rvLLM challenges vLLM in Rust

This is the right kind of vLLM challenger: less hand-wavy AI abstraction, more systems-level pressure on the serving stack where ops pain actually lives.

• –Near-parity at batch sizes 32-64 on H100 suggests the Rust port is credible, not just a benchmark vanity project
• –The ~50 MB container and 35-second source build are operational advantages that matter in CI, deployment, and reproducibility
• –The gap at batch 1 and batch 128 means “drop-in replacement” is still aspirational, especially for latency-sensitive and high-concurrency workloads
• –Explicit VRAM and GEMM controls, plus no-fallback kernel validation, will appeal to teams that care about predictable inference behavior
• –If rvLLM sustains these numbers, it competes with vLLM on maintainability and shipping simplicity, not just tokens/sec