Gemma 4 26B hangs vLLM startup

This looks like the usual gap between a polished model launch and production-ready distributed serving. Gemma 4 is efficient on paper, but the rough edges are in the orchestration layer, not the checkpoint size.

• –The reported `RayTaskError(ValueError)` points at Ray-backed startup and worker coordination, not generation quality
• –vLLM's Gemma 4 docs support the model, but the deployment path still has sharp edges around memory profiling and multimodal initialization
• –The docs explicitly recommend trimming multimodal work with `--limit-mm-per-prompt image=0 audio=0` for text-only workloads, which suggests startup memory accounting is still expensive
• –Quantization may reduce pressure, but it will not fix a broken distributed boot path if the failure is in profiling or engine selection
• –For adopters, this is a reminder that "supported" and "boringly deployable" are not the same thing yet