vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : sanity checks for access to logits

Browse source
This commit is contained in:
Jared Van Bortel 2023-11-30 14:36:02 -05:00
parent 1f5cd83275
commit f91707bbe1
1 changed files with 20 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

20
llama.cpp
View file

@ -1468,6 +1468,10 @@ struct llama_context {
// decode output (2-dimensional array: [n_tokens][n_vocab])
std::vector<float> logits;
#ifndef NDEBUG
// guard against access to unset logits
std::vector<bool> logits_valid;
#endif
bool logits_all = false;
// input embedding (1-dimensional array: [n_embd])
@ -5609,6 +5613,12 @@ static int llama_decode_internal(
{
auto & logits_out = lctx.logits;
#ifndef NDEBUG
auto & logits_valid = lctx.logits_valid;
logits_valid.clear();
logits_valid.resize(n_vocab);
#endif
if (batch.logits) {
logits_out.resize(n_vocab * n_tokens);
for (uint32_t i = 0; i < n_tokens; i++) {
@ -5616,13 +5626,22 @@ static int llama_decode_internal(
continue;
}
memcpy(logits_out.data() + (n_vocab*i), (float *) ggml_get_data(res) + (n_vocab*i), sizeof(float)*n_vocab);
#ifndef NDEBUG
logits_valid[i] = true;
#endif
}
} else if (lctx.logits_all) {
logits_out.resize(n_vocab * n_tokens);
memcpy(logits_out.data(), (float *) ggml_get_data(res), sizeof(float)*n_vocab*n_tokens);
#ifndef NDEBUG
std::fill(logits_valid.begin(), logits_valid.end(), true);
#endif
} else {
logits_out.resize(n_vocab);
memcpy(logits_out.data(), (float *) ggml_get_data(res) + (n_vocab*(n_tokens - 1)), sizeof(float)*n_vocab);
#ifndef NDEBUG
logits_valid[n_tokens - 1] = true;
#endif
}
}
@ -9465,6 +9484,7 @@ float * llama_get_logits(struct llama_context * ctx) {
}
float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
assert(ctx->logits_valid.at(i));
return ctx->logits.data() + i*ctx->model.hparams.n_vocab;
}