Fix shard finalization issue to prevent skipping non-LLM tail layers

[lvliang-intel]

Description

Root Cause

ShardWriter.finalize() uses get_lm_head_name() to identify the tied embedding head and skip it. For diffusion model like FLUX, this returns "proj_out" (the last leaf module). Combined with tie_word_embeddings defaulting to True when the attribute is absent from the model config, proj_out.weight and proj_out.bias were silently skipped and never written to disk.

This bug was latent in the old commit but never triggered because FLUX RTN quantization disabled low_cpu_mem_usage on that code path, so finalize() was never called. PR #1386 changed that branch to keep low_cpu_mem_usage=True, which activated the is_immediate_saving path and exposed the bug.

Fix

Change the default value of tie_word_embeddings from True to False, consistent with the same logic in utils.py:364.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please specify):

Related Issues

Fixes or relates to #

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.

[Copilot]

Pull request overview

Fixes a sharded-weight finalization bug where ShardWriter.finalize() could mistakenly treat the last leaf module (e.g., diffusion proj_out) as a tied LM head and silently skip writing its tensors when model.config.tie_word_embeddings is missing.

Changes:

• Change the default tie_word_embeddings assumption in ShardWriter.finalize() from True to False.
• Read model.config.tie_word_embeddings only when the attribute is explicitly present.

---

You can also share your feedback on Copilot code review. Take the survey.

[yiliu30]

LGTM