Skip to main content

ローカル音声チャットボット

note

この文書は AI によって翻訳されています。内容に不正確な点や改善すべき点がございましたら、文書下部のコメント欄または以下の Issue ページにてご報告ください。
https://github.com/Seeed-Studio/wiki-documents/issues

ローカル音声チャットボット : reComputerでRivaとLlama2をデプロイする

はじめに

人工知能技術が急速に進化する中で、音声インタラクションは人間とコンピュータの重要なインタラクションモードとなっています。特にスマートホーム、パーソナルアシスタント、カスタマーサービスサポートなどの分野では、音声チャットボットの需要が大幅に増加しています。しかし、既存の音声チャットボットの多くはクラウドコンピューティングサービスに依存しており、データプライバシーやネットワーク遅延に関する懸念が生じることがあります。

このプロジェクトでは、これらの問題に対処するために、ローカルで動作する音声チャットボットを構築することを目指しています。Nvidia RivaMeta Llama2を活用し、安全でプライバシーを保護し、迅速に応答する音声インタラクションシステムを開発しました。

前提条件

  • 16GB以上のメモリを搭載したJetsonデバイス。
  • ハードウェアデバイスは、jetpack 5.1.1オペレーティングシステムが事前にフラッシュされている必要があります。
  • スピーカーとマイク

注意: 私はすべての実験をJetson AGX Orin 32GB H01 Kitを使用して完了しましたが、メモリが少ないデバイスで小型モデルをロードすることも試すことができます。

始めるにあたって

ハードウェア接続

  • 音声入力/出力デバイスをreComputerに接続します。
  • reComputerの電源を入れ、正常なネットワークアクセスがあることを確認します。

Rivaサーバーのインストール

詳細についてはこちらを参照してください。

ステップ1. NVIDIA NGCにアクセスしてログインします。

ステップ2. NGC APIキーを取得します。

アカウント(右上隅) --> 設定 --> APIキーを取得 --> APIキーを生成 --> 確認

info

生成されたAPIキーを記録してください。

ステップ3. reComputerでNGCを設定します。

reComputerのターミナルを開きます(reComputerのデスクトップでショートカットキーCtrl+Alt+Tを使用してターミナルをすばやく開くか、別のコンピュータを使用してreComputerのターミナルにリモートアクセスできます)。以下のコマンドを順に入力します。

cd ~ && mkdir ngc_setup && cd ngc_setup
wget --content-disposition https://api.ngc.nvidia.com/v2/resources/nvidia/ngc-apps/ngc_cli/versions/3.36.0/files/ngccli_arm64.zip && unzip ngccli_arm64.zip
chmod u+x ngc-cli/ngc
echo "export PATH=\"\$PATH:$(pwd)/ngc-cli\"" >> ~/.bash_profile && source ~/.bash_profile
ngc config set

ターミナルのインタラクティブインターフェースで関連情報を入力します。

ステップ4. reComputerでRivaサーバーをインストールして実行します。

reComputerのターミナルで以下のコマンドを入力します。

cd ~ && mkdir riva_setup && cd riva_setup
ngc registry resource download-version nvidia/riva/riva_quickstart_arm64:2.13.1
cd riva_quickstart_v2.13.1

Vimを使用して設定ファイルを編集します。

vim config.sh

# キーボードの`A`キーを押して編集モードに入ります。
# 以下の指示に従って18行目と20行目を編集します。

# service_enabled_nlp=true --> service_enabled_nlp=false
# service_enabled_nmt=true --> service_enabled_nmt=false

# キーボードの`ESC`キーを押して編集モードを終了し、ショートカット`Shift+Z Z`を使用して編集内容を保存してエディタを閉じます。

編集後の設定ファイル:

config.sh
# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.

# GPU family of target platform. Supported values: tegra, non-tegra
riva_target_gpu_family="non-tegra"

# Name of tegra platform that is being used. Supported tegra platforms: orin, xavier
riva_tegra_platform="orin"

# Enable or Disable Riva Services
# For any language other than en-US: service_enabled_nlp must be set to false
service_enabled_asr=true
service_enabled_nlp=false
service_enabled_tts=true
service_enabled_nmt=false

# Configure translation services
# Text-to-Text translation (T2T):
# - service_enabled_nmt must be set to true
# - Uncomment desired model for source and target languages in models_nmt field
# Speech-to-Text translation (S2T):
# - service_enabled_asr, service_enabled_nmt must be set to true
# - Set language code of input speech in the asr_language_code field
# - Uncomment desired model for source and target languages in models_nmt field
# Speech-to-Speech translation (S2S):
# - service_enabled_asr, service_enabled_nmt, service_enabled_tts must be set to true
# - Set language code of input speech in the asr_language_code field
# - Uncomment desired model for source and target languages in models_nmt field
# - Set language code of output speech in the tts_language_code field

# Enable Riva Enterprise
# If enrolled in Enterprise, enable Riva Enterprise by setting configuration
# here. You must explicitly acknowledge you have read and agree to the EULA.
# RIVA_API_KEY=<ngc api key>
# RIVA_API_NGC_ORG=<ngc organization>
# RIVA_EULA=accept

# Language code to fetch ASR models of a specific language
# Supported language codes: ar-AR, en-US, en-GB, de-DE, es-ES, es-US, fr-FR, hi-IN, it-IT, ja-JP, ru-RU, ko-KR, pt-BR, zh-CN, es-en-US, ja-en-JP
# For multiple languages enter space separated language codes.
asr_language_code=("en-US")

# ASR acoustic model architecture
# Supported values are: conformer, conformer_xl (en-US + amd64 only), citrinet_1024, citrinet_256 (en-US + arm64 only), jasper (en-US + amd64 only), quartznet (en-US + amd64 only)
asr_acoustic_model=("conformer")

# ASR acoustic model architecture variant
# Supported values for the architecture are:
# conformer: unified(de-DE, ja-JP and zh-CN only), ml_cs(es-en-US only), unified_ml_cs(ja-en-JP only)
# For the default model, keep the field empty
asr_acoustic_model_variant=("")

# ASR decoder type to be used
# If you'd like to use greedy decoder for ASR instead of flashlight/os2s decoder then set the below $use_asr_greedy_decoder to true
use_asr_greedy_decoder=false

# Language code to fetch TTS models of a specific language
# Supported language codes: en-US, es-ES, it-IT, de-DE, zh-CN
# For multiple languages enter space separated language codes
tts_language_code=("en-US")

# Specify one or more GPUs to use
# specifying more than one GPU is currently an experimental feature, and may result in undefined behaviours.
gpus_to_use="device=0"

# Specify the encryption key to use to deploy models
MODEL_DEPLOY_KEY="tlt_encode"

# Locations to use for storing models artifacts
#
# If an absolute path is specified, the data will be written to that location
# Otherwise, a Docker volume will be used (default).
#
# riva_init.sh will create a `rmir` and `models` directory in the volume or
# path specified.
#
# RMIR ($riva_model_loc/rmir)
# Riva uses an intermediate representation (RMIR) for models
# that are ready to deploy but not yet fully optimized for deployment. Pretrained
# versions can be obtained from NGC (by specifying NGC models below) and will be
# downloaded to $riva_model_loc/rmir by `riva_init.sh`
#
# Custom models produced by NeMo or TLT and prepared using riva-build
# may also be copied manually to this location $(riva_model_loc/rmir).
#
# Models ($riva_model_loc/models)
# During the riva_init process, the RMIR files in $riva_model_loc/rmir
# are inspected and optimized for deployment. The optimized versions are
# stored in $riva_model_loc/models. The riva server exclusively uses these
# optimized versions.
riva_model_loc="riva-model-repo"

if [[ $riva_target_gpu_family == "tegra" ]]; then
riva_model_loc="`pwd`/model_repository"
fi

# The default RMIRs are downloaded from NGC by default in the above $riva_rmir_loc directory
# If you'd like to skip the download from NGC and use the existing RMIRs in the $riva_rmir_loc
# then set the below $use_existing_rmirs flag to true. You can also deploy your set of custom
# RMIRs by keeping them in the riva_rmir_loc dir and use this quickstart script with the
# below flag to deploy them all together.
use_existing_rmirs=false

# Ports to expose for Riva services
riva_speech_api_port="50051"

# NGC orgs
riva_ngc_org="nvidia"
riva_ngc_team="riva"
riva_ngc_image_version="2.13.1"
riva_ngc_model_version="2.13.0"

# Pre-built models listed below will be downloaded from NGC. If models already exist in $riva-rmir
# then models can be commented out to skip download from NGC

########## ASR MODELS ##########

models_asr=()

for lang_code in ${asr_language_code[@]}; do
modified_lang_code="${lang_code//-/_}"
modified_lang_code=${modified_lang_code,,}

decoder=""
if [ "$use_asr_greedy_decoder" = true ]; then
decoder="_gre"
fi

if [[ ${asr_acoustic_model_variant} != "" ]]; then
if [[ ${asr_acoustic_model} == "conformer" && ${asr_acoustic_model_variant} != "unified" && ${asr_acoustic_model_variant} != "ml_cs" && ${asr_acoustic_model_variant} != "unified_ml_cs" ]]; then
echo "Valid variants for Conformer are: unified, ml_cs and unified_ml_cs."
exit 1
elif [[ ${asr_acoustic_model} != "conformer" ]]; then
echo "Invalid variant for ${asr_acoustic_model}."
exit 1
fi
asr_acoustic_model_variant="_${asr_acoustic_model_variant}"
fi

if [[ ${asr_acoustic_model} == "conformer_xl" && ${lang_code} != "en-US" ]]; then
echo "Conformer-XL acoustic model is only available for language code en-US."
exit 1
fi

if [[ ${asr_acoustic_model_variant} == "_unified" && ${lang_code} != "de-DE" && ${lang_code} != "ja-JP" && ${lang_code} != "zh-CN" ]]; then
echo "Unified Conformer acoustic model is only available for language code de-DE, ja-JP and zh-CN."
exit 1
fi

if [[ ${asr_acoustic_model_variant} == "_ml_cs" && ${lang_code} != "es-en-US" ]]; then
echo "Multilingual Code Switch Conformer acoustic model is only available for language code es-en-US."
exit 1
fi

if [[ ${asr_acoustic_model_variant} == "_unified_ml_cs" && ${lang_code} != "ja-en-JP" ]]; then
echo "Unified Multilingual Code Switch Conformer acoustic model is only available for language code ja-en-JP."
exit 1
fi

if [[ $riva_target_gpu_family == "tegra" ]]; then

if [[ ${asr_acoustic_model} == "jasper" || \
${asr_acoustic_model} == "quartznet" || \
${asr_acoustic_model} == "conformer_xl" ]]; then
echo "Conformer-XL, Jasper and Quartznet models are not available for arm64 architecture"
exit 1
fi

if [[ ${asr_acoustic_model} == "citrinet_256" && ${lang_code} != "en-US" ]]; then
echo "For arm64 architecture, citrinet_256 acoustic model is only available for language code en-US."
exit 1
fi

models_asr+=(
### Streaming w/ CPU decoder, best latency configuration
"${riva_ngc_org}/${riva_ngc_team}/models_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_str:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"

### Offline w/ CPU decoder
# "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_ofl${decoder}:${riva_ngc_model_version}"
)
else

if [[ ${asr_acoustic_model} != "conformer" && \
${asr_acoustic_model} != "conformer_xl" && \
${asr_acoustic_model} != "citrinet_1024" && \
${asr_acoustic_model} != "jasper" && \
${asr_acoustic_model} != "quartznet" ]]; then
echo "For amd64 architecture, valid acoustic models are conformer, conformer_xl, citrinet_1024, jasper and quartznet."
exit 1
fi

if [[ (${asr_acoustic_model} == "jasper" || \
${asr_acoustic_model} == "quartznet") && \
${lang_code} != "en-US" ]]; then
echo "jasper and quartznet acoustic models are only available for language code en-US."
exit 1
fi

models_asr+=(
### Streaming w/ CPU decoder, best latency configuration
"${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_str${decoder}:${riva_ngc_model_version}"

### Streaming w/ CPU decoder, best throughput configuration
# "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_str_thr${decoder}:${riva_ngc_model_version}"

### Offline w/ CPU decoder
"${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_ofl${decoder}:${riva_ngc_model_version}"
)
fi

### Punctuation model
if [[ ${asr_acoustic_model_variant} != "_unified" && ${asr_acoustic_model_variant} != "_unified_ml_cs" ]]; then
pnc_lang=$(echo $modified_lang_code | cut -d "_" -f 1)
pnc_region=${modified_lang_code##*_}
modified_lang_code=${pnc_lang}_${pnc_region}
if [[ $riva_target_gpu_family == "tegra" ]]; then
models_asr+=(
"${riva_ngc_org}/${riva_ngc_team}/models_nlp_punctuation_bert_base_${modified_lang_code}:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
else
models_asr+=(
"${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_punctuation_bert_base_${modified_lang_code}:${riva_ngc_model_version}"
)
fi
fi
done

### Speaker diarization model
models_asr+=(
# "${riva_ngc_org}/${riva_ngc_team}/rmir_diarizer_offline:${riva_ngc_model_version}"
)

########## NLP MODELS ##########

if [[ $riva_target_gpu_family == "tegra" ]]; then
models_nlp=(
### Bert base Punctuation model
"${riva_ngc_org}/${riva_ngc_team}/models_nlp_punctuation_bert_base_en_us:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"

### BERT Base Intent Slot model for misty domain fine-tuned on weather, smalltalk/personality, poi/map datasets.
# "${riva_ngc_org}/${riva_ngc_team}/models_nlp_intent_slot_misty_bert_base:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"

### DistilBERT Intent Slot model for misty domain fine-tuned on weather, smalltalk/personality, poi/map datasets.
# "${riva_ngc_org}/${riva_ngc_team}/models_nlp_intent_slot_misty_distilbert:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
else
models_nlp=(
### Bert base Punctuation model
"${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_punctuation_bert_base_en_us:${riva_ngc_model_version}"

### BERT base Named Entity Recognition model fine-tuned on GMB dataset with class labels LOC, PER, ORG etc.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_named_entity_recognition_bert_base:${riva_ngc_model_version}"

### BERT Base Intent Slot model fine-tuned on weather dataset.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_intent_slot_bert_base:${riva_ngc_model_version}"

### BERT Base Question Answering model fine-tuned on Squad v2.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_question_answering_bert_base:${riva_ngc_model_version}"

### Megatron345M Question Answering model fine-tuned on Squad v2.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_question_answering_megatron:${riva_ngc_model_version}"

### Bert base Text Classification model fine-tuned on 4class (weather, meteorology, personality, nomatch) domain model.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_text_classification_bert_base:${riva_ngc_model_version}"
)
fi

########## TTS MODELS ##########

models_tts=()

for lang_code in ${tts_language_code[@]}; do
modified_lang_code="${lang_code//-/_}"
modified_lang_code=${modified_lang_code,,}

if [[ $riva_target_gpu_family == "tegra" ]]; then
if [[ ${lang_code} == "en-US" ]]; then
models_tts+=(
### These models have been trained with energy conditioning and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_en_us_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
# "${riva_ngc_org}/${riva_ngc_team}/models_tts_radtts_hifigan_en_us_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"

### This model uses the ARPABET for inference and training.
# "${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_en_us:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
elif [[ ${lang_code} == "zh-CN" ]]; then
models_tts+=(
### This model is multi-speaker with emotion and and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_zh_cn_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
else
### These models are single-speaker and use the International Phonetic Alphabet (IPA) for inference and training.
if [[ ${lang_code} != "de-DE" ]]; then
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_${modified_lang_code}_f_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
fi
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_${modified_lang_code}_m_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
fi
else
if [[ ${lang_code} == "en-US" ]]; then
models_tts+=(
### These models have been trained with energy conditioning and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_en_us_ipa:${riva_ngc_model_version}"
# "${riva_ngc_org}/${riva_ngc_team}/rmir_tts_radtts_hifigan_en_us_ipa:${riva_ngc_model_version}"

### This model uses the ARPABET for inference and training.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_en_us:${riva_ngc_model_version}"
)
elif [[ ${lang_code} == "zh-CN" ]]; then
models_tts+=(
### This model is multi-speaker with emotion and and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_zh_cn_ipa:${riva_ngc_model_version}"
)
else
### These models are single-speaker and use the International Phonetic Alphabet (IPA) for inference and training.
if [[ ${lang_code} != "de-DE" ]]; then
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_${modified_lang_code}_f_ipa:${riva_ngc_model_version}"
)
fi
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_${modified_lang_code}_m_ipa:${riva_ngc_model_version}"
)
fi
fi
done

######### NMT models ###############

# Models follow Source language _ One or more target languages model architecture
# Source or target language "any" means the model supports 32 languages mentioned in docs.
# e.g., rmir_nmt_de_en_24x6 is a German to English 24x6 bilingual model
# and rmir_megatronnmt_en_any_500m is a English to 32 languages megatron model

models_nmt=(
###### Bilingual models
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_de_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_es_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_zh_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_ru_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_fr_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_de_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_es_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_ru_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_zh_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_fr_en_24x6:${riva_ngc_model_version}"

###### Multilingual models
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_deesfr_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_deesfr_12x2:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_deesfr_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_deesfr_en_12x2:${riva_ngc_model_version}"

###### Megatron models
#"${riva_ngc_org}/${riva_ngc_team}/rmir_megatronnmt_any_en_500m:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_megatronnmt_en_any_500m:${riva_ngc_model_version}"
)

NGC_TARGET=${riva_ngc_org}
if [[ ! -z ${riva_ngc_team} ]]; then
NGC_TARGET="${NGC_TARGET}/${riva_ngc_team}"
else
team="\"\""
fi

# Specify paths to SSL Key and Certificate files to use TLS/SSL Credentials for a secured connection.
# If either are empty, an insecure connection will be used.
# Stored within container at /ssl/servert.crt and /ssl/server.key
# Optional, one can also specify a root certificate, stored within container at /ssl/root_server.crt
ssl_server_cert=""
ssl_server_key=""
ssl_root_cert=""

# define Docker images required to run Riva
image_speech_api="nvcr.io/${NGC_TARGET}/riva-speech:${riva_ngc_image_version}"

# define Docker images required to setup Riva
image_init_speech="nvcr.io/${NGC_TARGET}/riva-speech:${riva_ngc_image_version}-servicemaker"

# daemon names
riva_daemon_speech="riva-speech"
if [[ $riva_target_gpu_family != "tegra" ]]; then
riva_daemon_client="riva-client"
fi

同様の方法で /etc/docker/daemon.json を修正します。

sudo vim /etc/docker/daemon.json
# この行を追加 >> "default-runtime": "nvidia"

# キーボードで `ESC` を押して編集モードを終了し、ショートカット `Shift+Z Z` を使用して編集内容を保存し、エディタを閉じます。

sudo systemctl restart docker

編集後の設定ファイルは以下のようになります:

/etc/docker/daemon.json
{   
"default-runtime": "nvidia",
"runtimes": {
"nvidia": {
"path": "nvidia-container-runtime",
"runtimeArgs": []
}
}
}

以下のコマンドを使用して Riva を初期化し、起動します:

sudo bash riva_init.sh
sudo bash riva_start.sh
info

注意:このターミナルを閉じずに維持してください。

LLM のインストールと実行

インストールプロセスを簡略化するために、Dusty の jetson-containers プロジェクトを参照して text generation inference をインストールし、text generation inference を使用して Llama2 7B 大規模言語モデルをロードします。新しいターミナルを開き、以下のコマンドを実行してください。

cd ~
git clone https://github.com/dusty-nv/jetson-containers.git
cd jetson-containers
pip install -r requirements.txt
./run.sh $(./autotag text-generation-inference)
export HUGGING_FACE_HUB_TOKEN=<your hugging face token>
text-generation-launcher --model-id meta-llama/Llama-2-7b-chat-hf --port 8899
info

Hugging Face トークンは こちら から取得できます。

info

注意:このターミナルを閉じずに維持してください。

ローカルチャットボットデモをクローンして実行してみる

現在、少なくとも2つのターミナルが開いているはずです。一つは Riva サーバーを実行しており、もう一つは text generation inference サーバーを実行しています。次に、新しいターミナルを開いてデモを実行します。

cd ~
git clone https://github.com/yuyoujiang/Deploy-Riva-LLama-on-Jetson.git
cd Deploy-Riva-LLama-on-Jetson

# オーディオ入力/出力デバイスを確認します。
python3 local_chatbot.py --list-input-devices
python3 local_chatbot.py --list-output-devices

python3 local_chatbot.py --input-device <your device id> --output-device <your device id>
# 例: python3 local_chatbot.py --input-device 25 --output-device 30

効果のデモンストレーション

参考文献

技術サポート & 製品に関する議論

弊社製品をご利用いただきありがとうございます!製品の使用体験がスムーズになるよう、さまざまなサポートを提供しております。お客様の好みやニーズに合わせた複数のコミュニケーションチャネルをご用意しています。

Loading Comments...