root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-14 14:59:52 +00:00
Code Issues Actions 14 Packages Projects Releases Wiki Activity

llama : sanity checks for access to logits (#4274)

Browse Source 
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Jared Van Bortel 2023-12-15 22:16:15 -05:00 committed by GitHub
parent 88ae8952b6
commit 8a5be3bd58
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 22 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
llama.cpp
View File

@ -1505,6 +1505,10 @@ struct llama_context {
// decode output (2-dimensional array: [n_tokens][n_vocab]) // decode output (2-dimensional array: [n_tokens][n_vocab])
std::vector<float> logits; std::vector<float> logits;
#ifndef NDEBUG
// guard against access to unset logits
std::vector<bool> logits_valid;
#endif
bool logits_all = false; bool logits_all = false;
// input embedding (1-dimensional array: [n_embd]) // input embedding (1-dimensional array: [n_embd])
@ -6150,6 +6154,14 @@ static int llama_decode_internal(
{ {
auto & logits_out = lctx.logits; auto & logits_out = lctx.logits;
#ifndef NDEBUG
auto & logits_valid = lctx.logits_valid;
logits_valid.clear();
logits_valid.resize(n_tokens);
logits_out.clear();
#endif
if (batch.logits) { if (batch.logits) {
logits_out.resize(n_vocab * n_tokens); logits_out.resize(n_vocab * n_tokens);
for (uint32_t i = 0; i < n_tokens; i++) { for (uint32_t i = 0; i < n_tokens; i++) {
@ -6157,13 +6169,22 @@ static int llama_decode_internal(
continue; continue;
} }
memcpy(logits_out.data() + (n_vocab*i), (float *) ggml_get_data(res) + (n_vocab*i), sizeof(float)*n_vocab); 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) { } else if (lctx.logits_all) {
logits_out.resize(n_vocab * n_tokens); logits_out.resize(n_vocab * n_tokens);
memcpy(logits_out.data(), (float *) ggml_get_data(res), sizeof(float)*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 { } else {
logits_out.resize(n_vocab); logits_out.resize(n_vocab);
memcpy(logits_out.data(), (float *) ggml_get_data(res) + (n_vocab*(n_tokens - 1)), sizeof(float)*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
} }
} }
@ -10052,6 +10073,7 @@ float * llama_get_logits(struct llama_context * ctx) {
} }
float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { 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; return ctx->logits.data() + i*ctx->model.hparams.n_vocab;
} }