anamnesis 0.4.2

Parse any tensor format, recover any precision — framework-agnostic FP8/GPTQ/AWQ/BnB dequantization, NPZ parsing, and PyTorch .pth conversion for Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

# Tests

## Integration Tests

| File | Feature gate | What it tests |
|------|-------------|---------------|
| `cli.rs` | `cli` | CLI binary (`anamnesis`/`amn`): subcommand routing, argument parsing, output format, error messages, `--version` |
| `cross_validation.rs` | — | FP8 dequantization: bit-exact comparison against PyTorch on 7 real models (3 FP8 schemes × 3 scale dtypes) |
| `cross_validation_gptq.rs` | `gptq` | GPTQ dequantization: bit-exact comparison against PyTorch on 4 real models (2 quantizers × 2 bit widths) |
| `cross_validation_awq.rs` | `awq` | AWQ dequantization: bit-exact comparison against PyTorch on 2 real models (AutoAWQ GEMM, 4-bit) |
| `cross_validation_bnb.rs` | `bnb` | BitsAndBytes dequantization: bit-exact comparison against PyTorch on 4 real models (NF4, FP4, double-quant NF4, INT8) |
| `cross_validation_gguf.rs` | `gguf` | GGUF block-quant dequantization: bit-exact comparison against `gguf` Python package (`ggml-org` reference) on 10 kernels from 3 real models (SmolLM2-135M, TinyLlama-1.1B) |
| `cross_validation_pth.rs` | `pth` | PyTorch `.pth` parsing: byte-exact comparison against PyTorch on 3 AlgZoo models (RNN + Transformer, both newer and older ZIP prefix formats) |

## Fixtures

Each `fixtures/<scheme>_reference/` directory contains:

- `generate_<scheme>.py` — Python script that extracts a small slice from a real quantized model, dequantizes with PyTorch/bitsandbytes, and writes a binary fixture file
- `*.bin` — Binary fixture files containing packed input tensors + expected BF16 output

The fixture format is scheme-specific (documented in each generator script). All fixtures are committed to the repo so that `cargo test` works without downloading models or running Python.

| Directory | Models | Slice size | Generator |
|-----------|--------|-----------|-----------|
| `fp8_reference/` | 7 models (EXAONE, Qwen3, Llama, Ministral) | 256x256 | `generate_fp8.py` |
| `gptq_reference/` | 4 models (Falcon3, Llama-3.2) | 256x256 | `generate_gptq.py` |
| `awq_reference/` | 2 models (Llama-3.2, Falcon3) | 256x256 | `generate_awq.py` |
| `bnb_reference/` | 4 models (Llama-3.2 NF4/FP4/INT8) | 4096 elements (NF4/FP4), 256x256 (INT8) | `generate_bnb.py` |
| `npz_reference/` | Gemma Scope 2B SAE (small slice) | 5 F32 arrays | `generate_npz.py` |
| `gguf_reference/` | 3 models (SmolLM2-135M, TinyLlama-1.1B) | 65 536 elements (2 048 / 256 blocks) | `generate_gguf.py` |
| `pth_reference/` | 3 AlgZoo models (see below) | Full model (10–432 params) | `generate_pth_reference.py` |

### AlgZoo `.pth` Fixtures

The `.pth` fixtures are real models from [AlgZoo](https://github.com/alignment-research-center/alg-zoo)
(Alignment Research Center, MIT-0 license), downloaded from their public
GCS bucket (`gs://arc-ml-public/alg/zoo/`). All are F32, tiny (2–3.5 KB):

| Fixture | AlgZoo source | Architecture | ZIP prefix | Params |
|---------|--------------|-------------|-----------|--------|
| `algzoo_rnn_small.pth` | `2nd_argmax_2_2_1073741824_2` | RNN (3 tensors) | `archive/` (newer PyTorch) | 10 |
| `algzoo_transformer_small.pth` | `longest_cycle_2_3_1073741824_0` | Transformer (7 tensors) | `archive/` (newer PyTorch) | 50 |
| `algzoo_rnn_blog.pth` | `one_layer_16_hidden_94_acc_2nd_argmax` | RNN (3 tensors) | `{model_name}/` (older PyTorch) | 432 |

Reference JSON manifests are generated by `generate_pth_reference.py` (requires
`torch` and `safetensors`). The Rust tests compare tensor names, shapes, dtypes,
and **raw bytes** against the Python-generated reference (0 byte difference expected).

## Regenerating Fixtures

Fixtures only need regeneration if the dequantization formula changes. To regenerate:

1. Download test models via `hf-fm` (see generator scripts for model IDs)
2. Run the Python generator: `python tests/fixtures/<scheme>_reference/generate_<scheme>.py`
3. Commit the updated `.bin` files

## Running Tests

```bash
# All tests (requires all features)
cargo test --all-features

# Specific scheme
cargo test --features gptq --test cross_validation_gptq

# Release mode with AVX2 (for timing comparison)
RUSTFLAGS="-C target-cpu=native" cargo test --release --all-features -- --nocapture
```