model_patcher: support setter-backed WeightHooks safely

[liminfei-amd]

Summary

Use complete replacement semantics for setter-backed Hook weights and make
MaxSpeed caching an all-or-nothing HookGroup decision.

Problem

FP8 and INT8 modules expose setters that reject the Hook path's forced
inplace_update=True. Copy-only backup and restore also lose structured
quantization metadata.

The current per-weight cache decision can additionally publish a partial group
or accumulate the remaining entries on CPU.

Change

• Install the setter's complete return_weight=True result.
• Preserve copy/replacement mode across backup, cache, and restore.
• Preflight the full HookGroup cache requirement.
• Publish only complete caches; otherwise keep the group uncached.

Tests

The PR branch passes 15/15 focused tests covering FP8, INT8 ConvRot,
cache/restore, A -> B -> none, ordinary parameters, and low/high cache
budgets. All 15 reported GitHub checks are green. Its full CPU suite retains
only the same four environment failures as baseline.

A clean checkout of the combined test branch (#14413, #15532, and #15533)
passes 17/17 focused tests and the complete CPU suite
(1238 passed, 10 skipped).

On one physical gfx1100 GPU with PyTorch 2.9.1 / ROCm 7.2.1 and
comfy-kitchen==0.2.30:

• full public INT8 and FP8 Hook workflows each applied and restored all 264
  mapped weights and produced finite, nonblack output;
• two independent reduced MaxSpeed HookGroups completed 10 consecutive
  prompts in one service;
• the whole-group budget check selected the uncached fallback, every prompt
  ended with zero dynamic pins, and RSS/swap reached a stable plateau;
• logs contained no ERROR lora, unloaded-key, OOM, weight_scale, or
  inplace_update error.

Public Krea2 FP8 and INT8 Hook workflows also completed on the earlier
PyTorch 2.12 / ROCm 7.13 validation stack.

The #14382 reporter
independently confirmed that the combined branch at dad437146b works for
INT8, FP8, and BF16 on the original RX 7900 XTX / two-private-LoRA workflow.
This is integration evidence because that branch also contains #14413 and
#15533; the isolated setter/cache contracts are covered by the focused and
physical-GPU tests above.

Full-size independent HookGroups can still exceed the temporary
weight/GEMM-workspace capacity of a 16 GiB board. That hardware envelope is not
claimed by this change.

Follow-up to #14382. Depends on #14413. This draft branch temporarily includes
that prerequisite commit so the follow-up tests remain runnable before it
merges.

Addresses #15530.

---

This is an experimental, AI-generated code change, reviewed by AMD engineers before submission.

Made with Cursor