Configuration

The ModelConfig dataclass drives the whole engine: family defaults, head_dim / GQA derivation, quantization detection, and loading from config.json. It is family-dispatched on model_type (Qwen3, Llama 3.x, Mistral).

anllm.config.ModelConfig dataclass

Shared configuration for the supported causal-LM architectures.

This single dataclass covers both Qwen3 and the Llama family. Family specific defaults land in :meth:preferred_defaults, and any field absent from a checkpoint's config.json falls back to those defaults so that a Llama config.json (which omits Qwen3-only keys) still parses cleanly.

anllm.config.Qwen3Config dataclass

Bases: ModelConfig

Backwards-compatible alias for :class:ModelConfig.

Calls that construct Qwen3Config directly (tests included) keep working. Any fields Qwen3 doesn't specify fall through to the shared defaults, exactly as if built through the family-aware loader.