Instructions to use tcclaviger/Qwen3.8-27B-RFI8-AA with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tcclaviger/Qwen3.8-27B-RFI8-AA with Transformers:

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

pipe = pipeline("image-text-to-text", model="tcclaviger/Qwen3.8-27B-RFI8-AA")
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("tcclaviger/Qwen3.8-27B-RFI8-AA")
model = AutoModelForMultimodalLM.from_pretrained("tcclaviger/Qwen3.8-27B-RFI8-AA", 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 tcclaviger/Qwen3.8-27B-RFI8-AA with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tcclaviger/Qwen3.8-27B-RFI8-AA"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tcclaviger/Qwen3.8-27B-RFI8-AA",
		"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/tcclaviger/Qwen3.8-27B-RFI8-AA

SGLang

How to use tcclaviger/Qwen3.8-27B-RFI8-AA 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 "tcclaviger/Qwen3.8-27B-RFI8-AA" \
    --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": "tcclaviger/Qwen3.8-27B-RFI8-AA",
		"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 "tcclaviger/Qwen3.8-27B-RFI8-AA" \
        --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": "tcclaviger/Qwen3.8-27B-RFI8-AA",
		"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 tcclaviger/Qwen3.8-27B-RFI8-AA with Docker Model Runner:
```
docker model run hf.co/tcclaviger/Qwen3.8-27B-RFI8-AA
```

tcclaviger commited on 27 days ago

Commit

1d1f6eb

verified ·

1 Parent(s): 29d1385

README: full PPL matrix (A8/A16 x fp8/bf16 KV), styling, prose cleanup

Browse files

Files changed (1) hide show

README.md +13 -15

README.md CHANGED Viewed

@@ -4,7 +4,7 @@ license: apache-2.0
 pipeline_tag: image-text-to-text
 ---
-<div style="background: linear-gradient(135deg, #6a1b9a, #8e24aa); border: 2px solid #4a148c; border-radius: 12px; padding: 20px; color: #ffffff; margin-bottom: 24px;">
 ## RFI8 Activation-Aware Quant — read this first
@@ -12,7 +12,7 @@ This is an **RFI8 activation-aware quantization** of Qwen3.8-27B (8-bit weights,
 Hadamard-32 rotation, group-32 block-float scales).
 This checkpoint includes **calibrated KV-cache scales for both the main model and
-the MTP speculative-decoding draft model**. Serving with `--kv-cache-dtype fp8`
 is nearly lossless for quality, and brings long-context speed uplifts and KV
 cache space savings (roughly double the context in the same VRAM).
@@ -30,24 +30,22 @@ falls back to Triton reference kernels and runs substantially slower.
 | checkpoint | PPL |
 |---|---|
 | BF16 original | 6.9224 |
-| **this quant, W8A16 mode** (fp16 activations, `--kv-cache-dtype fp8`, MTP spec decode) | **6.9369** |
-| **this quant, W8A8 mode** (as shipped: int8 activations, `--kv-cache-dtype fp8`, MTP spec decode) | **6.9392** |
-Both quant rows were measured on the served model **with `--kv-cache-dtype fp8`
-active** — the perplexity computation is teacher-forced prefill, and prefill
-attention reads K/V back through the fp8 cache, so these numbers INCLUDE the
-fp8 KV quantization (with this checkpoint's calibrated scales) on the
-full-attention layers. This is the as-shipped serving configuration, not an
-idealized one. (bf16-KV numbers for A8 and A16, isolating pure weight/activation
-error, coming soon.)
 ### W8A8 vs W8A16
 The checkpoint ships in **W8A8** mode (int8 activations). A8 vs A16 is a
 quality/throughput tradeoff: **W8A16 (fp16 activations) is nearly
 indistinguishable from the BF16 checkpoint** in quality, at lower throughput
-than A8. To switch to W8A16, edit `config.json` and change the activation dtype
-from `int8` to `fp16`:
 ```json
 "quantization_config": {
@@ -61,8 +59,8 @@ from `int8` to `fp16`:
 }
 ```
-(That is the single `"act_dtype": "int8"` value under
-`quantization_config.config_groups.group_0.weights` — nothing else changes.)
 </div>

 pipeline_tag: image-text-to-text
 ---
+<div style="background: rgba(45, 10, 70, 0.5); border: 2px solid #4a148c; border-radius: 12px; padding: 20px; margin-bottom: 24px;">
 ## RFI8 Activation-Aware Quant — read this first
 Hadamard-32 rotation, group-32 block-float scales).
 This checkpoint includes **calibrated KV-cache scales for both the main model and
+the MTP speculative-decoding draft model**. Serving with --kv-cache-dtype fp8
 is nearly lossless for quality, and brings long-context speed uplifts and KV
 cache space savings (roughly double the context in the same VRAM).
 | checkpoint | PPL |
 |---|---|
 | BF16 original | 6.9224 |
+| this quant, W8A16, bf16 KV cache | 6.8978 |
+| this quant, W8A8, bf16 KV cache | 6.9034 |
+| this quant, W8A16, fp8 KV cache | 6.9369 |
+| **this quant, W8A8, fp8 KV cache (as shipped)** | **6.9392** |
+The fp8 KV rows include the KV-cache quantization, not just weight/activation
+error. All rows measured on the served model; the PPL tool is included in the
+docker image.
 ### W8A8 vs W8A16
 The checkpoint ships in **W8A8** mode (int8 activations). A8 vs A16 is a
 quality/throughput tradeoff: **W8A16 (fp16 activations) is nearly
 indistinguishable from the BF16 checkpoint** in quality, at lower throughput
+than A8. To switch to W8A16, edit config.json and change the activation dtype
+from int8 to fp16:
 ```json
 "quantization_config": {
 }
 ```
+(That is the single "act_dtype": "int8" value under
+quantization_config.config_groups.group_0.weights — nothing else changes.)
 </div>