开源 AI 食谱

加入 Hugging Face 社区

并获得增强型文档体验

协作模型、数据集和 Spaces

使用加速推理实现更快的示例

在文档主题之间切换

入门

简介

本食谱的目的是向您展示如何正确地对 TGI 进行基准测试。有关更多背景信息和解释，请先查看这篇热门博客。

设置

确保您的环境已安装 TGI；docker 是一个不错的选择。这些命令可以轻松地复制/粘贴到终端中，这可能更方便。不必强求使用 Jupyter。如果您只想测试一下，可以复制并使用 derek-thomas/tgi-benchmark-space。

TGI 启动器

>>> !text-generation-launcher --version

text-generation-launcher 2.2.1-dev0

下面我们可以看到 TGI 的不同设置。请务必阅读它们，并确定哪些设置对您的用例最重要。

以下是一些最重要的设置

--model-id
--quantize 量化可以节省内存，但不一定能提高速度
--max-input-tokens 这允许 TGI 优化预填充操作
--max-total-tokens 与上述结合，TGI 现在知道最大输入和输出 token 是多少
--max-batch-size 这让 TGI 知道它可以一次处理多少个请求。

最后 3 个设置共同提供了必要的限制，以优化您的用例。通过尽可能恰当地设置这些参数，您可以获得许多性能改进。

>>> !text-generation-launcher -h

Text Generation Launcher

[1m[4mUsage:[0m [1mtext-generation-launcher[0m [OPTIONS]

[1m[4mOptions:[0m
      [1m--model-id[0m 
          The name of the model to load. Can be a MODEL_ID as listed on  like `gpt2` or `OpenAssistant/oasst-sft-1-pythia-12b`. Or it can be a local directory containing the necessary files as saved by `save_pretrained(...)` methods of transformers [env: MODEL_ID=] [default: bigscience/bloom-560m]
      [1m--revision[0m 
          The actual revision of the model if you're referring to a model on the hub. You can use a specific commit id or a branch like `refs/pr/2` [env: REVISION=]
      [1m--validation-workers[0m 
          The number of tokenizer workers used for payload validation and truncation inside the router [env: VALIDATION_WORKERS=] [default: 2]
      [1m--sharded[0m 
          Whether to shard the model across multiple GPUs By default text-generation-inference will use all available GPUs to run the model. Setting it to `false` deactivates `num_shard` [env: SHARDED=] [possible values: true, false]
      [1m--num-shard[0m 
          The number of shards to use if you don't want to use all GPUs on a given machine. You can use `CUDA_VISIBLE_DEVICES=0,1 text-generation-launcher... --num_shard 2` and `CUDA_VISIBLE_DEVICES=2,3 text-generation-launcher... --num_shard 2` to launch 2 copies with 2 shard each on a given machine with 4 GPUs for instance [env: NUM_SHARD=]
      [1m--quantize[0m 
          Whether you want the model to be quantized [env: QUANTIZE=] [possible values: awq, eetq, exl2, gptq, marlin, bitsandbytes, bitsandbytes-nf4, bitsandbytes-fp4, fp8]
      [1m--speculate[0m 
          The number of input_ids to speculate on If using a medusa model, the heads will be picked up automatically Other wise, it will use n-gram speculation which is relatively free in terms of compute, but the speedup heavily depends on the task [env: SPECULATE=]
      [1m--dtype[0m 
          The dtype to be forced upon the model. This option cannot be used with `--quantize` [env: DTYPE=] [possible values: float16, bfloat16]
      [1m--trust-remote-code[0m
          Whether you want to execute hub modelling code. Explicitly passing a `revision` is encouraged when loading a model with custom code to ensure no malicious code has been contributed in a newer revision [env: TRUST_REMOTE_CODE=]
      [1m--max-concurrent-requests[0m 
          The maximum amount of concurrent requests for this particular deployment. Having a low limit will refuse clients requests instead of having them wait for too long and is usually good to handle backpressure correctly [env: MAX_CONCURRENT_REQUESTS=] [default: 128]
      [1m--max-best-of[0m 
          This is the maximum allowed value for clients to set `best_of`. Best of makes `n` generations at the same time, and return the best in terms of overall log probability over the entire generated sequence [env: MAX_BEST_OF=] [default: 2]
      [1m--max-stop-sequences[0m 
          This is the maximum allowed value for clients to set `stop_sequences`. Stop sequences are used to allow the model to stop on more than just the EOS token, and enable more complex "prompting" where users can preprompt the model in a specific way and define their "own" stop token aligned with their prompt [env: MAX_STOP_SEQUENCES=] [default: 4]
      [1m--max-top-n-tokens[0m 
          This is the maximum allowed value for clients to set `top_n_tokens`. `top_n_tokens` is used to return information about the the `n` most likely tokens at each generation step, instead of just the sampled token. This information can be used for downstream tasks like for classification or ranking [env: MAX_TOP_N_TOKENS=] [default: 5]
      [1m--max-input-tokens[0m 
          This is the maximum allowed input length (expressed in number of tokens) for users. The larger this value, the longer prompt users can send which can impact the overall memory required to handle the load. Please note that some models have a finite range of sequence they can handle. Default to min(max_position_embeddings - 1, 4095) [env: MAX_INPUT_TOKENS=]
      [1m--max-input-length[0m 
          Legacy version of [`Args::max_input_tokens`] [env: MAX_INPUT_LENGTH=]
      [1m--max-total-tokens[0m 
          This is the most important value to set as it defines the "memory budget" of running clients requests. Clients will send input sequences and ask to generate `max_new_tokens` on top. with a value of `1512` users can send either a prompt of `1000` and ask for `512` new tokens, or send a prompt of `1` and ask for `1511` max_new_tokens. The larger this value, the larger amount each request will be in your RAM and the less effective batching can be. Default to min(max_position_embeddings, 4096) [env: MAX_TOTAL_TOKENS=]
      [1m--waiting-served-ratio[0m 
          This represents the ratio of waiting queries vs running queries where you want to start considering pausing the running queries to include the waiting ones into the same batch. `waiting_served_ratio=1.2` Means when 12 queries are waiting and there's only 10 queries left in the current batch we check if we can fit those 12 waiting queries into the batching strategy, and if yes, then batching happens delaying the 10 running queries by a `prefill` run [env: WAITING_SERVED_RATIO=] [default: 0.3]
      [1m--max-batch-prefill-tokens[0m 
          Limits the number of tokens for the prefill operation. Since this operation take the most memory and is compute bound, it is interesting to limit the number of requests that can be sent. Default to `max_input_tokens + 50` to give a bit of room [env: MAX_BATCH_PREFILL_TOKENS=]
      [1m--max-batch-total-tokens[0m 
          **IMPORTANT** This is one critical control to allow maximum usage of the available hardware [env: MAX_BATCH_TOTAL_TOKENS=]
      [1m--max-waiting-tokens[0m 
          This setting defines how many tokens can be passed before forcing the waiting queries to be put on the batch (if the size of the batch allows for it). New queries require 1 `prefill` forward, which is different from `decode` and therefore you need to pause the running batch in order to run `prefill` to create the correct values for the waiting queries to be able to join the batch [env: MAX_WAITING_TOKENS=] [default: 20]
      [1m--max-batch-size[0m 
          Enforce a maximum number of requests per batch Specific flag for hardware targets that do not support unpadded inference [env: MAX_BATCH_SIZE=]
      [1m--cuda-graphs[0m 
          Specify the batch sizes to compute cuda graphs for. Use "0" to disable. Default = "1,2,4,8,16,32" [env: CUDA_GRAPHS=]
      [1m--hostname[0m 
          The IP address to listen on [env: HOSTNAME=r-derek-thomas-tgi-benchmark-space-geij6846-b385a-lont4] [default: 0.0.0.0]
  [1m-p[0m, [1m--port[0m 
          The port to listen on [env: PORT=80] [default: 3000]
      [1m--shard-uds-path[0m 
          The name of the socket for gRPC communication between the webserver and the shards [env: SHARD_UDS_PATH=] [default: /tmp/text-generation-server]
      [1m--master-addr[0m 
          The address the master shard will listen on. (setting used by torch distributed) [env: MASTER_ADDR=] [default: localhost]
      [1m--master-port[0m 
          The address the master port will listen on. (setting used by torch distributed) [env: MASTER_PORT=] [default: 29500]
      [1m--huggingface-hub-cache[0m 
          The location of the huggingface hub cache. Used to override the location if you want to provide a mounted disk for instance [env: HUGGINGFACE_HUB_CACHE=]
      [1m--weights-cache-override[0m 
          The location of the huggingface hub cache. Used to override the location if you want to provide a mounted disk for instance [env: WEIGHTS_CACHE_OVERRIDE=]
      [1m--disable-custom-kernels[0m
          For some models (like bloom), text-generation-inference implemented custom cuda kernels to speed up inference. Those kernels were only tested on A100. Use this flag to disable them if you're running on different hardware and encounter issues [env: DISABLE_CUSTOM_KERNELS=]
      [1m--cuda-memory-fraction[0m 
          Limit the CUDA available memory. The allowed value equals the total visible memory multiplied by cuda-memory-fraction [env: CUDA_MEMORY_FRACTION=] [default: 1.0]
      [1m--rope-scaling[0m 
          Rope scaling will only be used for RoPE models and allow rescaling the position rotary to accomodate for larger prompts [env: ROPE_SCALING=] [possible values: linear, dynamic]
      [1m--rope-factor[0m 
          Rope scaling will only be used for RoPE models See `rope_scaling` [env: ROPE_FACTOR=]
      [1m--json-output[0m
          Outputs the logs in JSON format (useful for telemetry) [env: JSON_OUTPUT=]
      [1m--otlp-endpoint[0m 
          [env: OTLP_ENDPOINT=]
      [1m--otlp-service-name[0m 
          [env: OTLP_SERVICE_NAME=] [default: text-generation-inference.router]
      [1m--cors-allow-origin[0m 
          [env: CORS_ALLOW_ORIGIN=]
      [1m--api-key[0m 
          [env: API_KEY=]
      [1m--watermark-gamma[0m 
          [env: WATERMARK_GAMMA=]
      [1m--watermark-delta[0m 
          [env: WATERMARK_DELTA=]
      [1m--ngrok[0m
          Enable ngrok tunneling [env: NGROK=]
      [1m--ngrok-authtoken[0m 
          ngrok authentication token [env: NGROK_AUTHTOKEN=]
      [1m--ngrok-edge[0m 
          ngrok edge [env: NGROK_EDGE=]
      [1m--tokenizer-config-path[0m 
          The path to the tokenizer config file. This path is used to load the tokenizer configuration which may include a `chat_template`. If not provided, the default config will be used from the model hub [env: TOKENIZER_CONFIG_PATH=]
      [1m--disable-grammar-support[0m
          Disable outlines grammar constrained generation. This is a feature that allows you to generate text that follows a specific grammar [env: DISABLE_GRAMMAR_SUPPORT=]
  [1m-e[0m, [1m--env[0m
          Display a lot of information about your runtime environment
      [1m--max-client-batch-size[0m 
          Control the maximum number of inputs that a client can send in a single request [env: MAX_CLIENT_BATCH_SIZE=] [default: 4]
      [1m--lora-adapters[0m 
          Lora Adapters a list of adapter ids i.e. `repo/adapter1,repo/adapter2` to load during startup that will be available to callers via the `adapter_id` field in a request [env: LORA_ADAPTERS=]
      [1m--usage-stats[0m 
          Control if anonymous usage stats are collected. Options are "on", "off" and "no-stack" Defaul is on [env: USAGE_STATS=] [default: on] [possible values: on, off, no-stack]
  [1m-h[0m, [1m--help[0m
          Print help (see more with '--help')
  [1m-V[0m, [1m--version[0m
          Print version

由于我们不需要命令具有交互性，因此可以直接从食谱中启动。

在这个食谱中，我们只使用默认设置，因为我们的目的是了解基准测试工具。

如果您在 Space 上运行，则需要更改这些参数，因为我们不想与 Spaces 服务器发生冲突

--hostname
--port

您可以根据自己的需要更改或删除它们。

>>> !RUST_BACKTRACE=1 \
>>> text-generation-launcher \
>>> --model-id astronomer/Llama-3-8B-Instruct-GPTQ-8-Bit \
>>> --quantize gptq \
>>> --hostname 0.0.0.0 \
>>> --port 1337

[2m2024-08-16T12:07:56.411768Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Args &#123;
    model_id: "astronomer/Llama-3-8B-Instruct-GPTQ-8-Bit",
    revision: None,
    validation_workers: 2,
    sharded: None,
    num_shard: None,
    quantize: Some(
        Gptq,
    ),
    speculate: None,
    dtype: None,
    trust_remote_code: false,
    max_concurrent_requests: 128,
    max_best_of: 2,
    max_stop_sequences: 4,
    max_top_n_tokens: 5,
    max_input_tokens: None,
    max_input_length: None,
    max_total_tokens: None,
    waiting_served_ratio: 0.3,
    max_batch_prefill_tokens: None,
    max_batch_total_tokens: None,
    max_waiting_tokens: 20,
    max_batch_size: None,
    cuda_graphs: None,
    hostname: "0.0.0.0",
    port: 1337,
    shard_uds_path: "/tmp/text-generation-server",
    master_addr: "localhost",
    master_port: 29500,
    huggingface_hub_cache: None,
    weights_cache_override: None,
    disable_custom_kernels: false,
    cuda_memory_fraction: 1.0,
    rope_scaling: None,
    rope_factor: None,
    json_output: false,
    otlp_endpoint: None,
    otlp_service_name: "text-generation-inference.router",
    cors_allow_origin: [],
    api_key: None,
    watermark_gamma: None,
    watermark_delta: None,
    ngrok: false,
    ngrok_authtoken: None,
    ngrok_edge: None,
    tokenizer_config_path: None,
    disable_grammar_support: false,
    env: false,
    max_client_batch_size: 4,
    lora_adapters: None,
    usage_stats: On,
}
[2m2024-08-16T12:07:56.411941Z[0m [32m INFO[0m [2mhf_hub[0m[2m:[0m Token file not found "/data/token"    
[2Kconfig.json [00:00:00] [████████████████████████] 1021 B/1021 B 50.70 KiB/s (0s)[2m2024-08-16T12:07:56.458451Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Model supports up to 8192 but tgi will now set its default to 4096 instead. This is to save VRAM by refusing large prompts in order to allow more users on the same hardware. You can increase that size using `--max-batch-prefill-tokens=8242 --max-total-tokens=8192 --max-input-tokens=8191`.
[2m2024-08-16T12:07:56.458473Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Default `max_input_tokens` to 4095
[2m2024-08-16T12:07:56.458480Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Default `max_total_tokens` to 4096
[2m2024-08-16T12:07:56.458487Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Default `max_batch_prefill_tokens` to 4145
[2m2024-08-16T12:07:56.458494Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Using default cuda graphs [1, 2, 4, 8, 16, 32]
[2m2024-08-16T12:07:56.458606Z[0m [32m INFO[0m [1mdownload[0m: [2mtext_generation_launcher[0m[2m:[0m Starting check and download process for astronomer/Llama-3-8B-Instruct-GPTQ-8-Bit
[2m2024-08-16T12:07:59.750101Z[0m [32m INFO[0m [2mtext_generation_launcher[0m[2m:[0m Download file: model.safetensors
^C
[2m2024-08-16T12:08:09.101893Z[0m [32m INFO[0m [1mdownload[0m: [2mtext_generation_launcher[0m[2m:[0m Terminating download
[2m2024-08-16T12:08:09.102368Z[0m [32m INFO[0m [1mdownload[0m: [2mtext_generation_launcher[0m[2m:[0m Waiting for download to gracefully shutdown

TGI 基准测试

现在让我们学习如何启动基准测试工具！

在这里，我们可以看到 TGI 基准测试的不同设置。

以下是一些比较重要的 TGI 基准测试设置

--tokenizer-name 这项是必需的，以便工具知道使用哪个分词器
--batch-size 这对于负载测试很重要。我们应该使用足够的数值来观察吞吐量和延迟的变化。请注意，在基准测试工具中，批次大小指的是虚拟用户数量。
--sequence-length 又名输入标记，它与您的用例需求保持一致很重要
--decode-length 又名输出标记，它与您的用例需求保持一致很重要
--runs 默认值为 10

💡 提示： 在探索阶段使用较低的 --runs 数值，但在最终确定时使用较高的数值，以获得更精确的统计数据

>>> !text-generation-benchmark -h

Text Generation Benchmarking tool

[1m[4mUsage:[0m [1mtext-generation-benchmark[0m [OPTIONS] [1m--tokenizer-name[0m 

[1m[4mOptions:[0m
  [1m-t[0m, [1m--tokenizer-name[0m 
          The name of the tokenizer (as in model_id on the huggingface hub, or local path) [env: TOKENIZER_NAME=]
      [1m--revision[0m 
          The revision to use for the tokenizer if on the hub [env: REVISION=] [default: main]
  [1m-b[0m, [1m--batch-size[0m 
          The various batch sizes to benchmark for, the idea is to get enough batching to start seeing increased latency, this usually means you're moving from memory bound (usual as BS=1) to compute bound, and this is a sweet spot for the maximum batch size for the model under test
  [1m-s[0m, [1m--sequence-length[0m 
          This is the initial prompt sent to the text-generation-server length in token. Longer prompt will slow down the benchmark. Usually the latency grows somewhat linearly with this for the prefill step [env: SEQUENCE_LENGTH=] [default: 10]
  [1m-d[0m, [1m--decode-length[0m 
          This is how many tokens will be generated by the server and averaged out to give the `decode` latency. This is the *critical* number you want to optimize for LLM spend most of their time doing decoding [env: DECODE_LENGTH=] [default: 8]
  [1m-r[0m, [1m--runs[0m 
          How many runs should we average from [env: RUNS=] [default: 10]
  [1m-w[0m, [1m--warmups[0m 
          Number of warmup cycles [env: WARMUPS=] [default: 1]
  [1m-m[0m, [1m--master-shard-uds-path[0m 
          The location of the grpc socket. This benchmark tool bypasses the router completely and directly talks to the gRPC processes [env: MASTER_SHARD_UDS_PATH=] [default: /tmp/text-generation-server-0]
      [1m--temperature[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: TEMPERATURE=]
      [1m--top-k[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: TOP_K=]
      [1m--top-p[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: TOP_P=]
      [1m--typical-p[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: TYPICAL_P=]
      [1m--repetition-penalty[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: REPETITION_PENALTY=]
      [1m--frequency-penalty[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: FREQUENCY_PENALTY=]
      [1m--watermark[0m
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: WATERMARK=]
      [1m--do-sample[0m
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: DO_SAMPLE=]
      [1m--top-n-tokens[0m 
          Generation parameter in case you want to specifically test/debug particular decoding strategies, for full doc refer to the `text-generation-server` [env: TOP_N_TOKENS=]
  [1m-h[0m, [1m--help[0m
          Print help (see more with '--help')
  [1m-V[0m, [1m--version[0m
          Print version

以下是一个示例命令。请注意，我重复添加感兴趣的批次大小，以确保基准测试工具使用所有批次大小。我还根据估计的用户活动，确定了哪些批次大小很重要。

⚠️ 警告： 请注意，TGI 基准测试工具设计为在终端中使用，而不是在 Jupyter 笔记本中使用。这意味着您需要在 Jupyter 终端选项卡中复制/粘贴该命令。我只是为了方便起见将其放在这里。

!text-generation-benchmark \
--tokenizer-name astronomer/Llama-3-8B-Instruct-GPTQ-8-Bit \
--sequence-length 70 \
--decode-length 50 \
--batch-size 1 \
--batch-size 2 \
--batch-size 4 \
--batch-size 8 \
--batch-size 16 \
--batch-size 32 \
--batch-size 64 \
--batch-size 128

< > 在 GitHub 上更新

←使用 LLM 作为评委清理现有的偏好数据集使用知识图增强 RAG 推理→