Instructions to use bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16 with Transformers:

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

pipe = pipeline("text-generation", model="bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16")
model = AutoModelForCausalLM.from_pretrained("bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16

SGLang

How to use bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16 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 "bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16 with Docker Model Runner:
```
docker model run hf.co/bullerwins/DeepSeek-V4-Flash-0731-AutoRound-W4A16
```

DeepSeek-V4-Flash-0731-AutoRound-W4A16 / AUTOROUND_QUANTIZATION.md

bullerwins

Add files using upload-large-folder tool

331f482 verified about 1 month ago

preview code

Raw

History Blame

1.51 kB

DeepSeek-V4-Flash-0731 AutoRound W4A16

Source: /mnt/llms/models/deepseek-ai/DeepSeek-V4-Flash-0731

Generated with AutoRound 0.15.0 (/home/ubuntuai/auto-round, commit 60b813cb) using model-free RTN:

auto-round /mnt/llms/models/deepseek-ai/DeepSeek-V4-Flash-0731 \
  --model_free \
  --scheme W4A16 \
  --ignore_layers compressor,indexer.weights_proj \
  --layer_config "{'wo_a':{bits:16}}" \
  --output_dir /mnt/llms/models/bullerwins/DeepSeek-V4-Flash-0731-AutoRound

The checkpoint uses symmetric INT4 weights with group size 128 and BF16 activations. Quality-sensitive token embeddings, LM head, MoE routers, attention compressors/indexer projections, and wo_a remain BF16. AutoRound quantized 35,672 eligible layers and preserved 242 layers.

block_name_to_quantize was removed from both quantization configurations after export. AutoRound emitted the source-level prefix layers, while vLLM instantiates these modules under model.layers; leaving the field caused vLLM to treat the MoE as unquantized. Explicit BF16 exclusions in extra_config remain intact.

The checkpoint was validated with vLLM using PP=3 on CUDA devices 0,2,6, VLLM_PP_LAYER_PARTITION=8,27,8, and 8 GiB CPU offload per rank. A non-thinking chat probe returned 323 for 17*19.

This is an RTN conversion rather than calibration-based AutoRound tuning. It follows Intel's published generation recipe for Intel/DeepSeek-V4-Flash-W4A16-AutoRound, updated with AutoRound 0.15's DeepSeek-V4 handling.