char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Xuan-Son Nguyenserver: add dedup-cache-models preset option (#27346)ee4c505a4

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
4.3 KiB121 linesraw
1#pragma once
2
3#include "hf-cache.h"
4
5#include <string>
6#include <vector>
7#include <functional>
8
9struct common_params_model;
10
11using common_header      = std::pair<std::string, std::string>;
12using common_header_list = std::vector<common_header>;
13
14struct common_download_progress {
15    std::string url;
16    size_t downloaded = 0;
17    size_t total      = 0;
18    bool cached       = false;
19};
20
21class common_download_callback {
22public:
23    virtual ~common_download_callback() = default;
24    virtual void on_start(const common_download_progress & p) = 0;
25    virtual void on_update(const common_download_progress & p) = 0;
26    virtual void on_done(const common_download_progress & p, bool ok) = 0;
27    virtual bool is_cancelled() const { return false; }
28};
29
30struct common_remote_params {
31    common_header_list headers;
32    long timeout  = 0;           // in seconds, 0 means no timeout
33    long max_size = 0;           // unlimited if 0
34};
35
36// get remote file content, returns <http_code, raw_response_body>
37std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);
38
39// split HF repo with tag into <repo, tag>, for example:
40// - "ggml-org/models:F16" -> <"ggml-org/models", "F16">
41// tag is optional and can be empty
42std::pair<std::string, std::string> common_download_split_repo_tag(const std::string & hf_repo_with_tag);
43
44// Result of common_list_cached_models
45struct common_cached_model_info {
46    std::string repo;
47    std::string tag;
48    std::string to_string() const {
49        return repo + ":" + tag;
50    }
51};
52
53// Options for common_download_file_single
54struct common_download_opts {
55    std::string bearer_token;
56    common_header_list headers;
57    bool offline = false;
58    bool download_mmproj  = false;
59    bool download_mtp     = false;
60    bool download_eagle3  = false;
61    bool download_dflash  = false;
62    bool download_dspark  = false;
63    common_download_callback * callback = nullptr;
64};
65
66struct common_download_task {
67    common_download_opts opts;
68    std::string url;
69    std::string local_path;
70    std::function<void()> on_done;
71    bool is_hf = false;
72
73    common_download_task() = default;
74    common_download_task(hf_cache::hf_file f,
75            const common_download_opts & opts,
76            std::function<void()> on_done = nullptr)
77        : opts(opts), url(f.url), local_path(f.local_path), on_done(on_done), is_hf(true) {}
78};
79
80void common_download_run_tasks(const std::vector<common_download_task> & tasks);
81
82// if url is a multi-part GGUF file, returns all parts, otherwise returns the single file
83std::vector<std::string> common_download_get_all_parts(const std::string & url);
84
85// returns list of cached models
86std::vector<common_cached_model_info> common_list_cached_models();
87
88// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
89// returns an empty string if the model is not present in the cache
90std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
91
92// download single file from url to local path
93// returns status code or -1 on error
94// skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
95int common_download_file_single(const std::string & url,
96                                const std::string & path,
97                                const common_download_opts & opts = {},
98                                bool skip_etag = false);
99
100// resolve and download model from Docker registry
101// return local path to downloaded model file
102std::string common_docker_resolve_model(const std::string & docker);
103
104// Remove a cached model from disk
105// input format: "user/model" or "user/model:tag"
106// - if tag is omitted, removes the entire repo cache directory
107// - if tag is present, removes only files matching that tag (and orphaned blobs)
108// returns true if anything was removed
109bool common_download_remove(const std::string & hf_repo_with_tag);
110
111struct common_download_hf_plan {
112    hf_cache::hf_file primary;
113    hf_cache::hf_files model_files;
114    hf_cache::hf_file mmproj;
115    hf_cache::hf_file mtp;
116    hf_cache::hf_file eagle3;
117    hf_cache::hf_file dflash;
118    hf_cache::hf_file dspark;
119    hf_cache::hf_file preset; // if set, only this file is downloaded
120};
121common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);