Instructions to use RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8 with Transformers:

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

pipe = pipeline("text-generation", model="RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8")
model = AutoModelForCausalLM.from_pretrained("RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8

SGLang

How to use RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8 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 "RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8" \
    --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": "RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8",
		"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 "RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8" \
        --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": "RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8 with Docker Model Runner:
```
docker model run hf.co/RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8
```

Fix kv_cache_scheme: DSV4 sparse attention requires FP8 KV cache

by nathans-redhat-com - opened 9 days ago

base: refs/heads/main

←

from: refs/pr/4

Discussion Files changed

-1

nathans-redhat-com

9 days ago

kv_cache_scheme is currently null, but DeepSeek-V4's sparse attention architecture requires FP8 KV cache.
vLLM hard-asserts kv_cache_dtype.startswith("fp8") in vllm/models/deepseek_v4/attention.py - the model fails to load without it.

The sgl-project FP8 variant (sgl-project/DeepSeek-V4-Flash-FP8) works because its quant_method: "fp8" triggers implicit FP8 KV cache inference. This checkpoint uses quant_method: "compressed-tensors" which does not, so the scheme must be declared explicitly.

Verified on an 8xH100 cluster — the model crashes at worker startup without this fix.

Fix kv_cache_scheme: DSV4 sparse attention requires FP8 KV cache3b13bba8

kylesayrs

Red Hat AI org 8 days ago

@nathans-redhat-com thanks for the suggestion, I’ll update the other models as well. Could you please add quant arg dynamic=True to avoid using static scales?

kylesayrs

Red Hat AI org 8 days ago

https://huggingface.co/RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8/commit/89c6e0b95031bc556c9b49e10e925bc1f8f16033

kylesayrs changed pull request status to closed 8 days ago

kylesayrs

Red Hat AI org 7 days ago

•

edited 7 days ago

@nathans-redhat-com I did not observe the behavior that you specified, I still needed to pass --kv_cache_dtype=fp8. The change has been reverted

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