77tps on the 5090

Serving Qwen3.5-27B-Text-NVFP4-MTP with vLLM 0.19 on an RTX 5090 achieved 77 tps.

• Context window: works up to 218,592 tokens
• Concurrent sessions: up to 2, though speed per session drops
• Issue: full 256k context does not work on vLLM 0.19
• Alternative: vLLM 0.17 works as guided, but tps is lower due to lack of optimization

Settings for reproduction were also provided.

• --gpu-memory-utilization 0.93
• --attention-backend flashinfer
• --performance-mode interactivity
• --kv-cache-dtype fp8_e4m3
• --max-num-seqs 2
• --quantization modelopt
• --enable-prefix-caching
• --tool-call-parser qwen3_coder

They also noted that vLLM's KV size calculation fix PR was key, and that while the model card's guide is useful, other models they tested didn't perform as well as expected, so they don't recommend them.