warranty

warranty reads the memory-error record an NVIDIA GPU keeps about itself and returns it in full, and is loadable through kernels. A GPU with ECC counts corrected and uncorrected errors by location, remaps failing DRAM rows into spare capacity, and records when a bank runs out of spares. nvidia-smi shows only part of this. The package reads all of it through NVML, scrubs a region of memory for a current measurement, classifies the card, and formats the result as an RMA ticket with a digest, or as an RMA request letter addressed to NVIDIA's chief executive.

Usage

from kernels import get_kernel

w = get_kernel("phanerozoic/warranty", version=1, trust_remote_code=True)

rep = w.report(0)            # counters, remap state, 2 GiB scrub, verdict
rep["verdict"]               # "healthy" ... "remap_exhausted"
print(w.claim(rep))          # ticket text, SHA-256 over the report
print(w.email(rep, sender="Your Name"))

python -m warranty.rma_email --sender "Your Name"        # prints the letter
python -m warranty.rma_email --json                      # prints the report

It reads counters and returns text. It does not write to the device and does not send anything. pynvml (nvidia-ml-py) is required; torch is used for the scrub when present.

API

Method

Counters come from nvmlDeviceGetMemoryErrorCounter over both scopes (volatile since the last reset, aggregate over the part's life), both kinds (corrected, uncorrected) and every location NVML defines; zero counts are omitted. Row remapping comes from nvmlDeviceGetRemappedRows (rows remapped for correctable and uncorrectable causes, whether a remap awaits a reset, and whether a remap has failed) and nvmlDeviceGetRowRemapperHistogram (banks grouped by how many spare rows remain). Parts that retire pages instead of remapping rows report through nvmlDeviceGetRetiredPages.

The scrub fills a region of device memory with a seeded pattern, reads it back, and counts mismatches, over three passes with the pattern inverted between them. The region is the smaller of scrub_gib and half of free memory.

The verdict is ordered by severity. A remap failure or a bank with no spare rows is remap_exhausted; any uncorrected count or scrub mismatch is uncorrectable; a remap or page retirement waiting on reset is remap_pending; any corrected count or correctable remap is degrading; otherwise healthy.

The digest is SHA-256 over the report serialized with sorted keys and no whitespace. claim and email embed it, so the text can be checked against the report.

Correctness

tests/test_warranty.py runs a fixed sample record through every verdict transition, checks that a scrub mismatch alone yields uncorrectable, that the digest in claim verifies and stops verifying once a counter is edited, that email carries the headers, salutation, counts, remap-failure text, digest, signature and attached ticket, and, when an NVML device is present, that report reads it and its digest verifies.

Requirements and limits

• pynvml (nvidia-ml-py). Without it, every entry point raises with the install hint.
• ECC counters exist only on parts with ECC enabled; consumer cards without ECC report empty counters and the verdict rests on the scrub.
• Row remapping is Ampere and later; page retirement is Volta and Turing. Each is reported when the part supports it and omitted otherwise.
• The scrub needs torch with CUDA and runs on the device with the same index; pass scrub_gib=0 to skip it.
• The letter is text. Sending it is up to you.

References

NVIDIA Management Library (NVML) API Reference, Device Queries: memory error counters, remapped rows, row remapper histogram, retired pages; NVIDIA GPU Memory Error Management (application note).

License

Apache-2.0.