adaptfm-quant-dflash/pyproject.toml at main · nota-github/adaptfm-quant-dflash · 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
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
# edgefm-eqc — Efficient Qwen Competition reproduction (single unified venv).
#
# Historically this repo used FIVE mutually-incompatible venvs. They are now
# unified into ONE, resolvable environment. The key facts that make a single
# venv possible (verified with `uv pip install --dry-run` against each stage's
# real, installed versions — NOT the upstream pyproject defaults):
#
#   * vllm 0.22.1   requires  transformers (!=5.0..5.4, !=5.5.0, >=4.56), torch==2.11.0
#   * modelopt 0.44 requires  transformers ~5.5.x  (caps the upper bound)
#   * DFlash train (SpecForge, HF backend, NO sglang) ran on torch 2.11 / tf 5.10.2
#   * lm-eval 0.4.11 (eval harness, HTTP client) has no torch/transformers pin
#
#   => the intersection: torch 2.11.0+cu130 / transformers 5.5.x / triton 3.6 /
#      vllm 0.22.1 / modelopt 0.44 satisfies EVERY stage at once.
#
# The transformers upper bound (<5.6) comes ONLY from modelopt (QAD packing /
# training). If you never re-quantize or re-train QAD, drop the `qad` extra and
# relax `transformers` to `>=5.6` for the newer runtime line.
#
# sglang is intentionally EXCLUDED: it hard-pins transformers==4.57.1 which
# conflicts with Qwen3.5 (needs 5.x). SpecForge trains on its `hf` backend, so
# sglang is not needed — install SpecForge itself with `--no-deps` (see README).
#
# Setup:
#   uv sync                 # runtime only (serve + data-gen + eval + DFlash-eval)
#   uv sync --all-extras    # + QAD training/packing + DFlash (SpecForge) training
#   uv pip install -e src/SpecForge --no-deps   # SpecForge training framework (no sglang)

[project]
name = "edgefm-eqc"
version = "0.1.0"
description = "Efficient Qwen Competition reproduction: QAD INT4 + DFlash speculative decoding on Qwen3.5-4B."
readme = "README.md"
requires-python = ">=3.12,<3.13"
license = "Apache-2.0"
authors = [
    { name = "Jaeyeon Kim", email = "jaeyeon.kim@nota.ai" },
    { name = "Bo-Kyeong Kim", email = "bokyeong.kim@nota.ai" },
    { name = "Jewon Lee", email = "je1lee.research@gmail.com" },
]

dependencies = [
    # --- torch stack (cu130 wheels; see [tool.uv.sources]) ---
    "torch==2.11.0",
    "torchvision==0.26.0",
    "torchaudio==2.11.0",

    # --- inference runtime (serving + data-gen + DFlash spec-decode eval) ---
    "vllm==0.22.1",
    # Upper bound is modelopt's; !=5.5.0 is vllm's. See header note.
    "transformers>=5.5.1,<5.6",
    "compressed-tensors==0.15.0.1",
    "flashinfer-python",

    # --- serving proxy (src/serving/: FastAPI -> spawns `vllm serve`) ---
    "fastapi",
    "uvicorn[standard]",
    "httpx",
    "pyyaml",

    # --- data generation (src/data_generator/) ---
    "openai",
    "pyarrow",
    "huggingface-hub",

    # --- eval harness (src/eval_harness/: quality gates over HTTP) ---
    # Gates are mmlu_pro / ifeval / gpqa_diamond only — none use math-verify, so
    # the grader's `[math]` extra is dropped: its antlr4-runtime==4.11 pin clashes
    # with modelopt->omegaconf's antlr4-runtime==4.9.*, and it changes no scores
    # for these three tasks. Add `math` back only if you score a math task.
    "lm-eval[api,ifeval]==0.4.11",
    "datasets",
    "accelerate",

    # --- common utilities used across scripts ---
    "numpy",
    "tqdm",
    "safetensors",

    # --- in-repo editable packages (see [tool.uv.sources]) ---
    "dflash",           # ./src/dflash_runtime — DFlash `method: dflash` for vllm spec-decode
    "eqc-vllm-plugin",  # ./src/vllm_plugin — W4A16 draft load + SWA plugin
]

[project.urls]
Repository = "https://github.com/nota-github/edgefm-eqc-final"

[project.optional-dependencies]
# Quantization-Aware Distillation: training + compressed-tensors packing (src/qad/).
# modelopt 0.44 is what caps transformers to <5.6 (see header).
qad = [
    "nvidia-modelopt[hf]>=0.44",
    "onnx>=1.16",
    "onnxruntime>=1.18",
    "bitsandbytes>=0.43",
    "flash-linear-attention",  # GatedDeltaNet kernels for Qwen3.5
    "causal-conv1d",           # causal conv kernels for Qwen3.5
    "deepspeed>=0.15",         # ZeRO-2 for seq-16384 full-vocab logits
    "wandb",
]

# DFlash drafter training (src/SpecForge/, `hf` target backend, NO sglang).
# SpecForge itself is installed separately: `uv pip install -e src/SpecForge --no-deps`.
train = [
    "yunchang",
    "tensorboard",
    "pydantic",
    "psutil",
    "ninja",
    "packaging",
    "wandb",
]

# This repo root is a collection of loose scripts/dirs, not an importable
# package — tell uv not to try to build/install the root itself.
[tool.uv]
package = false
# CUDA-only repo: resolve/lock for Linux only. vllm/deepspeed/bitsandbytes/triton
# ship no macOS wheels, so a universal lock would fail on the darwin split.
environments = ["sys_platform == 'linux'"]

# --- uv: local editable packages + PyTorch cu130 index ------------------------
[tool.uv.sources]
torch = { index = "pytorch-cu130" }
torchvision = { index = "pytorch-cu130" }
torchaudio = { index = "pytorch-cu130" }
dflash = { path = "src/dflash_runtime", editable = true }
eqc-vllm-plugin = { path = "src/vllm_plugin", editable = true }

[[tool.uv.index]]
name = "pytorch-cu130"
url = "https://download.pytorch.org/whl/cu130"
explicit = true