Instructions to use Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0")
model = AutoModelForMultimodalLM.from_pretrained("Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0

SGLang

How to use Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0 with Docker Model Runner:
```
docker model run hf.co/Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Gemma-4 E4B-it · AutoRound W4A16 (MLP + attention output)

A 4-bit weight-only compression of google/gemma-4-E4B-it covering the language-model decoder's MLP and attention-output projections, served under vLLM via the gptq_marlin kernel.

Role: fallback candidate for team Godspeed AI's Round 2 entry in the Resilient AI Challenge (image-to-text). The primary submission is g4e4-it-r2-awq-smoke-v0, which extends quantization to the full decoder (q/k/v/o + MLP via AWQ-Marlin) and adds a response-economy chat template. This repo is the conservative weights-only sibling: fewer modules touched, no template modification.

Model information

Property	Value
Base model	`google/gemma-4-E4B-it` (weights untouched outside quantized scope)
Quantized scope	`language_model.layers..mlp.{gate,up,down}_proj`, `.self_attn.o_proj`
Preserved at bf16	q/k/v projections, vision tower, audio tower, embeddings, lm_head, norms
Format	W4A16, group size 128, exported GPTQ, served as `gptq_marlin`
Algorithm	AutoRound 0.12.3 MLLM mode, RTN (`--iters 0`), asymmetric
Size on disk	10.6 GB (vs 16.0 GB bf16, −34%)
Training	None — no finetuning, distillation, or healing at any stage

Benchmark results

NVIDIA L4 (evaluator hardware), temperature=1.0, top_p=0.95, top_k=64, max_tokens=180:

Benchmark	BF16 base	This model	Δ energy
9-category image validation (caption, OCR, table, receipt, dashboard, chart, form, handwriting, diagram)	40/41 facts · 0.738 Wh	37/41 facts · ~0.36 Wh	−51%
50-row Indic 5×5 (translation, summarization, QA, transliteration, code-switch)	61/65 · 0.557 Wh	59/65 · 0.256 Wh	−54%

Reference reproduction on RTX A6000 (Ampere sm_86): −33% (image) and −52% (Indic) at equal or better facts. Per-category recovery stays above 80% in every measured category on both GPUs.

Usage

vllm serve Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0 --config vllm_config.yaml

The packaged vllm_config.yaml is self-contained (no infrastructure-specific parameters): gpu-memory-utilization 0.90, max-model-len 8192, dtype bfloat16, quantization gptq_marlin, trust-remote-code true.

Implementation notes

GPTQ format rejects mixed-precision shards inside vLLM's fused qkv_proj, which is why q/k/v stay bf16 in this artifact. The sister AWQ repo solves that with the AWQ-Marlin pack layout and quantizes the full decoder.
vLLM 0.20.2 prints a cosmetic Casting torch.float16 to torch.bfloat16 warning at load; behavior is correct.
Gemma 4's heterogeneous head dims (local 256 / global 512) force the TRITON_ATTN backend automatically.
Cold start to first /health ≈ 2–3 minutes (weight load + torch.compile); warm restarts ≈ 15 s.

Limitations

Image-understanding composition is slightly more sensitive to quantization than document OCR (worst single category ≈ 89% of base).
This artifact deliberately leaves ~15% additional energy savings on the table versus the primary submission in exchange for minimal-surface changes.

License

Apache 2.0, inherited from google/gemma-4-E4B-it.

Downloads last month: -

Safetensors

Model size

5B params

Tensor type

I32

BF16

Model tree for Shankara-A-S/g4e4-it-r2-w4a16-mlpo-v0

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Quantized

(367)

this model