Address PR review feedback for eagle_base_lora feature

- Move peft imports (LoraConfig, inject_adapter_in_model, LoraLayer) inside
  the methods that use them (_inject_base_lora, _set_base_lora_enabled) so
  peft is not a hard top-level dependency for all speculative decoding users
- Change eagle_base_lora_target_modules default from [] to None to avoid
  mutable default shared across config instances
- Tighten LoRA key filtering from "lora_A" in k to ".lora_A." in k to avoid
  false positives, and add fail-fast RuntimeError when no LoRA tensors found

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Ye Yu <yeyu@nvidia.com>

Author: yeyu-nvidia

modelopt/torch/export/plugins/hf_spec_export.py

@@ -195,7 +195,12 @@ def _export_lora(self, export_dir: Path, full_sd: dict):
         """Export base model LoRA adapter weights alongside the eagle module artifacts."""
         from peft import LoraConfig
 
-        lora_sd = {k: v for k, v in full_sd.items() if "lora_A" in k or "lora_B" in k}
+        lora_sd = {k: v for k, v in full_sd.items() if ".lora_A." in k or ".lora_B." in k}
+        if not lora_sd:
+            raise RuntimeError(
+                "No LoRA adapter tensors found in the model state dict. "
+                "Ensure eagle_base_lora=True and the model was converted with LoRA adapters."
+            )
         save_file(lora_sd, export_dir / "lora_adapter_model.safetensors")
 
         lora_config = LoraConfig(

modelopt/torch/speculative/config.py

@@ -129,11 +129,11 @@ class EagleConfig(ModeloptBaseConfig):
         description="LoRA alpha (scaling) for the base model adapters.",
     )
 
-    eagle_base_lora_target_modules: list = ModeloptField(
-        default=[],
+    eagle_base_lora_target_modules: list | None = ModeloptField(
+        default=None,
         description=(
             "List of module name patterns to apply LoRA to in the base model "
-            "(e.g. ['q_proj', 'v_proj']). Empty list uses peft defaults."
+            "(e.g. ['q_proj', 'v_proj']). None uses peft defaults."
         ),
     )
 

modelopt/torch/speculative/plugins/transformers.py

@@ -37,9 +37,6 @@
 import torch
 import transformers
 from packaging.version import Version
-from peft import LoraConfig
-from peft.mapping import inject_adapter_in_model
-from peft.tuners.lora import LoraLayer
 from torch import nn
 from torch.nn import CrossEntropyLoss
 from torch.nn.attention.flex_attention import BlockMask, create_block_mask
@@ -552,6 +549,9 @@ def _get_eagle_device(self):
 
     def _inject_base_lora(self):
         """Inject HF PEFT LoRA adapters into the base model in-place and unfreeze them."""
+        from peft import LoraConfig
+        from peft.mapping import inject_adapter_in_model
+
         target_modules = self.eagle_base_lora_target_modules or None
         lora_config = LoraConfig(
             r=self.eagle_base_lora_rank,
@@ -567,6 +567,8 @@ def _inject_base_lora(self):
 
     def _set_base_lora_enabled(self, enabled: bool) -> None:
         """Enable or disable LoRA adapters in the base model."""
+        from peft.tuners.lora import LoraLayer
+
         for module in self._base_model.modules():
             if isinstance(module, LoraLayer):
                 module.enable_adapters(enabled)