Card synced byte-identical from canada-quant/vllm-glm53-flash-sm121 main @ 1c142c95 (PR #6): 8 full-attention layers cost ~3x the KV per token, so 1M is validated only with the incoai drafter (use 262K / 8 GiB or 800K / 16 GiB); built-in MTP head recommended for long context / non-English (community data labelled); k=5 boots. Launcher: KV_CACHE_MEM default 3 GiB -> 8 GiB (3 GiB was sized for the incoai drafter and is too small for this drafter class); comment fixes.
29 days ago
card: recipe/raws link -> public vllm-glm53-flash-sm121/drafter (was a private repo)
