[Bug] qwen3.5推理TileLang依赖报错
### Checklist
- [x] 1. I have searched related issues but cannot get the expected help.
- [x] 2. The bug has not been fixed in the latest version.
- [x] 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
### Describe the bug
使用最新版lmdeploy docker v0.12.3推理Qwen/Qwen3.5-35B-A3B报错，与TileLang相关。
### Reproduction
root@xxxx:/# lmdeploy serve api_server Qwen/Qwen3.5-35B-A3B --server-port 10236 --backend pytorch --tool-call-parser qwen3
Downloading Model from https://www.modelscope.cn to directory: /root/.cache/modelscope/hub/models/Qwen/Qwen3.5-35B-A3B
2026-04-09 01:35:32,628 - modelscope - INFO - Target directory already exists, skipping creation.
2026-04-09 01:35:48,424 - lmdeploy - WARNING - transformers.py:22 - LMDeploy requires transformers version: [4.33.0 ~ 5.3.0], but found version: 5.5.0
Loading weights from safetensors: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 14/14 [00:14<00:00, 1.07s/it]
2026-04-09 01:36:14 [TileLang:tilelang.jit.kernel:INFO]: TileLang begins to compile kernel `causal_conv1d_fwd_main` with `out_idx=None`
Process mp_engine_proc:
Traceback (most recent call last):
File "/usr/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
self.run()
File "/usr/lib/python3.12/multiprocessing/process.py", line 108, in run
self._target(*self._args, **self._kwargs)
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/mp_engine/zmq_engine.py", line 109, in _mp_proc
engine = Engine.from_pretrained(
^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/engine.py", line 232, in from_pretrained
return cls(
^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/engine.py", line 151, in __init__
self.executor.init()
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/executor/base.py", line 255, in init
self.warmup()
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/executor/uni_executor.py", line 82, in warmup
self.model_agent.warmup()
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/model_agent/agent.py", line 485, in warmup
self._forward_impl(inputs)
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/model_agent/agent.py", line 1132, in _forward_impl
output = model_forward(
^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/engine/model_agent/agent.py", line 184, in model_forward
output = model(**input_dict)
^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/backends/cuda/graph_runner.py", line 235, in __call__
output = self.model(**kwargs)
^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/models/qwen3_5.py", line 1143, in forward
hidden_states, target_inputs_embeds = self.model(
^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/models/qwen3_5.py", line 1059, in forward
hidden_states = self.language_model(
^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/models/qwen3_5.py", line 975, in forward
hidden_states, residual = decoder_layer(
^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/models/qwen3_5.py", line 789, in forward
hidden_states = self.linear_attn(
^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/models/qwen3_5.py", line 552, in forward
mixed_qkv, conv_state = self.conv1d(mixed_qkv, conv_state, gated_delta_meta=gated_delta_meta)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/nn/gated_delta.py", line 330, in forward
return self.causal_conv1d_func(x, self.weight, self.bias, conv_state, gated_delta_meta=gated_delta_meta)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/lib/python3.12/contextlib.py", line 81, in inner
return func(*args, **kwds)
^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/nn/gated_delta.py", line 190, in __call__
return self.conv1d_func(x, weight, bias, conv_state, gated_delta_meta=gated_delta_meta)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/nn/gated_delta.py", line 129, in conv1d_func
out = self.causal_conv1d_fn(
^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/backends/cuda/causal_conv1d.py", line 26, in conv1d_fn
return self.causal_conv1d_fn(x,
^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/lmdeploy/pytorch/kernels/cuda/causal_conv1d.py", line 195, in causal_conv1d_fn
kernel = causal_conv1d_fwd(hidden_size, kernel_size, bias is not None, activation, dtype, x.stride(2),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/jit/__init__.py", line 440, in __call__
kernel = self.compile(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/jit/__init__.py", line 375, in compile
kernel_result = compile(
^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/jit/__init__.py", line 98, in compile
return cached(
^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/cache/__init__.py", line 74, in cached
return _dispatch_map[execution_backend].cached(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/cache/kernel_cache.py", line 264, in cached
kernel = JITKernel(
^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/jit/kernel.py", line 137, in __init__
adapter = self._compile_and_create_adapter(func, out_idx)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/jit/kernel.py", line 241, in _compile_and_create_adapter
artifact = tilelang.lower(
^^^^^^^^^^^^^^^
File "/opt/py3/lib/python3.12/site-packages/tilelang/engine/lower.py", line 246, in lower
PreLowerSemanticCheck(mod)
File "/opt/py3/lib/python3.12/site-packages/tilelang/engine/phase.py", line 132, in PreLowerSemanticCheck
tilelang.analysis.NestedLoopChecker()(mod)
File "/opt/py3/lib/python3.12/site-packages/tilelang/3rdparty/tvm/python/tvm/ir/transform.py", line 167, in __call__
return _ffi_transform_api.RunPass(self, mod)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "python/tvm_ffi/cython/function.pxi", line 929, in tvm_ffi.core.Function.__call__
File "<unknown>", line 0, in tvm::transform::Pass::operator()(tvm::IRModule) const
File "<unknown>", line 0, in tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
File "<unknown>", line 0, in tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
File "python/tvm_ffi/cython/function.pxi", line 1083, in tvm_ffi.core.tvm_ffi_callback
File "/opt/py3/lib/python3.12/site-packages/tilelang/analysis/nested_loop_checker.py", line 116, in pass_fn
_NestedLoopCheckVisitor().visit_stmt(func.body)
File "/opt/py3/lib/python3.12/site-packages/tilelang/3rdparty/tvm/python/tvm/runtime/support.py", line 159, in __init__
self._inst = cls(*args, **kwargs)
File "/opt/py3/lib/python3.12/site-packages/tilelang/3rdparty/tvm/python/tvm/runtime/support.py", line 197, in __setattr__
super(TVMDerivedObject, self).__setattr__(name, value)
AttributeError: '_NestedLoopCheckVisitor' object has no attribute '_inst'
### Environment
```Shell
docker: openmmlab/lmdeploy:v0.12.3-cu12.8
H200
```
### Error traceback
```Shell
```
1 条评论