root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-14 04:24:30 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

falcon-main.cpp : fixes

Browse Source
This commit is contained in:
klosax 2023-08-19 12:08:17 +02:00 committed by GitHub
parent dadf098b5a
commit 781bf2481f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 29 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
examples/gptneox-wip/falcon-main.cpp
View File

@ -378,6 +378,7 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
{
int keyidx;
// check model architecture kv
keyidx = gguf_find_key(ggufctx, "general.architecture");
if (keyidx != -1) {
if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "falcon") != 0) {
@ -389,6 +390,18 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
return false;
}
// check model tensor data layout kv
keyidx = gguf_find_key(ggufctx, "falcon.tensor_data_layout");
if (keyidx != -1) {
if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "jploski") != 0) {
fprintf(stdout, "%s: model tensor data layout not supported!\n", __func__);
return false;
}
} else {
fprintf(stdout, "%s: gguf model tensor data layout not found!\n", __func__);
return false;
}
}
// load hparams
@ -479,6 +492,9 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
vocab.token_to_id[word] = i;
vocab.id_to_token[i] = word;
if( vocab.id_to_token[i] == "\n" ) {
vocab.linefeed_id = i;
}
}
std::vector<std::pair<std::string, std::string>> bpe_merges;
@ -501,17 +517,19 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
vocab.populate_bpe_ranks(bpe_merges);
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.bos_token_id"); if( keyidx != -1 ) { vocab.special_bos_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); vocab.special_have_bos=true; }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.eos_token_id"); if( keyidx != -1 ) { vocab.special_eos_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); vocab.special_have_eos=true; }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.unknown_token_id"); if( keyidx != -1 ) { vocab.special_unk_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); vocab.special_have_unk=true; }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); vocab.special_have_sep=true; }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) { vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); vocab.special_have_pad=true; }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.bos_token_id"); if( keyidx != -1 ) { vocab.special_bos_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.eos_token_id"); if( keyidx != -1 ) { vocab.special_eos_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.unknown_token_id"); if( keyidx != -1 ) { vocab.special_unk_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) { vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
if( vocab.linefeed_id != -1 ) { fprintf(stdout, "%s: LF token = %d\n", __func__, vocab.linefeed_id ); }
if( vocab.special_have_bos ) { fprintf(stdout, "%s: bos token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
if( vocab.special_have_eos ) { fprintf(stdout, "%s: eos token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
if( vocab.special_have_unk ) { fprintf(stdout, "%s: unk token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
if( vocab.special_have_sep ) { fprintf(stdout, "%s: sep token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
if( vocab.special_have_pad ) { fprintf(stdout, "%s: pad token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
}
@ -1070,7 +1088,7 @@ int main(int argc, char ** argv) {
fflush(stdout);
// end of text token
if (vocab.special_have_eos && embd.back() == vocab.special_eos_id) {
if (vocab.special_eos_id != -1 && embd.back() == vocab.special_eos_id) {
break;
}
}