format, fix import

Author: dsikka

examples/quantization_w8a8_fp8/qwen3_vl_moe_fp8_example.py

@@ -1,4 +1,3 @@
-
 from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration
 
 from llmcompressor import oneshot
@@ -28,7 +27,7 @@
         "re:.*lm_head",
         "re:visual.*",
         "re:model.visual.*",
-        're:.*mlp.gate$',
+        "re:.*mlp.gate$",
     ],
 )
 
@@ -38,4 +37,4 @@
 # Save to disk in compressed-tensors format.
 SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-DYNAMIC"
 model.save_pretrained(SAVE_DIR)
-processor.save_pretrained(SAVE_DIR)
+processor.save_pretrained(SAVE_DIR)

src/llmcompressor/modeling/prepare.py

@@ -1,11 +1,21 @@
 import tqdm
 from compressed_tensors.utils import replace_module
+from loguru import logger
 from transformers import PreTrainedModel
 
 from llmcompressor.modeling.deepseek_v3 import replace as replace_deepseekv3
 from llmcompressor.modeling.llama4 import replace as replace_llama4
 from llmcompressor.modeling.qwen3_moe import replace as replace_Qwen3MoE
-from llmcompressor.modeling.qwen3_vl_moe import replace as replace_Qwen3VLMoE
+
+try:
+    from llmcompressor.modeling.qwen3_vl_moe import replace as replace_Qwen3VLMoE
+except ImportError:
+    logger.warning(
+        "Qwen3-VL-MoE support is not available. "
+        "Please ensure that you have the correct version of transformers installed."
+    )
+    replace_Qwen3VLMoE = None
+
 from llmcompressor.utils.helpers import patch_attr
 
 __all__ = ["replace_modules_for_calibration"]
@@ -14,9 +24,11 @@
 replacements = {
     "DeepseekV3MoE": replace_deepseekv3,
     "Llama4TextMoe": replace_llama4,
-    "Qwen3VLMoeTextSparseMoeBlock": replace_Qwen3VLMoE,
 }
 
+if replace_Qwen3VLMoE is not None:
+    replacements["Qwen3VLMoeTextSparseMoeBlock"] = replace_Qwen3VLMoE
+
 
 def replace_modules_for_calibration(
     model: PreTrainedModel,

src/llmcompressor/modeling/qwen3_vl_moe.py

@@ -14,11 +14,14 @@ def __init__(self, config, original):
         self.gate = wrap_gate(original.gate)
         self.experts = SequentialQwen3VLMoeTextExperts(config, original.experts)
 
+
 class SequentialQwen3VLMoeTextExperts(torch.nn.ModuleList):
     def __init__(self, config, original):
         self.num_experts = original.gate_up_proj.shape[0]
         with skip_weights_initialize():
-            super().__init__([Qwen3VLMoeTextMLP(config) for _ in range(self.num_experts)])
+            super().__init__(
+                [Qwen3VLMoeTextMLP(config) for _ in range(self.num_experts)]
+            )
 
         intermediate_size = original.down_proj.shape[1]