russell@unturf.com
81bef63b2e
transformers-0002: MOAD-0004 (CWE-312) regnet convert script logs HF_TOKEN verbatim transformers-0003: MOAD-0001 (CWE-407) convert_tokens_to_string O(T×S) list scan - marian, m2m_100, speech_to_text, siglip, gpt_sw3 all affected - all_special_tokens is list[str]; fix: cache set() before loop; 5x speedup vllm-0002: MOAD-0001 (CWE-407) Grok2Tokenizer O(N×V) dict.values() scan - decode() and convert_ids_to_tokens() use .values() view per token - sibling Mistral tokenizer already uses frozenset correctly - fix: add _special_token_ids frozenset at __init__; 10x speedup at N=2048, V=200 MOADs 0002/0003/0005 CLEAN for both repos
2.2 KiB
SCAN — vLLM — all 5 MOADs scanned 2026-03-31
Repository: https://github.com/vllm-project/vllm Version: HEAD (depth=1 clone) Language: Python, C++ (CUDA kernels)
MOAD-0001 (CWE-407): 2 DEFECTS FOUND
vllm-0001 (pre-existing, UNDF-2026-000000874)
vllm/lora/punica_wrapper/utils.py: lora_index_to_id.index(x) called per token in convert_mapping() — O(L) list.index() per token in a prompt-length loop. Patch: pre-build dict {lora_id: index} for O(1) lookup.
vllm-0002 (new)
vllm/tokenizers/grok2.py: decode() and convert_ids_to_tokens() use
token_id not in self._special_tokens.values() — dict.values() is a view,
membership scan is O(V) per token. V = number of special tokens in Grok-2.
Called per output token during streaming LLM inference.
The sibling Mistral tokenizer in the same codebase already uses
_special_token_ids_set: frozenset[int] correctly.
Patch: add self._special_token_ids = frozenset(self._special_tokens.values())
at init time and replace both .values() checks.
Speedup: ~10x at N=2048, V=200.
MOAD-0002 (Intertangle): CLEAN
vLLM has well-separated subsystems. The engine, scheduler, model runner, and workers communicate through clean interfaces (SchedulerOutput, ExecuteModelReq). No hot-path coupling through mutable global god objects found.
MOAD-0003 (Leaked Context): CLEAN
_current_stream_tls = threading.local() in vllm/utils/torch_utils.py tracks
the current CUDA stream per thread — this is CUDA device state (infrastructure),
not per-request identity. No request-scoped identity leaked through ThreadLocal.
contextvars.ContextVar("_NoInitOrTensorImpl.is_active") in tensorizer.py is
a task-scoped boolean flag for serialization, not per-request identity.
MOAD-0004 (Logged Secret): CLEAN
No environment variable secrets (API keys, tokens) logged verbatim in logger calls. vLLM's VLLM_API_KEY is read but not logged. MODELSCOPE_API_TOKEN is passed to client constructor, not printed.
MOAD-0005 (Thundering Herd): CLEAN
_CPU_MOE_LAYER_CACHE is written once per layer at model init (weakref.ref), read-only during inference. No concurrent cache miss+compute+set pattern. KV cache allocation uses the v1 scheduler which runs in a single async loop. Block pool allocation is single-threaded per scheduler step.