feat(0506): Layerwise export: API-driven, env-var-free, opt-in flag

[vbaddi]

Summary

Encapsulate the layerwise export+stitch+compile orchestration loop (previously a 200+ line example with monkey-patches and an LAYERWISE_EXPORT env var) behind a single layerwise=True flag on .compile() / .export().

What's new

- New flag: layerwise: bool = False (+ layerwise_window_size: int = 1) on:                                                                
- QEFFAutoModelForImageTextToText.compile() / .export()                                                                                 
- QEFFAutoModelForCausalLM.compile() / .export()                                                                                        
- No env vars. LAYERWISE_EXPORT is gone. Control is a process-local class flag toggled by an internal context manager.                    

Backward compatibility

• layerwise=False is the default.
• Verified: pytest tests/unit_test/models/test_model_quickcheck.py -n auto → 130 passed, 3 skipped (was 121 / 3 before this PR).

Tests added

1. 9 new tests covering windowing helpers, the supported/unsupported guard (parametrized over all 3 supported architectures),
2. env-var-not-leaked invariant, and the context manager's flag toggle.

---

Usage: Enable layerwise

 import torch                                                                                                                              
 from transformers import AutoConfig                                                                                                       
 from QEfficient import QEFFAutoModelForImageTextToText               

 config = AutoConfig.from_pretrained("Qwen/Qwen3-VL-235B-A22B-Instruct")

 model = QEFFAutoModelForImageTextToText.from_pretrained(             
     "Qwen/Qwen3-VL-235B-A22B-Instruct",
     attn_implementation="eager", kv_offload=True,
     config=config, torch_dtype=torch.float16,
 )                                                                    

 qpc = model.compile(                                                 
     batch_size=1, prefill_seq_len=1, ctx_len=4096,
     num_cores=16, num_devices=4,
     mxfp6_matmul=True, aic_enable_depth_first=True,
     skip_vision=True, split_retained_state_io=True,
     use_onnx_subfunctions=True, mos=1,
     layerwise=True,                         # opt-in
     layerwise_window_size=1,        # layers per window              
 )

Disable (default)

Just don't pass layerwise. Behavior is identical to before this PR.

  qpc = model.compile(batch_size=1, prefill_seq_len=1, ctx_len=4096, ...)

  CausalLM (qwen3_moe)                                                 

  from QEfficient import QEFFAutoModelForCausalLM

  model = QEFFAutoModelForCausalLM.from_pretrained("Qwen/Qwen3-235B-A22B-Instruct-2507")
  qpc = model.compile(                                                 
      prefill_seq_len=4, ctx_len=128, num_cores=16, num_devices=1,
      mxfp6_matmul=True, mxint8_kv_cache=True,
      layerwise=True, layerwise_window_size=1,
  )   

---

Test Plan

• pyenv activate qeff && python -m pytest -q tests/unit_test/models/test_model_quickcheck.py -n auto → 130 passed, 3 skipped

[abhishek-singh591]

Backward compatibility is not yet guaranteed for Qwen 3.5 and Qwen3-VL-MoE. PR #1043 addresses this issue for Qwen3-VL, but Qwen 3.5 will still need additional modeling changes. We skipped the Qwen 3.5 unit tests, which is why all tests are currently passing.