mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-06-27 18:04:59 +08:00
wassname
70aa6aa96b
Fire the paper sweep as independent H100/A100-80 containers instead of serial pueue runs. One Volume caches model + svd + out/; train.py runs unmodified (torch 2.7 + Dao flash-attn wheel, code mounted at runtime). Verified: vanilla 60-step reproduces the local baseline. Skill at ~/.claude/skills/modal documents the patterns. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
69 lines
2.8 KiB
TOML
69 lines
2.8 KiB
TOML
[project]
|
|
name = "vgrout"
|
|
version = "0.1.0"
|
|
description = "vGROUT: vector gradient routing against reward hacking (Nanda's LeetCode benchmark)"
|
|
requires-python = ">=3.13,<3.14" # pinned cp313 wheels (causal-conv1d, flash-attn)
|
|
dependencies = [
|
|
"torch>=2.4",
|
|
# transformers>=4.58 has Qwen3.5 (model_type=qwen3_5, gated-delta-net).
|
|
# Per HF card: install from main if 4.58 not yet released. We pin to main
|
|
# via [tool.uv.sources] below; the version spec here is just a floor.
|
|
"transformers>=4.58.0.dev0",
|
|
"einops>=0.8",
|
|
"jaxtyping>=0.2",
|
|
"beartype>=0.18",
|
|
"loguru>=0.7",
|
|
"polars>=1.0",
|
|
"tabulate>=0.9",
|
|
"tyro>=0.8",
|
|
"tqdm>=4.66",
|
|
"numpy<2.0",
|
|
"datasets>=3.0",
|
|
"huggingface_hub>=0.24",
|
|
"wandb>=0.18",
|
|
"peft>=0.13",
|
|
"flash-linear-attention>=0.5.0",
|
|
# Qwen3.5's gated-delta-net fast path needs causal-conv1d's compiled CUDA
|
|
# kernel. The Dao-AILab repo publishes prebuilt wheels keyed by (cuda, torch,
|
|
# python, abi). The matching wheel for our cu12 + torch 2.8 + cp313 stack is
|
|
# pinned in [tool.uv.sources] so `uv sync` doesn't try to compile from source.
|
|
"causal-conv1d",
|
|
# Flash-attention for the regular self_attn blocks. v2.8.3 is the first
|
|
# release with Blackwell sm_120 kernels (consumer RTX PRO 6000). Pinned to
|
|
# mjun0812 prebuilds — see [tool.uv.sources] below.
|
|
"flash-attn",
|
|
"modal>=1.4.3",
|
|
]
|
|
|
|
[project.optional-dependencies]
|
|
gpu = [
|
|
"vllm>=0.10",
|
|
]
|
|
|
|
[tool.ruff.lint]
|
|
ignore = ["F722"] # jaxtyping shape strings
|
|
|
|
[build-system]
|
|
requires = ["setuptools>=68"]
|
|
build-backend = "setuptools.build_meta"
|
|
|
|
[tool.setuptools.packages.find]
|
|
where = ["src"]
|
|
|
|
[tool.uv]
|
|
exclude-newer = "2026-05-23"
|
|
|
|
[tool.uv.sources]
|
|
# Qwen3.5 (qwen3_5 model_type, gated-delta-net) lands in transformers main; pin
|
|
# until 4.58 release. v5.7.0 changelog note: "incorrect cached forward behavior
|
|
# in Qwen3.5's gated-delta-net linear attention" — fixed on main.
|
|
transformers = { git = "https://github.com/huggingface/transformers.git", rev = "main" }
|
|
# Prebuilt CUDA wheel for our exact stack: cu12 + torch 2.8 + cp313 + cxx11abi.
|
|
# Verified Blackwell sm_120 dispatch on the RTX PRO 6000. If torch/python is
|
|
# bumped, find the new match at https://github.com/Dao-AILab/causal-conv1d/releases.
|
|
causal-conv1d = { url = "https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.6.2.post1/causal_conv1d-1.6.2.post1+cu12torch2.8cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" }
|
|
# flash-attn 2.8.3 prebuilt for cu128 + torch 2.8 + cp313 (Blackwell sm_120). If
|
|
# torch/python is bumped, walk https://github.com/mjun0812/flash-attention-prebuild-wheels/releases
|
|
# for the matching tag string in the wheel filename.
|
|
flash-attn = { url = "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.16/flash_attn-2.8.3%2Bcu128torch2.8-cp313-cp313-linux_x86_64.whl" }
|