File size: 6,249 Bytes
8d11530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#include "llama.h"
#include <vector>
#include <string>
#include <cstring>
#include <cstdio>
#include <cstdlib>

// Global context for the loaded model
static llama_model* g_model = nullptr;
static llama_context* g_ctx = nullptr;
static llama_sampler* g_smpl = nullptr;

extern "C" {

// Initialize the model
bool init_model(const char* model_path) {
    llama_backend_init();
    
    llama_model_params model_params = llama_model_default_params();
    model_params.n_gpu_layers = 0; // CPU only for now
    g_model = llama_model_load_from_file(model_path, model_params);
    
    if (!g_model) return false;

    llama_context_params ctx_params = llama_context_default_params();
    ctx_params.n_ctx = 4096; // 256 per sequence for 16 users
    ctx_params.n_batch = 512;
    ctx_params.n_threads = 8;
    ctx_params.n_threads_batch = 8;
    ctx_params.n_seq_max = 16;
    g_ctx = llama_init_from_model(g_model, ctx_params);
    
    if (!g_ctx) return false;

    auto sparams = llama_sampler_chain_default_params();
    g_smpl = llama_sampler_chain_init(sparams);
    llama_sampler_chain_add(g_smpl, llama_sampler_init_greedy());

    return true;
}

// Helper to add a token to a batch manually
void batch_add(llama_batch & batch, llama_token id, llama_pos pos, const std::vector<llama_seq_id> & seq_ids, bool logits) {
    batch.token[batch.n_tokens] = id;
    batch.pos[batch.n_tokens]   = pos;
    batch.n_seq_id[batch.n_tokens] = seq_ids.size();
    for (size_t i = 0; i < seq_ids.size(); ++i) {
        batch.seq_id[batch.n_tokens][i] = seq_ids[i];
    }
    batch.logits[batch.n_tokens] = logits;
    batch.n_tokens++;
}

// Global state for streaming
static int g_count = 0;
static int g_step = 0;
static int g_max_tokens = 0;
static std::vector<std::string> g_responses;
static std::vector<bool> g_active;
static std::vector<int> g_n_pos;
static std::vector<int> g_logits_idx;
static std::vector<std::vector<llama_token>> g_all_tokens;
static llama_batch g_batch;
static const llama_vocab* g_vocab = nullptr;

// 1. Start a new batch (Prefill)
void start_batch(const char** prompts, int count, int max_tokens) {
    if (!g_ctx || count == 0) return;

    g_vocab = llama_model_get_vocab(g_model);
    g_count = count;
    g_max_tokens = max_tokens;
    g_step = 0;
    
    // Reset state
    g_responses.assign(count, "");
    g_active.assign(count, true);
    g_n_pos.assign(count, 0);
    g_logits_idx.assign(count, -1);
    g_all_tokens.clear();

    // Tokenize
    for (int i = 0; i < count; i++) {
        int n_prompt = -llama_tokenize(g_vocab, prompts[i], strlen(prompts[i]), NULL, 0, true, true);
        std::vector<llama_token> tokens(n_prompt);
        llama_tokenize(g_vocab, prompts[i], strlen(prompts[i]), tokens.data(), tokens.size(), true, true);
        g_all_tokens.push_back(tokens);
    }

    // Clear KV cache
    llama_memory_clear(llama_get_memory(g_ctx), true);

    // Init batch
    if (g_batch.token) llama_batch_free(g_batch); // Free if exists
    g_batch = llama_batch_init(4096, 0, 1); // Larger batch for safety

    // Prefill
    g_batch.n_tokens = 0;
    for (int i = 0; i < count; i++) {
        for (size_t j = 0; j < g_all_tokens[i].size(); j++) {
            bool is_last = (j == g_all_tokens[i].size() - 1);
            if (is_last) g_logits_idx[i] = g_batch.n_tokens;
            batch_add(g_batch, g_all_tokens[i][j], g_n_pos[i]++, { (llama_seq_id)i }, is_last);
        }
    }

    // Decode Prefill
    if (llama_decode(g_ctx, g_batch)) {
        fprintf(stderr, "Failed to decode prefill\n");
    }
}

// 2. Decode one step (Generate next token for all active sequences)
// Returns true if any sequence is still active, false if all done
bool decode_step(const char** results) {
    if (g_step >= g_max_tokens) return false;

    g_batch.n_tokens = 0;
    bool any_active = false;
    std::vector<int> next_logits_idx(g_count, -1);
    int current_batch_pos = 0;

    for (int i = 0; i < g_count; i++) {
        results[i] = nullptr; // Default to null (no new token this step if inactive)
        
        if (!g_active[i]) continue;

        // Sample
        llama_token id = llama_sampler_sample(g_smpl, g_ctx, g_logits_idx[i]);
        llama_sampler_accept(g_smpl, id);

        // Check EOG/limit
        if (llama_vocab_is_eog(g_vocab, id) || g_n_pos[i] >= 4096) { // Hard limit matches n_ctx
            g_active[i] = false;
            continue;
        }

        // Decode token to string
        static char buf[256]; // Static buffer for simplicity (not thread safe across batches, but fine here)
        int n = llama_token_to_piece(g_vocab, id, buf, sizeof(buf), 0, true);
        if (n < 0) {
            // error or empty
        } else {
             // Allocate new string for result to pass back to Python
             // Python side must handle this appropriately or we just copy
             // For ctypes `const char**`, we need a pointer that persists until next call
             // We can use a thread_local buffer or just strdup and let Python manage? 
             // Ideally Python frees, but to keep it simple we use a persistent buffer in C++?
             // No, let's just return a pointer to a static buffer is risky if multiple threads.
             // But we are single threaded engine key.
             // Safer: return strdup, Python bridge won't free it effortlessly though. 
             // Let's use a global vector of string buffers for the current step.
        }
        
        // Actually, we just need to return the piece. 
        // Let's rely on a persistent buffer for the current step.
        // We can just strdup and leak if we don't free? No.
        // Let's assume the Python side consumes it immediately.
        results[i] = strdup(buf); 

        next_logits_idx[i] = current_batch_pos++;
        batch_add(g_batch, id, g_n_pos[i]++, { (llama_seq_id)i }, true);
        any_active = true;
    }

    if (!any_active) return false;

    g_logits_idx = next_logits_idx;
    if (llama_decode(g_ctx, g_batch)) {
         return false;
    }

    g_step++;
    return true;
}

// Cleanup
void cleanup() {
    if (g_smpl) llama_sampler_free(g_smpl);
    if (g_ctx) llama_free(g_ctx);
    if (g_model) llama_model_free(g_model);
}

}