Instructions to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with Transformers:

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

pipe = pipeline("text-generation", model="stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S
# Run inference directly in the terminal:
llama cli -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S
# Run inference directly in the terminal:
llama cli -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S
# Run inference directly in the terminal:
./build/bin/llama cli -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Use Docker

docker model run hf.co/stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Llama new
LM Studio
Jan

vLLM

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

SGLang

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S 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 "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S" \
    --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": "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S" \
        --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": "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with Ollama:
```
ollama run hf.co/stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S
```
Unsloth Desktop

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with Docker Model Runner:
```
docker model run hf.co/stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S
```

Lemonade

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Run and chat with the model

lemonade run user.Step-3.5-Flash-GGUF-Q4_K_S-Q4_K_S

List all available models

lemonade list

Hermes Agent

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S:Q4_K_S" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

INT8 quantization for KVCache on DGX Spark/GB10

by JDWarner - opened Feb 3

Discussion

JDWarner

Feb 3

Per the model card:
"On NVIDIA DGX Spark, the Step 3.5 Flash achieves a generation speed of 20 tokens per second; by integrating the INT8 quantization technology for KVCache, it supports an extended context window of up to 256K tokens, thus delivering long text processing capabilities on par with cloud-based inference."

Unless I missed something (possible!), this does not seem to have instructions or breadcrumbs in the model card. The provided start-up command for this GGUF on the Spark seems to limit context to 16k. Could you please provide some guidance on how to use INT8 KVCache with this Int4 GGUF on the DGX Spark? Thanks!

apohelios

StepFun org Feb 4

Sorry I’m a bit late to this discussion.

Would you mind trying the Spark usage example from this page and checking whether it works for you?
https://github.com/stepfun-ai/Step-3.5-Flash/blob/main/llama.cpp/docs/step3.5-flash.md

JDWarner

Feb 6

Thanks for the link, this is great documentation. It seems the -ctk q8_0 and -ctv q8_0 options are essential for the Int8 KVCache.

Per the guide, thus far I have held to or a little below 200k context rather than push it further. I am not currently forcing MMQ with the build flag -DGGML_CUDA_FORCE_MMQ=ON, favoring performance over a little bit of extra context window, and seeing 10-20% higher throughput for low-medium context (up to 24 t/s in the 2-8k context range).

The linked guide suggests the environment variable GGML_CUDA_ENABLE_UNIFIED_MEMORY=1, which I'm unsure is relevant for the DGX Spark. Does this have any function on the Spark, where memory is already inherently unified?

apohelios

StepFun org Feb 9

Thanks for the link, this is great documentation. It seems the -ctk q8_0 and -ctv q8_0 options are essential for the Int8 KVCache.

Per the guide, thus far I have held to or a little below 200k context rather than push it further. I am not currently forcing MMQ with the build flag -DGGML_CUDA_FORCE_MMQ=ON, favoring performance over a little bit of extra context window, and seeing 10-20% higher throughput for low-medium context (up to 24 t/s in the 2-8k context range).

The linked guide suggests the environment variable GGML_CUDA_ENABLE_UNIFIED_MEMORY=1, which I'm unsure is relevant for the DGX Spark. Does this have any function on the Spark, where memory is already inherently unified?

GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. On my DGX Spark tests, a ~256k context would crash without this setting, but with GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 it can run (but maybe hurts performance).

Nathan-Maine

Mar 31

Adding data from a similar KV cache quantization test on DGX Spark (GB10).

Tested llama.cpp --cache-type-k q4_0 --cache-type-v q4_0 vs q8_0 vs f16 on Nemotron-3-Nano-30B-A3B at 128K context. The q8_0 approach mentioned in this thread is the right call — q4_0 has a devastating performance cliff at 64K+ context (282.7 → 21.3 tok/s prompt processing, -92.5% due to dequant overhead).

q8_0 stays within 5% of f16 speed at all context lengths while halving KV cache size. On Spark's 128GB unified memory, that's the sweet spot.

Full benchmark data with tables and methodology: https://www.linkedin.com/pulse/i-benchmarked-kv-cache-quantization-my-dgx-spark-heres-nathan-maine-szxtc

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