vLLM guide unlocks AWQ on Blackwell GPUs

This reads like the kind of hard-won operator knowledge that often matters more than the official compatibility table: not a new feature announcement, but a practical path through the current kernel gaps on bleeding-edge NVIDIA GPUs.

• –`awq_marlin` appears to be the right vLLM quantization path for AWQ weights on newer hardware, while `TRITON_ATTN` covers the attention side where FlashAttention still lacks SM_120 support.
• –The guide is especially useful because it targets WSL2 on Windows, where CUDA, PyTorch, and driver mismatches can make a seemingly model-specific failure look like a platform bug.
• –The latency numbers are helpful as sanity checks, but they’re anecdotal rather than a controlled benchmark, so readers should still validate throughput and stability on their own stack.
• –The Gemma 2 note is a good reminder that serving success and chat-template correctness are separate issues; a model can load cleanly and still fail at the frontend prompt layer.
• –For AI infra folks, the takeaway is simple: Blackwell support is starting to work in practice, but it still depends on picking the exact kernels vLLM currently prefers.