char-fork/llama.cpp

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

Adrien Gallouëtcommon : only load backends when required (#22290)bf76ac77b

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
4.4 KiB125 linesraw
1// Test for state restore with fragmented KV cache
2// This tests the fix for: https://github.com/ggml-org/llama.cpp/issues/17527
3// The issue was that state restore required contiguous KV cache slots,
4// which fails when the cache is fragmented.
5//
6// The fix changes find_slot(ubatch, true) to find_slot(ubatch, false)
7// in state_read_meta(), allowing non-contiguous slot allocation.
8
9#include "arg.h"
10#include "common.h"
11#include "llama.h"
12
13#include <vector>
14#include <cstdio>
15#include <cstring>
16
17int main(int argc, char ** argv) {
18    common_params params;
19
20    params.sampling.seed = 1234;
21    params.kv_unified = true;
22    params.n_parallel = 3;
23    params.n_ctx = 256;
24
25    common_init();
26
27    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
28        return 1;
29    }
30
31    // init
32
33    ggml_backend_load_all();
34
35    common_init_result_ptr llama_init = common_init_from_params(params);
36
37    llama_model * model = llama_init->model();
38    llama_context * ctx = llama_init->context();
39
40    if (model == nullptr || ctx == nullptr) {
41        fprintf(stderr, "%s : failed to init\n", __func__);
42        return 1;
43    }
44
45    GGML_UNUSED(model);
46
47    // tokenize prompt
48    std::vector<llama_token> tokens(70, 1);
49
50    // interleave the 3 sequences:
51    // 01201230123...
52    llama_batch batch = llama_batch_init(params.n_parallel*tokens.size(), 0, 1);
53    for (size_t i = 0; i < tokens.size(); i++) {
54        for (int s = 0; s < params.n_parallel; ++s) {
55            common_batch_add(batch, tokens[i], i, {s}, false);
56        }
57    }
58    batch.logits[batch.n_tokens - 1] = true;
59
60    if (llama_decode(ctx, batch)) {
61        fprintf(stderr, "%s : failed to decode seq 0\n", __func__);
62        return 1;
63    }
64
65    fprintf(stderr, "%s : processed prompt on seq 0, 1, 2 (%zu tokens each)\n", __func__, tokens.size());
66
67    // Save state of seq 1
68    std::vector<uint8_t> seq_state(llama_state_seq_get_size(ctx, 1));
69    const size_t ncopy = llama_state_seq_get_data(ctx, seq_state.data(), seq_state.size(), 1);
70    if (ncopy != seq_state.size()) {
71        fprintf(stderr, "%s : failed to save seq 1 state\n", __func__);
72        return 1;
73    }
74    fprintf(stderr, "%s : saved seq 1 state, %zu bytes\n", __func__, ncopy);
75
76    // clear seq 1 to create a "hole" in the KV cache (fragmentation)
77    // 0.20.20.20.2....
78    llama_memory_t mem = llama_get_memory(ctx);
79    llama_memory_seq_rm(mem, 1, -1, -1);
80    fprintf(stderr, "%s : cleared seq 1 to create fragmentation\n", __func__);
81
82    // Now the cache has holes where seq 1 was
83    // This creates fragmentation - there's no contiguous block large enough
84    // for the seq 1 state if we only look for contiguous slots
85
86    // Restore seq 1 state into seq 1 (should work with non-contiguous allocation)
87    // We use seq 1 since it's a valid sequence ID (0 to n_parallel-1)
88    // Before the fix, this would fail with "failed to find available cells in kv cache"
89    const size_t nset = llama_state_seq_set_data(ctx, seq_state.data(), seq_state.size(), 1);
90    if (nset != seq_state.size()) {
91        fprintf(stderr, "%s : FAILED to restore seq state into fragmented cache (got %zu, expected %zu)\n",
92                __func__, nset, seq_state.size());
93        fprintf(stderr, "%s : This is the bug - state restore fails with fragmented KV cache\n", __func__);
94        llama_batch_free(batch);
95        return 1;
96    }
97    fprintf(stderr, "%s : restored state into seq 1, %zu bytes\n", __func__, nset);
98
99    // Verify we can decode with the restored state
100    // Generate one token to verify the restored state is usable
101    auto sparams = llama_sampler_chain_default_params();
102    llama_sampler * smpl = llama_sampler_chain_init(sparams);
103    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sampling.seed));
104
105    auto next_token = llama_sampler_sample(smpl, ctx, -1);
106    auto next_token_str = common_token_to_piece(ctx, next_token);
107
108    common_batch_clear(batch);
109    common_batch_add(batch, next_token, (int)tokens.size(), {1}, true);
110
111    if (llama_decode(ctx, batch)) {
112        fprintf(stderr, "%s : failed to decode with restored state\n", __func__);
113        llama_sampler_free(smpl);
114        llama_batch_free(batch);
115        return 1;
116    }
117
118    fprintf(stderr, "%s : successfully decoded with restored state, generated: '%s'\n", __func__, next_token_str.c_str());
119    fprintf(stderr, "%s : SUCCESS - state restore works with fragmented KV cache\n", __func__);
120
121    llama_sampler_free(smpl);
122    llama_batch_free(batch);
123
124    return 0;
125}