Skip to content

feat: allow deepseek_v3 architecture to use Kimi's bpe pattern #6653

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
biswapanda merged 2 commits into from
Feb 26, 2026
Merged

feat: allow deepseek_v3 architecture to use Kimi's bpe pattern #6653

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
734d407
feat: allow deepseek_v3 architecture to use BPE
biswapanda Feb 26, 2026
ee943d4
add comment
biswapanda Feb 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 7 additions & 2 deletions lib/llm/src/tokenizers/tiktoken.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,10 +150,15 @@ fn detect_bpe_pattern(directory: &Path) -> Result<&'static str> {
})?;

match model_type.as_str() {
"kimi" | "kimi_k2" | "kimi_k25" => Ok(KIMI_PATTERN),
// baseten-admin/Kimi-2.5-text-nvfp4-v3 model has model_type: "deepseek_v3" in its config.json
// because Kimi K2.5 is built on the DeepSeek V3 architecture.
// it still ships the Kimi tiktoken tokenizer file, so the KIMI_PATTERN BPE regex is the
// correct pattern to use. No pure DeepSeek V3 model uses tiktoken.model files
// (they use tokenizer.json instead) so this match is safe.
"kimi" | "kimi_k2" | "kimi_k25" | "deepseek_v3" => Ok(KIMI_PATTERN),
Comment thread
biswapanda marked this conversation as resolved.
Comment thread
biswapanda marked this conversation as resolved.
_ => Err(Error::msg(format!(
"Unsupported tiktoken model_type '{model_type}'. \
Currently supported: kimi, kimi_k2, kimi_k25. \
Currently supported: kimi, kimi_k2, kimi_k25, deepseek_v3. \
To add a new model type, extend detect_bpe_pattern() in tokenizers/tiktoken.rs \
with the appropriate BPE regex pattern. \
Alternatively, provide a tokenizer.json (HuggingFace format) instead."
Expand Down
Loading