1. # Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
2. #
3. # NVIDIA CORPORATION and its licensors retain all intellectual property
4. # and proprietary rights in and to this software, related documentation
5. # and any modifications thereto.  Any use, reproduction, disclosure or
6. # distribution of this software and related documentation without an express
7. # license agreement from NVIDIA CORPORATION is strictly prohibited.
8.  
9. # GPU family of target platform. Supported values: tegra, non-tegra
10. riva_target_gpu_family="tegra"
11.  
12. # Name of tegra platform that is being used. Supported tegra platforms: orin, xavier
13. riva_tegra_platform="orin"
14.  
15. # Enable or Disable Riva Services
16. service_enabled_asr=true
17. service_enabled_nlp=true
18. service_enabled_tts=true
19.  
20. # Enable Riva Enterprise
21. # If enrolled in Enterprise, enable Riva Enterprise by setting configuration
22. # here. You must explicitly acknowledge you have read and agree to the EULA.
23. # RIVA_API_KEY=<ngc api key>
24. # RIVA_API_NGC_ORG=<ngc organization>
25. # RIVA_EULA=accept
26.  
27. # Language code to fetch models of a specify language
28. # Currently only ASR supports languages other than English
29. # Supported language codes: en-US, en-GB, de-DE, es-US, ru-RU, zh-CN, hi-IN, fr-FR, ko-KR, pt-BR
30. # for any language other than English, set service_enabled_nlp and service_enabled_tts to False
31. # for multiple languages enter space separated language codes.
32. language_code=("en-US")
33.  
34. # ASR acoustic model architecture
35. # Supported values are: conformer, citrinet_1024, citrinet_256 (en-US + arm64 only), jasper (en-US + amd64 only), quartznet (en-US + amd64 only)
36. asr_acoustic_model=("conformer")
37.  
38. # Specify one or more GPUs to use
39. # specifying more than one GPU is currently an experimental feature, and may result in undefined behaviours.
40. gpus_to_use="device=0"
41.  
42. # Specify the encryption key to use to deploy models
43. MODEL_DEPLOY_KEY="tlt_encode"
44.  
45. # Locations to use for storing models artifacts
46. #
47. # If an absolute path is specified, the data will be written to that location
48. # Otherwise, a docker volume will be used (default).
49. #
50. # riva_init.sh will create a `rmir` and `models` directory in the volume or
51. # path specified.
52. #
53. # RMIR ($riva_model_loc/rmir)
54. # Riva uses an intermediate representation (RMIR) for models
55. # that are ready to deploy but not yet fully optimized for deployment. Pretrained
56. # versions can be obtained from NGC (by specifying NGC models below) and will be
57. # downloaded to $riva_model_loc/rmir by `riva_init.sh`
58. #
59. # Custom models produced by NeMo or TLT and prepared using riva-build
60. # may also be copied manually to this location $(riva_model_loc/rmir).
61. #
62. # Models ($riva_model_loc/models)
63. # During the riva_init process, the RMIR files in $riva_model_loc/rmir
64. # are inspected and optimized for deployment. The optimized versions are
65. # stored in $riva_model_loc/models. The riva server exclusively uses these
66. # optimized versions.
67. riva_model_loc="riva-model-repo"
68.  
69. if [[ $riva_target_gpu_family == "tegra" ]]; then
70.     riva_model_loc="`pwd`/model_repository"
71. fi
72.  
73. # The default RMIRs are downloaded from NGC by default in the above $riva_rmir_loc directory
74. # If you'd like to skip the download from NGC and use the existing RMIRs in the $riva_rmir_loc
75. # then set the below $use_existing_rmirs flag to true. You can also deploy your set of custom
76. # RMIRs by keeping them in the riva_rmir_loc dir and use this quickstart script with the
77. # below flag to deploy them all together.
78. use_existing_rmirs=false
79.  
80. # Ports to expose for Riva services
81. riva_speech_api_port="50051"
82.  
83. # NGC orgs
84. riva_ngc_org="nvidia"
85. riva_ngc_team="riva"
86. riva_ngc_image_version="2.7.0"
87. riva_ngc_model_version="2.7.0"
88.  
89. # Pre-built models listed below will be downloaded from NGC. If models already exist in $riva-rmir
90. # then models can be commented out to skip download from NGC
91.  
92. ########## ASR MODELS ##########
93.  
94. models_asr=()
95.  
96. for lang_code in ${language_code[@]}; do
97.     modified_lang_code="${lang_code/-/_}"
98.     modified_lang_code=${modified_lang_code,,}
99.     if [[ $riva_target_gpu_family  == "tegra" ]]; then
100.  
101.       if [[ ${asr_acoustic_model} == "jasper" || \
102.             ${asr_acoustic_model} == "quartznet" ]]; then
103.           echo "Jasper and Quartznet models are not available for arm64 architecture"
104.           exit 1
105.       fi
106.  
107.       if [[ ${asr_acoustic_model} == "citrinet_256" && ${lang_code} != "en-US" ]]; then
108.         echo "For arm64 architecture, citrinet_256 acoustic model is only available for language code en-US."
109.         exit 1
110.       fi
111.       models_asr+=(
112.           "${riva_ngc_org}/${riva_ngc_team}/models_asr_${asr_acoustic_model}_${modified_lang_code}_str:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
113.       )
114.     else
115.  
116.       if [[ ${asr_acoustic_model} != "conformer" && \
117.             ${asr_acoustic_model} != "citrinet_1024" && \
118.             ${asr_acoustic_model} != "jasper" && \
119.             ${asr_acoustic_model} != "quartznet" ]]; then
120.         echo "For amd64 architecture, valid acoustic models are conformer, citrinet_1024, jasper and quartznet."
121.         exit 1
122.       fi
123.  
124.       if [[ (${asr_acoustic_model} == "jasper" || \
125.             ${asr_acoustic_model} == "quartznet") && \
126.             ${lang_code} != "en-US" ]]; then
127.         echo "jasper and quartznet acoustic models are only available for language code en-US."
128.         exit 1
129.       fi
130.       #Setting default Conformer Mandarin models to greedy decoder due to high high latency in os2s.
131.       decoder=""
132.       if [[ ${asr_acoustic_model} == "conformer" && ${lang_code} == "zh-CN" ]]; then
133.         decoder="_gre"
134.       fi
135.       models_asr+=(
136.       ### Streaming w/ CPU decoder, best latency configuration
137.           "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}_${modified_lang_code}_str${decoder}:${riva_ngc_model_version}"
138.  
139.       ### Streaming w/ CPU decoder, best throughput configuration
140.       #    "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}_${modified_lang_code}_str_thr${decoder}:${riva_ngc_model_version}"
141.  
142.       ### Offline w/ CPU decoder,
143.           "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}_${modified_lang_code}_ofl${decoder}:${riva_ngc_model_version}"
144.       )
145.     fi
146.  
147.     ### Punctuation model
148.     if [[ "${lang_code}"  == "en-US" || "${lang_code}" == "de-DE" || "${lang_code}" == "hi-IN" || "${lang_code}" == "fr-FR" || "${lang_code}" == "es-US" || "${lang_code}" == "zh-CN" || "${lang_code}" == "en-GB" ]]; then
149.       if [[ $riva_target_gpu_family == "tegra" ]]; then
150.         models_asr+=(
151.             "${riva_ngc_org}/${riva_ngc_team}/models_nlp_punctuation_bert_base_${modified_lang_code}:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
152.         )
153.       else
154.         models_asr+=(
155.             "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_punctuation_bert_base_${modified_lang_code}:${riva_ngc_model_version}"
156.         )
157.       fi
158.     fi
159.  
160. done
161. ########## NLP MODELS ##########
162.  
163. if [[ $riva_target_gpu_family == "tegra" ]]; then
164.   models_nlp=(
165.   ### BERT Base Intent Slot model for misty domain fine-tuned on weather, smalltalk/personality, poi/map datasets.
166.       "${riva_ngc_org}/${riva_ngc_team}/models_nlp_intent_slot_misty_bert_base:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
167.  
168.   ### DistilBERT Intent Slot model for misty domain fine-tuned on weather, smalltalk/personality, poi/map datasets.
169.   #    "${riva_ngc_org}/${riva_ngc_team}/models_nlp_intent_slot_misty_distilbert:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
170.   )
171. else
172.   models_nlp=(
173.   ### Bert base Punctuation model
174.       "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_punctuation_bert_base_en_us:${riva_ngc_model_version}"
175.  
176.   ### BERT base Named Entity Recognition model fine-tuned on GMB dataset with class labels LOC, PER, ORG etc.
177.       "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_named_entity_recognition_bert_base:${riva_ngc_model_version}"
178.  
179.   ### BERT Base Intent Slot model fine-tuned on weather dataset.
180.       "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_intent_slot_bert_base:${riva_ngc_model_version}"
181.  
182.   ### BERT Base Question Answering model fine-tuned on Squad v2.
183.       "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_question_answering_bert_base:${riva_ngc_model_version}"
184.  
185.   ### Megatron345M Question Answering model fine-tuned on Squad v2.
186.   #    "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_question_answering_megatron:${riva_ngc_model_version}"
187.  
188.   ### Bert base Text Classification model fine-tuned on 4class (weather, meteorology, personality, nomatch) domain model.
189.       "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_text_classification_bert_base:${riva_ngc_model_version}"
190.   )
191. fi
192.  
193. ########## TTS MODELS ##########
194.  
195. if [[ $riva_target_gpu_family == "tegra" ]]; then
196.   models_tts=(
197.      "${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_en_us:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
198.   )
199. else
200.   models_tts=(
201.     "${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_en_us:${riva_ngc_model_version}"
202.   )
203. fi
204.  
205. NGC_TARGET=${riva_ngc_org}
206. if [[ ! -z ${riva_ngc_team} ]]; then
207.   NGC_TARGET="${NGC_TARGET}/${riva_ngc_team}"
208. else
209.   team="\"\""
210. fi
211.  
212. # Specify paths to SSL Key and Certificate files to use TLS/SSL Credentials for a secured connection.
213. # If either are empty, an insecure connection will be used.
214. # Stored within container at /ssl/servert.crt and /ssl/server.key
215. # Optional, one can also specify a root certificate, stored within container at /ssl/root_server.crt
216. ssl_server_cert=""
217. ssl_server_key=""
218. ssl_root_cert=""
219.  
220. # define docker images required to run Riva
221. image_speech_api="nvcr.io/${NGC_TARGET}/riva-speech:${riva_ngc_image_version}-l4t-aarch64"
222.  
223. # define docker images required to setup Riva
224. image_init_speech="nvcr.io/${NGC_TARGET}/riva-speech:${riva_ngc_image_version}-servicemaker-l4t-aarch64"
225.  
226. # daemon names
227. riva_daemon_speech="riva-speech"
228. if [[ $riva_target_gpu_family != "tegra" ]]; then
229.     riva_daemon_client="riva-client"
230. fi
231.