I tested that dense works normally, but moe throws an error.

Fix Plan

Fix Name: Update Model Import

The issue is caused by an incorrect model import. The correct model name is Qwen3VLMoeForConditionalGeneration instead of Qwen3VLMoeForConditionalGeneration.

Step-by-Step Solution

1. Update Model Import

Replace the incorrect model import with the correct one:

from transformers import Qwen3VLMoeForConditionalGeneration, AutoProcessor

model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8", dtype="auto", device_map="auto",
)

2. Verify Model Import

Verify that the model import is correct by checking the model's attributes and methods:

print(model.config)
print(model.device_map)

3. Test Model

Test the model with a sample input to ensure it's working correctly:

input_ids = torch.tensor([[1, 2, 3]])
outputs = model(input_ids)
print(outputs)

Example Use Case

from transformers import Qwen3VLMoeForConditionalGeneration, AutoProcessor

model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8", dtype="auto", device_map="auto",
)

input_ids = torch.tensor([[1, 2, 3]])
outputs = model(input_ids)
print(outputs)

Verification

Verify that the fix worked by checking the model's output and ensuring it's correct.