Menu
Checklist for adding support for a new GGUF model architecture (or a variant of an existing one) to SharpInference.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
| name | new-arch |
| description | Checklist for adding support for a new GGUF model architecture (or a variant of an existing one) to SharpInference. |
First check it isn't already covered: many "new" models reduce to an existing arch (e.g. Ornith-1.0 is qwen35/qwen35moe — see CLAUDE.md). Inspect the GGUF before writing any code:
dotnet run --project src/SharpInference.Cli -c Release -- list-metadata -m model.gguf
dotnet run --project src/SharpInference.Cli -c Release -- list-tensors -m model.gguf
docs/<arch>-plan.md before coding; docs/qwen35moe-plan.md
is the template (tensor layout, hparams, forward-pass math, phased milestones).
This is an architect-agent task.src/SharpInference.Core/ModelGraph.cs: arch name dispatch, hparam
mapping, tensor-name wiring. Tokenizer + chat template usually come free via
Microsoft.ML.Tokenizers + JinjaChatTemplate; add a ToolCallAdapter entry
only if the model family has its own tool-call wire format.IForwardPass implementation whenever the math
allows: dense → ForwardPass/CudaForwardPass/GpuForwardPass; MoE →
HybridForwardPass/CudaHybridForwardPass; hybrid Gated-DeltaNet →
HybridGdnForwardPass/CudaHybridGdnForwardPass/VulkanHybridGdnForwardPass.
A genuinely new block type means new kernels in all backends you support —
scope CPU-first, GPU after parity.parity-check skill and
scripts/generate-reference-logits.ps1).scripts/download-model.ps1 preset (extend the ValidateSet),
CLAUDE.md architecture list, README.parity-check skill, levels 2 and 4). New-arch work is not done until greedy
tokens match llama.cpp at temp 0 on at least one real model.Plan doc and design → architect agent. Steps 2–5 → implementer agent(s), one
per step, sequential (each depends on the previous). Validation → test-runner
plus the parity-check skill in a background agent.
Drive the architect→workhorse flow — turn an approved plan into delegated Sonnet implementation with test and review gates, keeping the main (Opus/Fable) thread as coordinator.
Work a GitHub issue end-to-end — fetch it, verify the claim, turn it into an architect spec, and drive the delegated implementation flow with correct issue references in branches and commits.
Cross-check SharpInference numerics against llama.cpp and run accuracy gates. Use when output quality or numerics are in question — new architecture bring-up, kernel or quantization changes, KV-compression work, "model produces garbage" reports.
CLI run recipes for specific models and modes — VibeThinker, Ornith-1.0, Gemma 4 vision, image generation + upscaling, perplexity gating, whole-turn JSON-schema structured output, DSpark speculative decoding, download-model presets. Use when actually running inference for one of these.
Required workflow when adding, editing, or removing Vulkan GLSL shader consts — the committed precompiled SPIR-V table must be regenerated or tests fail on drift.