docs: document model, quantization and upscale changes from recent PRs

[Pfannkuchensack]

Summary

Several recent merged PRs changed behaviour that the docs either did not describe or described incorrectly. The FP8 Storage page was the most out of date: it still said Wan ignores the setting, that the toggle is "silently a no-op" for quantized formats, that the exit hook casts back to fp8, and that only CUDA qualifies. This PR is stacked on #377 and uses its new page layout.

User guide:

• FP8 Storage & Compute:
  • The table of models the setting applies to is rebuilt from fp8_storage_support(), and a new section explains where the toggle is shown.
  • Wan honours the setting, except Comfy fp8_scaled files (Offer FP8 Storage only where it does something #350).
  • The exit hook restores the stored weights instead of casting back (fix(fp8): follow-ups from the fp8 stack review #330).
  • Covers the CUDA/XPU device probe, and warns about MXFP8 auto-enable (Inpainting now implemented in development branch InvokeAI#380).
  • UI labels corrected.
• Models introduction:
  • New overview of the quantized formats (GGUF, NF4, SDNQ, int8_convrot, nvfp4, fp8, MXFP8).
  • How models are identified: standalone 16-channel VAEs (Identify standalone 16-channel VAEs as FLUX.1 or SD3 #364) and Klein vs. Klein Base (Share the int8 install sequence, load int8 Qwen3 encoders, fix the Klein Base heuristic #280).
  • Offline encoders (Bundle encoder tokenizers and configs instead of fetching them from HuggingFace #324).
  • Ideogram 4: GGUF encoder (feat(krea2): load the Qwen3-VL encoder from a llama.cpp GGUF #317) and second-branch residency (One object for the checkpoint prefix, and Ideogram 4 stops filling the card #315).
  • New ERNIE-Image section (feat(ernie-image): support single-file transformers #270).
• Krea-2:
  • A table of transformer builds (NVFP4, GGUF, int8, fp8, MXFP8).
  • Qwen3-VL GGUF encoders (feat(krea2): load the Qwen3-VL encoder from a llama.cpp GGUF #317).
  • INVOKE_KREA2_SDPA_BACKEND and its side effects (perf(krea2): rank cuDNN first for SDPA, with a fallback list and a probe #172, perf(krea2): keep the K/V heads unexpanded where a fused kernel takes them #174, fix: close 34 findings from the PR 171-194 review #241).
  • LoRA on quantized builds.
• New Upscaling page for the Upscale panel, including FLUX.1 (Drive the Upscale widget from an architecture table, and add FLUX.1 #363). It covers the panel sections, what each model family needs, and that the FLUX.1 detail pass is full-frame, with only its VAE tiled.
• Low-VRAM mode: new section on tiled VAE decode and encode. It covers force_tiled_decode and which model families read it, the node fields, and the out-of-memory retry (feat(vae): tiled decode for the FLUX.1 autoencoder #171, Load the FLUX.1 VAE as a diffusers AutoencoderKL, and tile its encode #355).
• SDNQ comparison table: adds int8, nvfp4 and fp8 checkpoints; LoRA on fp8 goes through the sidecar patcher.
• System requirements: new Ideogram 4 and ERNIE-Image rows; int8/NVFP4 figures for Z-Image, Krea-2 and Klein.

Contributor docs:

• New model integration guide:
  • Prefix stripping for detection and for loaders.
  • Reserving once: install_int8_convrot_layers makes its own reservation.
  • reject_quantized_side_channel.
  • Declaring FP8 Storage support with Unimplemented/NotApplicable.
  • Adding vendored encoder assets to package-data.
• Removed development/Guides/models.mdx: it duplicated the integration guide with webv1 paths. Its URL now redirects to the guide.

Config descriptions:

• force_tiled_decode no longer says "as of this release"; it lists the families that read it and says it also tiles encodes.
• fp8_compute covers ROCm GPUs that pass the matmul probe, and no longer credits itself alone with the VRAM saving.
• docs/src/generated/settings.json regenerated.

Related Issues / Discussions

• Stacked on docs: reorganize for v7 with prefix-ordered sidebar and new Users Guide pages #377; merge that first.
• MXFP8 auto-enable: FP8 Storage is switched on at install for MXFP8 checkpoints, where it costs quality InvokeAI#9620. The FP8 Storage page describes the current behaviour and its workaround until that is fixed.
• Documents behaviour from feat(vae): tiled decode for the FLUX.1 autoencoder #171, perf(krea2): rank cuDNN first for SDPA, with a fallback list and a probe #172, perf(krea2): keep the K/V heads unexpanded where a fused kernel takes them #174, feat(quantization): share the int8_convrot scheme and read it for Krea-2 and Z-Image #194, fix: close 34 findings from the PR 171-194 review #241, feat(quantization): keep Comfy nvfp4 checkpoints packed #266, feat(ernie-image): support single-file transformers #270, Share the int8 install sequence, load int8 Qwen3 encoders, fix the Klein Base heuristic #280, Decode MXFP8 block scales instead of refusing them #283, One object for the checkpoint prefix, and Ideogram 4 stops filling the card #315, feat(krea2): load the Qwen3-VL encoder from a llama.cpp GGUF #317, Bundle encoder tokenizers and configs instead of fetching them from HuggingFace #324, fix(fp8): follow-ups from the fp8 stack review #330, Offer FP8 Storage only where it does something #350, Load the FLUX.1 VAE as a diffusers AutoencoderKL, and tile its encode #355, Drive the Upscale widget from an architecture table, and add FLUX.1 #363 and Identify standalone 16-channel VAEs as FLUX.1 or SD3 #364.

QA Instructions

From docs/:

• pnpm build: passes, "All internal links are valid." Also passes with DEPLOY_TARGET=ghpages.
• pnpm run check-deploy-output: passes for both targets.
• pnpm run check-redirects: 88 redirect targets validated.
• uv run --no-sync python ../scripts/generate_docs_json.py: settings.json matches the committed file.

From the repo root:

• uv tool run ruff@0.11.2 check and format --check on config_default.py: clean.
• uv run --no-sync pytest tests/test_config.py: 48 passed.

Checked in a local preview: the Upscaling, FP8 Storage, Krea-2, Models introduction and System Requirements pages, including tables, the sidebar position of the new page, and anchors.

Every factual claim was checked against the code on this branch: loader declarations, node fields and titles, UI strings in en.json, the webv2 upscale graph, and starter model entries. Figures come from the measurements in the linked PRs.

Note for anyone running pnpm run generate-docs-data: it calls uv run without --no-sync. That re-syncs the virtualenv without extras and replaces an accelerator-specific torch build. This PR does not change the script.

Review

Findings resolved:

• The Upscale page first described the FLUX.1 detail pass as tiled and listed the wrong contents for the panel's Advanced section. It now matches the graph and the widget.
• force_tiled_decode was described as applying to every VAE. It now lists the families that read it.
• The FP8 table missed SD3, CogView4, SDXL Refiner, ERNIE-Image, LTX-2, all FLUX.2 variants and SD2 ControlNet. It is now rebuilt from fp8_storage_support().
• The fp8 rows in the SDNQ table said LoRAs patch directly. Layers holding fp8 weights always use sidecar patching.
• Krea-2 sizes mixed transformer-only and total figures. The table is now transformer-only, and encoder sizes are listed separately.
• Contributor guide: prefix stripping for detection versus loaders, and double reservation when using install_int8_convrot_layers.

Remaining limitations:

• Ideogram 4 with a GGUF 8B encoder was not measured on hardware, so the page makes no fit claim for it.
• The RAM minimums in the new system-requirements rows are estimates, like the rest of that table.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)