Instructions to use bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 with Transformers:

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

pipe = pipeline("image-text-to-text", model="bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16")
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)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16")
model = AutoModelForMultimodalLM.from_pretrained("bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16", 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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16",
		"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/bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16

SGLang

How to use bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 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 "bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16" \
    --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": "bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16",
		"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 "bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16" \
        --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": "bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16",
		"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 bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 with Docker Model Runner:
```
docker model run hf.co/bjonor/Swift-Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16
```

bjonor commited on 20 days ago

Commit

21055aa

verified ·

1 Parent(s): 2fd81f4

Model card: link code repo, SergiioB cookbook (correct patch attribution) and reference artifact

Browse files

Files changed (1) hide show

README.md +16 -7

README.md CHANGED Viewed

@@ -55,11 +55,12 @@ models below continue to apply (see [License](#license-and-attribution)).
 | Source revision | `048328f4059015b63f860a453bf94834af0db683` |
 | Calibration | `HuggingFaceH4/ultrachat_200k` `train_sft[:256]`, truncated to 2048 tokens |
 | Calibration revision | `8049631c405ae6576f93f445c6b8166f76f5505a` |
-| Repository | quantization recipe + verification: https://github.com/BjornNordblom/intel-arc-b70-quant |
 The `quantize_config.json` is field-for-field identical to the community
-reference artifact `SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16` except for
-the quant-time-only `meta.offload_to_disk` flag. Note that the reference artifact
 quantizes the **base** Qwen3.8-27B; this repository quantizes the **Swift**
 fine-tune.
@@ -112,7 +113,8 @@ Notes for this base model on XPU:
 - **MTP draft must be built unquantized.** The checkpoint flags this via the
   `dynamic` exclusion, but the XPU build tested here also needs the draft layer
-  built without `quant_config` (upstream patch used by the recipe repo:
   `B70_MTP_BF16_DRAFT=1` gate plus a small metadata patch for the
   max-model-length boundary).
 - **`vllm-xpu-kernels` < 0.1.14.1 has a mixed-batch limitation**: batching
@@ -124,6 +126,11 @@ Notes for this base model on XPU:
   backport used by the recipe repo (a single-function change in `vllm/_xpu_ops.py`).
 - `--kv-cache-dtype fp8` is a serving choice, not part of the checkpoint.
 ### Transformers
 GPTQ checkpoints need a GPTQ-capable loader (`gptqmodel` or `auto-gptq`) and
@@ -139,7 +146,7 @@ MTP 3 speculative tokens, fp8 KV cache):
 |---|---|
 | Quantization contract (`verify_quant.py`) | PASS — bits 4, group 128, sym, `desc_act=false`, 15 MTP tensors preserved, 333 vision tensors, 400 modules quantized, `lm_head` untouched |
 | Tiny-model smoke test + endpoint/streaming test | PASS |
-| Decode, p512/g128, median of 5 | 58.8 tok/s (this artifact) vs 60.4 tok/s (`SergiioB/…` reference quant) |
 | MTP draft acceptance | 3.38 / 4 tokens accepted (59.5%) vs 3.43 (60.9%) for the reference quant |
 | Solo TTFT / decode (short prompt) | ~0.95 s / ~59 tok/s |
 | Concurrency | 12-request storm with 4 × ~10k-token prefills: 12/12 HTTP 200, no engine failure, MTP acceptance ~62% (with the mixed-batch fix above) |
@@ -247,8 +254,10 @@ Addendum for this quantization (the uploader's modification notice):
 - UkisAI for the Swift fine-tune and the licence terms above.
 - Alibaba Cloud / Qwen for Qwen3.8-27B (Apache-2.0).
 - The `gptqmodel` project for the quantizer.
-- Intel's Arc Pro B70 XPU inference cookbook for the serving patches (MTP BF16
-  draft, GDN boundary handling, mixed-batch split-dispatch backport).
 ## Model card contact

 | Source revision | `048328f4059015b63f860a453bf94834af0db683` |
 | Calibration | `HuggingFaceH4/ultrachat_200k` `train_sft[:256]`, truncated to 2048 tokens |
 | Calibration revision | `8049631c405ae6576f93f445c6b8166f76f5505a` |
+| Code | quantization, verification and Intel-XPU serving recipe: [BjornNordblom/intel-arc-b70-quant](https://github.com/BjornNordblom/intel-arc-b70-quant) |
 The `quantize_config.json` is field-for-field identical to the community
+reference artifact
+[`SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16`](https://huggingface.co/SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16)
+except for the quant-time-only `meta.offload_to_disk` flag. Note that the reference artifact
 quantizes the **base** Qwen3.8-27B; this repository quantizes the **Swift**
 fine-tune.
 - **MTP draft must be built unquantized.** The checkpoint flags this via the
   `dynamic` exclusion, but the XPU build tested here also needs the draft layer
+  built without `quant_config` (upstream patch used by [the recipe
+  repo](https://github.com/BjornNordblom/intel-arc-b70-quant):
   `B70_MTP_BF16_DRAFT=1` gate plus a small metadata patch for the
   max-model-length boundary).
 - **`vllm-xpu-kernels` < 0.1.14.1 has a mixed-batch limitation**: batching
   backport used by the recipe repo (a single-function change in `vllm/_xpu_ops.py`).
 - `--kv-cache-dtype fp8` is a serving choice, not part of the checkpoint.
+Full reproduction path — pinned environments, `quant_swift.py` /
+`verify_quant.py`, `launch.sh` (MTP and vision flags), the XPU patches and the
+benchmark harness — is in the
+[recipe repository](https://github.com/BjornNordblom/intel-arc-b70-quant).
 ### Transformers
 GPTQ checkpoints need a GPTQ-capable loader (`gptqmodel` or `auto-gptq`) and
 |---|---|
 | Quantization contract (`verify_quant.py`) | PASS — bits 4, group 128, sym, `desc_act=false`, 15 MTP tensors preserved, 333 vision tensors, 400 modules quantized, `lm_head` untouched |
 | Tiny-model smoke test + endpoint/streaming test | PASS |
+| Decode, p512/g128, median of 5 | 58.8 tok/s (this artifact) vs 60.4 tok/s ([`SergiioB/…`](https://huggingface.co/SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16) reference quant) |
 | MTP draft acceptance | 3.38 / 4 tokens accepted (59.5%) vs 3.43 (60.9%) for the reference quant |
 | Solo TTFT / decode (short prompt) | ~0.95 s / ~59 tok/s |
 | Concurrency | 12-request storm with 4 × ~10k-token prefills: 12/12 HTTP 200, no engine failure, MTP acceptance ~62% (with the mixed-batch fix above) |
 - UkisAI for the Swift fine-tune and the licence terms above.
 - Alibaba Cloud / Qwen for Qwen3.8-27B (Apache-2.0).
 - The `gptqmodel` project for the quantizer.
+- SergiioB's [intel-arc-pro-b70-inference-cookbook](https://github.com/SergiioB/intel-arc-pro-b70-inference-cookbook)
+  for the Intel Arc Pro B70 XPU serving patches (MTP BF16 draft, GDN boundary
+  handling, mixed-batch split-dispatch backport); the copies vendored in the
+  recipe repo are MIT, Copyright (c) 2026 SergiioB.
 ## Model card contact