Checksum Mismatch Tanks GGUF Throughput

This is a reminder that “the model got slower” is often the wrong first diagnosis; file integrity can fail before your runtime does.

• –Corrupted weights can look like an inference regression, especially when throughput falls off a cliff without any config change
• –Checksumming downloaded models should be part of the default debugging flow for local LLMs, not an afterthought
• –The risk is higher when models are manually transformed or modified, because a bad conversion can quietly poison the artifact
• –The practical fix is simple: compare hashes before blaming quantization, kernels, or hardware