nvidia-specdecode-al-doctor
Fail-closed speculative-decoding AL/D doctor for ThumbGate. Checks speedup ≤ AL/(1+ρD), attention D=128/G-1, tile alignment, and maps to checkpoint-speculative-decoding-acceptance. Steal from NVIDIA co-design blog — never clone TensorRT/EAGLE/Model-Optimizer. Slash: /nvidia-specdecode-al-doctor.
Install
npx skillscat add igorganapolsky/thumbgate/nvidia-specdecode-al-doctor Install via the SkillsCat registry.
The skill validates speculative decoding parameters and claimed speedup against the derived bound, attention size, tile alignment, and checkpoint acceptance criteria, rejecting configurations that violate these constraints. It is appropriate when verifying that draft‑length, draft‑depth ratio, and speedup claim satisfy the required theoretical limits before running ThumbGate.
NVIDIA Speculative-Decoding AL/D Doctor
When
Speculative decoding, MTP, EAGLE, draft length, accept length, claimed speedup, SPEED-Bench AL talk.
Do
npx thumbgate nvidia-specdecode-al-doctor \
--speculative-decoding \
--accept-length=AL \
--draft-length=D \
--draft-depth-ratio=ρ \
--claimed-speedup=X \
--jsonPair with:
npx thumbgate deepseek-v4-runtime-guardrails \
--speculative-decoding --accept-length=AL --draft-length=D \
--draft-depth-ratio=ρ --claimed-speedup=X --cache-coherence-eval --jsonNever
- Claim speedup without measured AL
- Ship TensorRT-LLM / Model-Optimizer / EAGLE training clones
- Treat missing AL as a pass