[Bugfix] Fail when a diffusion LoRA adapter binds no layer

[Hiro208]

DiffusionLoRAManager.add_adapter() returns True and activation completes even
when the adapter's target_modules match no wrappable engine-side module. The
output is identical to the base model and the only trace is a DEBUG log, so a
mistyped target_modules, or a diffusers-named checkpoint against a
differently-named engine layout, goes undetected. #7190 has the full breakdown.

vLLM core raises on this mismatch in LoRAModel.from_local_checkpoint. The
in-memory path the diffusion stack uses (from_lora_tensors) has no equivalent
check, so nothing catches it.

The bookkeeping was already present but conditional: lora_names_by_id was only
populated when the pipeline supplies a _validate_diffusion_lora_binding hook,
so pipelines without that hook recorded nothing. This makes it unconditional and
raises from _bind_adapter_weights when bound_lora_names is empty, naming the
expected and received module lists. _activate_adapter already wraps the bind in
a try/except that resets every wrapper and re-raises, so the rollback path is
unchanged.

Verified on CPU:

• with the fix, tests/diffusion/lora/test_lora_manager.py is 25 passed
• reverting only manager.py and keeping the new test gives 1 failed, with the
  log showing Activating adapter: id=3 completing silently — the bug this closes

Closes #7190

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/diffusion_model_integration.md, docs/design/module/diffusion/index.md.

Module owners: @RuixiangMa @wtomin @david6666666

Routing: @RuixiangMa via module of the changed files, module named in the PR description, semantic router; @wtomin via module of the changed files, module named in the PR description, semantic router; @david6666666 via module of the changed files, module named in the PR description

@Hiro208, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[Hiro208]

Self-review:

• lora_names_by_id was only built when the pipeline supplies
  _validate_diffusion_lora_binding, so _record_bound was a no-op on every pipeline
  without that hook and bound_lora_names stayed empty no matter what actually bound.
  Making it unconditional is what lets the guard mean anything.
• The guard raises from _bind_adapter_weights, which _activate_adapter already wraps
  in a try/except that calls _reset_lora_layers() and re-raises, so a rejected adapter
  leaves every wrapper inactive. No new failure path.
• The message uses self._expected_lora_modules, the same set _load_adapter passes to
  vLLM core as expected_lora_modules, so the expected/received wording lines up with
  what core reports on the checkpoint path.
• Ran the whole file on CPU: 25 passed. Reverting only manager.py and keeping the new
  test gives 1 failed, with the log showing Activating adapter: id=3 completing
  silently — so the test pins the bug, not the fix.
• The packed and sublayer branches already call _record_bound, so partially matching
  adapters are unaffected; the existing packed / suspend-resume tests still pass.

[congw729]

LGTM! This looks good to merge once CI passes.

[Hiro208]

@congw729 Rebased onto main to retrigger the Read the Docs build — the previous
one was cancelled and left the required check stuck. RTD is green now; could you
re-apply ready so Buildkite runs on the new head? Thanks.

[NickCao]

LGTM