diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md index 49d402ba..3720774d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml index 08fc103c..059dc06d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md index 00fd6923..e130b44e 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml index 06d00182..57e6ea8d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md index 9d5e46b8..cb98d456 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml index f95eb18b..2914624c 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md index f4b4585e..cc08ab9c 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md @@ -168,6 +168,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml index c0afa0b0..cd1bc5d8 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md index 8fe64d72..c95b995f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md @@ -168,6 +168,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml index 07cee6c5..39d95bb3 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/README.md index e53fa535..e815a6b0 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/README.md @@ -180,6 +180,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/config.yaml index 82a31a26..35d13306 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-babelscape-wikineural-multilingual-ner/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/README.md index 9169b4dc..a6aa5613 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/README.md @@ -180,6 +180,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/config.yaml index a5c164dd..fc5e1859 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-dslim-bert-base-ner-uncased/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md index 349a6af9..2fa929b2 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml index f995b6e6..32da85ee 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md index 1a067245..9f9bdf49 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml index 5526ab54..043d03a3 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md index 91770d5c..62300c21 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml index 5d201f46..e865f119 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md index adb9d922..6dfa391f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml index c27ee069..5e5265db 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/README.md index c79527f0..7c3b3d1f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/README.md @@ -180,6 +180,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/config.yaml index 12670237..0c317229 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-lcampillos-roberta-es-clinical-trials-ner/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md index 7671ee6a..d6a7c5f4 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml index a1ac5872..3f390cc8 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md index 51ccd44c..1a7e898e 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml index de83f65e..58c00adf 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md index 17f64854..2a00b938 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml index 481f6336..b33af703 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md index 751e5c09..cd3e6504 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml index e54df7d7..651e646c 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md index de54a8b4..805810ef 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml index 6741abc9..c0c152e0 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md index fcea68db..4d29170d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml index 5055c160..11898eb3 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md index b2b9f1f5..9e31c4fe 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml index bd676606..506a8173 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md index 3233c22d..2f18ee1a 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml index e03d912f..1032b415 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md index 954dc8fd..92bbbcfd 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml index 6d709999..1255d441 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder_bert diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md index 18f7e265..58f24d2d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml index 7b45b33c..d3b8c18d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md index 7c5af841..b06c5080 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml index f5919a8f..be34073f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md index 9787f758..432b790f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml index d6162268..4691b5bc 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md index fde3603e..e34ca554 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml index 80ce0092..98f19caa 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md index 5c479a4c..bfeb314b 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml index 6ec3d19f..973fa2e4 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md index 3525c725..7534540c 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md @@ -168,6 +168,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml index 0e0ef388..5cce56bd 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md index e437588f..e6c2743c 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md @@ -168,6 +168,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml index 78b8a92b..8f09b428 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md index d65b060d..ae4e141b 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml index 365fb6a6..5cfd33a6 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md index a9d3c3f1..90096f4a 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml index 151d24a5..6817eb4f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md index 442d6e97..9d8dcdf1 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml index 2bc0e145..d228f14b 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md index d89fde45..257c4d22 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml index 3ec46322..f66c903f 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md index 77fa774f..b71e7e6d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml index 54a55f9f..b6d4f069 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md index a422be56..8b7688e8 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml index af2ccf72..6c16d03e 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md index 24988153..60b0d8c6 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml index 74b3bd80..1455f2fe 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md index 19aea075..68b4cf08 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md @@ -168,6 +168,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml index b283fb6f..ef9408e9 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md index 33d1983c..ea28aa4b 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml index 881343a1..305c4ed9 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md index 3f0227fb..d056728e 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml index 1a97785c..7625c4b7 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md index 02e60bd4..5dbf9139 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml index a80a3478..4bd58afd 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md index dbfef732..4923a35a 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml index 9ebddcf5..0753ddb1 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md index 913dbeed..d68e4a58 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml index 21b3da96..31959573 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md index c0d510c2..9c9a6338 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml index f7b30d76..e9f4c8c5 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md index b5797c63..25bac0b0 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml index de310dbe..053e46ee 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md index 459cb3ce..e6849b74 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml index 0d85c1a2..af70fc56 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md index e6ad83e0..87d15e0b 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml index a15f890d..7ac64c3c 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md index a1d7d7d2..14460ffb 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml index cb254db4..5e92b3ca 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md index ef8e1b81..fb788e48 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md @@ -162,6 +162,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml index ccd3c42d..047d8020 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md index d5bf004d..c7a8dc08 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml index 966d5a3c..424cb4d7 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md index 39899697..ecc4d620 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml index 43afb372..6846daae 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml @@ -14,6 +14,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md index e85501e4..dff42d48 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml index 20d35c63..74d2d609 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md index f30fa8fd..d1c0ca00 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md @@ -161,6 +161,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml index 5e490a51..887cc982 100644 --- a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml @@ -11,6 +11,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: encoder diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md index 0f2df2f5..0fd853a8 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml index b9d714ce..fe6c8fdb 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md index 388de80f..8233aaa2 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml index 8f9a8bc3..ba178548 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md index 2a588375..ff1f9092 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md @@ -151,6 +151,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml index 4695f8ba..8dcf2c15 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md index a4a9f741..99629bb0 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml index 9d864be6..07b47707 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md index 0f07efb6..963cde9d 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml index ff8e1215..f71bc477 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md index aba79f02..d34715a5 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml index a4f95eb8..77b4e000 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md index 08d12366..e78274cd 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml index 779a9d14..e95a46bd 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md index 4b4cae48..760fcc70 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml index 5b2b418d..8f11df4b 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md index 90f58826..b00386ae 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml index fc22eaf7..47f29af1 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md index f7ee35ec..879e9418 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml index d52a470a..e81a3510 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md index ce82ab19..8a5a75f2 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md @@ -151,6 +151,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: checkpoint_repository: diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml index be29a840..c306e665 100644 --- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml @@ -15,6 +15,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: inference_stack: v2 build: diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md index ad96330a..280f6736 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml index 296c65dd..8769473f 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md index 67bf8b5a..7be78ecf 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml index 43b6dc35..53be9de7 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md index 80b69da8..478710b6 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml index d051babb..f01c22b4 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md index dccff493..6fe874b5 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml index 5a29f297..e2e8fcce 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md index 1b14d9bf..edb2edd8 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml index 72a8349b..07118a33 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md index f0f5f39c..5c1d37ee 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml index c9666a8f..6ab7a25a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md index 44d49ec9..3f4a1b7a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml index 010a4e7f..8938f93c 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md index a3b93f92..39968d58 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml index 85829cea..16b04abb 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md index f15bc82b..f6c4e2d3 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml index 9006e0ee..54cb195a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md index 36ff5723..a6538a0a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml index 47a5b3b1..32c58d08 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md index e70711dc..1066c219 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml index 8d395a47..8d05bfcb 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md index 51edf882..a3d68624 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml index dcfb1d8a..7a0bbc1b 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md index aee9a97c..087c4adb 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml index da463d4a..c6277532 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md index 13d8b2b6..7cfd75fe 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml index 215dd074..400dc93c 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md index c401138e..8ef6d719 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml index be35731d..6ca3898f 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md index f35cf47b..5c263d36 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml index 972fad2c..ed52517b 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md index 282dd9fe..2ebbee7c 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml index eec4a13d..8df423f5 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md index b6e3eff9..6ec17bd2 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml index 97449aeb..aa8eec78 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md index f11d5700..8923ea09 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml index f07e9d02..567c0190 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md index c11eafc0..06956268 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml index 7438254e..46d68592 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md index a4e7260d..73e5c2a4 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml index f7b8b2cf..08b1b328 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md index 850c58dc..2135c108 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml index 40f51570..1e6e9f05 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md index 2fa13067..a7557ec2 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml index 3e61aa0c..3213c168 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md index 723d2136..9b5a3a3e 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml index 9a2e2a2b..85e686a1 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md index 885c5b9e..8213354a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml index c767998f..17f0c542 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md index c3a55f19..04536f94 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml index 994438d5..62ddddb1 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md index 0f5ae98a..c5ea8b1a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml index 5e919d60..4268c3be 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md index a37b059c..b9222eb7 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md @@ -149,6 +149,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml index 94709b28..5a835cc5 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md index 265a4731..674411ef 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml index f9999435..0a33001d 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md index c8834a83..4709d0c4 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml index e809dd7c..ee143f48 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md index 007468ce..4ff9b887 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml index 8ef63567..c52a7254 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md index 41f70537..7c4e10fb 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml index 5dfcd2a1..62b9bcdf 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md index dc5f1caa..d51db312 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md +++ b/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md @@ -150,6 +150,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml index 3548da52..35d3318a 100644 --- a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml @@ -16,6 +16,14 @@ resources: cpu: '1' memory: 10Gi use_gpu: true +runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 + is_websocket_endpoint: false + transport: + kind: http trt_llm: build: base_model: decoder diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md index ea445cc6..307ac7dd 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml index bf026acf..3ed9b548 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md index af132166..e2abb13d 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml index 4b5904a8..08f139e7 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md index 06bdaf04..553cda8b 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml index 98e3c0c9..9047bb88 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md index 2b6005a5..613cea37 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md @@ -198,6 +198,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml index daa6fa81..d3f40379 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml @@ -36,6 +36,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md index 600b1a9d..04bb73d1 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md @@ -198,6 +198,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml index 188e5838..4f5a5f08 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml @@ -36,6 +36,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md index 58ed299c..621194c1 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml index 9250697e..0c9d1b25 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md index ca746073..688e2270 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml index 98746c8d..62845585 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md index 6bad1f84..72ae5e09 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml index e68e8c95..a458d3b8 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md index d92a9723..8a0ef89b 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml index ac49be15..c1643bfc 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md index 07400264..5532ba9f 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml index 57b066b8..f78e31eb 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md index 03a9dceb..319ea30c 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml index 60998c34..8412479b 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md index f98793c0..58d8c187 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml index b035c847..84831a04 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md index 166582a3..973677dd 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml index 3834cc1d..5e59f88b 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md index 676b96dc..ef1b6331 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml index 050819f1..ef9caa67 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md index 4facc589..7ca4fcd6 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md +++ b/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md @@ -191,6 +191,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml index 54fb71b4..43deb810 100644 --- a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml +++ b/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml @@ -31,6 +31,10 @@ resources: memory: 2Gi use_gpu: true runtime: + health_checks: + restart_threshold_seconds: 30 + startup_threshold_seconds: 1800 + stop_traffic_threshold_seconds: 30 is_websocket_endpoint: false predict_concurrency: 32 transport: diff --git a/11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py b/11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py index ea62f49e..4d27ecbf 100644 --- a/11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py +++ b/11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py @@ -168,6 +168,13 @@ def make_truss_config(self, dp: "Deployment") -> TrussConfig: memory="10Gi", ), model_name=dp.model_nickname, + runtime=dict( + health_checks=dict( + startup_threshold_seconds=1800, + restart_threshold_seconds=30, + stop_traffic_threshold_seconds=30, + ) + ), ) @@ -265,6 +272,11 @@ def make_truss_config(self, dp: "Deployment") -> TrussConfig: model_name=dp.model_nickname, runtime=dict( predict_concurrency=32, + health_checks=dict( + startup_threshold_seconds=1800, + restart_threshold_seconds=30, + stop_traffic_threshold_seconds=30, + ), ), ) @@ -341,6 +353,13 @@ def make_truss_config(self, dp): ), model_name=dp.model_nickname, trt_llm=self.trt_config, + runtime=dict( + health_checks=dict( + startup_threshold_seconds=1800, + restart_threshold_seconds=30, + stop_traffic_threshold_seconds=30, + ) + ), ) @@ -417,6 +436,13 @@ def make_truss_config(self, dp): ), model_name=dp.model_nickname, trt_llm=self.trt_config, + runtime=dict( + health_checks=dict( + startup_threshold_seconds=1800, + restart_threshold_seconds=30, + stop_traffic_threshold_seconds=30, + ) + ), )