topk, bmm, max ops - Multiple dispatch failed for 'torch.ops.aten.size'; all __torch_dispatch__ handlers returned NotImplemented:

[rahulvijayaraghavan]

🐛 Describe the bug

Following error is observed for out variants of topk, bmm and max ops:
Multiple dispatch failed for 'torch.ops.aten.size'; all torch_dispatch handlers returned NotImplemented:

Error logs

E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1] fake tensor raised TypeError
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1] Traceback (most recent call last):
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py", line 862, in __torch_dispatch__
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     return self.dispatch(func, types, args, kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py", line 1207, in dispatch
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     return self._cached_dispatch_impl(func, types, args, kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py", line 940, in _cached_dispatch_impl
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     output = self._dispatch_impl(func, types, args, kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py", line 1424, in _dispatch_impl
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     r = func(*args, **kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_ops.py", line 600, in __call__
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     return self_._op(*args, **kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/utils/_stats.py", line 20, in wrapper
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     return fn(*args, **kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py", line 547, in __torch_dispatch__
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     return func(*args, **kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   File "/usr/local/lib/python3.10/dist-packages/torch/_ops.py", line 600, in __call__
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]     return self_._op(*args, **kwargs)
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1] TypeError: Multiple dispatch failed for 'torch.ops.aten.size'; all __torch_dispatch__ handlers returned NotImplemented:
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1] 
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1]   - tensor subclass <class 'torch._subclasses.fake_tensor.FakeTensor'>
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1] 
E0327 08:40:40.192000 137790200328192 torch/_subclasses/fake_tensor.py:864] [0/1] For more information, try re-running with TORCH_LOGS=not_implemented
---------------------------------------------------------------------------
TorchRuntimeError                         Traceback (most recent call last)
[<ipython-input-3-db886c9eeb8c>](https://localhost:8080/#) in <cell line: 12>()
     10 
     11 x = torch.arange(1., 8.)
---> 12 opt_model(x, 3, out=(values, indices))
     13 # Multiple dispatch failed for 'torch.ops.aten.size'; all __torch_dispatch__ handlers returned NotImplemented:

34 frames
[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/eval_frame.py](https://localhost:8080/#) in _fn(*args, **kwargs)
    407             prior = set_eval_frame(callback)
    408             try:
--> 409                 return fn(*args, **kwargs)
    410             finally:
    411                 set_eval_frame(prior)

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in catch_errors(frame, cache_entry, frame_state)
    937         with compile_lock, _disable_current_modes():
    938             # skip=1: skip this frame
--> 939             return callback(frame, cache_entry, hooks, frame_state, skip=1)
    940 
    941     catch_errors._torchdynamo_orig_callable = callback  # type: ignore[attr-defined]

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in _convert_frame(frame, cache_entry, hooks, frame_state, skip)
    800         counters["frames"]["total"] += 1
    801         try:
--> 802             result = inner_convert(
    803                 frame, cache_entry, hooks, frame_state, skip=skip + 1
    804             )

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in _convert_frame_assert(frame, cache_entry, hooks, frame_state, skip)
    398         )
    399 
--> 400         return _compile(
    401             frame.f_code,
    402             frame.f_globals,

[/usr/lib/python3.10/contextlib.py](https://localhost:8080/#) in inner(*args, **kwds)
     77         def inner(*args, **kwds):
     78             with self._recreate_cm():
---> 79                 return func(*args, **kwds)
     80         return inner
     81 

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in _compile(code, globals, locals, builtins, compiler_fn, one_graph, export, export_constraints, hooks, cache_size, frame, frame_state, compile_id, skip)
    684         fail_user_frame_lineno: Optional[int] = None
    685         try:
--> 686             guarded_code = compile_inner(code, one_graph, hooks, transform)
    687             return guarded_code
    688         except (

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in time_wrapper(*args, **kwargs)
    263             with torch.profiler.record_function(f"{key} (dynamo_timed)"):
    264                 t0 = time.time()
--> 265                 r = func(*args, **kwargs)
    266                 time_spent = time.time() - t0
    267             compilation_time_metrics[key].append(time_spent)

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in compile_inner(code, one_graph, hooks, transform)
    539             CompileContext.get().attempt = attempt
    540             try:
--> 541                 out_code = transform_code_object(code, transform)
    542                 break
    543             except exc.RestartAnalysis as e:

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/bytecode_transformation.py](https://localhost:8080/#) in transform_code_object(code, transformations, safe)
   1034     propagate_line_nums(instructions)
   1035 
-> 1036     transformations(instructions, code_options)
   1037     return clean_and_assemble_instructions(instructions, keys, code_options)[1]
   1038 

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in _fn(*args, **kwargs)
    163         cleanup = setup_compile_debug()
    164         try:
--> 165             return fn(*args, **kwargs)
    166         finally:
    167             cleanup.close()

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py](https://localhost:8080/#) in transform(instructions, code_options)
    501         try:
    502             with tracing(tracer.output.tracing_context), tracer.set_current_tx():
--> 503                 tracer.run()
    504         except exc.UnspecializeRestartAnalysis:
    505             speculation_log.clear()

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py](https://localhost:8080/#) in run(self)
   2114 
   2115     def run(self):
-> 2116         super().run()
   2117 
   2118     def match_nested_cell(self, name, cell):

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py](https://localhost:8080/#) in run(self)
    837             try:
    838                 self.output.push_tx(self)
--> 839                 while self.step():
    840                     pass
    841             except BackendCompilerFailed:

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py](https://localhost:8080/#) in step(self)
    751 
    752         try:
--> 753             self.dispatch_table[inst.opcode](self, inst)
    754             return not self.output.should_exit
    755         except ReturnValueOp:

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py](https://localhost:8080/#) in wrapper(self, inst)
    479                 return handle_graph_break(self, inst, speculation.reason)
    480             try:
--> 481                 return inner_fn(self, inst)
    482             except Unsupported as excp:
    483                 if self.generic_context_manager_depth > 0:

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py](https://localhost:8080/#) in CALL_FUNCTION_KW(self, inst)
   1253         kwargs = dict(zip(argnames, kwargs_list))
   1254         assert len(kwargs) == len(argnames)
-> 1255         self.call_function(fn, args, kwargs)
   1256 
   1257     def LOAD_METHOD_SUPER(self, inst):

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/symbolic_convert.py](https://localhost:8080/#) in call_function(self, fn, args, kwargs)
    691         if inner_fn and callable(inner_fn) and is_forbidden(inner_fn):
    692             raise AssertionError(f"Attempt to trace forbidden callable {inner_fn}")
--> 693         self.push(fn.call_function(self, args, kwargs))
    694 
    695     def inline_user_function_return(self, fn, args, kwargs):

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/variables/torch.py](https://localhost:8080/#) in call_function(self, tx, args, kwargs)
    743                     fn_ = getattr(torch, torch_sym_op)
    744 
--> 745             tensor_variable = wrap_fx_proxy(
    746                 tx=tx,
    747                 proxy=tx.output.create_proxy(

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/variables/builder.py](https://localhost:8080/#) in wrap_fx_proxy(tx, proxy, example_value, subclass_type, **options)
   1339     }
   1340     if subclass_type is None:
-> 1341         return wrap_fx_proxy_cls(target_cls=TensorVariable, **kwargs)
   1342     else:
   1343         result = wrap_fx_proxy_cls(target_cls=TensorWithTFOverrideVariable, **kwargs)

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/variables/builder.py](https://localhost:8080/#) in wrap_fx_proxy_cls(target_cls, tx, proxy, example_value, subclass_type, **options)
   1424         # only allow_non_graph_fake in this instance because we handle the non-fake
   1425         # cases properly below.
-> 1426         example_value = get_fake_value(proxy.node, tx, allow_non_graph_fake=True)
   1427 
   1428     # Handle recursive calls here

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in get_fake_value(node, tx, allow_non_graph_fake)
   1744             unimplemented(f"TypeError {node.target}: {cause}")
   1745 
-> 1746         raise TorchRuntimeError(str(e)).with_traceback(e.__traceback__) from None
   1747 
   1748     if not allow_non_graph_fake:

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in get_fake_value(node, tx, allow_non_graph_fake)
   1676     try:
   1677         with tx.fake_mode, enable_python_dispatcher():
-> 1678             ret_val = wrap_fake_exception(
   1679                 lambda: run_node(tx.output, node, args, kwargs, nnmodule)
   1680             )

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in wrap_fake_exception(fn)
   1207 def wrap_fake_exception(fn):
   1208     try:
-> 1209         return fn()
   1210     except UnsupportedFakeTensorException as e:
   1211         from .exc import unimplemented

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in <lambda>()
   1677         with tx.fake_mode, enable_python_dispatcher():
   1678             ret_val = wrap_fake_exception(
-> 1679                 lambda: run_node(tx.output, node, args, kwargs, nnmodule)
   1680             )
   1681     except Unsupported:

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in run_node(tracer, node, args, kwargs, nnmodule)
   1812             unimplemented(make_error_message(e), from_exc=e)
   1813         except Exception as e:
-> 1814             raise RuntimeError(make_error_message(e)).with_traceback(
   1815                 e.__traceback__
   1816             ) from e

[/usr/local/lib/python3.10/dist-packages/torch/_dynamo/utils.py](https://localhost:8080/#) in run_node(tracer, node, args, kwargs, nnmodule)
   1794         try:
   1795             if op == "call_function":
-> 1796                 return node.target(*args, **kwargs)
   1797             elif op == "call_method":
   1798                 return getattr(args[0], node.target)(*args[1:], **kwargs)

[/usr/local/lib/python3.10/dist-packages/torch/utils/_stats.py](https://localhost:8080/#) in wrapper(*args, **kwargs)
     18             simple_call_counter[fn.__qualname__] = 0
     19         simple_call_counter[fn.__qualname__] = simple_call_counter[fn.__qualname__] + 1
---> 20         return fn(*args, **kwargs)
     21     return wrapper

[/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py](https://localhost:8080/#) in __torch_dispatch__(self, func, types, args, kwargs)
    860         ), func
    861         try:
--> 862             return self.dispatch(func, types, args, kwargs)
    863         except TypeError:
    864             log.exception("fake tensor raised TypeError")

[/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py](https://localhost:8080/#) in dispatch(self, func, types, args, kwargs)
   1205 
   1206         if self.cache_enabled:
-> 1207             return self._cached_dispatch_impl(func, types, args, kwargs)
   1208         else:
   1209             return self._dispatch_impl(func, types, args, kwargs)

[/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py](https://localhost:8080/#) in _cached_dispatch_impl(self, func, types, args, kwargs)
    938 
    939         if output is unassigned:
--> 940             output = self._dispatch_impl(func, types, args, kwargs)
    941 
    942         return output

[/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py](https://localhost:8080/#) in _dispatch_impl(self, func, types, args, kwargs)
   1422         try:
   1423             with in_kernel_invocation_manager(self):
-> 1424                 r = func(*args, **kwargs)
   1425         except NotImplementedError as not_implemented_error:
   1426             return maybe_run_unsafe_fallback(not_implemented_error)

[/usr/local/lib/python3.10/dist-packages/torch/_ops.py](https://localhost:8080/#) in __call__(self_, *args, **kwargs)
    598         # use `self_` to avoid naming collide with aten ops arguments that
    599         # are named "self". This way, all the aten ops can be called by kwargs.
--> 600         return self_._op(*args, **kwargs)
    601 
    602     def __hash__(self):

[/usr/local/lib/python3.10/dist-packages/torch/utils/_stats.py](https://localhost:8080/#) in wrapper(*args, **kwargs)
     18             simple_call_counter[fn.__qualname__] = 0
     19         simple_call_counter[fn.__qualname__] = simple_call_counter[fn.__qualname__] + 1
---> 20         return fn(*args, **kwargs)
     21     return wrapper

[/usr/local/lib/python3.10/dist-packages/torch/_subclasses/fake_tensor.py](https://localhost:8080/#) in __torch_dispatch__(cls, func, types, args, kwargs)
    545 
    546         with fake_mode:  # type: ignore[attr-defined]
--> 547             return func(*args, **kwargs)
    548 
    549     @staticmethod

[/usr/local/lib/python3.10/dist-packages/torch/_ops.py](https://localhost:8080/#) in __call__(self_, *args, **kwargs)
    598         # use `self_` to avoid naming collide with aten ops arguments that
    599         # are named "self". This way, all the aten ops can be called by kwargs.
--> 600         return self_._op(*args, **kwargs)
    601 
    602     def __hash__(self):

TorchRuntimeError: Failed running call_function <built-in method topk of type object at 0x7d5191f98460>(*(FakeTensor(..., size=(s0,)), 3), **{'out': (FakeTensor(..., size=(3,)), FakeTensor(..., size=(3,), dtype=torch.int64))}):
Multiple dispatch failed for 'torch.ops.aten.size'; all __torch_dispatch__ handlers returned NotImplemented:

  - tensor subclass <class 'torch._subclasses.fake_tensor.FakeTensor'>

For more information, try re-running with TORCH_LOGS=not_implemented

from user code:
   File "<ipython-input-3-db886c9eeb8c>", line 2, in topk_func
    torch.topk(input, k, out=out)

Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information


You can suppress this exception and fall back to eager by setting:
    import torch._dynamo
    torch._dynamo.config.suppress_errors = True

Minified repro

topk

def topk_func(input, k, out):
  torch.topk(input, k, out=out)

values = torch.empty(3)
indices = torch.empty(3, dtype=torch.long)
opt_model = torch.compile(topk_func)

x = torch.arange(1., 6.)
opt_model(x, 3, out=(values, indices))

x = torch.arange(1., 8.)
opt_model(x, 3, out=(values, indices))

bmm

def bmm_func(input, mat, out):
  torch.bmm(input, mat, out=out)

opt_model = torch.compile(bmm_func)

inp1 = torch.randn(10, 3, 4)
mat1 = torch.randn(10, 4, 5)
out1 = torch.empty(10, 3, 5)
opt_model(inp1, mat1, out1)

inp2 = torch.randn(12, 4, 5)
mat2 = torch.randn(12, 5, 6)
out2 = torch.empty(12, 4, 6)
opt_model(inp2, mat2, out2)

max

def max_func(input, out):
  torch.max(input, 0, keepdim=True, out=out)

max = torch.empty(1)
max_indices = torch.empty(1, dtype=torch.long)

opt_model = torch.compile(max_func)

inp1 = torch.randn(4)
opt_model(inp1, out=(max, max_indices))

inp2 = torch.randn(5)
opt_model(inp2, out=(max, max_indices))

Versions

PyTorch version: 2.4.0.dev20240326+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.27.9
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.1.58+-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.6
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       x86_64
CPU op-mode(s):                     32-bit, 64-bit
Address sizes:                      46 bits physical, 48 bits virtual
Byte Order:                         Little Endian
CPU(s):                             2
On-line CPU(s) list:                0,1
Vendor ID:                          GenuineIntel
Model name:                         Intel(R) Xeon(R) CPU @ 2.20GHz
CPU family:                         6
Model:                              79
Thread(s) per core:                 2
Core(s) per socket:                 1
Socket(s):                          1
Stepping:                           0
BogoMIPS:                           4399.99
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm rdseed adx smap xsaveopt arat md_clear arch_capabilities
Hypervisor vendor:                  KVM
Virtualization type:                full
L1d cache:                          32 KiB (1 instance)
L1i cache:                          32 KiB (1 instance)
L2 cache:                           256 KiB (1 instance)
L3 cache:                           55 MiB (1 instance)
NUMA node(s):                       1
NUMA node0 CPU(s):                  0,1
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Mitigation; PTE Inversion
Vulnerability Mds:                  Vulnerable; SMT Host state unknown
Vulnerability Meltdown:             Vulnerable
Vulnerability Mmio stale data:      Vulnerable
Vulnerability Retbleed:             Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Vulnerable
Vulnerability Spectre v1:           Vulnerable: __user pointer sanitization and usercopy barriers only; no swapgs barriers
Vulnerability Spectre v2:           Vulnerable, IBPB: disabled, STIBP: disabled, PBRSB-eIBRS: Not affected
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Vulnerable

Versions of relevant libraries:
[pip3] numpy==1.25.2
[pip3] torch==2.4.0.dev20240326+cpu
[pip3] torchaudio==2.2.1+cu121
[pip3] torchdata==0.7.1
[pip3] torchsummary==1.5.1
[pip3] torchtext==0.17.1
[pip3] torchvision==0.17.1+cu121
[pip3] triton==2.2.0
[conda] Could not collect

cc @ezyang @msaroufim @bdhirsh @anijain2305 @zou3519 @chauhang @eellison

[ezyang]

Yeah there's some sort of problem here. It's also the reason why test_make_fx_symbolic_exhaustive_nn_functional_fractional_max_pool3d_cpu_float32 is xfailed

[ezyang]

What looks like is happening is that we're attempting to run the underlying meta for topk (eg), and somehow we're redispatching to fake tensor mode (!!!)