|
| 1 | +# Best Practice for ERNIE-4.5-21B-A3B |
| 2 | +## Environmental Preparation |
| 3 | +### 1.1 Hardware requirements |
| 4 | +The minimum number of GPUs required to deploy `ERNIE-4.5-21B-A3B` on the following hardware for each quantization is as follows: |
| 5 | +| | WINT8 | WINT4 | FP8 | |
| 6 | +|-----|-----|-----|-----| |
| 7 | +|H800 80GB| 1 | 1 | 1 | |
| 8 | +|A800 80GB| 1 | 1 | / | |
| 9 | +|H20 96GB| 1 | 1 | 1 | |
| 10 | +|L20 48GB| 1 | 1 | 1 | |
| 11 | +|A30 40GB| 2 | 1 | / | |
| 12 | +|A10 24GB| 2 | 1 | / | |
| 13 | + |
| 14 | +**Tips:** |
| 15 | +1. To modify the number of deployment GPUs, specify `--tensor-parallel-size 2` in starting command. |
| 16 | +2. For hardware not listed in the table, you can estimate whether it can be deployed based on the GPU memory. |
| 17 | + |
| 18 | +### 1.2 Install fastdeploy and prepare the model |
| 19 | +- Installation: Before starting the deployment, please ensure that your hardware environment meets the following conditions: |
| 20 | +``` |
| 21 | +GPU Driver >= 535 |
| 22 | +CUDA >= 12.3 |
| 23 | +CUDNN >= 9.5 |
| 24 | +Linux X86_64 |
| 25 | +Python >= 3.10 |
| 26 | +``` |
| 27 | +For SM 80/90 GPU（A30/A100/H100/） |
| 28 | +``` |
| 29 | +# Install stable release |
| 30 | +python -m pip install fastdeploy-gpu -i https://www.paddlepaddle.org.cn/packages/stable/fastdeploy-gpu-80_90/ --extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple |
| 31 | +
|
| 32 | +# Install latest Nightly build |
| 33 | +python -m pip install fastdeploy-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/fastdeploy-gpu-80_90/ --extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple |
| 34 | +``` |
| 35 | +For SM 86/89 GPU（A10/4090/L20/L40) |
| 36 | +``` |
| 37 | +# Install stable release |
| 38 | +python -m pip install fastdeploy-gpu -i https://www.paddlepaddle.org.cn/packages/stable/fastdeploy-gpu-86_89/ --extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple |
| 39 | +
|
| 40 | +# Install latest Nightly build |
| 41 | +python -m pip install fastdeploy-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/fastdeploy-gpu-86_89/ --extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple |
| 42 | +``` |
| 43 | +For detail, please refer to [Fastdeploy Installation](../get_started/installation/README.md). |
| 44 | + |
| 45 | +- Model Download，**Please note that models with Paddle suffix need to be used for Fastdeploy**： |
| 46 | + - Just specify the model name（e.g. `baidu/ERNIE-4.5-21B-A3B-Paddle`）to automatically download. The default download path is `~/` (i.e. the user's home directory). You can also modify the default download path by configuring the environment variable `FD_MODEL_CACHE` |
| 47 | + - If affected by network or other factors, you can also download the model through [huggingface](https://huggingface.co/), [modelscope](https://www.modelscope.cn/home), etc., and specify the model path when starting service |
| 48 | + |
| 49 | +## Start the Service |
| 50 | + |
| 51 | +Start the service by following command: |
| 52 | +```bash |
| 53 | +python -m fastdeploy.entrypoints.openai.api_server \ |
| 54 | + --model baidu/ERNIE-4.5-21B-A3B-Paddle \ |
| 55 | + --tensor-parallel-size 1 \ |
| 56 | + --quantization wint4 \ |
| 57 | + --max-model-len 32768 \ |
| 58 | + --kv-cache-ratio 0.75 \ |
| 59 | + --max-num-seqs 128 |
| 60 | +``` |
| 61 | +- `--quantization`: indicates the quantization strategy used by the model. Different quantization strategies will result in different performance and accuracy of the model. |
| 62 | +- `--max-model-len`: Indicates the maximum number of tokens supported by the currently deployed service. The larger the value, the longer the context length the model can support, but the more GPU memory is occupied, which may affect the concurrency. |
| 63 | +- `--kv-cache-ratio`: Indicates that KVCache blocks are distributed to the Prefill stage and the Decode stage according to the kv_cache_ratio ratio. Improper settings may result in insufficient KVCache blocks in a certain stage, thus affecting performance. If the service management global block is enabled, this setting is not required. |
| 64 | + |
| 65 | +For more parameter meanings and default settings, see [FastDeploy Parameter Documentation](../parameters.md)。 |
| 66 | + |
| 67 | +### 2.2 Advanced: How to get better performance |
| 68 | +#### 2.2.1 Correctly set parameters that match the application scenario |
| 69 | +Evaluate average input length, average output length, and maximum context length |
| 70 | +- Set max-model-len according to the maximum context length. For example, if the average input length is 1000 and the output length is 30000, then it is recommended to set it to 32768 |
| 71 | +- **Enable the service management global block** |
| 72 | + |
| 73 | +``` |
| 74 | +export ENABLE_V1_KVCACHE_SCHEDULER=1 |
| 75 | +``` |
| 76 | + |
| 77 | +#### 2.2.2 Prefix Caching |
| 78 | +**Idea:** The core idea of Prefix Caching is to avoid repeated calculations by caching the intermediate calculation results of the input sequence (KV Cache), thereby speeding up the response speed of multiple requests with the same prefix. For details, refer to [prefix-cache](../features/prefix_caching.md) |
| 79 | + |
| 80 | +**How to enable:** |
| 81 | +Add the following lines to the startup parameters, where `--enable-prefix-caching` enables prefix caching, and `--swap-space` enables CPU cache in addition to GPU cache. The size is GB and should be adjusted according to the actual situation of the machine. |
| 82 | +``` |
| 83 | +--enable-prefix-caching |
| 84 | +--swap-space 50 |
| 85 | +``` |
| 86 | + |
| 87 | +#### 2.2.3 Chunked Prefill |
| 88 | +**Idea:** This strategy is adopted to split the prefill stage request into small-scale sub-chunks, and execute them in batches mixed with the decode request. This can better balance the computation-intensive (Prefill) and memory-intensive (Decode) operations, optimize GPU resource utilization, reduce the computational workload and memory usage of a single Prefill, thereby reducing the peak memory usage and avoiding the problem of insufficient memory. For details, please refer to [Chunked Prefill](../features/chunked_prefill.md) |
| 89 | + |
| 90 | +**How to enable:** Add the following lines to the startup parameters |
| 91 | +``` |
| 92 | +--enable-chunked-prefill |
| 93 | +``` |
| 94 | + |
| 95 | +#### 2.2.4 MTP (Multi-Token Prediction) |
| 96 | +**Idea:** |
| 97 | +By predicting multiple tokens at once, the number of decoding steps is reduced to significantly speed up the generation speed, while maintaining the generation quality through certain strategies. For details, please refer to [Speculative Decoding](../features/speculative_decoding.md)。 |
| 98 | + |
| 99 | +**How to enable:** |
| 100 | +Add the following lines to the startup parameters |
| 101 | +``` |
| 102 | +--speculative-config '{"method": "mtp", "num_speculative_tokens": 1, "model": "${path_to_mtp_model}"}' |
| 103 | +``` |
| 104 | + |
| 105 | +#### 2.2.5 CUDAGraph |
| 106 | +**Idea:** |
| 107 | +CUDAGraph is a GPU computing acceleration technology provided by NVIDIA. It achieves efficient execution and optimization of GPU tasks by capturing CUDA operation sequences into a graph structure. The core idea of CUDAGraph is to encapsulate a series of GPU computing and memory operations into a re-executable graph, thereby reducing CPU-GPU communication overhead, reducing kernel startup latency, and improving overall computing performance. |
| 108 | + |
| 109 | +**How to enable:** |
| 110 | +Add the following lines to the startup parameters |
| 111 | +``` |
| 112 | +--use-cudagraph |
| 113 | +``` |
| 114 | +Notes: |
| 115 | +1. Usually, no additional parameters need to be set, but CUDAGraph will generate some additional memory overhead, which may need to be adjusted in some scenarios with limited memory. For detailed parameter adjustments, please refer to [GraphOptimizationBackend](../parameters.md) for related configuration parameter descriptions |
| 116 | +2. When CUDAGraph is enabled, only single-card inference is supported, that is, `--tensor-parallel-size 1` |
| 117 | +3. When CUDAGraph is enabled, it is not supported to enable `Chunked Prefill` and `Prefix Caching` at the same time |
| 118 | + |
| 119 | +#### 2.2.6 Rejection Sampling |
| 120 | +**Idea:** |
| 121 | +Rejection sampling is to generate samples from a proposal distribution that is easy to sample, avoiding explicit sorting to increase the sampling speed, which has a significant improvement on small-sized models. |
| 122 | + |
| 123 | +**How to enable:** |
| 124 | +Add the following environment variables before starting |
| 125 | +``` |
| 126 | +export FD_SAMPLING_CLASS=rejection |
| 127 | +``` |
| 128 | + |
| 129 | +## FAQ |
| 130 | +If you encounter any problems during use, you can refer to [FAQ](./FAQ.md). |
0 commit comments