PaddleFormers Backend

The PaddleFormers backend is FastDeploy's fallback mechanism, enabling rapid deployment of PaddleFormers-compatible models without waiting for native FastDeploy implementations.

Installation

Install PaddleFormers from source with paddlefleet (which automatically installs PaddlePaddle):

git clone https://github.com/PaddlePaddle/PaddleFormers.git
cd PaddleFormers

# Choose based on your CUDA version
# CUDA 12.6
pip install -e '.[paddlefleet]' --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu126/
# CUDA 12.9
pip install -e '.[paddlefleet]' --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu129/
# CUDA 13.0
pip install -e '.[paddlefleet]' --extra-index-url https://www.paddlepaddle.org.cn/packages/stable/cu130/

Note: If packages are not found, try adding --index-url https://mirrors.ustc.edu.cn/pypi/simple

For more options, refer to the PaddleFormers Official Installation Guide.

Quick Start

Online Serving Mode

python -m fastdeploy.entrypoints.openai.api_server \
    --model Qwen/Qwen3-4B \
    --host 0.0.0.0 \
    --port 8582 \
    --engine-worker-queue-port 18582 \
    --metrics-port 28582 \
    --max-model-len 32768 \
    --max-num-seqs 256 \
    --kv-cache-ratio 0.75 \
    --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.9 \
    --model-impl paddleformers

Test:

curl -X POST "http://0.0.0.0:8582/v1/chat/completions" \
-H "Content-Type: application/json" \
-d '{"messages": [{"role": "user", "content": "Write a haiku about programming"}]}'

Offline Inference Mode

from fastdeploy import LLM, SamplingParams

llm = LLM(
    model="Qwen/Qwen3-4B",
    model_impl="paddleformers",
    max_model_len=32768,
)

sampling_params = SamplingParams(max_tokens=4096, temperature=0.7)
messages = [[{"role": "user", "content": "Write a short poem about artificial intelligence"}]]
outputs = llm.chat(messages, sampling_params)

for output in outputs:
    print(output.outputs.text)

Parameter Reference

Parameter	Value	Description
--model-impl	auto	Default. Prefers native implementation, falls back to PaddleFormers if unavailable
--model-impl	fastdeploy	Uses native implementation only, raises error if unavailable
--model-impl	paddleformers	Forces use of PaddleFormers backend

Support Status

Model Type	Status
Dense Text Generation (Qwen3/Llama3/ERNIE)	✅ Supported
VLM (Vision-Language Models)	In Development
MOE (Mixture of Experts)	In Development

Optimization	Status
Tensor Parallel (TP)	✅
CUDA Graph	✅
Prefix Caching	✅
Chunked Prefill	✅
QKV/Gate+Up Fusion	✅

Limitations

Not Yet Supported: Expert Parallel (EP), Quantized Inference (INT4/INT8), Speculative Decoding

Performance Note: The PaddleFormers backend outperforms native PaddleFormers but is slightly slower than native FastDeploy implementations.