Ernie4.5 VL Model Usage Guide¶

This guide describes how to run ERNIE-4.5-VL-28B-A3B-PT and ERNIE-4.5-VL-424B-A47B-PT with native BF16.

Installing vLLM¶

CUDA¶

ERNIE-4.5-VL support was recently added to vLLM main branch and is not yet available in any official release:

uv venv --python 3.12 --seed
source .venv/bin/activate
uv pip install -U vllm --torch-backend auto

AMD ROCm: MI300X/MI325X/MI355X¶

uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/

⚠️ The vLLM wheel for ROCm is compatible with Python 3.12, ROCm 7.0, and glibc >= 2.35. If your environment is incompatible, please use docker flow in vLLM

Running Ernie4.5-VL¶

Serving Ernie4.5-VL Model on H100 GPUs¶

NOTE: torch.compile and cuda graph are not supported due to the heterogeneous expert architecture. (vision and text experts)

# 28B model 80G*1 GPU
vllm serve baidu/ERNIE-4.5-VL-28B-A3B-PT \
  --trust-remote-code

# 424B model 140G*8 GPU with native BF16
vllm serve baidu/ERNIE-4.5-VL-424B-A47B-PT \
  --trust-remote-code \
  --tensor-parallel-size 8

If you only want to test the functionality and only have 8×80G GPU, you can use the --cpu-offload-gb parameter to offload part of the weights to CPU memory, and additionally use FP8 online quantization to further reduce GPU memory.

# 424B model 80G*8 GPU with FP8 quantization and CPU offloading
vllm serve baidu/ERNIE-4.5-VL-424B-A47B-PT \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --quantization fp8 \
  --cpu-offload-gb 50

If your single node GPU memory is insufficient, native BF16 deployment may require multi nodes, multi node deployment reference vLLM doc to start ray cluster. Then run vllm on the master node

# 424B model 80G*16 GPU with native BF16
vllm serve baidu/ERNIE-4.5-VL-424B-A47B-PT \
  --trust-remote-code \
  --tensor-parallel-size 16

Serving Ernie4.5-VL Model on MI300X/MI325X/MI355X GPUs¶

Run the vLLM online serving on AMD GPUs using the command below:

VLLM_ROCM_USE_AITER=1 \
SAFETENSORS_FAST_GPU=1 \
vllm serve baidu/ERNIE-4.5-VL-28B-A3B-PT \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.9 \
    --disable-log-requests \
    --trust-remote-code

Benchmarking¶

For benchmarking, only the first vllm bench serve after service startup to ensure it is not affected by prefix cache

# Prompt-heavy benchmark (8k/1k)
vllm bench serve \
  --model baidu/ERNIE-4.5-VL-28B-A3B-PT \
  --host 127.0.0.1 \
  --port 8200 \
  --dataset-name random \
  --random-input-len 8000 \
  --random-output-len 1000 \
  --request-rate 10 \
  --num-prompts 16 \
  --ignore-eos \
  --trust-remote-code

Benchmark Configurations¶

Test different workloads by adjusting input/output lengths:

Prompt-heavy: 8000 input / 1000 output
Decode-heavy: 1000 input / 8000 output
Balanced: 1000 input / 1000 output

Test different batch sizes by changing --num-prompts, e.g., 1, 16, 32, 64, 128, 256, 512

Expected Output¶

============ Serving Benchmark Result ============
Successful requests:                     16        
Request rate configured (RPS):           10.00     
Benchmark duration (s):                  61.27     
Total input tokens:                      128000    
Total generated tokens:                  16000     
Request throughput (req/s):              0.26      
Output token throughput (tok/s):         261.13    
Total Token throughput (tok/s):          2350.19   
---------------Time to First Token----------------
Mean TTFT (ms):                          15663.63  
Median TTFT (ms):                        21148.69  
P99 TTFT (ms):                           22147.85  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          38.74     
Median TPOT (ms):                        38.90     
P99 TPOT (ms):                           39.92     
---------------Inter-token Latency----------------
Mean ITL (ms):                           43.27     
Median ITL (ms):                         36.35     
P99 ITL (ms):                            236.49    
==================================================