Add MPS support to the gemma node

[Zac (zboyles)]

Adds MPS (Apple Silicon) support to the Gemma text-encoder node.

Related to #302 (gemma text-encoder portion of MPS support)

Changes

• fork_rng device guard — torch.random.fork_rng(devices=[...]) only
  accepts CUDA devices; on MPS it raises. The device list is now passed only
  when the model is on CUDA. Seeded generation on CUDA is unchanged.
• Respect caller-provided dtype — _LTXVGemmaTextEncoderModel.__init__
  previously hardcoded torch.bfloat16, overwriting the dtype argument;
  it's now only the default. With this, clip_dtype flows through
  ltxv_gemma_clip into the model and the load_text_embeddings_pipeline
  loaders.
• float16 on MPS — MPS has incomplete bfloat16 support, so clip_dtype
  falls back to float16 when the torch device is MPS (CUDA keeps
  bfloat16). This also keeps the new torch.autocast(dtype=model.dtype)
  context in _enhance on a dtype MPS actually supports.

Rebased onto current master (single commit, one file, +11/−2). An earlier
revision also propagated dtype into the embeddings connectors; that's no
longer needed — the load_text_embeddings_pipeline refactor on master
already threads it through.

Testing status

• Logic verified against PyTorch's fork_rng device constraints; the CUDA
  path is preserved unchanged.
• I don't have the models locally for an end-to-end run — verification on
  Apple Silicon welcome.

[José Lugo (josephlugo)]

Can anyone with permissions review this please? M3 Ultra Mac Studio support is needed!

[Anton Bukov (k06a)]

Implemented similar fix here: Comfy-Org/ComfyUI#12809

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR improves device/dtype handling for the Gemma encoder path, aiming to make generation RNG forking safer across backends and to avoid unsupported dtypes on MPS.

Changes:

• Adjusted torch.random.fork_rng(...) usage to avoid passing a CUDA device on non-CUDA backends.
• Made dtype actually configurable (instead of always forcing bfloat16) and propagated it into embeddings connector loaders.
• Added MPS-specific dtype selection (fallback to float16) when loading the model.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Zac (zboyles)]

Implemented similar fix here: Comfy-Org/ComfyUI#12809

Thanks, #12809 is complementary rather than overlapping, it fixes device placement (text encoders now land on MPS instead of CPU). This PR fixes the node itself once it's there (fork_rng raises on non-CUDA devices, and MPS needs float16 over bfloat16). With #12809 merged, Mac users now hit these paths by default, so I've rebased this onto current master and trimmed it to the minimal functional diff.

[Zac (zboyles)]

Can anyone with permissions review this please? M3 Ultra Mac Studio support is needed!

José Lugo (@josephlugo) I've rebased the PR, would you be able to run it on your M3 Ultra and report back?