Troubleshooting

View as Markdown

Start with CloudWatch Logs for the endpoint — the model container streams its logs there, and most invoke failures explain themselves in the log line for that request.

Deployment issues

Usually a region mismatch or a permissions gap. Confirm the model package ARN is for the same region as your client, that you’re subscribed to the listing, and that your role has AmazonSageMakerFullAccess and AWSMarketplaceManageSubscriptions.

Almost always a service quota for the GPU instance type. Check the failure reason on the endpoint, then request the instance quota in the Service Quotas console and retry.

Invoke errors

Every non-2xx arrives as a boto3 ModelError (HTTP 424) — read the real status in OriginalStatusCode and the body in OriginalMessage (see Error handling on SageMaker). Check the Content-Type: real-time Saaras v3 needs multipart/form-data; boundary=…, Bulbul v3 needs application/json, and Vision needs the document’s MIME type. Confirm model is saaras:v3 or bulbul:v3, not a package version like saaras:v3.1.

Audio over 30 s on Saaras v3 REST, Bulbul v3 text over 2500 chars (real-time), a Vision sync document over ~5 pages (or over the 500-page hard cap) or one that can’t finish inside the 60 s InvokeEndpoint window, or a payload over ~6 MB — move to streaming (Saaras v3 / Bulbul v3) or async / batch (Vision). See Limitations.

Streaming accepts only 8 kHz or 16 kHz, mono, 16-bit PCM. Resample the audio and confirm the sample_rate query parameter matches the stream.

Latency & throttling

At capacity the container sheds fast with 429 (Saaras v3) or 503 (Bulbul v3, Sarvam Vision) carrying error.code: "service_overloaded" and a body retry_after — this is expected backpressure and is retryable. Back off for retry_after seconds with jitter. It is distinct from SageMaker’s own ThrottlingException (an invocation-rate limit on the AWS API, not the model). See Error handling on SageMaker.

The endpoint is saturated. Add instances or enable autoscaling, move to a larger instance, or raise per-instance concurrency (see Configure & tune). Watch ModelLatency and InvocationsPerInstance in CloudWatch.

Switch spiky workloads to an async endpoint with scale-to-zero, so you don’t pay for idle instances.

Still stuck? Contact developer@sarvam.ai with your endpoint name, region, and the relevant CloudWatch log lines.