0.1.0

What's Changed

• reinstate cuda rmsnorm (much faster in fp16/awq) + ct2 enc/dec config by @vince62s in #167
• [patch] remove dummy_load, move gpu_ranks warning out of TrainingConfig by @francoishernandez in #168
• fix batch inference by @vince62s in #169
• Code clean-ups by @vince62s in #171
• 120 columns makes more sense on modern screens by @vince62s in #176
• refactor transformer decoder and revamp the left padding attention mask by @vince62s in #178
• Major refactoring of convert HF by @francoishernandez in #156
• [patch] handle self_attn_backend edge case by @francoishernandez in #180
• hotfix post #178 by @vince62s in #181
• fix update vocab param loading by @vince62s in #184
• remove verbosity at validation/scoring by @vince62s in #185
• [patch] Add missing is_train kwarg in tokenize_id by @francoishernandez in #187
• Hugging face dataset streaming support by @vince62s in #177
• misc fixes by @vince62s in #192
• Gemma2 support by @francoishernandez in #160
• [convert_HF] handle special tokens defined in tokenizer_config.json by @francoishernandez in #196
• patch max_length handling in tokenize_id by @francoishernandez in #197

Full Changelog: 0.0.3...0.1.0