[Bug] qwen vl broken with qwen35 megatron bridge

Bug Description

(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler_update_weights_mixin.py", line 95, in update_weights_from_tensor [repeated 6x across cluster]
(SGLangEngine pid=87820)     success, message = worker.update_weights_from_tensor(recv_req) [repeated 6x across cluster]
(SGLangEngine pid=87820)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [repeated 12x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/managers/tp_worker.py", line 159, in update_weights_from_tensor [repeated 6x across cluster]
(SGLangEngine pid=87820)     success, message = self.model_runner.update_weights_from_tensor( [repeated 6x across cluster]
(SGLangEngine pid=87820)                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 1437, in update_weights_from_tensor [repeated 6x across cluster]
(SGLangEngine pid=87820)     return self._update_weights_from_flattened_bucket( [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 1488, in _update_weights_from_flattened_bucket [repeated 6x across cluster]
(SGLangEngine pid=87820)     self.model.load_weights(reconstructed_tensors) [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/models/qwen3_vl_moe.py", line 288, in load_weights [repeated 6x across cluster]
(SGLangEngine pid=87820)     load_fused_expert_weights( [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/models/qwen3_vl_moe.py", line 162, in load_fused_expert_weights [repeated 6x across cluster]
(SGLangEngine pid=87820)     weight_loader( [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/layers/moe/fused_moe_triton/layer.py", line 633, in weight_loader [repeated 6x across cluster]
(SGLangEngine pid=87820)     self._weight_loader_physical( [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/layers/moe/fused_moe_triton/layer.py", line 663, in _weight_loader_physical [repeated 6x across cluster]
(SGLangEngine pid=87820)     self._weight_loader_impl( [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/layers/moe/fused_moe_triton/layer.py", line 859, in _weight_loader_impl [repeated 6x across cluster]
(SGLangEngine pid=87820)     self._load_model_weight_or_group_weight_scale( [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/layers/moe/fused_moe_triton/layer.py", line 367, in _load_model_weight_or_group_weight_scale [repeated 6x across cluster]
(SGLangEngine pid=87820)   File "/sgl-workspace/sglang/python/sglang/srt/layers/moe/fused_moe_triton/layer.py", line 459, in _load_w13 [repeated 6x across cluster]
(SGLangEngine pid=87820)     expert_data.copy_(loaded_weight) [repeated 6x across cluster]
(SGLangEngine pid=87820) RuntimeError: The size of tensor a (2048) must match the size of tensor b (1536) at non-singleton dimension 1 [repeated 6x across cluster]
(SGLangEngine pid=87820) [2026-03-13 05:17:06] SIGQUIT received. signum=None, frame=None. It usually means one child failed.

Steps to Reproduce

When using https://github.com/coding-famer/Megatron-Bridge-slime/tree/qwen35, qwen3 vl is broken.
with image dataset

Expected Behavior

no

Actual Behavior

no

Environment

• slime version:
• Python version:
• PyTorch version:
• CUDA/ROCm version:
• GPU type and count:
• OS:
• SGLang version (if relevant):
• Megatron-LM version (if relevant):

Logs

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and my issue is not addressed there.
• I have searched for existing issues and this is not a duplicate.
• I have provided a minimal, reproducible example.

Start with python/sglang/srt/models/qwen3_vl_moe.py at load_weights and load_fused_expert_weights, then trace the dimension handling into layers/moe/fused_moe_triton/layer.py at _load_w13. Reproduce the qwen3 VL image-dataset path using the linked Megatron Bridge branch and compare the reconstructed tensor dimensions; done means weight loading completes without the reported mismatch.