vLLM issue #58241: cutlass_scaled_mm_supports_block_fp8() returns False on SM89 Ada despite native FP8 tensor cores, silently routing block-FP8 checkpoints to Marlin

Reporter ran a capability bisect over compute capabilities 80/86/89/90/100 on an NVIDIA RTX 6000 Ada (SM 8.9), vLLM 0.29.0, CUDA 13.4, model Qwen/Qwen3.8-27B-FP8 (128x128 block scales). cutlass_fp8_supported() (non-block) returns True on the same card. Single reporter for the capability finding; the repo-wide pattern of consumer-card exclusion is corroborated by independent PRs #56085 (RTX 4090 block-FP8 configs) and #54372 (RTX 4090 fused-MoE configs) landing in the same period.