refactor(export): split unified_export_hf into layered modules

.agents/skills/ptq/references/unsupported-models.md

@@ -224,7 +224,7 @@ quant_cfg["quant_cfg"]["*vision*"] = {"enable": False}
 quant_cfg["quant_cfg"]["*multi_modal_projector*"] = {"enable": False}
 ```
 
-**Known VLM export issue**: The export step (`requantize_resmooth_fused_llm_layers` in `unified_export_hf.py`) may try to run a dummy forward pass on the full VLM instead of the language model backbone. This currently only handles Nemotron VLMs. If hit, patch the export to use `is_multimodal_model()` for the VLM check instead of model-specific string matching.
+**Known VLM export issue**: The export step (`requantize_resmooth_fused_llm_layers` in `hf_export_prep.py`) may try to run a dummy forward pass on the full VLM instead of the language model backbone. This currently only handles Nemotron VLMs. If hit, patch the export to use `is_multimodal_model()` for the VLM check instead of model-specific string matching.
 
 ## Pattern 5: FP8 Checkpoint Handling
 

modelopt/torch/export/hf_export_handlers.py

@@ -22,6 +22,7 @@
 
 from modelopt.torch.quantization.utils import fsdp2_aware_weight_update
 
+from .hf_weight_export import _export_quantized_weight
 from .layer_utils import get_expert_linear_names, is_quantlinear, set_expert_quantizer_amax
 from .model_config import QUANTIZATION_NONE
 from .moe_utils import _export_fused_experts
@@ -41,10 +42,6 @@ def _export_weight(
     ctx: ExportContext,
     weight_name: str = "weight",
 ) -> None:
-    # Imported lazily to avoid a cycle: unified_export_hf imports this module to
-    # install the built-in handlers while retaining this legacy helper's import path.
-    from .unified_export_hf import _export_quantized_weight
-
     _export_quantized_weight(module, ctx.dtype, weight_name, _tied_cache=ctx.tied_cache)
 
 

modelopt/torch/export/model_utils.py

@@ -210,8 +210,8 @@ def _reorder_canonical_first(state_dict: dict, model: nn.Module) -> dict:
     name to scope the reorder to DiffusionGemma; other tied
     encoder-decoder models that ship dict-style ``_tied_weights_keys``
     can be added to the allowlist here. Mirrors the ``model_type``
-    dispatch used for the Whisper and Nemotron-VL branches elsewhere
-    in ``unified_export_hf.py``.
+    dispatch used for the Whisper and Nemotron-VL branches in
+    ``hf_export_prep.py``.
     """
     model_type = type(model).__name__.lower()
     if "diffusiongemma" not in model_type and "diffusion_gemma" not in model_type:

modelopt/torch/export/moe_utils.py

@@ -22,6 +22,8 @@
 import torch
 import torch.nn as nn
 
+from .hf_weight_export import _export_quantized_weight
+
 
 def _alias_per_expert_subtree_from_prior(module: nn.Module, prior: nn.Module, n: int) -> None:
     """Build per-expert subtree on ``module`` by aliasing ``prior``'s packed buffers.
@@ -114,7 +116,6 @@ def _export_fused_experts(
     ``_export_transformers_checkpoint``) and scoped to one export
     invocation; when ``None`` the corresponding alias step is skipped.
     """
-    from modelopt.torch.export.unified_export_hf import _export_quantized_weight
     from modelopt.torch.quantization.plugins.huggingface import _get_fused_expert_intermediate_dim
 
     n = module.num_experts

modelopt/torch/export/plugins/vllm_fakequant_hf.py

@@ -35,9 +35,9 @@
 from modelopt.torch.quantization.utils.layerwise_calib import LayerActivationCollector
 from modelopt.torch.utils import get_unwrapped_name, safe_save
 
+from ..hf_export_prep import collect_shared_input_modules
 from ..layer_utils import get_experts_list, is_moe
 from ..quant_utils import get_quantization_format
-from ..unified_export_hf import collect_shared_input_modules
 
 __all__ = [
     "export_hf_vllm_fq_checkpoint",

modelopt/torch/export/registry.py

@@ -23,7 +23,7 @@
 
 Preparation and export use separate registries because they have independent matching
 precedence. Registering a handler for a new module type replaces what previously required
-editing if/elif chains inside ``unified_export_hf.py``.
+editing if/elif chains inside ``hf_weight_export.py``.
 """
 
 from collections.abc import Callable