Is this page helpful?

Troubleshoot NVIDIA NIM for Nemotron Parse#

Use this documentation to troubleshoot NVIDIA NIM for Nemotron Parse.

list-model-profiles Command Is Not Available#

Parse NIM has no notion of model profile selection. No profile selection step is required, and the list-model-profiles command is not available. For details, refer to Model Profiles and Kernel Selection.

The NIM Fails to Start#

Confirm that you accepted the container license terms in the NGC Catalog, that the NVIDIA driver and Container Toolkit are current, and that the selected GPU is listed in the support matrix.

If the NIM can’t download the model weights, confirm the following:

  • You passed the credentials for your download provider: NGC_API_KEY for NGC, or HF_TOKEN for Hugging Face. A complete, verified local checkpoint does not require credentials. For details, refer to Model Download Providers.

  • The host directory mounted at /model is writable by the user that runs the container (-u $(id -u)).

  • On an air-gapped host, the directory mounted at /model contains the parse subdirectory with the supported checkpoint staged in advance.

Startup checks the checkpoint files and fails if a nonempty file is corrupt or belongs to another revision. Stage the supported checkpoint in a clean directory and update the mount. The NIM does not replace incompatible nonempty files automatically.

For details, refer to Launch the NIM.

The NIM Fails with an Out-of-Memory Error#

Confirm that the selected GPU is listed in the support matrix. Decoder capacity increases GPU memory use. Reduce NIM_PIPELINE_MAX_BATCH_SIZE; the minimum is 1. An explicit value overrides the capacity selected by NIM_PERFORMANCE_MODE. For details, refer to Configure NVIDIA NIM for Nemotron Parse.

A Request Returns HTTP 415#

The /v1/parse endpoint accepts only rendered page images in PNG or JPEG format. PDF and SVG files return HTTP 415. Also make sure that the document part declares an image media type, for example type=image/png. The NIM checks the declared media type, not the file contents. To process a PDF, render each page to an image first. For details, refer to Parse a PDF.

A Request Returns HTTP 422#

Check that the image is nonempty and can be decoded. Both source dimensions must be at least 17 pixels. A valid PNG or JPEG below this minimum also returns HTTP 422. For details, refer to Parse a PDF.

The Markdown Output Is Incomplete#

If the response has finish_reason: "length", generation reached the output limit. If your request used fewer than 9,000 tokens, increase options.decoding.max_output_tokens or omit options to use 9,000.

If the output still ends with length at 9,000 tokens, there is no continuation or resume option. Try improving the page image readability, or crop it into smaller regions and parse each region separately. Keep each region within the input bounds and check the combined output for missing or repeated content. A stop response means the model generated an end-of-sequence token; it does not verify extraction completeness.

A Request Returns HTTP 503 or 504#

For /v1/parse, HTTP 503 means that the queue is full, fail-fast admission rejected the request, or a worker is unavailable. Retry after the interval in Retry-After, and reduce client concurrency. With NIM_SERVER_FAIL_FAST=1, requests are also rejected when reserved and active work reaches NIM_PIPELINE_MAX_BATCH_SIZE, even before the queue is full. Increasing NIM_SERVER_MAX_QUEUE_SIZE allows more waiting requests but does not increase decoder capacity.

HTTP 504 means that queue wait and inference exceeded NIM_SERVER_REQUEST_TIMEOUT_S. This deadline starts after upload and multipart parsing. Reduce client concurrency or choose a work timeout appropriate for the workload, up to 3,600 seconds.

A health endpoint can also return HTTP 503. Its body has object: "health.response", message: "unavailable", and ready: false or live: false; this is a failed health check, not a queue-overload response.