openjev on P300x2

Serves AlexWortega/openjev, a Qwen3.5-4B NLI cross-encoder, as a vLLM pooling model on four Blackhole chips with 4-way tensor parallelism. The repo holds tt-metal and vllm-tt-plugin patches (classifier head, pooled prefill traces, pooling support), a tt-inference-server catalog entry, eval scripts and results: MNLI-500 accuracy 0.892 on TT vs 0.896 on CPU bf16, and about 32 ms per input up to 128 tokens. Includes a zero-shot Flappy Bird demo driven by /classify calls, with videos and a browser replay viewer. Weights are not duplicated; they load from the upstream checkpoint.