mac-cloud-smoke

SSH（本机）

~/.ssh/config：

sshconfig

Host autodl
  HostName connect.bjb1.seetacloud.com
  User root
  Port 25458
  IdentityFile ~/.ssh/autodl_ed25519
  StrictHostKeyChecking no
  UserKnownHostsFile /dev/null

用法：

•ssh autodl
•rsync -avP xxx autodl:~/

AutoDL 一键只做公钥：mac-cloud-smoke/autodl.sh

云端 uv（通用）

bash

UV=/root/miniconda3/bin/uv
PY=/root/miniconda3/bin/python
$UV venv -p $PY --system-site-packages ~/venv
. ~/venv/bin/activate

HF 慢就开 mirror：

bash

export HF_ENDPOINT=https://hf-mirror.com
export HF_HOME=$HOME/hf

HF 还是慢：本机先下到 HF_HOME=...，再 rsync -avP hf/ autodl:~/hf/，云端强制 *_OFFLINE=1。

prefix-linear-attention: based_fda smoke（JRT 360M）

坑：

•transformers 不会展开 ~：checkpoint/tokenizer 路径用绝对路径

bash

mkdir -p ~/pla-logs
cd ~/prefix-linear-attention/lm-eval-harness
export PYTHONPATH=~/prefix-linear-attention

export TORCH_COMPILE_DISABLE=1 TORCHDYNAMO_DISABLE=1
export LD_LIBRARY_PATH=/root/miniconda3/lib/python3.12/site-packages/torch/lib:${LD_LIBRARY_PATH:-}
export WANDB_DISABLED=true
export PLA_FUTURE_SEED=0

python -m lm_eval \
  --model jrt_lm \
  --model_args checkpoint_name=/root/hf-local/JRT-360M-30B,tokenizer=/root/hf-local/gpt2,arch=JRT \
  --tasks based_fda --decode_mode default_left_pad \
  --batch_size 1 --limit 50 \
  --output_path ~/pla-logs/baseline_limit50.json

Future-Seed（只 prefill/context）：

bash

export PLA_FUTURE_SEED=1 PLA_FUTURE_SEED_ALPHA=1.0 PLA_FUTURE_SEED_LAYER_START=0
python -m lm_eval \
  --model jrt_lm \
  --model_args checkpoint_name=/root/hf-local/JRT-360M-30B,tokenizer=/root/hf-local/gpt2,arch=JRT \
  --tasks based_fda --decode_mode default_left_pad \
  --batch_size 1 --limit 50 \
  --output_path ~/pla-logs/future_seed_a1_limit50.json

当前结果（2026-02-07, limit50）：

•baseline contains=0.24
•+Future-Seed(alpha=1.0) contains=0.20

flash-attn（4090 / torch2.5.1+cu124）最终坑

•pip install flash-attn 基本必炸：用 sdist build wheel
•需要 3 个 wheel：flash_attn + dropout_layer_norm + fused_dense_lib
•4090 需要 sm89；dropout_layer_norm 默认不带，要手动加
•build 容易 OOM：MAX_JOBS=12
•runtime 缺 libc10.so：加 LD_LIBRARY_PATH=...torch/lib
•
torch2.5.1 没 torch.library.wrap_triton：补丁
- •找到路径：python -c 'import flash_attn.ops.triton as t; print(t.__file__)'
- •改 .../site-packages/flash_attn/ops/triton/__init__.py：

python

import torch
if not hasattr(torch.library, "wrap_triton"):
    torch.library.wrap_triton = lambda fn: fn

swanlab（不用 wandb）

bash

$UV pip install -U swanlab
export SWANLAB_API_KEY=...
swanlab login --api-key "$SWANLAB_API_KEY"