Opening as draft until vllm v0.17.0 is released. This release will include multiple optimizations for gpt-oss:

• Enable https://huggingface.co/amd/gpt-oss-120b-w-mxfp4-a-fp8 moe quant: [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales vllm-project/vllm#30357 (10-15%)
• RoPE+KVCache fusion: [ROCm] AITER fused RoPE+KVCache vllm-project/vllm#33443 (3-4%)
• Enable AITER RoPE: [ROCm]: Enable customop and rope+kvcache fusion for AITER RoPE vllm-project/vllm#35180 (1%)

Some other updates:

• gpt-oss CK moe backend: [ROCm][Quantization] Add Composable Kernel (CK) backend support for M… vllm-project/vllm#34301 (Not used in this PR since Triton W4A8 is better, but was previously the default MOE on MI355 before [AMD] Update AMD MI300X, MI325X, MI355X GPT-OSS vLLM images to v0.16.0 #806)
• Attention backends can no longer be passed in as env vars in vLLM, update to --attention-backend: [Deprecation] Remove deprecated environment variables vllm-project/vllm#32812
• FULL_AND_PIECEWISE is already the default, no need to specify it

Author: https://github.com/Rohan138