feat(weight-free): weight-free ONNX export for causal LMs (dynamo path)

[amarquic]

Summary

This PR adds weight-free ONNX export support for causal LMs on top of Sharvari's dynamo export branch.

Weight-free export exports the ONNX graph with no embedded weights. Weights are stored in the original safetensors checkpoint and loaded by the QAIC compiler at compile time via weight_spec.json. This enables:

• Near-zero RAM during export (model on meta device)
• Reuse of the same ONNX across dtype/quantization variants
• Faster re-compilation without re-exporting

Key changes

New infrastructure:

• exporter/weight_free.py — core weight-free dynamo export: meta model trace, promotes initializers to external data, builds weight_spec.json
• exporter/weight_spec.py — WeightSpec format: maps ONNX initializer names → safetensors checkpoint keys
• exporter/checkpoint_transforms.py — checkpoint preparation pipeline: MoE expert stacking, dtype conversion, MXFP4 dequant, auto .bin→safetensors conversion

Base class wiring:

• base/modeling_qeff.py — _export(): use_weight_free_export flag, meta device guard, tmp ONNX path, weight_spec embedding in ONNX metadata
• transformers/models/modeling_auto.py — QEFFAutoModelForCausalLM: _checkpoint_transforms, CCL Dim fix, use_weight_free_export through compile chain

Model-level fixes for dynamo tracing:

• Llama, Qwen3, Phi3, Qwen3-MoE, GLM4-MoE, Grok-1, GPT-OSS, GPT2, Falcon — device alignment, meta device guards, rotary embedding fixes
• cache_utils.py — device=position_ids.device on all torch.arange calls

Infrastructure fixes:

• base/onnx_transforms.py — RewriteUnsupportedOpsTransform (SplitToSequence→Split, CastLike→Cast)
• utils/export_utils.py — _RetainedState rename guard for dynamo path
• utils/torch_patches.py — temporarily_enable/disable_nested_compile_regions
• customop/ — backward stubs, CtxGatherCB shape fix, GemmaRMSNorm runtime offset

Examples:

• examples/text_generation/weight_free_export_verify.py — single-model smoke-test
• examples/text_generation/continuous_batching_weightfree.py — weight-free + CB
• examples/performance/compute_context_length/weight_free_ccl_verify.py — weight-free + CCL

Validated models

Qwen3-30B-A3B, Qwen3-MoE, GPT-OSS, GPT2, Gemma-2-2B, Gemma-7B, Llama-3.2-1B, Phi3, GLM4-MoE

Usage

from accelerate import init_empty_weights
with init_empty_weights():
    meta_model = AutoModelForCausalLM.from_config(config)
model = QEFFAutoModelForCausalLM(meta_model, pretrained_model_name_or_path=model_name)
model.compile(use_dynamo=True, use_weight_free_export=True, ctx_len=4096, ...)

Test commands run

python3 examples/text_generation/weight_free_export_verify.py --model_name Qwen/Qwen3-30B-A3B-Instruct-2507
python3 examples/text_generation/weight_free_export_verify.py --model_name google/gemma-2-2b
python3 examples/text_generation/weight_free_export_verify.py --model_name meta-llama/Llama-3.2-1B
python3 examples/text_generation/continuous_batching_weightfree.py --model_name Qwen/Qwen3-30B-A3B-Instruct-2507