Fix _is_cached class/instance mismatch for multi-transformer pipelines

[jiangyon-amd]

Summary

CachedAdapter.create_context() sets _is_cached on block_adapter.pipe.__class__,
but sets _context_manager on the instance:

block_adapter.pipe._context_manager = context_manager       # instance level
...
block_adapter.pipe.__class__._is_cached = True               # class level

For pipelines with more than one transformer that don't have a real pipe object
(a BlockAdapter(transformer=..., ...) call per transformer, as one would do for a
dual/multi-DiT model), each transformer gets wrapped in its own fresh
FakeDiffusionPipeline() instance -- but all of those instances share the same
class. Once the first transformer's setup flips the class-level _is_cached flag,
BlockAdapter.is_cached() -- ultimately a plain getattr(pipe, "_is_cached", False),
which can't distinguish an instance attribute from an inherited class attribute --
reports every sibling instance as already cached too, even though it never got its
own _context_manager. create_context() then returns early, and the caller's
assert hasattr(block_adapter.pipe, "_context_manager") fails for the second
transformer.

Fix: only take the early-return/skip path when this pipe instance actually already
carries its own _context_manager.

Repro / verification

Added tests/api/test_multi_transformer_is_cached.py, which calls enable_cache on
two independent transformer-only BlockAdapters (mirroring what a dual-DiT model
does) and asserts the second one gets its own context manager.

• Against current main: fails with

  AssertionError
  ...cache_adapter.py:439: in collect_unified_blocks
      assert hasattr(block_adapter.pipe, "_context_manager")
  

  preceded by the log line Pipeline has been already cached, skip creating cache context again. for the second transformer -- confirming the class-level flag leaked from the first.
• With this fix applied: passes, and the second transformer has its own distinct _context_manager.

Notes

Found while integrating a dual-DiT pipeline (video + audio branches, each wrapped
via the transformer-only API) with cache-dit.

[Copilot]

Pull request overview

Fixes a caching-state leak that occurs when multiple transformer-only BlockAdapters each create their own FakeDiffusionPipeline() instance but share the same pipeline class, causing class-level _is_cached to incorrectly short-circuit cache setup for sibling instances.

Changes:

• Tighten CachedAdapter.create_context()’s early-return condition so it only skips when the current pipe instance already has a context manager.
• Add a regression test covering two independent transformer-only adapters to ensure both get distinct context managers.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
tests/api/test_multi_transformer_is_cached.py	Adds regression coverage for multi-transformer transformer-only caching with shared FakeDiffusionPipeline class.
src/cache_dit/caching/cache_adapters/cache_adapter.py	Prevents premature early-return in create_context() when _is_cached is inherited from the pipe class but the instance lacks _context_manager.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[DefTruth]

The pre-commit ci failed, please init and run pre-commit