### What this PR does / why we need it? 1. update expected accuracy for DeepSeek-V2-Lite 2. add batch size ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? Accuracy CI passed - vLLM version: v0.10.2 - vLLM main: https://github.com/vllm-project/vllm/commit/838d7116ba59db528647b29f0d000742f4af9d4b Signed-off-by: hfadzxy <starmoon_zhang@163.com>