vLLM update brings DeepSeek V4.1-Flash to eight A100 GPUs

Merged on November 16th, the decode path adds A100 support. The contributor reports 488.8 P90 tokens per second per user on H100 against an InferenceX line of 323.6, using different configurations.

By · Published · Updated

An open server chassis holds two large finned components and bundled cables, with green indicator lights and an “Rw” logo on the rear panel.

Merged on November 16th, the decode path adds A100 support. The contributor reports 488.8 P90 tokens per second per user on H100 against an InferenceX line of 323.6, using different configurations.