| Alias | The served model name, which /v1/models lists and clients send as model. A recipe’s run.alias; the recipes use the repo id. |
| Card | A GPU model a recipe is written for, with an id such as p100. A recipe’s [card] table says which GPUs match: compute capability, memory and driver. See Recipe file. |
| Champion | A change that won a hill-climb campaign: it was faster than the campaign’s best so far, under that campaign’s guards. A champion only has to beat the previous one, so it may not contain every earlier winning change. |
| Compute capability | NVIDIA’s version number for a GPU architecture. The Tesla P100 is 6.0 (sm_60). nvidia-smi reports it as compute_cap. |
| Context | The number of tokens the server keeps for one conversation: prompt and output together. The recipes allocate 65,536 tokens. --context changes it and makes the run untuned. |
| Decode | Generating output tokens one at a time, after the prompt has been read. Decode speed is in tokens per second (tok/s). On a P100 it is limited mostly by how fast the weights can be read from memory. |
| Digest | The sha256 of a container image’s manifest, such as sha256:35409ff9…. A digest names exactly one image; a tag can be moved. Recipes pin images by digest. |
| Draft | A recipe with status = "draft": it has not passed the release gate. pull, run and serve refuse it unless --allow-draft. |
| Engine | The inference engine: a fork of llama.cpp tuned for the P100, in engine/. It runs as llama-server inside the engine image. |
| Engine release | A tagged engine version, engine-vX.Y.Z, such as engine-v0.1.0: the exact source in engine/ at that tag, and its build configuration. Versioned separately from the launcher. |
| Engine switch | A GGML_CUDA_* environment variable that turns on one of the engine’s P100 kernels or changes its setting. A recipe sets them in [run.env], exactly as measured. See Engine switches. |
| Hill climb | An automated campaign in which AI agents propose changes to the engine, and a judge they cannot change measures each one and keeps only the ones that are faster and pass the guards. The engine-v0.1.0 decode climb made 131 judged attempts and kept 28 improvements. |
| Image | One build of an engine release for one GPU architecture, published as ghcr.io/iammrduncan/tensorless-engine:<X.Y.Z>-<arch>. See Engine image. |
| Judge | The measuring program of a hill climb. It builds the candidate, runs it on the card, times it from outside the engine and checks its numerics. Agents cannot see or change it. |
| Launch target | A model you can launch: one released recipe. See Launch targets. |
| Launcher | The tensorless command. It holds no inference code; it reads recipes, fetches and checks the image and the model, and runs the container. |
| MTP | Multi-token prediction: a form of speculative decoding in which a small draft head proposes tokens that the model then checks. Not in the v0.1.0 recipes. |
| P0k+19 | The tensor mix of Qwen3.8-27B-PN3: which tensors use pn3t, which use pn4t, and which keep standard types. See PN weight format. |
| Pascal | NVIDIA’s GPU architecture of 2016, which includes the Tesla P100. CUDA 13 dropped it, and the R580 driver branch is its last. |
| PN | The P100-native block weight format of this engine, PNX-FORMAT-v0. Every block is 34 bytes and decodes exactly in fp32. See PN weight format. |
| pn3t | The 3-bit PN type, ggml type 124: 80 weights in 34 bytes, 3.40 bits per weight. |
| pn4t | The 4-bit PN type, ggml type 125: 64 weights in 34 bytes, 4.25 bits per weight. |
| Prefill | Reading the prompt before the first output token. Prefill speed is in prompt tokens per second. |
| Q3Floor | Short for Qwen3.8-27B-UD-IQ3_XXS-Q3Floor: Unsloth’s UD-IQ3_XXS quantization of Qwen3.8-27B with its 91 sub-3-bit tensors raised to IQ3_XXS, a 3-bit floor, and its token embedding replaced by the Q3_K one from Unsloth’s UD-IQ4_XS. |
| Receipt | What the release gate measured for a recipe, with the scope of each number and the image digest and model hash it was measured on. See Receipts. |
| Recipe | One model on one card, recipes/<org>/<repo>/<card>.toml: the image digest, the model’s repository, revision, file and sha256, the engine switches and server flags, the hardware requirements and the measured numbers. See Recipe file. |
| Release candidate | An engine image built for measuring, tagged X.Y.Z-rcN-<arch>. When it passes the release gate, its digest gets the release tag without a rebuild. |
| Release gate | The checks a candidate image must pass for every recipe on every card before it is released: build checks, numerics against the previous release, speed, long-context decode and prefill, and quality checks. |
| Released | A recipe with status = "released", whose image was measured on its card. pull, run and serve accept it with no extra flag. |
| Repo id | A Hugging Face repository name, org/repo, such as iamMrDuncan/Qwen3.8-27B-PN3-GGUF. tensorless names models by repo id, and matches it without case. |
| Short context | The decode measurement at a short prompt: 256 greedy tokens after a prompt of about 600 tokens in the research receipts. Its receipt key is short. |
| Short name | A recipe’s optional model.name, such as pn3 or q3floor, which works anywhere a repo id does. |
| sm_60 | The CUDA target for compute capability 6.0, the Tesla P100. Written sm60 in image tags and recipes. |
| TTFT | Time to first token: from sending a request to receiving the first generated token. It includes prefill. |
| UD | Unsloth Dynamic, Unsloth’s name for its mixed-precision quantizations, such as UD-IQ3_XXS. |
| Unpublished | A recipe whose revision, sha256 or image digest is all zeros, or whose size_bytes is 0. It cannot be pulled or run, even with --allow-draft. |
| Untuned | A run that differs from what the recipe measured: another --context, extra arguments after --, or another @REVISION. The launcher labels it, and the receipt’s numbers do not apply. |