[perf] Disable every offload path on unified memory, unblocking MiniMax H3 generation on one GB10

[KyleNeverGivesUp]

Purpose

CPU offload reduces accelerator pressure only when host and accelerator memory
are separate physical pools. On GB10, Jetson-class integrated CUDA devices, and
Apple silicon, moving a component to the host does not free the underlying pool
and can increase peak memory through transient copies.

This PR applies one worker-local unified-memory policy to every component
offload mode while preserving explicit FSDP inference sharding.

Changes

• Classify the selected device after each worker binds its local rank; the
  driver does not initialize CUDA or classify every worker from device zero.
• Clear the five host-placement modes on unified memory:
  dit_cpu_offload, dit_layerwise_offload,
  text_encoder_cpu_offload, image_encoder_cpu_offload, and
  vae_cpu_offload.
• Defer inference conflict normalization until device classification. Unified
  CUDA therefore clears layerwise offload before resolving the conflict and
  preserves an explicit use_fsdp_inference=True; discrete CUDA retains the
  historical layerwise-offload precedence.
• Cache classification per arguments object and device ID while still clearing
  any offload flag that a later caller re-enables.
• Apply the same policy to direct pipeline construction and role-aware encoder
  loading, including explicit CPU targets.
• Keep pin_cpu_memory outside the policy because it also controls output
  transfers rather than component placement.
• Collect the CPU-safe platform and pipeline tests in the unit lane and document
  the automatic policy in the offloading guide.

Dependency

This PR is rebased onto current main after merged #1710 and #1714. Its final
branch contains only the unified-memory extension and maintainer fix after the
reviewed loader/worker foundation. Final head:
6fb6af9fabb633b235be67e81288da1558fd81f1.

Verification

focused policy, loader, pipeline, and worker tests: 45 passed
post-#1714 loader/platform integration checks: 13 passed
pinned MiniMax-H3 packing/media/scheduler and pipeline checks: 37 passed
related LingBot/Z-Image coverage: 18 passed, 4 explicit opt-in skips
exact unit lane: 988 passed, 7 skipped
pre-commit and diff checks: passed

A real NVIDIA GB10 policy probe started from production defaults with explicit
FSDP inference enabled. It classified the device as unified memory, cleared all
five host-offload flags, and preserved FSDP. The combined #1710/#1714/#1715
stack was also exercised by the later full FastH3 lifecycle run used to validate
#1761.

Checklist

• Tests cover each flag, worker-local classification, caching, and FSDP
• Applicable pre-commit hooks pass
• General offloading documentation is updated
• Memory behavior was validated on GB10

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5e06f62897

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[mergify]

⚠️ PR title format required

Your PR title must start with a type tag in brackets. Examples:

• [feat] Add new model support
• [bugfix] Fix VAE tiling corruption
• [refactor] Restructure training pipeline
• [perf] Optimize attention kernel
• [ci] Update test infrastructure
• [infra] Add activation trace hooks
• [docs] Add inference guide
• [misc] Clean up configs
• [new-model] Port Flux2 to FastVideo
• [skill] Add add-model agent skill

Valid tags: feat, feature, bugfix, fix, refactor, perf, ci, infra, doc, docs, misc, chore, kernel, new-model, skill, skills

Please update your PR title and the merge protection check will pass automatically.

[SolitaryThinker]

/merge

[SolitaryThinker]

/test full

[SolitaryThinker]

/merge

[SolitaryThinker]

/test full

[SolitaryThinker]

/test fastcheck