Skip to content

Ernie4.5 Text Model Usage Guide

This guide describes how to run ERNIE-4.5-21B-A3B-PT and ERNIE-4.5-300B-A47B-PT with native BF16.

Installing vLLM

Note: transformers >= 4.54.0 and vllm >= 0.10.1

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

Running Ernie4.5

# 21B model 80G*1 GPU
vllm serve baidu/ERNIE-4.5-21B-A3B-PT

# 300B model 80G*8 GPU with vllm FP8 online quantification
vllm serve baidu/ERNIE-4.5-300B-A47B-PT \
  --tensor-parallel-size 8 \
  --gpu-memory-utilization=0.95 \
  --quantization fp8

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 

# 300B model 80G*16 GPU with native BF16
vllm serve baidu/ERNIE-4.5-300B-A47B-PT \
  --tensor-parallel-size 16

Running Ernie4.5 MTP

# 21B MTP model 80G*1 GPU
vllm serve baidu/ERNIE-4.5-21B-A3B-PT \
  --speculative-config '{"method": "ernie_mtp","model": "baidu/ERNIE-4.5-21B-A3B-PT","num_speculative_tokens": 1}'

# 300B MTP model 80G*8 GPU with vllm FP8 online quantification
vllm serve baidu/ERNIE-4.5-300B-A47B-PT \
  --tensor-parallel-size 8 \
  --gpu-memory-utilization=0.95 \
  --quantization fp8 \
  --speculative-config '{"method": "ernie_mtp","model": "baidu/ERNIE-4.5-300B-A47B-PT","num_speculative_tokens": 1}'

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 

# 300B MTP model 80G*16 GPU with native BF16
vllm serve baidu/ERNIE-4.5-300B-A47B-PT \
  --tensor-parallel-size 16 \
  --speculative-config '{"method": "ernie_mtp","model": "baidu/ERNIE-4.5-300B-A47B-PT","num_speculative_tokens": 1}'

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-4.5-21B-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

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):                  18.65     
Total input tokens:                      127952    
Total generated tokens:                  16000     
Request throughput (req/s):              0.86      
Output token throughput (tok/s):         857.78    
Total Token throughput (tok/s):          7717.46   
---------------Time to First Token----------------
Mean TTFT (ms):                          876.28    
Median TTFT (ms):                        910.42    
P99 TTFT (ms):                           1596.48   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          16.84     
Median TPOT (ms):                        16.86     
P99 TPOT (ms):                           18.11     
---------------Inter-token Latency----------------
Mean ITL (ms):                           16.84     
Median ITL (ms):                         15.49     
P99 ITL (ms):                            20.69     
==================================================