir);
|
|
const auto now = std::filesystem::file_time_type::clock::now();
|
|
for (const auto & de : std::filesystem::directory_iterator(dir)) {
|
|
const std::string path = de.path().string();
|
|
if (!de.is_regular_file()) {
|
|
continue;
|
|
}
|
|
// left by a writer that was killed
|
|
if (path.find(std::string(DISK_CACHE_EXT) + ".tmp") != std::string::npos) {
|
|
if (now - de.last_write_time() > std::chrono::hours(1)) {
|
|
std::filesystem::remove(de.path());
|
|
}
|
|
continue;
|
|
}
|
|
if (de.path().extension() != DISK_CACHE_EXT) {
|
|
continue;
|
|
}
|
|
disk_cache_io io;
|
|
disk_cache_header hdr;
|
|
server_prompt_disk_entry e;
|
|
std::vector<disk_cache_ckpt> ckpts;
|
|
uint64_t n_state = 0;
|
|
const int res = disk_cache_open(path, key_hash, io, hdr, e.tokens, ckpts, n_state);
|
|
if (res < 0) {
|
|
SRV_WRN("disk cache: '%s' is truncated or corrupt, removing it\n", path.c_str());
|
|
std::filesystem::remove(de.path());
|
|
n_bad++;
|
|
continue;
|
|
}
|
|
if (res == 0) {
|
|
n_other++;
|
|
continue;
|
|
}
|
|
e.path = path;
|
|
for (const auto & c : ckpts) {
|
|
e.ckpt_n_tokens.push_back(c.n_tokens);
|
|
}
|
|
entries.push_back(std::move(e));
|
|
}
|
|
} catch (const std::exception & err) {
|
|
SRV_WRN("disk cache: error reading '%s': %s\n", dir.c_str(), err.what());
|
|
}
|
|
|
|
SRV_INF("disk cache: '%s', %zu prompts of this model, %zu other files, %zu bad files removed, limit %.0f MiB\n",
|
|
dir.c_str(), entries.size(), n_other, n_bad, limit_size / (1024.0 * 1024.0));
|
|
}
|
|
|
|
size_t server_prompt_disk_cache::n_reuse(const llama_tokens & tokens, const std::vector<int64_t> & ckpt_n_tokens, const llama_tokens & tokens_new, bool can_trim) {
|
|
size_t lcp = 0;
|
|
while (lcp < tokens.size() && lcp < tokens_new.size() && tokens[lcp] == tokens_new[lcp]) {
|
|
lcp++;
|
|
}
|
|
if (lcp == tokens.size() || can_trim) {
|
|
return lcp;
|
|
}
|
|
size_t res = 0;
|
|
for (const int64_t n : ckpt_n_tokens) {
|
|
if (n <= (int64_t) lcp) {
|
|
res = std::max(res, (size_t) n);
|
|
}
|
|
}
|
|
return res;
|
|
}
|
|
|
|
int server_prompt_disk_cache::find(const llama_tokens & tokens_new, size_t n_have, size_t n_min, bool can_trim) const {
|
|
int best = -1;
|
|
size_t n_best = 0;
|
|
for (size_t i = 0; i < entries.size(); i++) {
|
|
const size_t n = n_reuse(entries[i].tokens, entries[i].ckpt_n_tokens, tokens_new, can_trim);
|
|
if (n >= n_have + n_min && n > n_best) {
|
|
best = (int) i;
|
|
n_best = n;
|
|
}
|
|
}
|
|
return best;
|
|
}
|
|
|
|
bool server_prompt_disk_cache::has(const llama_tokens & tokens) {
|
|
for (const auto & e : entries) {
|
|
std::error_code ec;
|
|
if (e.tokens == tokens && std::filesystem::exists(e.path, ec)) {
|
|
disk_cache_touch(e.path);
|
|
return true;
|
|
}
|
|
}
|
|
return false;
|
|
}
|
|
|
|
void server_prompt_disk_cache::remove(int i) {
|
|
std::error_code ec;
|
|
std::filesystem::remove(entries[i].path, ec);
|
|
entries.erase(entries.begin() + i);
|
|
}
|
|
|
|
bool server_prompt_disk_cache::load(int i, server_prompt & prompt, std::vector<uint8_t> & data, bool has_mtmd) {
|
|
const std::string path = entries[i].path;
|
|
|
|
disk_cache_io io;
|
|
disk_cache_header hdr;
|
|
llama_tokens tokens;
|
|
std::vector<disk_cache_ckpt> ckpts;
|
|
uint64_t n_state = 0;
|
|
std::list<common_prompt_checkpoint> checkpoints;
|
|
|
|
bool ok = disk_cache_open(path, key_hash, io, hdr, tokens, ckpts, n_state) == 1 && tokens == entries[i].tokens;
|
|
if (ok) {
|
|
try {
|
|
data.resize(n_state);
|
|
io.read(data.data(), data.size());
|
|
for (const auto & c : ckpts) {
|
|
auto & cur = checkpoints.emplace_back();
|
|
cur.id_task = -1; // not created by a task: may fail to load, then the prompt is re-processed
|
|
cur.n_tokens = c.n_tokens;
|
|
cur.pos_min = c.pos_min;
|
|
cur.pos_max = c.pos_max;
|
|
cur.data_tgt.resize(c.size);
|
|
io.read(cur.data_tgt.data(), cur.data_tgt.size());
|
|
}
|
|
} catch (const std::bad_alloc &) {
|
|
SRV_WRN("disk cache: out of memory reading '%s'\n", path.c_str());
|
|
data.clear();
|
|
return false;
|
|
}
|
|
ok = io.ok && io.hash == hdr.hash_payload;
|
|
io.drop_cache();
|
|
}
|
|
if (!ok) {
|
|
SRV_WRN("disk cache: '%s' is missing, truncated or corrupt, removing it\n", path.c_str());
|
|
data.clear();
|
|
remove(i);
|
|
return false;
|
|
}
|
|
|
|
disk_cache_touch(path);
|
|
|
|
prompt.tokens = server_tokens(tokens, has_mtmd);
|
|
prompt.checkpoints = std::move(checkpoints);
|
|
|
|
return true;
|
|
}
|
|
|
|
bool server_prompt_disk_cache::save(const server_prompt & prompt, const std::vector<uint8_t> & data) {
|
|
const llama_tokens & tokens = prompt.tokens.get_tokens();
|
|
|
|
std::vector<disk_cache_ckpt> ckpts;
|
|
for (const auto & c : prompt.checkpoints) {
|
|
ckpts.push_back({ c.n_tokens, c.pos_min, c.pos_max, c.data_tgt.size() });
|
|
}
|
|
|
|
disk_cache_header hdr = {
|
|
/*.magic =*/ DISK_CACHE_MAGIC,
|
|
/*.version =*/ DISK_CACHE_VERSION,
|
|
/*.key_hash =*/ key_hash,
|
|
/*.n_tokens =*/ tokens.size(),
|
|
/*.n_ckpt =*/ ckpts.size(),
|
|
/*.size_payload =*/ disk_cache_payload_size(tokens.size(), ckpts, data.size()),
|
|
/*.hash_payload =*/ 0,
|
|
};
|
|
const uint64_t size_file = sizeof(hdr) + hdr.size_payload;
|
|
|
|
char name[64];
|
|
snprintf(name, sizeof(name), "%016llx-%016llx", (unsigned long long) key_hash,
|
|
(unsigned long long) disk_cache_hash_bytes(tokens.data(), tokens.size()*sizeof(llama_token)));
|
|
const std::string path = (std::filesystem::path(dir) / name).string() + DISK_CACHE_EXT;
|
|
const std::string path_tmp = path + ".tmp";
|
|
|
|
// entries contained in the new one are obsolete: removed once it is written, not counted for the limit
|
|
const auto is_obsolete = [&](const server_prompt_disk_entry & e) {
|
|
return e.path != path && e.tokens.size() <= tokens.size() && std::equal(e.tokens.begin(), e.tokens.end(), tokens.begin());
|
|
};
|
|
std::vector<std::string> obsolete;
|
|
for (const auto & e : entries) {
|
|
if (is_obsolete(e)) {
|
|
obsolete.push_back(e.path);
|
|
}
|
|
}
|
|
|
|
if (limit_size > 0 && size_file > limit_size) {
|
|
SRV_WRN("disk cache: prompt state of %.1f MiB exceeds the limit of %.1f MiB, not saved\n", size_file / (1024.0 * 1024.0), limit_size / (1024.0 * 1024.0));
|
|
return false;
|
|
}
|
|
|
|
try {
|
|
// evict the least recently used files of every model in dir
|
|
struct file_info {
|
|
std::filesystem::path path;
|
|
uint64_t size;
|
|
std::filesystem::file_time_type time;
|
|
};
|
|
std::vector<file_info> files;
|
|
uint64_t size_total = 0;
|
|
for (const auto & de : std::filesystem::directory_iterator(dir)) {
|
|
if (de.is_regular_file() && de.path().extension() == DISK_CACHE_EXT && de.path().string() != path &&
|
|
std::find(obsolete.begin(), obsolete.end(), de.path().string()) == obsolete.end()) {
|
|
files.push_back({ de.path(), de.file_size(), de.last_write_time() });
|
|
size_total += files.back().size;
|
|
}
|
|
}
|
|
std::sort(files.begin(), files.end(), [](const file_info & a, const file_info & b) { return a.time < b.time; });
|
|
for (size_t i = 0; limit_size > 0 && size_total + size_file > limit_size && i < files.size(); i++) {
|
|
SRV_INF("disk cache: size limit, removing '%s' (%.1f MiB)\n", files[i].path.string().c_str(), files[i].size / (1024.0 * 1024.0));
|
|
std::filesystem::remove(files[i].path);
|
|
size_total -= files[i].size;
|
|
for (size_t j = 0; j < entries.size(); j++) {
|
|
if (entries[j].path == files[i].path.string()) {
|
|
entries.erase(entries.begin() + j);
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
const uint64_t avail = std::filesystem::space(dir).available;
|
|
if (avail < size_file + DISK_CACHE_MIN_FREE) {
|
|
SRV_WRN("disk cache: only %.1f MiB free in '%s', prompt state of %.1f MiB not saved\n", avail / (1024.0 * 1024.0), dir.c_str(), size_file / (1024.0 * 1024.0));
|
|
return false;
|
|
}
|
|
} catch (const std::exception & err) {
|
|
SRV_WRN("disk cache: error reading '%s': %s\n", dir.c_str(), err.what());
|
|
return false;
|
|
}
|
|
|
|
bool ok;
|
|
{
|
|
disk_cache_io io;
|
|
io.f = fopen(path_tmp.c_str(), "wb");
|
|
ok = io.f != nullptr && fwrite(&hdr, 1, sizeof(hdr), io.f) == sizeof(hdr);
|
|
if (ok) {
|
|
const uint64_t n_state = data.size();
|
|
io.write(tokens.data(), tokens.size()*sizeof(llama_token));
|
|
io.write(ckpts.data(), ckpts.size()*sizeof(disk_cache_ckpt));
|
|
io.write(&n_state, sizeof(n_state));
|
|
io.write(data.data(), data.size());
|
|
for (const auto & c : prompt.checkpoints) {
|
|
io.write(c.data_tgt.data(), c.data_tgt.size());
|
|
}
|
|
hdr.hash_payload = io.hash;
|
|
ok = io.ok && fseek(io.f, 0, SEEK_SET) == 0 && fwrite(&hdr, 1, sizeof(hdr), io.f) == sizeof(hdr) && fflush(io.f) == 0;
|
|
#if defined(__linux__)
|
|
// write back now, so the pages can be dropped and the rename never exposes missing data
|
|
ok = ok && fdatasync(fileno(io.f)) == 0;
|
|
io.drop_cache();
|
|
#endif
|
|
}
|
|
if (io.f != nullptr && fclose(io.f) != 0) {
|
|
ok = false;
|
|
}
|
|
io.f = nullptr;
|
|
if (!ok) {
|
|
SRV_WRN("disk cache: failed to write '%s': %s, not saved\n", path_tmp.c_str(), strerror(errno));
|
|
}
|
|
}
|
|
std::error_code ec;
|
|
if (ok) {
|
|
std::filesystem::rename(path_tmp, path, ec);
|
|
ok = !ec;
|
|
if (!ok) {
|
|
SRV_WRN("disk cache: failed to rename '%s': %s, not saved\n", path_tmp.c_str(), ec.message().c_str());
|
|
}
|
|
}
|
|
if (!ok) {
|
|
std::filesystem::remove(path_tmp, ec);
|
|
return false;
|
|
}
|
|
|
|
for (size_t i = 0; i < entries.size(); ) {
|
|
const auto & e = entries[i];
|
|
if (e.path == path) {
|
|
entries.erase(entries.begin() + i);
|
|
} else if (is_obsolete(e)) {
|
|
SRV_INF("disk cache: removing '%s', a prefix of the new prompt\n", e.path.c_str());
|
|
remove((int) i);
|
|
} else {
|
|
i++;
|
|
}
|
|
}
|
|
|
|
server_prompt_disk_entry e;
|
|
e.path = path;
|
|
e.tokens = tokens;
|
|
for (const auto & c : ckpts) {
|
|
e.ckpt_n_tokens.push_back(c.n_tokens);
|
|
}
|
|
entries.push_back(std::move(e));
|
|
|
|
return true;
|
|
}
|