- Forward logit_bias and logprobs natively to llama.cpp
- Support n>1 completions with seed increment for diversity
- Fix logprobs returning empty dict when not requested
- Use config.eos_token_id_list for all EOS tokens as stop conditions
(fixes models like Llama-3 that define multiple EOS token IDs)
- Load vision/draft models before main model so autosplit accounts
for their VRAM usage
- Fix loss computation in ExLlamav3_HF: use cache across chunks so
sequences longer than 2048 tokens get correct perplexity values
- Stop llama-server subprocess on model unload instead of relying on GC
- Fix tool_calls[].index being string instead of int in API responses
- Omit tool_calls key from API response when empty per OpenAI spec
- Prevent division by zero when micro_batch_size > batch_size in training
- Copy sampler_priority list before mutating in ExLlamaV3
- Normalize presence/frequency_penalty names for ExLlamaV3 sampler sorting
- Restore original chat_template after training instead of leaving it mutated
The tool call response parser only handled JSON-based formats, causing
tool_calls to always be empty for models that use non-JSON formats.
Add parsers for three additional tool call formats:
- Qwen3.5: <tool_call><function=name><parameter=key>value</parameter>
- Mistral/Devstral: functionName{"arg": "value"}
- GPT-OSS: <|channel|>commentary to=functions.name<|message|>{...}
Also fix multi-turn tool conversations crashing with Jinja2
UndefinedError on tool_call_id by preserving tool_calls and
tool_call_id metadata through the chat history conversion.
If --user-data-dir is not set, auto-detect: use ../user_data when
./user_data doesn't exist, making it easy to share user data across
portable builds by placing it one folder up.
Instead of two separate paths (format files vs Chat Template), all
instruction training now uses apply_chat_template() with assistant-only
label masking. Users pick a Jinja2 template from the dropdown or use the
model's built-in chat template — both work identically.
- Remove pre-padding from tokenize() and tokenize_conversation()
- Collate function now right-pads each batch to the longest sequence
- Set tokenizer padding_side to "right" (standard for training)
- Remove dead natural_keys import
- Reduces wasted compute on batches with short sequences
- Aligns with axolotl/unsloth approach
