rebase

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

Author: kylesayrs

src/llmcompressor/modeling/__init__.py

@@ -14,6 +14,7 @@
 from .llama4 import SequentialLlama4TextMoe  # noqa: F401
 from .qwen3_moe import CalibrationQwen3MoeSparseMoeBlock  # noqa: F401
 from .qwen3_vl_moe import CalibrateQwen3VLMoeTextSparseMoeBlock  # noqa: F401
+from .qwen3_next_moe import CalibrationQwen3NextSparseMoeBlock  # noqa: F401
 # TODO: add granite4, Qwen3Next
 
 from .fuse import *

src/llmcompressor/modeling/moe_context.py

@@ -133,7 +133,7 @@ def moe_calibration_context(
         yield
     finally:
         # Step 2: Restore non-permanent modules
-        for name, (_original, replacement) in replaced.items():
+        for name, (original, replacement) in replaced.items():
             if not replacement.is_permanent:
                 restored = replacement.restore(original)
                 model.set_submodule(name, restored)

src/llmcompressor/modeling/qwen3_vl_moe.py

@@ -15,6 +15,8 @@ class CalibrateQwen3VLMoeTextSparseMoeBlock(MoECalibrationModule):
     experts.
     """
 
+    is_permanent = True
+
     def __init__(
         self,
         original: OriginalQwen3VLMoeTextSparseMoeBlock,
@@ -82,6 +84,9 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         next_states = next_states.reshape(batch_size, sequence_length, hidden_dim)
         return next_states, router_logits
 
+    def restore(self, original: torch.nn.Module) -> torch.nn.Module:
+        return original
+
 
 class SequentialQwen3VLMoeTextExperts(torch.nn.ModuleList):
     def __init__(self, config, original):