Add B200 workload variants for all H200-only models - #12

Open

khluu wants to merge 8 commits into

mainfrom

worktree-add-b200-workloads

Commits on May 28, 2026

Add B200 workload variants for all H200-only models

khluu
and
claude
committed

Commits on May 29, 2026

Use deep_gemm MoE backend for B200 workloads

khluu
and
claude
committed
Switch B200 workloads to NVFP4 models and match vLLM recipes

khluu
and
claude
committed
Fix JSON serve_args quoting for kimi/minimax B200 workloads

khluu
and
claude
committed
docs: note serve_args word-split is unquoted (no quoted/space-bearing JSON)

khluu
and
claude
committed
Fix deepseek_v3_2 B200: drop forced FLASHINFER_MLA backend

khluu
and
claude
committed
Cache B200 models on /raid instead of root disk

khluu
and
claude
committed

Commits on Jun 1, 2026

Fix kimi attention-config arg and native vllm-bench pandas dep

khluu
and
claude
committed