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

perplexity : no need for decode_helper

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-01-18 13:57:58 +02:00
parent 4351c4632d
commit af309010ab
1 changed files with 1 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

30
examples/perplexity/perplexity.cpp
View file

@ -442,33 +442,6 @@ static std::vector<float> hellaswag_evaluate_tokens(
return result; return result;
} }
// decode in batches of ctx_params.n_batch tokens
static bool decode_helper(llama_context * ctx, llama_batch & batch, int32_t n_batch) {
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
return false;
}
}
return true;
}
static void hellaswag_score(llama_context * ctx, const gpt_params & params) { static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
// Calculates hellaswag score (acc_norm) from prompt // Calculates hellaswag score (acc_norm) from prompt
// //
@ -589,7 +562,6 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
const int n_vocab = llama_n_vocab(llama_get_model(ctx)); const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_ctx = llama_n_ctx(ctx); const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
GGML_ASSERT(params.n_batch >= n_ctx && "HellaSwag currently requires n_batch >= n_ctx"); GGML_ASSERT(params.n_batch >= n_ctx && "HellaSwag currently requires n_batch >= n_ctx");
@ -648,7 +620,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
llama_kv_cache_clear(ctx); llama_kv_cache_clear(ctx);
if (!decode_helper(ctx, batch, n_batch)) { if (llama_decode(ctx, batch) != 0) {
fprintf(stderr, "%s: llama_decode() failed\n", __func__); fprintf(stderr, "%s: llama_decode() failed\n", __func__);
return; return;
} }