Lance-3B-Video AWQ INT4

4-bit AWQ-quantized variant of bytedance-research/Lance (the Lance_3B_Video checkpoint), the unified multimodal model for image and video generation, editing, and understanding.

File-size reduction: 28.4 GB → 6.0 GB (4.7×) Inference VRAM (LLM only, bf16 activations): ~14 GB → ~7 GB

What was quantized

Lance is not a standard HuggingFace transformers model. It's a custom PreTrainedModel whose decoder is a Qwen2.5-VL with parallel Mixture-of-Tasks experts: every transformer layer has both mlp/self_attn.* (understanding path) AND mlp_moe_gen/self_attn.*_moe_gen (generation expert). Tokens route through one or the other depending on type (text vs VAE latent).

504 Linear modules quantized (360 with AWQ-style activation-aware scale fusion, 144 with plain per-group min-max because they have no clean fuse target — o_proj and down_proj).

Calibration

The two-task mix is the minimum that exercises both MoT paths — pure understanding leaves the _moe_gen weights without activation data, and AWQ falls back to plain min-max for those.

Quantization recipe

• 4-bit asymmetric per-group (group_size = 128, packed two-nibbles-per-byte)
• AWQ alpha grid search ∈ {0.0, 0.05, …, 1.0} per fusion group, MSE-minimizing on synthetic Gaussian inputs scaled by per-channel mean-abs activation
• Scale s normalised so geomean(s) ≈ 1 to avoid drift, then fused as norm.weight /= s and consumer.weight *= s

Fusion groups per decoder layer:

• input_layernorm → [q_proj, k_proj, v_proj]
• input_layernorm_moe_gen → [q_proj_moe_gen, k_proj_moe_gen, v_proj_moe_gen]
• post_attention_layernorm → [mlp.gate_proj, mlp.up_proj]
• post_attention_layernorm_moe_gen → [mlp_moe_gen.gate_proj, mlp_moe_gen.up_proj]

o_proj and down_proj are post-nonlinearity — no clean scalar fuse — quantized with plain per-group min-max.

File layout

Lance_3B_Video-AWQ-INT4/
├── awq_state_dict.safetensors   # 6.02 GB: packed INT4 + bf16 pass-through
├── awq_meta.json                # per-weight scheme + group_size + shape
└── README.md

awq_state_dict.safetensors keys per quantized linear:

• <name>.qweight — uint8 [out, in // 2], two INT4 nibbles per byte
• <name>.scales — bf16 [out, in // group_size]
• <name>.zeros — uint8 [out, in // group_size]
• <name>.bias — bf16 [out] (if the original linear had one)

Plus all pass-through (non-quantized) tensors at bf16, including the AWQ-rescaled RMSNorm weights — load these in place of the originals.

How to use

The Lance source ships a custom Lance PreTrainedModel (in github.com/bytedance/Lance). vLLM / TensorRT-LLM / AutoAWQ do NOT support it directly, since it's not a standard HF model. The cleanest path is a runtime swap-in:

import torch
from safetensors import safe_open

# 1) Build Lance normally (e.g. via the official inference_lance.py)
# 2) Swap every targeted nn.Linear in `language_model.*` for WQLinearINT4
# 3) Stream the AWQ packed buffers from awq_state_dict.safetensors

# A reference WQLinearINT4 module + swap-in loader is at:
#   https://github.com/Reza2kn/lance-quant/blob/main/patches/quantized_linear.py

A full reproduction (calibration scripts + the WQLinearINT4 implementation + a run_quant_eval.py that loads this checkpoint into Lance) is at: https://github.com/Reza2kn/lance-quant

For convenience, we also ship a ComfyUI custom node that wires this checkpoint into Lance's six tasks (t2i, t2v, image_edit, video_edit, x2t_image, x2t_video).

Quality

Side-by-side on Lance's bundled x2t_image example (6 cases):

Comparison vs naïve min-max INT4 (no calibration), same setup:

The proper AWQ calibration is essential — naïve INT4 destroys outputs because of activation outliers in certain weight columns.

Limitations & known caveats

• Inference runtime is slow with pure-PyTorch dequantization (~10× the bf16 baseline) because the swap-in WQLinearINT4 does a tensor-by-tensor dequant on every forward. To get production throughput you'll want a fused INT4 GEMM kernel (e.g. via Triton, marlin, or exllamav2). The companion repo will add this.
• This is a v1 calibration with a small set (17 samples across 2 tasks). Adding image_edit, t2v, video_edit, x2t_video would likely close the small gap on case 04.
• The bundled Qwen2.5-VL-ViT (in the original checkpoint) is kept at bf16 here. A separate quantization of the ViT is on the roadmap.
• Wan2.2_VAE.pth is not part of this checkpoint — fetch it from the original repo (it's used unmodified, kept in bf16/fp32 since diffusion VAEs are sensitive to low-bit quantization).

Reproduction

# Clone Lance source + download original weights (~57 GB)
git clone https://github.com/bytedance/Lance.git
cd Lance && bash setup_env.sh
huggingface-cli download bytedance-research/Lance --local-dir downloads

# Clone this repo's quantization toolkit
git clone https://github.com/Reza2kn/lance-quant.git
cd lance-quant/scripts

# 1) Calibrate (collect activation statistics)
python awq_calibrate_single.py --task x2t_image \
    --model_path ../../Lance/downloads/Lance_3B_Video \
    --vit_path ../../Lance/downloads/Qwen2.5-VL-ViT \
    --example_json ../../Lance/config/examples/x2t_image_example.json \
    --out ../calib/x2t_image_stats.pt

python awq_calibrate_single.py --task t2i \
    --model_path ../../Lance/downloads/Lance_3B_Video \
    --vit_path ../../Lance/downloads/Qwen2.5-VL-ViT \
    --example_json ../../Lance/config/examples/t2i_example.json \
    --out ../calib/t2i_stats.pt --num_timesteps 2

# 2) Merge stats + apply AWQ
python awq_merge_stats.py --inputs ../calib/x2t_image_stats.pt ../calib/t2i_stats.pt --out ../calib/act_stats.pt
python awq_apply.py --src ../../Lance/downloads/Lance_3B_Video/model.safetensors \
    --stats ../calib/act_stats.pt --out ../models/Lance_3B_Video-AWQ-INT4

# 3) Evaluate
python run_quant_eval.py --task x2t_image \
    --model_path ../../Lance/downloads/Lance_3B_Video \
    --awq_dir ../models/Lance_3B_Video-AWQ-INT4

Citation

If you use this quantized checkpoint, please also cite the original Lance paper:

@misc{lance2026,
  title  = {Lance: Unified Multimodal Modeling by Multi-Task Synergy},
  author = {Fengyi Fu and Mengqi Huang and Shaojin Wu and others},
  year   = {2026}
}

License

Apache 2.0, inherited from the base model.