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

speculative : refactor and add a simpler example (#10362)

Browse source 
* speculative : refactor and add a simpler example

ggml-ci

* speculative : clean-up and add comments and TODOs [no ci]

* speculative : manage context in common_speculative

ggml-ci

* speculative : simplify

ggml-ci

* speculative : simplify (cont)

ggml-ci

* speculative : add --draft-min CLI arg

* speculative : minor fixup

* make : build fixes

* speculative : do not redraft previous drafts

ggml-ci

* speculative : fix the draft sampling

ggml-ci

* speculative : fix compile warning

* common : refactor args

ggml-ci

* common : change defaults [no ci]

* common : final touches

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-11-25 09:58:41 +02:00 • committed by GitHub
parent cce5a90075
commit d9d54e498d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
28 changed files with 1028 additions and 326 deletions
Download patch file Download diff file Expand all files Collapse all files

76
common/common.cpp
View file

@ -536,12 +536,12 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf << "\n" << std::to_string(i)
<< ":token '" << detokenized << "'"
<< ":pos " << std::to_string(batch.pos[i])
<< ":n_seq_id " << std::to_string(batch.n_seq_id[i])
<< ":seq_id " << std::to_string(batch.seq_id[i][0])
<< ":logits " << std::to_string(batch.logits[i]);
buf << "\n" << std::to_string(i)
<< ", token '" << detokenized << "'"
<< ", pos " << std::to_string(batch.pos[i])
<< ", n_seq_id " << std::to_string(batch.n_seq_id[i])
<< ", seq_id " << std::to_string(batch.seq_id[i][0])
<< ", logits " << std::to_string(batch.logits[i]);
}
buf << " ]";
@ -925,9 +925,9 @@ struct common_init_result common_init_from_params(common_params & params) {
common_lora_adapters_apply(lctx, iparams.lora_adapters);
}
if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
params.sparams.ignore_eos = false;
params.sampling.ignore_eos = false;
}
if (params.warmup) {
@ -1490,6 +1490,66 @@ void common_batch_add(
batch.n_tokens++;
}
//
// Token utils
//
size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
return i;
}
size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
// check for empty sequences
if (a.empty() || b.empty()) {
return 0;
}
// get the lengths of the input sequences
size_t a_len = a.size();
size_t b_len = b.size();
// initialize the maximum length of the longest common subsequence (LCS)
size_t max_length = 0;
// use two rows instead of a 2D matrix to optimize space
std::vector<size_t> prev_row(b_len + 1, 0);
std::vector<size_t> curr_row(b_len + 1, 0);
// iterate through the elements of a
for (size_t i = 1; i <= a_len; i++) {
// iterate through the elements of b
for (size_t j = 1; j <= b_len; j++) {
// if elements at the current positions match
if (a[i - 1] == b[j - 1]) {
// if it's the first element of either sequences, set LCS length to 1
if (i == 1 || j == 1) {
curr_row[j] = 1;
} else {
// increment LCS length by 1 compared to the previous element
curr_row[j] = prev_row[j - 1] + 1;
}
// update max_length if necessary
if (curr_row[j] > max_length) {
max_length = curr_row[j];
}
} else {
// reset LCS length if elements don't match
curr_row[j] = 0;
}
}
// update the previous row for the next iteration
prev_row = curr_row;
}
// return the maximum length of the LCS
return max_length;
}
//
// Vocab utils
//