Benchmark Future Coverage

Benchmark coverage still needed

MoE cache benchmarks indexes the current public results. The table below lists remaining test gaps without placeholder performance claims.

Gap	Why it matters	Minimum useful test
Qwen fused gate_up quant layouts	Exercises a different physical bank layout	NP1 and NP4, stock/cache0/positive cache, close VRAM
Qwen NVFP4 with large block-scale banks	Tests slot-bound auxiliary banks	32K or 64K prompt plus 512 decode
BF16 MoE weights	Separates quantized-kernel behavior from cache-control behavior	NP1 prefill/decode and NP4 aggregate decode
LFM Q5 positive-cache NP4	Current build fails closed	Repeat only after grouped-layout support changes
Nemotron parallel grouped execution	NP1 ungated grouped execution now passes	NP4 stock/cache0/cache72 with close VRAM
Multi-GPU	Validates per-device cache ownership and transfer behavior	Same model and prompt on one GPU and two GPUs
Physical lower-VRAM GPUs	The 4/8/12 GB tiers are measured process ceilings on a 16 GB GPU	Repeat the published geometry on physical 4/8/12 GB cards and record free VRAM
MTP and other drafting modes	Target-only results do not measure draft acceptance or extra state	Target-only vs drafting with acceptance and peak VRAM

New rows should always publish the exact server launch command, model link and hash, prompt/output geometry, sampling, raw TTFT/prefill/decode/wall numbers, and peak VRAM. A speed ratio should be highlighted only when the VRAM envelopes are close.