| # EXLLM8 binary format v1 |
|
|
| All integer fields are little-endian. |
|
|
| ## File header |
|
|
| | Field | Type | Value | |
| |---|---|---| |
| | magic | 8 bytes | `EXLLM8\0\0` | |
| | format_version | uint32 | `1` | |
| | tensor_count | uint32 | number of serialized tensors | |
|
|
| ## Tensor record |
|
|
| Each tensor is stored as: |
|
|
| 1. `name_len`: uint16 |
| 2. `ndim`: uint8 |
| 3. `qtype`: uint8 |
| 4. `name`: `name_len` UTF-8 bytes |
| 5. `dims`: `ndim` × uint32 |
| 6. `scale_count`: uint32 |
| 7. `data_bytes`: uint32 |
| 8. `scales`: `scale_count` × float32 |
| 9. `data`: `data_bytes` bytes |
|
|
| ### qtype 1 — row-wise int8 |
|
|
| Used for 2D matrices. If a weight is shaped `[rows, cols]`, one float32 symmetric scale is stored per row and the matrix data is signed int8. |
|
|
| Approximate reconstruction: |
|
|
| ```text |
| weight[row, col] = int8_value[row, col] * scale[row] |
| ``` |
|
|
| Quantized values are limited to `[-127, 127]`. |
|
|
| ### qtype 2 — float16 |
|
|
| Used for the small RMSNorm vectors. `scale_count` is zero and `data` is IEEE-754 binary16. |
|
|
| ## Weight tying |
|
|
| `lm_head.weight` is not stored. It aliases `tok.weight`. The companion manifest declares this alias explicitly. |
|
|
| ## Deployment note |
|
|
| The float scales are an interchange representation. A SH4/no-FPU deployment should translate them into the fixed-point scale representation chosen by the C inference kernel. |
|
|
