Crate spark_model Copy item path Source pub use vision_item::VisionItem ;engine Inference engine — generate loop for a single request. factory Model factory: builds the right model from config + weights. forward Vendor-agnostic per-layer forward orchestration. layer Composable transformer layer traits (SDD). layers lora Startup-static PEFT LoRA adapter: remap/validate/pack into the
fixed-address rank-padded pool. v0 = one adapter, slot 0, always on. mistral_loader Weight loader for Mistral Small 4 (MLA + MoE architecture). model Generic transformer model. mtp_layout Checkpoint-level detection of MTP / next-token-prediction weights. precision_schedule Per-layer + per-tensor precision overrides (C.3, 2026-04-25). preflight Pre-flight weight-store / config consistency checks. quant_format Weight-quantization format abstraction. seq_state_reserve Per-sequence device state the serve must RESERVE, computed before the model exists. speculative Speculative decoding abstraction (SDD). ssm_reserve SSOT for the Phase-C decode-rollback ring depth. tp_shard Tensor-parallel weight sharding helpers. traits Model trait (SDD: single trait, multiple implementations possible). video_decode_ffmpeg Frame extraction for real-world containers, via ffmpeg. video_preprocess Video → patch tensor, the temporal sibling of crate::vision_preprocess vision_item The unit of vision input handed to the model. vision_preprocess CPU-side image preprocessing for Qwen3-VL vision inputs. weight_loader Weight loading traits and per-model loader implementations. weight_map Weight name mapping from HuggingFace safetensors to typed layer structures. model_type_ships_vanilla_norm_weights The dispatch predicate itself, on the bare model_type, so it is unit-testable
without constructing a full ModelConfig. requires_single_chunk_prefill Must chunked prefill run as a SINGLE chunk for this model? ships_vanilla_norm_weights True when the checkpoint ships HF-vanilla RMSNorm weights — i.e. the norm
weight is used as out = x * w / rms, not Qwen3-Next’s offset-from-1
out = x * (1 + w) / rms.