per-sample media budgets in processor; per-sample MRoPE position ids for batched vision inputs

per-sample media budgets in processor; per-sample MRoPE position ids for batched vision inputs

sync offline system prompts with training data: no_thinking -> no system message, deep_thinking -> <think>/<answer> tags

Align tokenizer with the standard (add silence/response tokens, embed chat_template)

fix: apply query RoPE in cross-attention when reusing cached vision KV during decode

Align processor with MOSS-VL-Instruct-0708 (Fast image/video processor)

Upload README.md with huggingface_hub

Update final v3 quantization benchmarks

Add files using upload-large-folder tool

Add files using upload-large-folder tool

initial commit