feat(weight-free): weight-free ONNX export for causalLM

[quic-amitraj]

Summary

Adds weight-free ONNX export for causal LM models on top of the dynamo export infrastructure.

Weight-free export builds the model on meta device (no weights in RAM), exports ONNX graph structure only, and lets the QAIC compiler load weights directly from the original safetensors checkpoint at compile time via weight_spec.json.

New files

• QEfficient/exporter/weight_free/ — core weight-free export, weight spec format, checkpoint prep pipeline (MoE expert stacking, dtype
  conversion, .bin→safetensors auto-convert)

• examples/text_generation/weight_free/ — single-model smoke test, CB weight-free smoke test

  Key changes

• modeling_qeff.py — use_weight_free_export flag; reorder example_inputs/dynamic_shapes by forward() signature so torch.export
  binds them correctly; skip SplitTensorsTransform; embed weight_spec.json as ONNX metadata; symlink prepared checkpoint next to ONNX

• modeling_auto.py — use_weight_free_export wired through export() and compile(); auto-enable dynamo=True

• onnx_transforms.py — disable unsupported transforms for dynamo path

• export_utils.py — separate dynamo/TorchScript subfunction setup paths

• torch_patches.py — temporarily_enable/disable_nested_compile_regions()

• core.py — dtype suffix in prepared checkpoint dir; prune fake initializers after meta-device export

  Model fixes

• codegen — fix embed_positions device/dtype for weight-free subfunction tracing

• glm4_moe — logits.float() to fix runtime buffer mismatch; rotary cos/sin device sync; _rotary_dim attribute; MoE expert placeholder
  dtype

• granitemoe, mixtral, phi3, qwen3_moe — rotary cos/sin device/dtype sync for meta-device tracing; MoE expert placeholder dtype

  • continuous_batching.py — fix use_dynamo=True → dynamo=True

  Tests

  pip install -e .[test]
  pytest tests/weight_free/ -m "not on_qaic" -n auto -v
  

---