[WIP] Eurus

[AlexPiche]

Tests

• Reproduce large DeepSeek run

Changes

• Handle tensor and pipeline parallelism more gracefully.
• Easily configurable task template and system prompt.
• Train on the PRIME RL dataset.
• Remove base_url: str | list[str] for base_url: str in the LLM class.
• Comment out symbolic_equal since it was hanging and it is not compatible with our current version of hydra.
• Improved and clean stats.
• Implemented a is_debug_mode() in VLLMServiceManager
• Add multiprocess support for dataset.map in finetune/data.py
• Rename datasets to datasets_list to avoid conflict with the datasets library.
• Normalize the loss to be over sequences and not tokens, i.e. replace masked_mean by masked_sum.
• Refactor load_datasets.
• Do not count the token of the prompt at every call.
• Use the vllm tokenization for llm_call.prompt_length_tokens and llm_call.output_length_tokens
• Assert that LLM does not produce more than cfg.llm.parameters.max_tokens
• Truncate from the left long prompts

[rizar]

Start looking pretty good! you can merge once you address all comments.