Error when exporting LiquidAI/LFM2.5-1.2B-Instruct #17439
Description
🐛 Describe the bug
I tried to export LiquidAI/LFM2.5-1.2B-Instruct. When input_ids was set to a fixed size, e.g. 128 export works as well as model inference. However, when added dynamic shape into this script (regular export, just applied monkey patch to remove data dependent flow and made function return a tensor instead of a complex type):
import torch
import torch.nn as nn
from transformers import AutoModelForCausalLM
from transformers.models.lfm2.modeling_lfm2 import Lfm2ShortConv
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
from executorch.exir import to_edge_transform_and_lower
from torch.export import Dim
def apply_mask_to_padding_states(hidden_states, attention_mask):
if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
dtype = hidden_states.dtype
hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
return hidden_states
def patched_slow_forward(
self,
x: torch.Tensor,
past_key_values = None,
cache_position: torch.LongTensor | None = None,
attention_mask: torch.Tensor | None = None,
):
seqlen = x.shape[1]
x = apply_mask_to_padding_states(x, attention_mask)
BCx = self.in_proj(x).transpose(-1, -2)
B, C, x = BCx.chunk(3, dim=-2)
Bx = B * x
if past_key_values is not None:
conv_state = nn.functional.pad(Bx, (self.L_cache - Bx.shape[-1], 0))
past_key_values.conv_cache[self.layer_idx].copy_(conv_state)
conv_out = self.conv(Bx)[..., :seqlen]
y = C * conv_out
y = y.transpose(-1, -2).contiguous()
y = self.out_proj(y)
return y
class ExportWrapper(nn.Module):
def __init__(self, model):
super().__init__()
self.model = model
def forward(self, input_ids):
outputs = self.model(input_ids=input_ids, use_cache=False)
return (outputs.logits,)
print("Applying monkey patch...")
original_slow_forward = Lfm2ShortConv.slow_forward
Lfm2ShortConv.slow_forward = patched_slow_forward
try:
print("Loading model...")
model = AutoModelForCausalLM.from_pretrained(
"LiquidAI/LFM2.5-1.2B-Instruct",
trust_remote_code=True,
torch_dtype=torch.float32
)
model.eval()
wrapper = ExportWrapper(model)
print("Preparing inputs...")
sample_inputs = (
torch.randint(1, model.config.vocab_size, (1, 128), dtype=torch.long),
)
dynamic_shapes = {
"input_ids": {
1: Dim(
"tokens", min=1, max=2048
),
}
}
print("Starting export (this captures the graph)...")
exported_program = torch.export.export(wrapper, sample_inputs, dynamic_shapes=dynamic_shapes)
print("Lowering to ExecuTorch (XNNPACK)...")
et_program = to_edge_transform_and_lower(
exported_program,
partitioner=[XnnpackPartitioner()]
).to_executorch()
output_filename = "LFM2.5-1.2B-Instruct.pte"
with open(output_filename, "wb") as f:
f.write(et_program.buffer)
print(f"✅ Success! Exported to {output_filename}")
finally:
Lfm2ShortConv.slow_forward = original_slow_forward
print("Restored original model methods.")I got the following error:
---------------------------------------------------------------------------
ConstraintViolationError Traceback (most recent call last)
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:1933, in _export_to_aten_ir_make_fx(mod, fake_args, fake_kwargs, fake_params_buffers, constant_attrs, produce_guards_callback, transform)
1932 try:
-> 1933 produce_guards_callback(gm)
1934 except (ConstraintViolationError, ValueRangeError) as e:
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:2080, in _non_strict_export.<locals>._produce_guards_callback(gm)
2079 def _produce_guards_callback(gm):
-> 2080 return produce_guards_and_solve_constraints(
2081 fake_mode=fake_mode,
2082 gm=gm,
2083 dynamic_shapes=dynamic_shapes,
2084 equalities_inputs=equalities_inputs,
2085 original_signature=original_signature,
2086 )
File /opt/miniconda3/lib/python3.13/site-packages/torch/_export/non_strict_utils.py:603, in produce_guards_and_solve_constraints(fake_mode, gm, dynamic_shapes, equalities_inputs, original_signature)
602 if constraint_violation_error:
--> 603 raise constraint_violation_error
File /opt/miniconda3/lib/python3.13/site-packages/torch/_export/non_strict_utils.py:565, in produce_guards_and_solve_constraints(fake_mode, gm, dynamic_shapes, equalities_inputs, original_signature)
564 try:
--> 565 shape_env.produce_guards(
566 placeholders,
567 sources,
568 input_contexts=input_contexts,
569 equalities_inputs=equalities_inputs,
570 ignore_static=False,
571 )
572 except ConstraintViolationError as e:
File /opt/miniconda3/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py:5315, in ShapeEnv.produce_guards(self, *args, **kwargs)
5311 """
5312 Like produce_guards_verbose, but only returns the non-verbose python guard expressions
5313 (no verbose guards produced.)
5314 """
-> 5315 return self.produce_guards_verbose(*args, **kwargs, langs=("python",))[0].exprs
File /opt/miniconda3/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py:6056, in ShapeEnv.produce_guards_verbose(self, placeholders, sources, source_ref, guards, input_contexts, equalities_inputs, _simplified, ignore_static, langs)
6055 err = "\n".join(error_msgs)
-> 6056 raise ConstraintViolationError(
6057 f"Constraints violated ({debug_names_str})! "
6058 'For more information, run with TORCH_LOGS="+dynamic".\n'
6059 f"{err}"
6060 )
6061 elif len(warn_msgs) > 0:
ConstraintViolationError: Constraints violated (tokens)! For more information, run with TORCH_LOGS="+dynamic".
- Not all values of tokens = L['input_ids'].size()[1] in the specified range tokens <= 2048 satisfy the generated guard min(64*L['input_ids'].size()[1], 256*L['input_ids'].size()[1]) == (64*L['input_ids'].size()[1]).
During handling of the above exception, another exception occurred:
UserError Traceback (most recent call last)
Cell In[5], line 89
87 print("Starting export (this captures the graph)...")
88 # Export the WRAPPER, not the raw model
---> 89 exported_program = torch.export.export(wrapper, sample_inputs, dynamic_shapes=dynamic_shapes)
91 print("Lowering to ExecuTorch (XNNPACK)...")
92 et_program = to_edge_transform_and_lower(
93 exported_program,
94 partitioner=[XnnpackPartitioner()]
95 ).to_executorch()
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/__init__.py:311, in export(mod, args, kwargs, dynamic_shapes, strict, preserve_module_call_signature, prefer_deferred_runtime_asserts_over_guards)
309 new_msg = str(e) + "\n\n" + draft_export_msg
310 e.args = (new_msg,)
--> 311 raise e
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/__init__.py:277, in export(mod, args, kwargs, dynamic_shapes, strict, preserve_module_call_signature, prefer_deferred_runtime_asserts_over_guards)
270 raise ValueError(
271 "Exporting a ScriptModule is not supported. "
272 "Maybe try converting your ScriptModule to an ExportedProgram "
273 "using `TS2EPConverter(mod, args, kwargs).convert()` instead."
274 )
276 try:
--> 277 return _export(
278 mod,
279 args,
280 kwargs,
281 dynamic_shapes,
282 strict=strict,
283 preserve_module_call_signature=preserve_module_call_signature,
284 pre_dispatch=True,
285 prefer_deferred_runtime_asserts_over_guards=prefer_deferred_runtime_asserts_over_guards,
286 )
287 except Exception as e:
288 draft_export_msg = (
289 "The error above occurred when calling torch.export.export. If you would "
290 "like to view some more information about this error, and get a list "
291 "of all other errors that may occur in your export call, you can "
292 "replace your `export()` call with `draft_export()`."
293 )
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:1271, in _log_export_wrapper.<locals>.wrapper(*args, **kwargs)
1265 if hasattr(e, "partial_fx_graph"):
1266 print(
1267 e.partial_fx_graph,
1268 file=sys.stderr,
1269 )
-> 1271 raise e
1272 finally:
1273 _EXPORT_FLAGS = None
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:1237, in _log_export_wrapper.<locals>.wrapper(*args, **kwargs)
1235 try:
1236 start = time.time()
-> 1237 ep = fn(*args, **kwargs)
1238 end = time.time()
1239 log_export_usage(
1240 event="export.time",
1241 metrics=end - start,
1242 flags=_EXPORT_FLAGS,
1243 **get_ep_stats(ep),
1244 )
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/exported_program.py:124, in _disable_prexisiting_fake_mode.<locals>.wrapper(*args, **kwargs)
121 @functools.wraps(fn)
122 def wrapper(*args, **kwargs):
123 with unset_fake_temporarily():
--> 124 return fn(*args, **kwargs)
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:2377, in _export(mod, args, kwargs, dynamic_shapes, strict, preserve_module_call_signature, pre_dispatch, prefer_deferred_runtime_asserts_over_guards)
2369 # NOTE Export training IR rollout
2370 # Old export calls export._trace(pre_dispatch=True)
2371 # and there are still lot of internal/OSS callsites that
(...) 2374 # export_training_ir_rollout_check returns True in OSS
2375 # while internally it returns False UNLESS otherwise specified.
2376 if pre_dispatch and export_training_ir_rollout_check():
-> 2377 ep = _export_for_training(
2378 mod,
2379 args,
2380 kwargs,
2381 dynamic_shapes,
2382 strict=strict,
2383 preserve_module_call_signature=preserve_module_call_signature,
2384 prefer_deferred_runtime_asserts_over_guards=prefer_deferred_runtime_asserts_over_guards,
2385 )
2386 dtrace_structured("exported_program", payload_fn=lambda: str(ep))
2387 return ep
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:1271, in _log_export_wrapper.<locals>.wrapper(*args, **kwargs)
1265 if hasattr(e, "partial_fx_graph"):
1266 print(
1267 e.partial_fx_graph,
1268 file=sys.stderr,
1269 )
-> 1271 raise e
1272 finally:
1273 _EXPORT_FLAGS = None
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:1237, in _log_export_wrapper.<locals>.wrapper(*args, **kwargs)
1235 try:
1236 start = time.time()
-> 1237 ep = fn(*args, **kwargs)
1238 end = time.time()
1239 log_export_usage(
1240 event="export.time",
1241 metrics=end - start,
1242 flags=_EXPORT_FLAGS,
1243 **get_ep_stats(ep),
1244 )
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/exported_program.py:124, in _disable_prexisiting_fake_mode.<locals>.wrapper(*args, **kwargs)
121 @functools.wraps(fn)
122 def wrapper(*args, **kwargs):
123 with unset_fake_temporarily():
--> 124 return fn(*args, **kwargs)
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:2185, in _export_for_training(mod, args, kwargs, dynamic_shapes, strict, preserve_module_call_signature, prefer_deferred_runtime_asserts_over_guards)
2181 from torch._subclasses.fake_tensor import fake_tensor_tls
2183 fake_tensor_tls.non_strict_export_fake_tensor_tracker.clear()
-> 2185 export_artifact = export_func(
2186 mod=mod,
2187 args=args,
2188 kwargs=kwargs,
2189 dynamic_shapes=dynamic_shapes,
2190 preserve_module_call_signature=preserve_module_call_signature,
2191 orig_in_spec=orig_in_spec,
2192 prefer_deferred_runtime_asserts_over_guards=prefer_deferred_runtime_asserts_over_guards,
2193 _to_aten_func=_export_to_aten_ir_make_fx,
2194 )
2196 # If we are tracing with fake inputs, it is expected to
2197 # see fake tensor constants.
2198 if not strict and not has_ambient_mode:
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:2116, in _non_strict_export(mod, args, kwargs, dynamic_shapes, preserve_module_call_signature, orig_in_spec, prefer_deferred_runtime_asserts_over_guards, _to_aten_func)
2097 with (
2098 fake_mode,
2099 _NonStrictTorchFunctionHandler(),
2100 tracing(tx),
2101 torch._dynamo.config.patch(dynamo_config),
2102 ):
2103 with (
2104 _fakify_script_objects(mod, fake_args, fake_kwargs, fake_mode) as (
2105 patched_mod,
(...) 2114 # _to_aten_func is _export_to_aten_ir when using the default non-strict export
2115 # We need to pass positional args correctly
-> 2116 aten_export_artifact = _to_aten_func(
2117 patched_mod,
2118 new_fake_args,
2119 new_fake_kwargs,
2120 fake_params_buffers,
2121 new_fake_constant_attrs,
2122 produce_guards_callback=_produce_guards_callback,
2123 transform=_tuplify_outputs,
2124 )
2125 # aten_export_artifact.constants contains only fake script objects, we need to map them back
2126 aten_export_artifact.constants = {
2127 fqn: map_fake_to_real[obj] if isinstance(obj, FakeScriptObject) else obj
2128 for fqn, obj in aten_export_artifact.constants.items()
2129 }
File /opt/miniconda3/lib/python3.13/site-packages/torch/export/_trace.py:1935, in _export_to_aten_ir_make_fx(mod, fake_args, fake_kwargs, fake_params_buffers, constant_attrs, produce_guards_callback, transform)
1933 produce_guards_callback(gm)
1934 except (ConstraintViolationError, ValueRangeError) as e:
-> 1935 raise UserError(UserErrorType.CONSTRAINT_VIOLATION, str(e)) # noqa: B904
1937 return _produce_aten_artifact(
1938 gm=gm,
1939 mod=mod,
(...) 1945 fake_params_buffers=fake_params_buffers,
1946 )
UserError: Constraints violated (tokens)! For more information, run with TORCH_LOGS="+dynamic".
- Not all values of tokens = L['input_ids'].size()[1] in the specified range tokens <= 2048 satisfy the generated guard min(64*L['input_ids'].size()[1], 256*L['input_ids'].size()[1]) == (64*L['input_ids'].size()[1]).
The error above occurred when calling torch.export.export. If you would like to view some more information about this error, and get a list of all other errors that may occur in your export call, you can replace your `export()` call with `draft_export()`.
Look that this error message makes no sense:
Not all values of tokens = L['input_ids'].size()[1] in the specified range tokens <= 2048 satisfy the generated guard min(64*L['input_ids'].size()[1], 256*L['input_ids'].size()[1]) == (64*L['input_ids'].size()[1]).
Assuming that ['input_ids'].size() values are non-negative this condition is always true.
Versions
Collecting environment information...
PyTorch version: 2.10.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: macOS 26.2 (arm64)
GCC version: Could not collect
Clang version: 17.0.0 (clang-1700.6.3.2)
CMake version: version 4.1.0
Libc version: N/A
Python version: 3.13.5 | packaged by Anaconda, Inc. | (main, Jun 12 2025, 11:23:37) [Clang 14.0.6 ] (64-bit runtime)
Python platform: macOS-26.2-arm64-arm-64bit-Mach-O
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A
CPU:
Apple M3 Pro
Versions of relevant libraries:
[pip3] executorch==1.1.0
[pip3] numpy==2.4.2
[pip3] pytorch_tokenizers==1.1.0
[pip3] torch==2.10.0
[pip3] torchao==0.15.0
[conda] executorch 1.1.0 pypi_0 pypi
[conda] numpy 2.3.5 pypi_0 pypi
[conda] pytorch-tokenizers 1.1.0 pypi_0 pypi
[conda] torch 2.10.0 pypi_0 pypi
[conda] torchao 0.15.0 pypi_0 pypi
[conda] torchvision 0.25.0 pypi_0 pypi