ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama : recognize IBM Granite 3.3 FIM tokens (#12988)

Browse source 
The Granite's FIM tokens are very similar to Qwen's; it's just that
they use underscore instead of a dash. So <fim_middle> for example
instead of <fim-middle>.

Opening up tokenizer_config.json in ibm-granite/granite-3.3-8b-base
shows:

```
    "<fim_prefix>",
    "<fim_middle>",
    "<fim_suffix>",
    "<fim_pad>",
    ...
    "<reponame>",
```
This commit is contained in:
Mikko Juola 2025-04-17 01:37:05 -07:00 • committed by GitHub
parent 12b17501e6
commit 971f245b3b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 5 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

5
src/llama-vocab.cpp
View file

@ -1841,6 +1841,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false if (false
|| t.first == "<|fim_prefix|>" // Qwen || t.first == "<|fim_prefix|>" // Qwen
|| t.first == "<fim-prefix>" || t.first == "<fim-prefix>"
|| t.first == "<fim_prefix>" // Granite
|| t.first == "<｜fim▁begin｜>" // DeepSeek || t.first == "<｜fim▁begin｜>" // DeepSeek
|| t.first == "<PRE>" || t.first == "<PRE>"
|| t.first == "▁<PRE>" // CodeLlama || t.first == "▁<PRE>" // CodeLlama
@ -1859,6 +1860,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false if (false
|| t.first == "<|fim_suffix|>" // Qwen || t.first == "<|fim_suffix|>" // Qwen
|| t.first == "<fim-suffix>" || t.first == "<fim-suffix>"
|| t.first == "<fim_suffix>" // Granite
|| t.first == "<｜fim▁hole｜>" // DeepSeek || t.first == "<｜fim▁hole｜>" // DeepSeek
|| t.first == "<SUF>" || t.first == "<SUF>"
|| t.first == "▁<SUF>" // CodeLlama || t.first == "▁<SUF>" // CodeLlama
@ -1877,6 +1879,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false if (false
|| t.first == "<|fim_middle|>" // Qwen || t.first == "<|fim_middle|>" // Qwen
|| t.first == "<fim-middle>" || t.first == "<fim-middle>"
|| t.first == "<fim_middle>" // Granite
|| t.first == "<｜fim▁end｜>" // DeepSeek || t.first == "<｜fim▁end｜>" // DeepSeek
|| t.first == "<MID>" || t.first == "<MID>"
|| t.first == "▁<MID>" // CodeLlama || t.first == "▁<MID>" // CodeLlama
@ -1895,6 +1898,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false if (false
|| t.first == "<|fim_pad|>" // Qwen || t.first == "<|fim_pad|>" // Qwen
|| t.first == "<fim-pad>" || t.first == "<fim-pad>"
|| t.first == "<fim_pad>" // Granite
|| t.first == "<PAD>" || t.first == "<PAD>"
) { ) {
special_fim_pad_id = t.second; special_fim_pad_id = t.second;
@ -1913,6 +1917,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<|repo_name|>" || t.first == "<|repo_name|>"
|| t.first == "<fim-repo>" || t.first == "<fim-repo>"
|| t.first == "<REPO>" || t.first == "<REPO>"
|| t.first == "<reponame>" // Granite
) { ) {
special_fim_rep_id = t.second; special_fim_rep_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) { if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {