diff --git a/aml/deployment.yaml b/aml/deployment.yaml index 88d05c307..9b90c2e6b 100644 --- a/aml/deployment.yaml +++ b/aml/deployment.yaml @@ -8,7 +8,7 @@ environment_variables: MODEL_ID: bigscience/bloom NUM_SHARD: 8 environment: - image: db4c2190dd824d1f950f5d1555fbadf0.azurecr.io/text-generation-inference:sha-941cd42 + image: db4c2190dd824d1f950f5d1555fbadf0.azurecr.io/text-generation-inference:sha-4e685d9 inference_config: liveness_route: port: 80 @@ -25,14 +25,14 @@ request_settings: max_concurrent_requests_per_instance: 256 liveness_probe: initial_delay: 200 - timeout: 20 + timeout: 60 period: 60 success_threshold: 1 - failure_threshold: 5 + failure_threshold: 2 readiness_probe: initial_delay: 200 - timeout: 20 + timeout: 60 period: 60 success_threshold: 1 - failure_threshold: 5 + failure_threshold: 2 instance_count: 1