File size: 6,179 Bytes
1fa479d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
#include "llama-expert-cache.h"

#include <cstdlib>
#include <cstring>
#include <cassert>
#include <algorithm>

#ifdef __APPLE__
#include <fcntl.h>
#include <unistd.h>
#endif

#ifdef __linux__
#include <fcntl.h>
#include <unistd.h>
#endif

#ifdef _WIN32
#include <io.h>
#include <windows.h>
#endif

// Page size for aligned allocation (matches Apple Silicon and most x86)
static constexpr size_t ALLOC_ALIGNMENT = 4096;

static size_t align_up(size_t val, size_t alignment) {
    return (val + alignment - 1) & ~(alignment - 1);
}

llama_expert_cache::llama_expert_cache(size_t max_bytes)
    : max_bytes_(max_bytes)
    , used_bytes_(0)
    , stats_{0, 0, 0, 0, max_bytes} {
}

llama_expert_cache::~llama_expert_cache() {
    for (auto & [key, entry] : cache_) {
        free_aligned(entry.data, entry.size_bytes);
    }
    cache_.clear();
    lru_order_.clear();
    used_bytes_ = 0;
}

void * llama_expert_cache::alloc_aligned(size_t size) {
    size_t alloc_size = align_up(size, ALLOC_ALIGNMENT);
#ifdef _WIN32
    void * ptr = _aligned_malloc(alloc_size, ALLOC_ALIGNMENT);
#else
    void * ptr = nullptr;
    int ret = posix_memalign(&ptr, ALLOC_ALIGNMENT, alloc_size);
    if (ret != 0) {
        ptr = nullptr;
    }
#endif
    return ptr;
}

void llama_expert_cache::free_aligned(void * ptr, size_t /*size*/) {
    if (!ptr) return;
#ifdef _WIN32
    _aligned_free(ptr);
#else
    free(ptr);
#endif
}

void * llama_expert_cache::load_from_disk(const llama_expert_disk_info & info) {
    void * buf = alloc_aligned(info.size_bytes);
    if (!buf) return nullptr;

#ifdef _WIN32
    // Windows: use _lseeki64 + _read or ReadFile
    _lseeki64(info.fd, info.file_offset, SEEK_SET);
    size_t remaining = info.size_bytes;
    char * dst = (char *)buf;
    while (remaining > 0) {
        int chunk = (int)std::min(remaining, (size_t)INT_MAX);
        int n = _read(info.fd, dst, chunk);
        if (n <= 0) {
            free_aligned(buf, info.size_bytes);
            return nullptr;
        }
        dst += n;
        remaining -= n;
    }
#else
    // POSIX: use pread for thread-safe positional read (no seek mutex needed)
    size_t remaining = info.size_bytes;
    char * dst = (char *)buf;
    off_t offset = (off_t)info.file_offset;
    while (remaining > 0) {
        ssize_t n = pread(info.fd, dst, remaining, offset);
        if (n <= 0) {
            free_aligned(buf, info.size_bytes);
            return nullptr;
        }
        dst += n;
        offset += n;
        remaining -= n;
    }
#endif

    return buf;
}

void llama_expert_cache::evict_until_free(size_t needed) {
    while (used_bytes_ + needed > max_bytes_ && !lru_order_.empty()) {
        // Evict least recently used (back of list)
        auto evict_key = lru_order_.back();
        lru_order_.pop_back();

        auto it = cache_.find(evict_key);
        if (it != cache_.end()) {
            used_bytes_ -= it->second.size_bytes;
            free_aligned(it->second.data, it->second.size_bytes);
            cache_.erase(it);
            stats_.evictions++;
        }
    }
}

void * llama_expert_cache::ensure(const llama_expert_key & key,
                                   const llama_expert_disk_info & disk_info) {
    std::lock_guard<std::mutex> lock(mutex_);

    // Check cache
    auto it = cache_.find(key);
    if (it != cache_.end()) {
        // Hit: move to front of LRU
        stats_.hits++;
        lru_order_.erase(it->second.lru_it);
        lru_order_.push_front(key);
        it->second.lru_it = lru_order_.begin();
        return it->second.data;
    }

    // Miss: load from disk
    stats_.misses++;

    size_t alloc_size = align_up(disk_info.size_bytes, ALLOC_ALIGNMENT);

    // Evict until we have space
    evict_until_free(alloc_size);

    // Load from disk (this does I/O while holding the lock —
    // acceptable for now, can be optimized with async prefetch later)
    void * data = load_from_disk(disk_info);
    if (!data) {
        return nullptr;
    }

    // Insert into cache
    lru_order_.push_front(key);
    llama_expert_entry entry;
    entry.key = key;
    entry.data = data;
    entry.size_bytes = alloc_size;
    entry.lru_it = lru_order_.begin();
    cache_[key] = entry;
    used_bytes_ += alloc_size;
    stats_.bytes_used = used_bytes_;

    return data;
}

std::pair<void *, bool> llama_expert_cache::get_or_alloc(
        const llama_expert_key & key, size_t size_bytes) {
    std::lock_guard<std::mutex> lock(mutex_);

    // Check cache
    auto it = cache_.find(key);
    if (it != cache_.end()) {
        stats_.hits++;
        lru_order_.erase(it->second.lru_it);
        lru_order_.push_front(key);
        it->second.lru_it = lru_order_.begin();
        return {it->second.data, true};  // hit
    }

    // Miss
    stats_.misses++;

    size_t alloc_size = align_up(size_bytes, ALLOC_ALIGNMENT);
    evict_until_free(alloc_size);

    void * data = alloc_aligned(alloc_size);
    if (!data) {
        return {nullptr, false};
    }

    lru_order_.push_front(key);
    llama_expert_entry entry;
    entry.key = key;
    entry.data = data;
    entry.size_bytes = alloc_size;
    entry.lru_it = lru_order_.begin();
    cache_[key] = entry;
    used_bytes_ += alloc_size;
    stats_.bytes_used = used_bytes_;

    return {data, false};  // miss — caller must fill
}

void llama_expert_cache::touch(const llama_expert_key & key) {
    std::lock_guard<std::mutex> lock(mutex_);
    auto it = cache_.find(key);
    if (it != cache_.end()) {
        lru_order_.erase(it->second.lru_it);
        lru_order_.push_front(key);
        it->second.lru_it = lru_order_.begin();
    }
}

bool llama_expert_cache::contains(const llama_expert_key & key) const {
    std::lock_guard<std::mutex> lock(mutex_);
    return cache_.find(key) != cache_.end();
}

llama_expert_cache_stats llama_expert_cache::get_stats() const {
    std::lock_guard<std::mutex> lock(mutex_);
    auto s = stats_;
    s.bytes_used = used_bytes_;
    s.bytes_capacity = max_bytes_;
    return s;
}

void llama_expert_cache::reset_stats() {
    std::lock_guard<std::mutex> lock(mutex_);
    stats_.hits = 0;
    stats_.misses = 0;
    stats_.evictions = 0;
}