Qwen3.5-27B on RTX 5090 served via vLLM @ 77 tps

After maxing out my cursor $20 sub and zai $10 sub for this month, I have resorted to a local llm setup. Got good outcome on RTX5090 running Qwen3.5 27B and achieved very good tps. Context window at 218k. It can even run 2 concurrent sessions with this config although per session speed drops as expected. For some reason i can't get it to work at full context window of 256k on vllm 0.19, it works on vllm 0.17 per the guide below but tps suffers as 0.17 doesn't have many of the optimization that v