Implementation:Sail sg LongSpec Pip Requirements Install

Knowledge Sources	LongSpec
Domains	DevOps, Environment_Management
Last Updated	2026-02-14 05:00 GMT

Overview

Concrete tool for installing LongSpec dependencies via pip requirements files and configuring WandB authentication for experiment tracking.

Description

The environment setup uses standard pip install -r requirements.txt with two separate requirements files:

• longspec/train/requirements.txt (22 dependencies): Full training stack including DeepSpeed, Fairscale, PEFT, Liger Kernel
• longspec/test/requirements.txt (7 dependencies): Minimal inference dependencies

Additionally, WandB login is required for training experiment tracking.

This is an External Tool Doc — it documents CLI tools rather than Python APIs.

Usage

Run once before starting training or inference.

Code Reference

Source Location

• Repository: LongSpec
• File (Training): longspec/train/requirements.txt
• Lines: L1-22
• File (Inference): longspec/test/requirements.txt
• Lines: L1-7
• File (Launch script): longspec/train/train.sh
• Lines: L1-2

Signature

# Training environment setup:
pip install -r longspec/train/requirements.txt
pip install deepspeed  # Installed separately
wandb login            # Authenticate experiment tracking

# Inference environment setup:
pip install -r longspec/test/requirements.txt

Import

# N/A - this is a CLI tool, not a Python import

I/O Contract

Inputs

Name	Type	Required	Description
Python environment	System	Yes	Python 3.8+ with pip
CUDA toolkit	System	Yes	CUDA-compatible GPU and drivers for flash_attn and triton
WandB API key	Credential	Yes (training)	Weights & Biases authentication token

Outputs

Name	Type	Description
Installed packages	System	All pinned dependencies available for import
WandB auth	Config file	~/.netrc or similar WandB credential store

Usage Examples

Training Setup

# Create and activate virtual environment
python -m venv longspec_env
source longspec_env/bin/activate

# Install training dependencies
pip install -r longspec/train/requirements.txt

# Install DeepSpeed (separate due to build requirements)
pip install deepspeed

# Configure experiment tracking
wandb login

Key Training Dependencies

accelerate==1.0.1
apex==0.9.10dev
bitsandbytes==0.45.5
datasets==2.19.1
fairscale==0.4.13
flash_attn==2.6.3
liger_kernel==0.3.1
omegaconf==2.3.0
peft==0.13.2
torch==2.6.0
transformers==4.51.1
triton==3.2.0
wandb==0.19.11

Inference Setup

pip install -r longspec/test/requirements.txt

Related Pages

Implements Principle

• Principle:Sail_sg_LongSpec_Environment_Setup