Instructions to use cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8 with Transformers:

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

pipe = pipeline("image-text-to-text", model="cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8")
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("cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8")
model = AutoModelForMultimodalLM.from_pretrained("cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8")
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 cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8",
		"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/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8

SGLang

How to use cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8 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 "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8" \
    --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": "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8",
		"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 "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8" \
        --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": "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8",
		"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 cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8 with Docker Model Runner:
```
docker model run hf.co/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT8
```

Compatible with vLLM on Ampere?

by HenkTenk - opened 29 days ago

Discussion

HenkTenk

29 days ago

I can't get this to start on my 4x3090 setup.
Am I missing something, is int8 even suppoted on Ampere?

vLLM logs:

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870] ValueError: Failed to find a kernel that can implement the WNA16 linear layer. Reasons: 

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870] CutlassW4A8LinearKernel requires capability 90, current compute  capability is 86

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870] MacheteLinearKernel requires capability 90, current compute  capability is 86

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870]  AllSparkLinearKernel cannot implement due to: Zero points currently not supported by AllSpark

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870]  MarlinLinearKernel cannot implement due to: Quant type (uint8) not supported by  Marlin, supported types are: [ScalarType.uint4]

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870]  ConchLinearKernel cannot implement due to: conch-triton-kernels is not installed, please install it via `pip install conch-triton-kernels` and try again!

(Worker_TP0 pid=224) ERROR 05-13 15:50:14 [multiproc_executor.py:870]  ExllamaLinearKernel cannot implement due to: Quant type (uint8) not supported by Exllama, supported types are: [ScalarType.uint4b8, ScalarType.uint8b128]

Ccatsf

21 days ago

I'm getting the same error on 2 x 3090

kezzyhu

19 days ago

so not support on blackwell?

qenme

18 days ago

Same issue here. I don't see why this shouldn't work on ampere or blackwell. @cyankiwi

Ccatsf

16 days ago

I did some experiments and this model cannot run on Ampere due to a kernel gap:

The model uses asymmetric INT8 quantization with zero points. vLLM tried every available kernel and all rejected it:

- Cutlass / Machete — require sm_90+ (Ampere), not Turing
- Marlin — only supports uint4, not uint8
- Exllama — supports uint8b128 (symmetric) but not uint8 (asymmetric with zero points)
- AllSpark — does not support zero points (also broken on sm_120 / Blackwell)
- Conch — supports zero points natively but requires pip install conch-triton-kernels; even when installed, it was not significantly faster than unquantized FP16 inference

Workarounds like forcing symmetric mode (self.symmetric = True patch) also hit a secondary bug: FLASHINFER does not support int8_per_token_head KV cache. Switching to TRITON_ATTN avoids that but doesn't solve the kernel gap.

@cpatonn : The model was quantized as asymmetric INT8 with zero points — vLLM has no fast kernel for on pre-Ampere hardware. Could you re-quantize as symmetric INT8 (no zero points) for Exllama/Marlin compatibility? I don't expect vLLM team to add upstream support.

krzysztofma

14 days ago

Any updates? @ccatsf - you said Conch - supports it , but it’s not significanty faster- ok but maybe it is on Ada or Blackwell ? Also , it still takes only half VRAM.
If it’s not good solution, then Maybe just use a FP8 quant or some Q8_0 (apparently Exllama supports it?) ?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment