[Question] INT4 scripts failed during `process_weights_after_loading ()`

Your Question

I try to reproduce the script run-qwen3-30B-A3B-int4.sh.

I make the following steps:

1. quantize model to INT4 with

python tools/convert_hf_to_int4.py \
--input-dir ${MY_MODEL_DIR}/Qwen/Qwen3-30B-A3B \
--output-dir ${MY_MODEL_DIR}/Qwen/Qwen3-30B-A3B-INT4 \
--data-dir ${MY_DATA_DIR}/Salesforce/wikitext/wikitext-2-raw-v1`

2. convert model to torch_dist with

source "./scripts/models/qwen3-4B.sh"
python tools/convert_hf_to_torch_dist.py \
${MODEL_ARGS[@]} \
--hf-checkpoint ${MY_MODEL_DIR}/Qwen/Qwen3-30B-A3B \
--save ${MY_MODEL_DIR}/Qwen/Qwen3-30B-A3B_torch_dist

3. run the script run-qwen3-30B-A3B-int4.sh with CKPT_ARGS changed

CKPT_ARGS=(
   --hf-checkpoint ${MY_MODEL_DIR}/Qwen/Qwen3-30B-A3B-INT4/
   --ref-load ${MY_MODEL_DIR}/Qwen/Qwen3-30B-A3B_torch_dist/
)

Here is the log:
Update weights: 95%|█████████▍| 109/115 [00:35<00:01, 4.72it/s]
Update weights: 96%|█████████▌| 110/115 [00:35<00:01, 4.57it/s]
Update weights: 97%|█████████▋| 111/115 [00:35<00:00, 4.66it/s]
Update weights: 97%|█████████▋| 112/115 [00:36<00:00, 3.68it/s]
Update weights: 98%|█████████▊| 113/115 [00:36<00:00, 4.53it/s]
(SGLangEngine pid=179232) [2026-03-02 15:17:25] Scheduler hit an exception: Traceback (most recent call last):
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/managers/scheduler.py", line 2955, in run_scheduler_process
(SGLangEngine pid=179232) scheduler.event_loop_overlap()
(SGLangEngine pid=179232) File "/usr/local/lib/python3.11/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
(SGLangEngine pid=179232) return func(*args, **kwargs)
(SGLangEngine pid=179232) ^^^^^^^^^^^^^^^^^^^^^
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/managers/scheduler.py", line 1121, in event_loop_overlap
(SGLangEngine pid=179232) self.process_input_requests(recv_reqs)
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/managers/scheduler.py", line 1329, in process_input_requests
(SGLangEngine pid=179232) output = self._request_dispatcher(recv_req)
(SGLangEngine pid=179232) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/utils.py", line 507, in call
(SGLangEngine pid=179232) return fn(obj)
(SGLangEngine pid=179232) ^^^^^^^
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/managers/scheduler_update_weights_mixin.py", line 125, in post_process_weights
(SGLangEngine pid=179232) success, message = self.tp_worker.post_process_weights(recv_req)
(SGLangEngine pid=179232) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/managers/tp_worker.py", line 181, in post_process_weights
(SGLangEngine pid=179232) success, message = self.model_runner.post_process_weights(recv_req)
(SGLangEngine pid=179232) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/model_executor/model_runner.py", line 2480, in post_process_weights
(SGLangEngine pid=179232) quant_method.process_weights_after_loading(module)
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py", line 691, in process_weights_after_loading
(SGLangEngine pid=179232) layer.scheme.process_weights_after_loading(layer)
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py", line 300, in process_weights_after_loading
(SGLangEngine pid=179232) _transform_param(layer, self.w_q_name, transform_w_q)
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py", line 239, in _transform_param
(SGLangEngine pid=179232) new_param = fn(old_param)
(SGLangEngine pid=179232) ^^^^^^^^^^^^^
(SGLangEngine pid=179232) File "/root/sglang/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py", line 247, in transform_w_q
(SGLangEngine pid=179232) assert isinstance(x, BasevLLMParameter), f"w_q must be BasevLLMParameter, but got {type(x)}"
(SGLangEngine pid=179232) AssertionError: w_q must be BasevLLMParameter, but got <class 'torch.nn.parameter.Parameter'>
(SGLangEngine pid=179232)
(SGLangEngine pid=179232) [2026-03-02 15:17:25] SIGQUIT received. signum=None, frame=None. It usually means one child failed.
Update weights: 100%|██████████| 115/115 [00:36<00:00, 3.17it/s]

What I've Tried

N/A

Environment (if relevant)

export SGLANG_COMMIT="24c91001cf99ba642be791e099d358f4dfe955f5"
export MEGATRON_COMMIT="3714d81d418c9f1bca4594fc35f9e8289f652862"

The patch version I used is v0.5.7

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and FAQ and my question is not answered there.
• I have searched for existing issues and my question has not been asked before.

Reproduce run-qwen3-30B-A3B-int4.sh with the stated CKPT_ARGS, then trace the failure from python/sglang/srt/model_executor/model_runner.py into compressed_tensors.py and compressed_tensors_wNa16.py. Confirm the cause of the parameter-type assertion and make the failure reproducible with an actionable fix or usage guidance.