Not in the repository. Every number in the top-level README was measured with these GGUF files in this directory:
| file | used for |
|---|---|
Qwen3.8-27B-UD-Q4_K_M.gguf |
the dense decode number, the greedy byte-compare text, serving (unsloth's UD quant) |
MTP/mtp-Qwen3.8-27B-Q4_0.gguf |
its MTP draft head for speculative decode (--spec-type draft-mtp -md <this> -ngld 99; unsloth ships it beside the model) |
Qwen3.5-35B-A3B-Q4_K_M.gguf |
the MoE decode number |
Meta-Llama-3.1-8B-Instruct-Q8_0.gguf |
the dense Q8 reference row |
sd/ |
stable-diffusion.cpp models: SDXL-Turbo, SD 1.5 (+ LCM LoRA), Z-Image-Turbo (Q8 DiT + Qwen3-4B encoder + FLUX VAE) |
tools/serve.sh and tools/soak.sh look for the 27B and its MTP head here by these names; tools/nv_shim_step.sh bench|simple|spec|sd
take any path.