Module infer 

Chronos-2 inference engine.

Architecture: encoder-only with alternating TimeSelfAttention + GroupSelfAttention + FFN. Key differences from standard T5:

• Standard Llama RoPE (rotate_half = [-x[half:], x[:half]])
• T5-style RMSNorm (no bias, no mean subtraction)
• No attention scale (scale=1.0 in MHA)
• GroupSelfAttention for batch=1 reduces to position-wise v → o projection

Structs

ChronosModel

ChronosModelBuilder

Fluent constructor for ChronosModel: point it at a GGUF file and a config (from a parsed config.json via config_from), then call build.

InferConfig