Compatibility with TGI (Text Generation Inference) framework v1.4.3

#1
by hugging-face-infrax - opened

Hello Team, currently we've trying to deploy the jan-hq/komodo-7b-chat-adapter model. But it seems there's an issue.

Server and tools Spec

  • TGI: v1.4.3
  • Azure A100 (80 GB)

How to Reproduce

  1. Got access to the base model https://huggingface.co/Yellow-AI-NLP/komodo-7b-base so we can load the Peft model.

  2. Here's our docker-compose.yml file to running the inference framework.

services:
  llm:
    image: ghcr.io/huggingface/text-generation-inference:1.4.3
    container_name: llm
    command: >
      --model-id jan-hq/komodo-7b-chat-adapter
      --max-total-tokens 8192
      --max-input-length 4096
      --num-shard 1
      --max-top-n-tokens 1
      --max-best-of 1
      --trust-remote-code
      --disable-custom-kernels
      --max-stop-sequences 1
      --validation-workers 1
      --waiting-served-ratio 0
      --max-batch-total-tokens 8192
      --max-waiting-tokens 4096
      --cuda-memory-fraction 0.8
      --max-concurrent-requests 512
      --max-batch-prefill-tokens 8192
    volumes:
      - ./data:/data
    ports:
      - 8080:80
    shm_size: '1gb'
    environment:
      - "HUGGING_FACE_HUB_TOKEN=${TOKEN}"
    restart: always
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:80/health"]
      interval: 30s
      timeout: 45s
      start_period: 180s
    deploy:
      resources:
        reservations:
          devices:
          - driver: nvidia
            count: all
            capabilities: [gpu]
  1. Check the logs and we've got a warning like this.
on.rs:159: Warning: Token 'gubernur' was expected to have ID '34993' but was given ID 'None'
2024-03-20T05:42:17.368924Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'seiring' was expected to have ID '34994' but was given ID 'None'
2024-03-20T05:42:17.368927Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Imam' was expected to have ID '34995' but was given ID 'None'
2024-03-20T05:42:17.368929Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'pengurus' was expected to have ID '34996' but was given ID 'None'
2024-03-20T05:42:17.368932Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Premier' was expected to have ID '34997' but was given ID 'None'
2024-03-20T05:42:17.368934Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'teknik' was expected to have ID '34998' but was given ID 'None'
2024-03-20T05:42:17.368937Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Lombok' was expected to have ID '34999' but was given ID 'None'
2024-03-20T05:42:17.368939Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'penerimaan' was expected to have ID '35000' but was given ID 'None'
2024-03-20T05:42:17.368941Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Nah' was expected to have ID '35001' but was given ID 'None'
2024-03-20T05:42:17.368944Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Cabang' was expected to have ID '35002' but was given ID 'None'
2024-03-20T05:42:17.368946Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'berikan' was expected to have ID '35003' but was given ID 'None'
2024-03-20T05:42:17.368949Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Perhubungan' was expected to have ID '35004' but was given ID 'None'
2024-03-20T05:42:17.368953Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Gunakan' was expected to have ID '35005' but was given ID 'None'
2024-03-20T05:42:17.368956Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'Turki' was expected to have ID '35006' but was given ID 'None'
2024-03-20T05:42:17.368958Z  WARN tokenizers::tokenizer::serialization: /usr/local/cargo/registry/src/index.crates.io-6f17d22bba15001f/tokenizers-0.15.2/src/tokenizer/serialization.rs:159: Warning: Token 'fans' was expected to have ID '35007' but was given ID 'None'

Could you help if there's something missing from the configuration? Thank you

Sign up or log in to comment