vLLM reports B300 DEP8 MTP3 validation and sparse prefill latency gains

vLLM reports that its sparse prefill path on B300 DEP8 now runs the indexer logits kernel 43% faster and the top-k kernel 84% faster, cutting overall prefill latency by 2.2%.

The measurements use MTP3, FP8 KV, a 60K-token input and the MegaMoE path, and the change also removes 7 runtime JIT warmups and caps chunks at 4,096 tokens.

In a separate post, vLLM says MTP3 scored 0.9667 on AIME25 against a 0.9500 baseline, while GPQA came in at 0.9242 versus 0.9268 and OCRBench at 0.895 versus 0.896. vLLM says two validation rounds on the tested build recorded zero request errors, and that final-release validation is not yet complete.