Panel Fit Lab

Use Fit lab to inspect an existing GGUF, plan a smaller model for your chosen instance, calibrate it, reproduce a published layout, verify measured quality, and save an instance profile. The original source model is preserved. An estimate is not a memory measurement or a quality verdict.

Open Fit and choose Open guided Fit lab: explain, plan, recipes and measured quality. Choose Guided tasks (/rust/fit/lab/guides) to run either workflow one checked step at a time. The recipe task accepts published recipe JSON, imports it through the existing recipe route and lets you undo that specific import using its generated recipe ID. Manual build and verification steps remain explicit, confirmed actions. The native pages are Explain and plan (/rust/fit/lab), Recipes and builds (/rust/fit/lab/recipes) and Quality and refusals (/rust/fit/lab/quality). Mutating controls require administrator privileges; other roles see a read-only explanation. The separate model-surgery pages are not part of these workflows.

Fit a model to this GPU

1. On Explain and plan, open 1 · Check readiness and Prerequisites and sources. Check that the original source and conversion tool are available. If they are missing, import the source and install a compatible converter first.
2. Under 2 · Explain any GGUF, enter Model GGUF path, optionally choose Full-precision source, and press Explain. If no original source matches, inspect the tensor-format report and find the original revision before trying to replicate it.
3. Under 3 · Fit to a device, review Device budget for the selected instance. Choose Original model, Selected instance, Base format (auto or mix), Keep free (MiB), Context (0 = current), Output format (auto or base), optional MTP format and Embedding format, and a Goal (with optional additional goals). Press Plan. The result includes saved plan IDs; candidate sizes and device fit remain estimates.
4. Under 4 · Review a plan, pick Saved plan and press Check plan. Inspect disk capacity and warnings. For a calibrated recipe, first complete 5 · Calibrate importance matrix: choose Source GGUF, Selected instance, Reference (Q8_0 or current instance) and optional Calibration paths (comma-separated), plus Max tokens and CPU threads. Confirm, then press Start importance matrix. Alternatively supply Matrix path and Source GGUF to Import calibrated matrix, which validates compatibility before import.
5. Return to 4 · Review a plan. Select Saved plan, enter Output file name, choose Importance matrix if necessary, and enter CPU threads. Check the confirmation box and press Build plan. Check Active Fit job for read-only, automatically refreshed progress; Cancel Fit job requests cancellation and does not guarantee a clean partial artifact.
6. Open Quality and refusals. Under 1 · Verify on this device, choose Candidate GGUF, Selected instance, Budget (quick or standard) and Curve depths; confirm and press Verify on device. Serving may be interrupted by measurement.
7. Under 2 · Measure KL divergence, supply Candidate GGUF, Baseline GGUF, Perplexity binary, optional Held-out corpus (blank = bundled sample), Context tokens, Chunks, Repeats, and Max perplexity increase. Confirm and press Measure KLD and perplexity. Check Evaluation history for completed measured values, stability and the quality policy verdict. A Fit quality index is not KLD: when KLD has not been measured, it is not available for comparison.
8. Under 3 · Save candidate profile, enter Instance, Candidate GGUF path and Profile name. Press Save model profile, then deliberately load and save the profile in Parameters. Saving a profile does not switch the running instance.

Replicate a published recipe

1. Open Recipes and builds, expand Saved recipes and formats and read importance matrix requirements. When you have recipe JSON, use Import recipe JSON; when you have a supported published GGUF link, use Published GGUF URL and Import layout. Both save a recipe, not a quantized model.
2. For a local published GGUF, first use Explain with its matching full-precision source. Review the model's tensor sections and the generated recipe. Different source revisions can produce different weights even with the same recipe.
3. Under 3 · Check and replicate, select Saved recipe, Full-precision source, optional Base mix (blank = recipe) and Calibration matrix, Output file name and CPU threads. Press Check recipe; review tensor agreement, warnings and disk availability. A mismatched or absent matrix can change the layout.
4. Confirm the write and press Build recipe. The job creates a new artifact; do not remove the original source. In 4 · Export or remove, use Recipe ID and Show recipe JSON to inspect the saved recipe. Delete saved recipe permanently removes a saved recipe, not a built-in one.
5. Verify the new artifact against the source under Quality and refusals using a held-out corpus and measured KL divergence. Publish measurement conditions alongside the recipe to make any comparisons meaningful.

Strict refusal check

Under 4 · Strict refusal check, select Instance, Refusal-prone prompts, Ordinary prompts, and optionally Include reasoning mode. Confirm and choose Check refusals. Read Strict refusal verdicts rather than treating keyword matches or empty replies as strict refusals. The native page polls active check progress without sending any write action; refresh to see completed history. Stop refusal check requests a stop between prompts.

Controls and tooltips

The following text matches the Help markers shown next to native controls and the attached classic-tab selectors. All operations reuse the existing native API; a form confirmation is not a permission bypass.

Additional native controls

Converter selection and measurements: On Explain and plan, select a conversion tool with Use converter; the choice applies to subsequent Fit jobs. Measure formats requires a specific selected instance, a source file, depths and at least one format. Confirm first: the measurement workflow interrupts the selected instance and restores it afterward. Measured format cards are shown above the measurement form. Saved plans can be opened with Show plan details; the job details form reads the same status owner used by the classic tab.

Other model types and source maintenance: Under Convert projector, image or audio weights, select Kind, Input file, Format, optional Rules and an Output file name, then confirm. Check upstream revisions performs a metadata check only. Fetch source downloads and converts full-precision model weights, and Delete Fit artifact permanently removes a selected generated file; each requires confirmation. Review available disk space and the reserve before any destructive or resource-intensive operation.

Other engine formats: Review Supported families and finished artifacts, then choose the full-precision source, device and recipe-specific options for ONNX, compressed-tensor or AutoRound outputs. Estimate factory resources reads planning results; Build factory artifact starts work only after confirmation. On Quality and refusals, Check output inspects a finished artifact and Measure fidelity compares it to its source using a selected number of chunks. A format check and a fidelity result answer different questions.

Quant Forge: The Quant Forge section on Explain and plan is backed by the existing native forms. Start with Disk estimate, then select Native RCO/GSQ only after reviewing the result and confirming. Use Refresh job snapshot, View saved spec, Verify receipt and GGUF, Resume or Request cancellation to manage a run. The advanced controls inspect the source tensor inventory, write an allocation manifest, validate the output and compare measurements through the quality gate. A storage envelope is an estimate, and a receipt is evidence to verify rather than a prediction of quality.

Paired verification: Under Verify original and candidate together, select both existing files and confirm. This runs the same verification workload for each. Review the returned measured evidence before saving a profile.

Common failure checks

• Source missing: check the original local GGUF before trying to plan or build.
• Model revision or tensor mismatch: use the exact matching source; imported layouts describe formats, not a promise of identical output bytes.
• Not enough disk space: free space or select a smaller candidate and run Check plan or Check recipe again.
• Calibration unavailable: rebuild without it only when the recipe permits; quality and tensor layout can change.
• No recorded KLD: run the separate held-out evaluation. Do not label a prediction or format estimate as a measurement.
• Permission denied: ask an administrator to complete build, profile and refusal actions.

Guided task definitions include Fit a model to this GPU and Replicate a published recipe. Open the Fit lab pages for the actual work and confirmations. No hardware result or quality value is supplied as a default.