GRACE

Gated Relational Alignment via Confidence-based Distillation
for Quantization-Aware Training of Vision-Language Models

Official PyTorch implementation of GRACE, accepted at ICML 2026.

📄 Paper  |  🤗 Models  |  📦 Training Data

📖 Table of Contents

• 📊 Results
• 🤗 Model Zoo
• 📁 Repository Layout
• ⚙️ Environment Setup
• 🗂️ Data Preparation
• 🚀 Training
• 📈 Evaluation
• 📦 Deployment (INT4 / AWQ Inference)
• 📝 Citation
• 🙏 Acknowledgements
• ⭐ Star History
• 📜 License

GRACE is a quantization-aware training (QAT) framework for vision-language models. The goal is simple to state: push a small student VLM down to 4-bit weights and still keep almost all of the accuracy you would have had at full precision. The trick is that GRACE trains quantization and distillation together, instead of quantizing a model after the fact, so the student learns to be a good low-bit model from the start.

Three pieces make this work:

• GDKD: confidence-gated Decoupled Knowledge Distillation (TCKD + NCKD), where the trade-off β is adapted online by an Information-Bottleneck controller.
• RCKA: Relational Centered Kernel Alignment on penultimate-layer visual tokens, which aligns the student's relational geometry to the teacher's.
• Group-wise LSQ QAT: learned per-group weight scales (W4 / W8, group size 128) on the LLM and the MLP projector, with the ViT kept frozen.

Training optimizes

L_total = L_CE  +  β · L_GDKD  +  ω · L_RCKA

where β is driven by the IB controller (τ, η) and ω is warmed up linearly. The defaults follow the values reported in the paper (Table 6); see finetune_qwen3vl_2b_grace.slurm for the full hyper-parameter list.

The reference implementation in this repo applies GRACE to Qwen3-VL-2B-Instruct (student) distilled from Qwen3-VL-8B-Instruct (teacher).

---

📊 Results

Comparison on 7 VLM benchmarks. The 8B model is the distillation teacher (reference upper bound); every GRACE-Qwen3 variant is a 2B student. The best result among the 2B Qwen3-VL models is in bold.

We release GRACE on Qwen3-VL here because it is the most current backbone and gives a fairer, up-to-date point of comparison, with the vanilla Qwen3-VL-2B-Instruct as the baseline. The paper itself reports GRACE on LLaVA-1.5 and Qwen2-VL, and we additionally release the LLaVA-1.5 W4G128 (INT4) checkpoint from the paper in the model zoo below.

Model	Params	Precision	HallB	MMBench	ScienceQA	AI2D	MMMU	SEED	MMStar	Avg
Qwen3-VL-8B (teacher, ref.)	8B	BF16	61.1	84.5	85.0	85.7	69.6	77.5	70.9	76.3
Qwen3-VL-2B (baseline)	2B	BF16	51.4	78.4	81.4	76.9	53.4	71.2	58.3	67.3
Qwen3-VL-2B-GRACE	2B	BF16	66.9	86.4	86.2	81.3	72.1	76.7	67.3	76.7
Qwen3-VL-2B-GRACE (W8G128)	2B	INT8	66.1	85.5	85.3	80.4	71.3	75.9	66.5	75.9
Qwen3-VL-2B-GRACE (W4G128)	2B	INT4	65.4	84.6	84.3	79.5	70.5	75.1	65.8	75.0

GRACE lifts the Qwen3-VL-2B baseline by +9.4 avg and matches or slightly exceeds the 8B teacher on average (76.7 vs. 76.3) at roughly a quarter of the parameters. The W4G128 (INT4) model still retains 98% of the BF16 average.

---

🤗 Model Zoo

Model	Backbone	Bits	Group	Checkpoint description	HF Hub
Qwen3-VL-2B-GRACE-BF16	Qwen3-VL-2B	bf16	n/a	Full-precision GRACE checkpoint; used as the student initialization for the W8/W4 Qwen3-VL runs.	ForeverBlue/Qwen3-VL-2B-GRACE-BF16
Qwen3-VL-2B-GRACE-W8G128	Qwen3-VL-2B	int8	128	INT8 QAT checkpoint with group size 128; high-retention quantized Qwen3-VL student.	ForeverBlue/Qwen3-VL-2B-GRACE-W8G128
Qwen3-VL-2B-GRACE-W4G128	Qwen3-VL-2B	int4	128	INT4 QAT checkpoint with group size 128; compact Qwen3-VL release retaining about 98% of the BF16 average.	ForeverBlue/Qwen3-VL-2B-GRACE-W4G128
Qwen3-VL-2B-GRACE-W4G128-AWQ	Qwen3-VL-2B	int4	128	Real AWQ-packed (qweight/qzeros/scales) deployment build of the W4G128 student; drop-in INT4 inference via the GRACE deploy loader.	ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ
LLaVA-1.5-7B-GRACE-W4G128	LLaVA-1.5-7B	int4	128	INT4 QAT checkpoint from the GRACE paper with learned scales; released for reproducing the LLaVA-1.5 experiments.	ForeverBlue/LLaVA-1.5-7B-GRACE-W4G128
LLaVA-1.5-7B-GRACE-W4G128-AWQ	LLaVA-1.5-7B	int4	128	Real AWQ-packed deployment build of the LLaVA-1.5 W4G128 checkpoint; loads through the GRACE / LLaVA-1.5 codebase.	ForeverBlue/LLaVA-1.5-7B-GRACE-W4G128-AWQ

The BF16 Qwen3-VL checkpoint is the full-precision GRACE student that we use to initialize the W8 and W4 Qwen3-VL runs. The LLaVA-1.5 W4G128 checkpoint corresponds to the paper setting and ships with the GRACE-specific QAT quantized weights, so you can reproduce the INT4 LLaVA experiments directly. The two *-AWQ repositories are the real 4-bit packed builds for deployment; see Deployment for how to run them or repack a checkpoint yourself.

Quick load:

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

ckpt = "ForeverBlue/Qwen3-VL-2B-GRACE-W4G128"
model = Qwen3VLForConditionalGeneration.from_pretrained(
    ckpt, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(ckpt)

---

📁 Repository Layout

.
├── qwen-vl-finetune/        # Training + INT4 deployment for Qwen3-VL
│   ├── qwenvl/
│   │   ├── data/            # Dataset registry + LLaVA-style loader
│   │   ├── quantize/        # QAT→AWQ packing + WQLinear_GEMM (Qwen3-VL)
│   │   └── train/
│   │       ├── train_qwen.py        # plain BF16 SFT
│   │       ├── train_qwen_qat.py    # group-wise LSQ QAT
│   │       ├── train_qwen_grace.py  # GRACE = QAT + GDKD + RCKA
│   │       ├── qat_modules.py       # LSQ fake-quant + save hooks
│   │       └── grace_modules.py     # GDKD, RCKA, IB controller
│   └── scripts/
│       ├── finetune_qwen3vl_2b_bf16.slurm   # BF16 SFT (baseline)
│       ├── finetune_qwen3vl_2b_sft.slurm    # BF16 SFT (alt config)
│       ├── finetune_qwen3vl_2b_qat.slurm    # QAT only (ablation)
│       ├── finetune_qwen3vl_2b_grace.slurm  # GRACE
│       └── deploy_awq_qwen.py               # pack & run real INT4 inference
├── evaluation/              # lmms-eval driver + per-benchmark configs
├── deployment/              # LLaVA-1.5 tree: QAT training + AWQ INT4 packing/inference
│   ├── llava/quantize/      # QAT to AWQ conversion + WQLinear_GEMM kernels
│   ├── scripts/deploy_awq_llava.py            # pack & run real INT4 inference
│   └── scripts/v1_5/finetune_qat.{sh,slurm}   # LLaVA-1.5 QAT launchers
├── qwen-vl-utils/           # Qwen3-VL multi-modal preprocessing helpers
├── cookbooks/               # Qwen3-VL inference / capability demos
├── docker/                  # CUDA 12.8 image for web demo
├── web_demo_mm.py           # Multi-modal Gradio demo
├── assets/                  # README figures (architecture, icon)
└── requirements.txt         # Pinned versions for the qwen3vl venv

---

⚙️ Environment Setup

GRACE was trained on 16 × A100 (64 GB) GPUs across 4 nodes. The reference SLURM scripts pin the host toolchain in their module load block:

Component	Version
CUDA driver / runtime (host)	12.3
GCC	12.2.0
Python	3.11
PyTorch	2.5.1 (cu121 wheels, forward-compatible with the CUDA 12.3 driver)
flash-attn	2.7.2.post1
DeepSpeed	0.15.4 (ZeRO-2)
transformers	5.9.0
accelerate	1.13.0

A frozen export of the full virtual environment is in requirements.txt.

1. Clone the repository

git clone https://github.com/ForeverBlue816/GRACE.git
cd GRACE

2. Build the venv from scratch

# (a) System modules. This is the Leonardo example; adapt it to your cluster.
module purge
module load profile/deeplrn
module load cuda/12.3
module load gcc/12.2.0

# (b) Create and activate the venv
python3.11 -m venv "${HOME}/qwen3vl"
source "${HOME}/qwen3vl/bin/activate"
pip install -U pip wheel setuptools

# (c) PyTorch + CUDA runtime (cu121 wheels)
pip install torch==2.5.1 torchvision==0.20.1 \
    --index-url https://download.pytorch.org/whl/cu121

# (d) Everything else, pinned to the released training env
pip install -r requirements.txt

# (e) flash-attn. Build this AFTER torch is installed, or the build will fail.
pip install flash-attn==2.7.2.post1 --no-build-isolation

# (f) Local utility package (image / video preprocessing for Qwen3-VL)
pip install -e qwen-vl-utils/

3. Pre-stage the model weights

Download the teacher and student weights on the login node (or any internet-reachable host) before launching a compute job:

huggingface-cli download Qwen/Qwen3-VL-2B-Instruct \
    --local-dir "${SCRATCH_ROOT}/Qwen3-VL-2B-Instruct"
huggingface-cli download Qwen/Qwen3-VL-8B-Instruct \
    --local-dir "${SCRATCH_ROOT}/Qwen3-VL-8B-Instruct"

---

🗂️ Data Preparation

GRACE is trained on the two ShareGPT4V annotation files (LLaVA-style schema, image + conversations[from/value]):

#	Annotation JSON	Size	Hugging Face
1	sharegpt4v_instruct_gpt4-vision_cap100k.json	134 MB	Lin-Chen/ShareGPT4V
2	sharegpt4v_mix665k_cap23k_coco-ap9k_lcs3k_sam9k_div2k.json	1.2 GB	Lin-Chen/ShareGPT4V

sharegpt4v_mix665k_* is the main SFT mix used in the paper, while sharegpt4v_instruct_gpt4-vision_cap100k.json is the original GPT-4V caption set.

1. Download the annotation JSONs

export SHAREGPT4V_ROOT=/path/to/ShareGPT4V
mkdir -p "${SHAREGPT4V_ROOT}"

huggingface-cli download Lin-Chen/ShareGPT4V \
    sharegpt4v_instruct_gpt4-vision_cap100k.json \
    sharegpt4v_mix665k_cap23k_coco-ap9k_lcs3k_sam9k_div2k.json \
    --repo-type dataset \
    --local-dir "${SHAREGPT4V_ROOT}"

2. Download the image archives

The ShareGPT4V annotations point at images under ${SHAREGPT4V_ROOT}/data/. Download and unpack the sources below. Only the ones the JSONs actually reference are required.

Source	URL
LAION-CC-SBU-558K	images.zip
COCO	train2017.zip
SAM (subset)	segment-anything-downloads, files 000000~000050.tar. For SFT-only you can use the 9k subset here.
GQA	images.zip
OCR-VQA	download script, saving all images as .jpg.
TextVQA	train_val_images.zip
Visual Genome	part1, part2
WebData (academic use only)	drive folder

3. Final directory layout

${SHAREGPT4V_ROOT}/
├── sharegpt4v_instruct_gpt4-vision_cap100k.json
├── sharegpt4v_mix665k_cap23k_coco-ap9k_lcs3k_sam9k_div2k.json
└── data/
    ├── llava/llava_pretrain/images/
    ├── coco/train2017/
    ├── sam/images/
    ├── gqa/images/
    ├── ocr_vqa/images/
    ├── textvqa/train_images/
    ├── vg/VG_100K/
    ├── vg/VG_100K_2/
    ├── share_textvqa/images/
    ├── web-celebrity/images/
    ├── web-landmark/images/
    └── wikiart/images/

The dataset registry that resolves these paths lives in qwen-vl-finetune/qwenvl/data/init.py, which reads SHAREGPT4V_ROOT from the environment.

---

🚀 Training

All three recipes are launched with sbatch. The full reference run uses 16 × A100 (64 GB) across 4 nodes, DeepSpeed ZeRO-2, BF16, and an effective batch size of 512. Adjust --nodes, PER_DEVICE_BATCH, and GRAD_ACCUM to fit your own cluster.

1. BF16 SFT baseline

Optional, and also the source of our released *-BF16 checkpoint:

sbatch qwen-vl-finetune/scripts/finetune_qwen3vl_2b_bf16.slurm

2. QAT-only baseline

This is the ablation of GRACE without distillation:

# W4 G128 (default)
sbatch qwen-vl-finetune/scripts/finetune_qwen3vl_2b_qat.slurm

# W8 G128
sbatch --export=ALL,QAT_BITS=8 qwen-vl-finetune/scripts/finetune_qwen3vl_2b_qat.slurm

3. GRACE (full method)

# W4 G128: produces ForeverBlue/Qwen3-VL-2B-GRACE-W4G128
sbatch qwen-vl-finetune/scripts/finetune_qwen3vl_2b_grace.slurm

# W8 G128: produces ForeverBlue/Qwen3-VL-2B-GRACE-W8G128
sbatch --export=ALL,QAT_BITS=8 qwen-vl-finetune/scripts/finetune_qwen3vl_2b_grace.slurm

Common environment-variable overrides (apply to all scripts):

Variable	Default	Description
SHAREGPT4V_ROOT	PATH_TO_SHAREGPT4V_ROOT	Root of the ShareGPT4V tree.
DATASETS	sharegpt4v_mix665k	Comma-separated; a %NN suffix downsamples.
MODEL_NAME_OR_PATH	${SCRATCH_ROOT}/Qwen3-VL-2B-Instruct	Student initialization.
TEACHER_MODEL_PATH	${SCRATCH_ROOT}/Qwen3-VL-8B-Instruct	GRACE only.
QAT_BITS / QAT_GROUP_SIZE	4 / 128	LSQ fake-quant config.
OUTPUT_DIR	${CKPT_ROOT}/${RUN_NAME}	Auto-resumes from the latest checkpoint-*.

GRACE-specific knobs (defaults follow the paper):

Variable	Default	Meaning
DKD_TEMPERATURE	2.0	KD temperature T.
DKD_ALPHA / DKD_BETA	1.0 / 4.0	TCKD / NCKD weights.
RCKA_WEIGHT	3.0	ω for L_RCKA.
RCKA_LAYER	-2	Hidden-state index used by RCKA.
IB_TAU / IB_ETA	3.0 / 0.003	IB controller target / step size.
IB_BETA_INIT/MIN/MAX	0.5 / 0.1 / 1.0	β schedule bounds.
RCKA_WARMUP_STEPS	400	Linear warmup for RCKA.

---

📈 Evaluation

We score every checkpoint with lmms-eval. The per-benchmark configs live under evaluation/.

# Example: evaluate the released W4 checkpoint on ScienceQA
sbatch --export=ALL,MODEL=ForeverBlue/Qwen3-VL-2B-GRACE-W4G128 \
       evaluation/ScienceQA/eval_scienceqa.slurm

---

📦 Deployment — INT4 / AWQ Inference

Both the Qwen3-VL and LLaVA-1.5 GRACE checkpoints ship as deployable real INT4 builds. A GRACE QAT checkpoint stores BF16 weights projected onto the INT4 grid, together with a qat_quantized_weights.bin sidecar that holds the learned per-group scales. For deployment we pack the quantized language-model layers into genuine 4-bit AutoAWQ tensors (qweight, qzeros, scales) that are compatible with AWQ-style INT4 GEMM kernels.

This packing step is bit-exact with respect to the learned integer weight codes: the INT4 codes are unchanged, and only the per-group scales are stored in FP16. The two backbones share the same conversion math but use different runtimes:

• Qwen3-VL packs through qwen-vl-finetune/qwenvl/quantize/ and runs in the qwen3vl training environment (transformers 5.9).
• LLaVA-1.5 packs through deployment/llava/quantize/ and runs in a separate transformers==4.37.2 environment.

Installing the fused INT4 kernels (pip install autoawq-kernels) in either environment enables a speedup; without them the model still runs through a correct, slower pure-PyTorch dequantization path.

How the conversion works (symmetric LSQ-QAT → asymmetric AWQ)

GRACE QAT is symmetric signed per group: code q ∈ [-8, 7], a per-group scale s, no zero point, so the dequantized weight is W = q · s (groups of group_size = 128 along the input dim). AWQ's GEMM kernel is asymmetric unsigned: W = scales · (q_awq − zeros) with q_awq ∈ [0, 15]. The two line up exactly with a constant zero-point:

zeros  = 8  (= 2^(bits-1))
scales = s  (= exp(log_scale), stored FP16)
q_awq  = q + 8 ∈ [0, 15]
⇒  scales · (q_awq − 8) = s · q = W      (no error beyond FP16 scale rounding)

Only the per-group scales change dtype; the integer codes are identical. The converters (qwenvl/quantize/qat_to_awq.py for Qwen3-VL, llava/quantize/qat_to_awq.py for LLaVA) treat the sidecar as the source of truth for which layers were quantized, pack each one, and (unless --no-verify) assert the max int-code mismatch is 0. Only the language-model linears (self_attn.{q,k,v,o}_proj and mlp.{gate,up,down}_proj across all decoder layers) are quantized; the vision tower, projector / merger, embeddings, lm_head, and norms stay in their original dtype.

Qwen3-VL-2B (AWQ INT4)

Repository	Stored format	Recommended use
Qwen3-VL-2B-GRACE-W4G128	BF16 weights on the INT4 grid, plus the qat_quantized_weights.bin sidecar	research / re-packing; the source for the conversion below
Qwen3-VL-2B-GRACE-W4G128-AWQ	real packed qweight / qzeros / scales	drop-in INT4 inference

The Qwen3-VL deployment uses the same qwen3vl training environment (no extra environment needed; qwen_vl_utils is already installed by the setup above). Run the commands from the qwen-vl-finetune/ directory.

Run the released AWQ model:

cd qwen-vl-finetune

# Download the packed checkpoint and print its local path
python - <<'PY'
from huggingface_hub import snapshot_download
print(snapshot_download("ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ"))
PY

# Run inference (substitute the path printed above for /path/to/...)
python scripts/deploy_awq_qwen.py \
    --load-packed /path/to/Qwen3-VL-2B-GRACE-W4G128-AWQ \
    --image ../deployment/images/chinaairlines.jpg \
    --query "Describe this image in detail." \
    --max-new-tokens 256

Or convert the BF16 QAT checkpoint → AWQ yourself:

# Load → pack to real INT4 → run → persist the packed model
python scripts/deploy_awq_qwen.py \
    --model-path /path/to/Qwen3-VL-2B-GRACE-W4G128 \
    --image ../deployment/images/chinaairlines.jpg \
    --query "Describe this image in detail." \
    --max-new-tokens 256 \
    --save-dir ./qwen3vl-2b-w4-awq-packed

This loads the checkpoint, reads its qat_quantized_weights.bin sidecar, swaps every quantized language-model linear (language_model.layers.* self_attn.{q,k,v,o}_proj and mlp.{gate,up,down}_proj) for an AWQ WQLinear_GEMM, verifies the packing is bit-exact (max int-code error = 0), runs inference, and writes the packed model plus awq_quantized_modules.json to --save-dir. Reload it instantly with --load-packed ./qwen3vl-2b-w4-awq-packed. The vision tower, merger, embeddings, lm_head, and norms stay BF16. For a fast LLM-only smoke test, drop --image and add --text-only.

Example image:

Example output from the Qwen3-VL-2B-GRACE-W4G128-AWQ model:

=================== OUTPUT ===================
The image captures a moment on an airport runway, where a China Airlines Boeing
777-300ER airplane is in the process of taxiing. The airplane, painted in a
striking combination of white and blue, is adorned with a pink flower design on
its tail, adding a touch of elegance to its appearance. The words "China Airlines"
and "Boeing" are prominently displayed on the side of the airplane, indicating its
affiliation and model.

The airplane is moving towards the right side of the image, suggesting it's either
preparing for takeoff or has just landed. The runway beneath it is a testament to
human engineering, with its smooth surface designed for the safe and efficient
movement of aircraft.

In the background, the airport's infrastructure is visible, including buildings and
other airport facilities. These structures, while not the main focus of the image,
provide context to the setting and the purpose of the airplane's journey.

Overall, the image presents a snapshot of modern aviation, showcasing the marvels
of technology and design that enable air travel.
==============================================

The INT4 GRACE student correctly reads the livery ("China Airlines"), identifies the Boeing 777-300ER, and grounds the description in the runway scene — at roughly a quarter of the BF16 weight footprint.

LLaVA-1.5-7B (AWQ INT4)

Repository	Stored format	Recommended use
LLaVA-1.5-7B-GRACE-W4G128	BF16 weights projected onto the INT4 grid, plus the qat_quantized_weights.bin sidecar	research, inspection, and re-packing experiments
LLaVA-1.5-7B-GRACE-W4G128-AWQ	real packed AWQ tensors: qweight, qzeros, scales	ready-to-run INT4 inference through the GRACE loader

For the 7B LLaVA-1.5 model the packing reduces the language-model weight footprint from ≈14.2 GB (BF16) to ≈4.6 GB, an approximately 3.1× reduction. The LLaVA-1.5 deployment code lives under deployment/: a vendored LLaVA-1.5 codebase extended with the GRACE QAT and AWQ loading utilities. It needs a dedicated transformers==4.37.2 environment, separate from the qwen3vl training environment. Run all commands in this section from the deployment/ directory.

1. Environment

The exact tested package versions are in deployment/requirements.txt for reproducibility.

cd deployment

# Create a fresh venv for the LLaVA-1.5 deployment stack.
# Do not reuse the qwen3vl training environment.
python3 -m venv ~/llava
source ~/llava/bin/activate

pip install -U pip
pip install -r requirements.txt
pip install -e . --no-deps

# Optional acceleration packages. Install these after PyTorch is available.
pip install flash-attn==2.5.8 --no-build-isolation
pip install autoawq-kernels

Tested on an NVIDIA A100 with a CUDA 12.x driver using the versions in requirements.txt. The pinned torch==2.1.2+cu121 package ships its own CUDA runtime, so a system CUDA toolkit and compatible compiler are only needed when building packages such as flash-attn or autoawq-kernels.

2. Run the released AWQ model

Download the packed INT4 checkpoint and run one-shot inference on the bundled chinaairlines.jpg example:

# Download the packed checkpoint and print its local path
python - <<'PY'
from huggingface_hub import snapshot_download
print(snapshot_download("ForeverBlue/LLaVA-1.5-7B-GRACE-W4G128-AWQ"))
PY

# Run inference (substitute the path printed above for /path/to/...)
python scripts/deploy_awq_llava.py \
    --load-packed /path/to/LLaVA-1.5-7B-GRACE-W4G128-AWQ \
    --image-file images/chinaairlines.jpg \
    --query "Please describe the scene in the picture in detail." \
    --conv-mode vicuna_v1 \
    --max-new-tokens 256

Example output:

=================== OUTPUT ===================
The image captures a moment at an airport, where a Boeing 787 Dreamliner, painted
in white and blue, is taxiing on the runway. The airplane is moving from the left
to the right of the frame, with its nose pointed towards the right side of the
image. The airplane is adorned with a pink flower on its tail, adding a touch of
color to the otherwise monochrome aircraft.

The background of the image provides a glimpse into the airport's infrastructure.
A control tower stands tall, overseeing the operations of the airport. A large
hangar is also visible, likely housing other aircraft or serving as a maintenance
facility.

The sky above is a clear blue, suggesting good weather conditions for flight. The
grass surrounding the runway is a vibrant green, indicating it might be spring or
summer. The overall scene is a typical day at an airport, with the Boeing 787
Dreamliner preparing for its next journey.
==============================================
[deploy] generated in 59.56s (4.3 tok/s nominal)
[deploy] peak GPU mem during generate: 5.91 GB

Here the packed INT4 checkpoint runs with a peak GPU memory footprint of about 5.9 GB. The 4.3 tokens/s figure is the pure-PyTorch dequantization fallback; installing autoawq-kernels switches on the fused INT4 kernels and substantially improves throughput.

--load-packed reuses the standard LLaVA architecture, tokenizer, and CLIP vision tower, then reconstructs the AWQ-packed modules listed in awq_quantized_modules.json and loads the packed INT4 tensors directly. No re-packing happens at inference time. The CLIP vision tower is resolved from the mm_vision_tower field in the model config; it defaults to openai/clip-vit-large-patch14-336, so the host needs either internet access on the first run or a valid local CLIP path.

Note: this checkpoint follows the original LLaVA-1.5 architecture with AWQ-packed weights, so it must be loaded through the GRACE deployment code. A plain from_pretrained call will not reconstruct the INT4 layers.

3. Convert a BF16 QAT checkpoint → AWQ yourself

To rebuild the AWQ checkpoint from the released QAT checkpoint (or to pack one you trained), load the BF16 fake-quant checkpoint, pack it into real 4-bit, and persist it with --save-dir:

# Download the QAT (fake-quant BF16 + sidecar) checkpoint
python - <<'PY'
from huggingface_hub import snapshot_download
print(snapshot_download("ForeverBlue/LLaVA-1.5-7B-GRACE-W4G128"))
PY

# Load fp16 → pack to real INT4 → run → persist the packed model
python scripts/deploy_awq_llava.py \
    --model-path /path/to/LLaVA-1.5-7B-GRACE-W4G128 \
    --image-file images/chinaairlines.jpg \
    --query "Please describe the scene in the picture in detail." \
    --conv-mode vicuna_v1 \
    --max-new-tokens 256 \
    --save-dir ./checkpoints/llava-w4-awq-packed

This loads the BF16 checkpoint, reads its qat_quantized_weights.bin sidecar, swaps every quantized LLM linear for an AWQ WQLinear_GEMM, verifies the packing is bit-exact, runs inference, and writes the packed model (plus awq_quantized_modules.json) to --save-dir. After that you can reload it instantly with --load-packed ./checkpoints/llava-w4-awq-packed. The --save-dir of one run is exactly the --load-packed of the next. For a fast LLM-only smoke test, drop --image-file and add --text-only.

4. Load it programmatically

import os, glob, json
from safetensors.torch import load_file
from llava.model.builder import load_pretrained_model
from llava.mm_utils import get_model_name_from_path
from llava.quantize import build_awq_skeleton

d = "/path/to/LLaVA-1.5-7B-GRACE-W4G128-AWQ"
meta = json.load(open(os.path.join(d, "awq_quantized_modules.json")))

tokenizer, model, image_processor, _ = load_pretrained_model(
    d, None, get_model_name_from_path(d), device_map="cuda", device="cuda"
)

# Replace the LLM linears with AWQ modules, then load the packed weights
build_awq_skeleton(
    model, meta["modules"], bits=meta["bits"],
    group_size=meta["group_size"], device="cuda"
)
sd = {}
for f in glob.glob(os.path.join(d, "*.safetensors")):
    sd.update(load_file(f))
prefixes = tuple(n + "." for n in meta["modules"])
model.load_state_dict(
    {k: v for k, v in sd.items() if k.startswith(prefixes)}, strict=False
)
model.eval()

---

📝 Citation

If GRACE or the released checkpoints help your research, please cite:

@inproceedings{chen2026gated,
  title     = {Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs},
  author    = {Chen, Yanlong and Habibian, Amirhossein and Benini, Luca and Li, Yawei},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2601.22709}
}

---

🙏 Acknowledgements

GRACE builds on the public Qwen3-VL release and the Qwen2.5-VL fine-tuning code. The ShareGPT4V training data comes from Lin-Chen/ShareGPT4V, and evaluation is powered by lmms-eval. A big thank-you to all of these communities.

Questions or contributions? Open an issue or a pull request, or reach out to Yanlong Chen at yanlchen@student.ethz.ch.

---

Star History

📜 License

This project is released under the Apache 2.0 license; see LICENSE for details. The Qwen3-VL base model weights are governed by their own license, and the ShareGPT4V images are restricted to academic use.

---

Built with ❤️ for efficient AI community