speculative : fix out-of-bounds read in ngram-map on prompt shrink

[o7si]

common_ngram_map_begin() invalidates stale entries using map.size_last_begin (the previous generation start), not size_begin (the new prompt length):

llama.cpp/common/ngram-map.cpp

Line 171 in 6ed481e

if (key.key_idx >= map.size_last_begin) {

When a slot is reused for a shorter prompt, size_begin < size_last_begin, so keys with key_idx in [size_begin, size_last_begin) are kept and later read out of bounds in common_ngram_map_draft():

llama.cpp/common/ngram-map.cpp

Line 346 in 6ed481e

if (inp[map.keys[i].key_idx + j] != key_tokens[j]) {

Since a key n-gram is size_key tokens long, use size_begin and remove a key when key_idx + size_key > size_begin.

Related issue:

• Assertion '__n < this->size()' failed. #23889