offline-packing-env-vars
OfficialEnsure safe, padding-free packing in training.
Software Engineering#debugging#env-vars#packing#training-pipeline#cu_lengths#llava-onevision2#two-stage-gating
AuthorEvolvingLMMs-Lab
Version1.0.0
Installs0
System Documentation
What problem does it solve?
This bilingual guide explains how to configure and debug the OFFLINE_PACKING_BMR and OFFLINE_PACKED_DATA environment variables that control LLaVA-OneVision2 training-side packing, clarifying why both must be enabled together to achieve padding-free attention.
Core Features & Use Cases
- Gate duality: documents the two orthogonal gating steps (Data-layer BMR and Batch-layer data packing) and their impact on cu_lengths and sub-sample encoding.
- VQA gate deprecation warning: highlights that OFFLINE_PACKING_VQA is dead code and should not be relied upon.
- Practical guidance: provides end-to-end workflow, validation steps, and diagnostics for real-world training pipelines using packed shards.
Quick Start
Set both OFFLINE_PACKING_BMR=1 and OFFLINE_PACKED_DATA=1 in your training shell, ensure MBS=1, and run your packed dataset training.
Dependency Matrix
Required Modules
None requiredComponents
Standard package💻 Claude Code Installation
Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.
Please help me install this Skill: Name: offline-packing-env-vars Download link: https://github.com/EvolvingLMMs-Lab/LLaVA-OneVision-2/archive/main.zip#offline-packing-env-vars Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
Agent Skills Search Helper
Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.