MakazhanAlpamys/Soup v0.10.8

on GitHub

What's New

Bug Fixes

GRPO — chat template + batch size (CPU)

• Fixed ValueError / RuntimeError: expanded size of tensor (8) must match existing size (0) — trl's GRPOTrainer calls apply_chat_template() on conversational prompts, but models without a chat template caused a crash. Now a default template is set automatically.
• Fixed generation_batch_size must be divisible by num_generations — batch_size is now ensured to be >= num_generations.

PPO — dataset tokenization

• Fixed ValueError: You should supply an encoding that includes input_ids, but you provided ['prompt_text'] — datasets are now tokenized (input_ids + attention_mask) before passing to trl's experimental PPOTrainer.

Tests

• 666 tests passing
• 5 new tests for chat_template and tokenization fixes

Install / Upgrade

pip install --upgrade soup-cli