server: refactor, add server_slot_sched, prompt running hash (#32154)

* server: refactor, add server_slot_sched, prompt running hash

* use sha256 hash
This commit is contained in:
Xuan-Son Nguyen authored and GitHub committed 2026-11-15 09:37:17 +01:00
1 parent 58a4a6b49c
commit 429460dc0e
11 files changed
+790 -474

No files matched your search

+77
View File
@@ -156,6 +156,82 @@ static void test_server_tokens_clone(testing & t) {
t.assert_true("get_text_tokens", cl.get_text_tokens() == llama_tokens({1, 2, 3, 4}));
}
static void test_server_tokens_prefix_hashes(testing & t) {
const size_t n_blk = server_tokens::hash_block_size;
auto make_tokens = [](size_t n) {
llama_tokens res(n);
for (size_t i = 0; i < n; i++) {
res[i] = 100 + (llama_token) i;
}
return res;
};
{
server_tokens a(make_tokens(3 * n_blk + 10), false);
t.assert_equal("full blocks only", (size_t) 3, a.get_prefix_hashes().size());
t.assert_equal("same tokens", 3 * n_blk, a.get_common_prefix_lower(a.clone().get_prefix_hashes()));
t.assert_equal("no hashes", (size_t) 0, a.get_common_prefix_lower({}));
// the exact value is 2 * n_blk + 5
llama_tokens toks = make_tokens(3 * n_blk + 10);
toks[2 * n_blk + 5] = 7;
server_tokens b(toks, false);
t.assert_equal("diverge in block 2", 2 * n_blk, a.get_common_prefix_lower(b.get_prefix_hashes()));
t.assert_equal("exact, diverge in block 2", 2 * n_blk + 5, a.get_common_prefix(b));
t.assert_equal("exact, same tokens", 3 * n_blk + 10, a.get_common_prefix(a.clone()));
t.assert_equal("b is shorter", 1 * n_blk, a.get_common_prefix_lower(server_tokens(make_tokens(n_blk + 3), false).get_prefix_hashes()));
toks[0] = 7;
t.assert_equal("diverge in block 0", (size_t) 0, a.get_common_prefix_lower(server_tokens(toks, false).get_prefix_hashes()));
}
{
server_tokens a(make_tokens(3 * n_blk), false);
const auto hashes = a.get_prefix_hashes();
server_tokens b = a.clone();
b.set_token(n_blk + 1, 7);
t.assert_equal("set_token", n_blk, b.get_common_prefix_lower(hashes));
b = a.clone();
b.keep_first(n_blk + 1);
for (size_t i = n_blk + 1; i < 3 * n_blk; i++) {
b.push_back(7);
}
t.assert_equal("keep_first + push_back", n_blk, b.get_common_prefix_lower(hashes));
b.clear();
t.assert_equal("clear", (size_t) 0, b.get_prefix_hashes().size());
}
{
// the image chunk starts in block 0 and ends in a later block
mtmd::input_chunks chunks(mtmd_test_create_input_chunks());
const mtmd_input_chunk * image = chunks[1];
const size_t n_image = mtmd_input_chunk_get_n_tokens(image);
const size_t i_image = n_blk - 4;
auto make_mtmd = [&](bool with_image, llama_token last) {
server_tokens res(make_tokens(i_image), true);
if (with_image) {
res.push_back(image);
}
while (res.size() < i_image + n_image) {
res.push_back(5);
}
res.push_back(last);
return res;
};
server_tokens a = make_mtmd(true, 1);
t.assert_true("image is longer than a block", n_image > n_blk);
t.assert_equal("exact value", i_image + n_image, a.get_common_prefix(make_mtmd(true, 2)));
t.assert_equal("block ends inside the image", i_image, a.get_common_prefix_lower(make_mtmd(true, 2).get_prefix_hashes()));
t.assert_equal("text in place of the image", (size_t) 0, a.get_common_prefix_lower(make_mtmd(false, 1).get_prefix_hashes()));
}
}
static void test_server_tokens(testing & t) {
t.test("basic", test_server_tokens_basic);
t.test("pos", test_server_tokens_pos);
@@ -163,6 +239,7 @@ static void test_server_tokens(testing & t) {
t.test("common_prefix", test_server_tokens_common_prefix);
t.test("serialize", test_server_tokens_serialize);
t.test("clone", test_server_tokens_clone);
t.test("prefix_hashes", test_server_tokens_prefix_hashes);
}
//