Bailing (Ling) Examples

This directory contains example scripts for Ling 2.0 MoE language models by inclusionAI.

Ling 2.0 uses a high-sparsity Mixture of Experts (MoE) architecture with sigmoid routing, QK-Norm, and Half RoPE.

Model	HF ID	Architecture	Params	Active Params
Ling-flash-2.0	inclusionAI/Ling-flash-2.0	MoE (256 experts, top-8)	100B	6.1B
Ling-flash-base-2.0	inclusionAI/Ling-flash-base-2.0	MoE (256 experts, top-8)	100B	6.1B
Ling-mini-2.0	inclusionAI/Ling-mini-2.0	MoE (256 experts, top-8)	16B	1.5B
Ling-mini-base-2.0	inclusionAI/Ling-mini-base-2.0	MoE (256 experts, top-8)	16B	1.5B

Workspace Configuration

All scripts use a WORKSPACE environment variable for the base directory. Default: /workspace.

export WORKSPACE=/your/custom/path

Directory structure:

• ${WORKSPACE}/models/ - Converted checkpoints
• ${WORKSPACE}/results/ - Training outputs

Checkpoint Conversion

See conversion.sh for checkpoint conversion examples.

Import HF → Megatron

python examples/conversion/convert_checkpoints.py import \
    --hf-model inclusionAI/Ling-flash-2.0 \
    --megatron-path ${WORKSPACE}/models/Ling-flash-2.0 \
    --trust-remote-code

Export Megatron → HF

python examples/conversion/convert_checkpoints.py export \
    --hf-model inclusionAI/Ling-flash-2.0 \
    --megatron-path ${WORKSPACE}/models/Ling-flash-2.0/iter_0000000 \
    --hf-path ${WORKSPACE}/models/Ling-flash-2.0-hf-export \
    --trust-remote-code

Round-trip Validation

python -m torch.distributed.run --nproc_per_node=8 \
    examples/conversion/hf_megatron_roundtrip_multi_gpu.py \
    --hf-model-id inclusionAI/Ling-flash-2.0 \
    --megatron-load-path ${WORKSPACE}/models/Ling-flash-2.0/iter_0000000 \
    --tp 1 --ep 8 \
    --trust-remote-code

Inference

See inference.sh for text generation with:

• Hugging Face checkpoint (inclusionAI/Ling-flash-2.0)
• Imported Megatron checkpoint (after conversion.sh import)
• Exported HF checkpoint (after conversion export)

The default parallelism for 8 GPUs is --tp 2 --ep 4. TP×PP×EP must equal --nproc_per_node.

Note: --tp 1 --ep 8 works for conversion round-trip but may cause issues during autoregressive inference with single-token batches (empty token dispatch to some EP ranks). Use --tp 2 --ep 4 for inference.

Note: All Ling 2.0 models use custom HuggingFace code, so --trust-remote-code is required for conversion and inference.