Crate combs_formats 

combs-formats

File-format adapter layer. The runtime and model crates never touch file formats directly; they go through the ModelSource trait (the LiteRT-LM ModelResources equivalent). Phase 1 ships the [safetensors] adapter (HuggingFace config.json + model.safetensors, mmap-backed, zero-copy views). GGUF / ONNX / litertlm adapters plug in here later by implementing the same trait.

Modules

quants

Reference CPU dequantizers for GGUF quant formats. These scalar implementations are the harmony reference that the fused GPU kernels in combs-models validate against: every kernel is tested against a portable reference.

Structs

AttentionPattern

Attention RoPE/scale settings that vary by layer type (Gemma2/3): pattern-th layers are “global” (full attention, rope_theta); the rest are “local” (sliding-window attention, rope_local_theta).

GgufSource

A parsed GGUF file.

ModelMetadata

Architecture + hyperparameter description of a model, format-agnostic.

QuantTensor

A quantized tensor’s packed bytes, exactly as stored in the file.

SafetensorsSource

ModelSource over a HuggingFace-format directory.

SamplerConfig

Default sampler parameters, typically from generation_config.json.

SectionInfo

A section directory entry.

TensorReader

A lazy view over one tensor’s raw bytes inside a ModelSource.

TfliteSource

A TFLite model file as a ModelSource (Gemma-family exports).

TokenizerSpec

Where to find the tokenizer and which special tokens it defines.

VisionConfig

Vision-encoder hyperparameters parsed from config.json::vision_config (plus top-level scale_factor / image_token_id).

Enums

Activation

MLP activation function, parsed from hidden_act/hidden_activation. The tanh-approximation family (gelu_pytorch_tanh, gelu_new, gelu_fast) all map to Activation::GeluTanh.

FormatError

Errors produced by format adapters.

QuantFormat

GGUF quant formats with a native device kernel (see combs-models).

RopeScaling

RoPE frequency scaling, parsed from HF rope_scaling (accepts both the modern rope_type and the legacy type key). Table math lives in combs-models::rope; this is parse-only.

TensorDtype

Element dtypes supported by the loaders.

Traits

ModelSource

The central adapter trait: a source of model weights + config, independent of the on-disk format (LiteRT-LM ModelResources equivalent).

Functions

ensure_tokenizer_json_from_spm

Ensures a tokenizer.json exists for a SentencePiece .model file, converting + caching it next to the source on first use. Returns the path (idempotent: an existing cache is reused).

litertlm_read_sections

Reads the section directory of a .litertlm file header.

open_model_source

Opens any supported model path: a .gguf file, or a directory in the HuggingFace safetensors layout. This is the single entry point the CLI, FFI and server use — format detection lives here.

spm_added_tokens

Parses a SentencePiece model and returns its special-token map (id → string), for TokenizerSpec.added_tokens.

Type Aliases

Result

Convenient result alias for this crate.