tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-29 04:35:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

imatrix : allow processing multiple chunks per batch

Browse Source 
* perplexity : simplify filling the batch
This commit is contained in:
Francis Couture-Harpin
2024-08-20 15:17:24 -04:00
parent 90db8146d5
commit bce54642c8
2 changed files with 75 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
examples/perplexity/perplexity.cpp
View File

@ -583,7 +583,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
int n_outputs = 0;
batch.n_tokens = 0;
// clear the batch
llama_batch_clear(batch);
for (int seq = 0; seq < n_seq_batch; seq++) {
int seq_start = batch_start + seq*n_ctx;
@ -596,16 +598,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
}
for (int k = 0; k < batch_size; ++k) {
const int idx = seq*n_ctx + k;
batch.token [idx] = tokens[seq_start + k];
batch.pos [idx] = j*n_batch + k;
batch.n_seq_id[idx] = 1;
batch.seq_id [idx][0] = seq;
batch.logits [idx] = batch.pos[idx] >= first ? 1 : 0;
n_outputs += batch.logits[idx] != 0;
llama_pos pos = j*n_batch + k;
llama_batch_add(batch, tokens[seq_start + k], pos, { seq }, pos >= first);
n_outputs += (int) (pos >= first);
}
batch.n_tokens += batch_size;
// restore the original token in case it was set to BOS
tokens[seq_start] = token_org;