Skip to content

loader_format_scope.md

Latest commit

 

History

History
69 lines (51 loc) · 2.89 KB

loader_format_scope.md

File metadata and controls

69 lines (51 loc) · 2.89 KB
Raw

Loader Format Scope

This document records the loader format decisions for the supported local file boundaries.

Safetensors Supported Subset

LIS supports a narrow safetensors subset:

  • local single-file path sources
  • 8-byte little-endian header length
  • JSON object header with top-level tensor entries
  • tensor fields: dtype, shape, and data_offsets
  • ignored __metadata__ object
  • supported dtypes: F32, F16, BF16, I32, and U32
  • supported shapes: rank 1 through LIS_TENSOR_MAX_RANK, nonzero dimensions, contiguous byte-size validation through tensor helpers
  • data offsets relative to the safetensors data section after the JSON header

Unsupported safetensors cases return explicit status values such as LIS_STATUS_FORMAT, LIS_STATUS_UNSUPPORTED_DTYPE, LIS_STATUS_UNSUPPORTED_SHAPE, LIS_STATUS_SHAPE_MISMATCH, or LIS_STATUS_IO.

Tensor names are accepted as simple unescaped JSON strings. Llama 3.x weight-name semantic validation is deferred until model mapping work defines the exact required tensor table.

The precision policy does not change safetensors parsing: F16 and BF16 bytes remain native in loaded tensor views. For HuggingFace-local Llama imports, the later mapping layer requires every mapped weight tensor to match config.weight_dtype; mixed per-tensor dtype artifacts are documented unsupported scope and are not normalized by the loader.

Llama Config Scope

LIS includes a narrow in-memory Llama-style JSON config parser. It maps the minimal fields needed by lis_model_metadata:

  • model_type
  • num_hidden_layers
  • hidden_size
  • intermediate_size
  • num_attention_heads
  • num_key_value_heads
  • head_dim
  • vocab_size
  • rope_theta
  • torch_dtype
  • max_position_embeddings

The parser accepts model_type values llama and llama3 as the initial Llama-style decoder-only boundary, but only for plain RoPE configs that use rope_theta. Configs containing rope_scaling or a non-default rope_type are rejected with LIS_STATUS_UNSUPPORTED because LIS does not implement those RoPE variants. GPT-2, Mistral, and GPT-OSS config parsing remain unsupported extension paths.

PyTorch Compatibility Scope

LIS does not implement PyTorch checkpoint loading or broad .pt, .pth, or .bin compatibility. Those paths are detected as LIS_MODEL_FORMAT_PYTORCH_UNSUPPORTED and return LIS_STATUS_UNSUPPORTED_FORMAT.

Future PyTorch-exported compatibility must define a specific export path, tensor naming convention, dtype scope, and comparison tests before implementation. The current implementation contains no PyTorch comparison cases or parity claims.

Save Support Decision

Save support is deferred. LIS implements no canonical save, safetensors save, or PyTorch-compatible export path.

Future save work must first choose and document one explicit semantic:

  • LIS canonical internal format save
  • limited safetensors save
  • limited PyTorch-compatible export