perplexity : fix integer overflow

[ggerganov]

fix #9779

There was an integer overflow in i*n_vocab:

frame #1: 0x0000000100403db8 llama-perplexity`process_logits(int, float const*, int const*, int, std::__1::vector<std::__1::thread, std::__1::allocator<std::__1::thread>>&, double&, double&, float*, float*)::$_0::operator()(this=0x0000600013e88068) const at perplexity.cpp:172:49
   169 	               break;
   170 	           }
   171 	           lock.unlock();
-> 172 	           const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
   173 	           const double v = -results.log_softmax;
   174 	           local_nll += v;
   175 	           local_nll2 += v*v;

(lldb) print i
(int) 14242
(lldb) print n_vocab
(int) 152064
(lldb) print i*n_vocab
(int) -2129271808

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ggerganov]

Using int64_t here is for n_chunk was incorrect. Pushing a fix

[JohannesGaessler]

IIRC the C standard only guarantees at least 16 bit for int.

[JohannesGaessler]

According to Wikpedia:

The standard integer size is platform-dependent. In C, it is denoted by int and required to be at least 16 bits. Windows and Unix systems have 32-bit ints on both 32-bit and 64-bit architectures.

[JohannesGaessler]

According to cppreference:

int — basic integer type. The keyword int may be omitted if any of the modifiers listed below are used. If no length modifiers are present, it's guaranteed to have a width of at least 16 bits. However, on 32/64 bit systems it is almost exclusively guaranteed to have width of at least 32 bits (see below).

[ggerganov]

I doubt this is a concern. We can safely assume that int will always be 32-bit