DataDependent Path Is Reported In RemoveRedundantRashapePattern1

[dayo09]

What?

While exporting multimodal model's RoPE, data dependent path error is reported.

--------------------------------------------------------------------------
GuardOnDataDependentSymNode               Traceback (most recent call last)
Cell In[20], line 5
      2 #name = 'model.visual.blocks.0.attn'
      3 #name = 'model.language_model.layers.0.self_attn'
      4 name = 'visual.rotary_pos_emb'
----> 5 sess.export_submodule_(name)
      7 #record = capture.records[name]
      8 
      9 # [ (name, type(value)) for (name, value) in record.items() ]
     10 # pp(record['kwargs']['past_key_values'])

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/vlmexport/export.py:28, in Session.export_submodule_(self, key)
     25 args = record['args']
     26 kwargs = record['kwargs']
---> 28 tico.convert(module, args=args, kwargs=kwargs).save(self._output_prefix / f'{key}.circle')

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/tico/utils/convert.py:323, in convert(mod, args, kwargs, dynamic_shapes, strict, config)
    318 with torch.no_grad():
    319     exported_program = export(
    320         mod, args, kwargs, dynamic_shapes=dynamic_shapes, strict=strict
    321     )
--> 323 circle_binary = convert_exported_module_to_circle(exported_program, config=config)
    325 return CircleModel(circle_binary)

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/tico/utils/convert.py:270, in convert_exported_module_to_circle(exported_program, config)
    223 # TODO Distinguish legalize and optimize
    224 circle_legalize = PassManager(
    225     passes=[
    226         FillMetaVal(),
   (...)    268     ]
    269 )
--> 270 circle_legalize.run(exported_program)
    272 # After this stage, ExportedProgram invariant is broken, i.e.,
    273 # graph can have a constant torch.tensor not lifted to a placeholder
    274 circle_legalize = PassManager(
    275     passes=[
    276         FillMetaVal(),
    277         CastMixedTypeArgs(preserve_ep_invariant=False),
    278     ]
    279 )

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/tico/utils/passes.py:65, in PassManager.run(self, exported_program)
     59 for _pass in self.passes:
     60     # Automatically update the signatures of the input and output.
     61     # https://github.com/pytorch/executorch/issues/4013#issuecomment-2187161844
     62     with exported_program.graph_module._set_replace_hook(
     63         exported_program.graph_signature.get_replace_hook()
     64     ):
---> 65         result = _pass.call(exported_program)
     66     modified = modified or result.modified
     67     if modified and self.strategy == PassStrategy.RESTART:

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/tico/utils/trace_decorators.py:63, in trace_graph_diff_on_pass.<locals>._call_traced.<locals>.wrapped(*args)
     61 assert isinstance(graph_module, torch.fx.GraphModule), type(graph_module)
     62 capture(graph_module.graph)
---> 63 ret = fn(*args)
     64 log(graph_module.graph, title=str(cls.__name__), recapture=False)
     65 return ret

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/tico/passes/remove_redundant_reshape.py:81, in RemoveRedundantReshapePattern1.call(self, exported_program)
     79 reshape1_input = reshape1_args.input
     80 # `(AxBxC) - aten.reshape` - (1xAxBxC)
---> 81 if [1] + list(extract_shape(reshape1_input)) != list(
     82     extract_shape(reshape1)
     83 ):
     84     continue
     86 ### permute

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/__init__.py:751, in SymBool.__bool__(self)
    750 def __bool__(self):
--> 751     return self.node.bool_()

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/sym_node.py:616, in SymNode.bool_(self)
    615 def bool_(self):
--> 616     return self.guard_bool("", 0)

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/sym_node.py:538, in SymNode.guard_bool(self, file, line)
    535 def guard_bool(self, file, line):
    536     # TODO: use the file/line for some useful diagnostic on why a
    537     # guard occurred
--> 538     r = self.evaluate()
    539     try:
    540         return bool(r)

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/sym_node.py:512, in SymNode.evaluate(self, size_oblivious)
    511 def evaluate(self, size_oblivious=False):
--> 512     return self.shape_env.evaluate_sym_node(self, size_oblivious)

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/symbolic_shapes.py:7223, in ShapeEnv.evaluate_sym_node(self, sym_node, size_oblivious, fallback_value)
   7218 """
   7219 Given a a SymNode, evaluates sym_node.expr, adding guards if necessary.
   7220 """
   7222 self._expr_sym_node_id = id(sym_node)
-> 7223 return self.evaluate_expr(
   7224     sym_node.expr,
   7225     sym_node.hint,
   7226     sym_node.fx_node,
   7227     size_oblivious,
   7228     fallback_value=fallback_value,
   7229 )

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/symbolic_shapes.py:7323, in ShapeEnv.evaluate_expr(self, orig_expr, hint, fx_node, size_oblivious, fallback_value, forcing_spec)
   7321 # Add extra state that evaluate_expr() depends on.
   7322 suppress_guards_tls = ShapeEnv._suppress_guards_tls()
-> 7323 return self._inner_evaluate_expr(
   7324     orig_expr,
   7325     hint,
   7326     fx_node,
   7327     size_oblivious,
   7328     forcing_spec,
   7329     suppress_guards_tls,
   7330     fallback_value,
   7331 )

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/recording.py:272, in record_shapeenv_event.<locals>.decorator.<locals>.wrapper(*args, **kwargs)
    265 try:
    266     if not shape_env.should_record_events or shape_env.is_recording:  # type: ignore[has-type]
    267         # If ShapeEnv is already recording an event, call the wrapped
    268         # function directly.
    269         #
    270         # NB: here, we skip the check of whether all ShapeEnv instances
    271         # are equal, in favor of a faster dispatch.
--> 272         return retlog(fn(*args, **kwargs))
    274     # Retrieve an instance of ShapeEnv.
    275     # Assumption: the collection of args and kwargs may not reference
    276     # different ShapeEnv instances.
    277     self = _extract_shape_env_and_assert_equal(args, kwargs)

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/symbolic_shapes.py:7346, in ShapeEnv._inner_evaluate_expr(self, orig_expr, hint, fx_node, size_oblivious, forcing_spec, _suppress_guards_tls, fallback_value)
   7333 @lru_cache(256)
   7334 @record_shapeenv_event(save_tracked_fakes=True, name="evaluate_expr")
   7335 def _inner_evaluate_expr(
   (...)   7343     fallback_value: Optional[bool] = None,
   7344 ) -> sympy.Basic:
   7345     try:
-> 7346         return self._evaluate_expr(
   7347             orig_expr,
   7348             hint,
   7349             fx_node,
   7350             size_oblivious,
   7351             fallback_value,
   7352             forcing_spec=forcing_spec,
   7353         )
   7354     except Exception as e:
   7355         if isinstance(e, GuardOnDataDependentSymNode):

File ~/opt/miniforge3/envs/Qwen2.5-VL/lib/python3.11/site-packages/torch/fx/experimental/symbolic_shapes.py:7570, in ShapeEnv._evaluate_expr(self, orig_expr, hint, fx_node, size_oblivious, fallback_value, forcing_spec)
   7566         if not size_oblivious:
   7567             size_oblivious_result = self._maybe_evaluate_static(
   7568                 expr, size_oblivious=True
   7569             )
-> 7570         raise self._make_data_dependent_error(
   7571             expr.xreplace(self.var_to_val),
   7572             expr,
   7573             size_oblivious_result=size_oblivious_result,
   7574             expr_sym_node_id=self._expr_sym_node_id,
   7575         )
   7576 else:
   7577     expr = new_expr

versions

• torch: '2.8.0+cu128'
• torchvision: '0.23.0+cu128'

[dayo09]

Progress update

Problematic target module: Qwen2_5_VisionRotaryEmbedding
Example input given: (tensor(30),)
Recommened input: (30,)

• When input is given as int, it compiled successfully. Otherwise, torch produces a guard with 'symint' type the content of the tensor assumed to be vary in range.

class Qwen2_5_VisionRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs

[dayo09]

Primitive integers are 'specialized' while tracing

https://docs.pytorch.org/docs/stable/export/programming_model.html#which-values-are-static-vs-dynamic

'specialized' means, treated as constant, non-variable.

• Tensor data : static
• Tensor shape: static/dynamic
• Python primitives: static
• The structure of list/tuple/dict/namedtuple: static (length, for example)
• The contained elements of list/tuple/dict/namedtuple: recursively applies the rule