specula/pyproject.toml at master · hotschmoe/specula · GitHub

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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
[project]
name = "specula"
version = "0.1.0"
description = "Speculative decoding research on Snapdragon X2 Elite Extreme"
requires-python = ">=3.12"
dependencies = [
    # HF model download & conversion (for future EAGLE-3 / DFlash GGUF conversion work)
    "huggingface-hub>=0.24",
    "transformers>=4.51",
    "safetensors>=0.4",
    "sentencepiece>=0.2",
    # Analysis & plotting
    "pandas>=2.2",
    "numpy>=2.0",
    "matplotlib>=3.9",
    "seaborn>=0.13",
    # Notebook
    "jupyterlab>=4.2",
    # Prompt corpus handling
    "datasets>=3.0",
    # HTTP for result uploads / sharing
    "httpx>=0.27",
]

[project.optional-dependencies]
# Only needed if we want to run z-lab's reference DFlash (CUDA) on a different
# machine, or if we run Transformers-based inference locally on CPU for
# ground-truth comparisons against llama.cpp.
reference = [
    "torch>=2.5",
    "accelerate>=0.34",
]

# Phase 5 NPU drafting path. Pinned per docs/npu_scoping.md section 4.
# ORT-QNN bundles a signed QAIRT stack, which is the only path that loads on
# retail Windows on ARM without Secure Boot tricks.
npu = [
    # 2.1.0 bundles QAIRT 2.45 (matches AI Hub default) but its
    # context-binary loader has unrecoverable bugs on the X2E94100
    # driver shipped with this machine — both the file-mapping retry
    # path and embed_mode=1 path crash the interpreter with no
    # traceback. See docs/npu_ort_qnn_version_match.md.
    # Workable path: stay on 1.24.4 (legacy built-in EP, QAIRT 2.42)
    # and pin AI Hub compiles to `--qairt_version 2.42` so the binary
    # matches what 1.24.4 can read.
    "onnxruntime-qnn==1.24.4",
    "qai-hub>=0.48",
    "onnx>=1.17",
    # Hugging Face tooling for downloading pre-exported ONNX + BPE tokenizer.
    # NOTE (historical): torch + optimum were initially excluded because
    # torch had no cp312 win_arm64 wheel. As of 2026-04-23 this is no
    # longer true — PyTorch 2.7+ ships win_arm64 cp312 wheels via
    # download.pytorch.org/whl/cpu/ (not on PyPI). The export toolchain
    # lives in the optional `npu-export` extra below; it's not pulled
    # into `npu` because the inference path doesn't need it.
    "huggingface-hub>=0.24",
    "tokenizers>=0.20",
]

# Local ONNX export pipeline on Windows-on-ARM (and x86, identical pins).
# Verified on the X2E 2026-04-23 — see docs/exporting_on_arm.md.
# Install with: uv pip install -e .[npu-export] \
#   --extra-index-url https://download.pytorch.org/whl/cpu
npu-export = [
    # torch 2.10.0+cpu is the latest with a cp312-cp312-win_arm64 wheel
    # at time of writing. 2.11.0 has no win_arm64 wheel yet.
    "torch==2.10.0",
    "optimum==2.1.0",
    "optimum-onnx==0.1.0",
    "transformers==4.57.6",
    "onnx==1.21.0",
    "onnx-graphsurgeon==0.6.1",
    # onnxruntime CPU-only (no QNN) — the export pipeline only needs
    # CPU-ORT for the cos-vs-source equivalence probes. Pinned to match
    # the QAIRT version on the inference side (see `npu` extra above).
    "onnxruntime==1.24.4",
]

[tool.uv]
# Default to CPU torch on Windows ARM (no CUDA, no ROCm on WoA).
# Reference implementations run on a different machine; Windows-ARM Python is
# for tooling, not inference.

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"