"torch_compile" still depends on Triton functionality #41
Description
I attempted to use linear_cross_entropy with my GTX 1070 (old, I know), but it fails with the following errors. These errors were unexpected, because the description for "torch_compile" mode seems to suggest that Triton support is not necessary:
"torch_compile = A highly optimized torch.compile implementation. This is typically the fastest but uses the most amount of memory. Good as a reference and for systems that don't support Triton."
I can run the same code on CPU, without any issues. It is only when I try to use my GPU that I get the following errors.
Just reporting this issue for posterity. I'm not really expecting a solution, but if there is an easy fix... I would greatly appreciate one!
File "/home/crow/repos/praxis/praxis/losses/cut_cross_entropy.py", line 15, in forward
loss = linear_cross_entropy(
embeddings,
...<5 lines>...
reduction="mean",
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/cut_cross_entropy/linear_cross_entropy.py", line 104, in linear_cross_entropy
return torch_compile_linear_cross_entropy(
e, c, targets, bias, ignore_index, softcap, reduction, shift
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/cut_cross_entropy/torch_compile.py", line 66, in torch_compile_linear_cross_entropy
loss = torch_compile_linear_cross_entropy_apply(
e,
...<5 lines>...
reduction=reduction,
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 574, in _fn
return fn(*args, **kwargs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1380, in __call__
return self._torchdynamo_orig_callable(
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
frame, cache_entry, self.hooks, frame_state, skip=1
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 547, in __call__
return _compile(
frame.f_code,
...<14 lines>...
skip=skip + 1,
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 986, in _compile
guarded_code = compile_inner(code, one_graph, hooks, transform)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 715, in compile_inner
return _compile_inner(code, one_graph, hooks, transform)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_utils_internal.py", line 95, in wrapper_function
return function(*args, **kwargs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 750, in _compile_inner
out_code = transform_code_object(code, transform)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/bytecode_transformation.py", line 1361, in transform_code_object
transformations(instructions, code_options)
~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 231, in _fn
return fn(*args, **kwargs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 662, in transform
tracer.run()
~~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 2868, in run
super().run()
~~~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 1052, in run
while self.step():
~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 962, in step
self.dispatch_table[inst.opcode](self, inst)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 3048, in RETURN_VALUE
self._return(inst)
~~~~~~~~~~~~^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 3033, in _return
self.output.compile_subgraph(
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
self,
^^^^^
...<2 lines>...
),
^^
)
^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/output_graph.py", line 1101, in compile_subgraph
self.compile_and_call_fx_graph(
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
tx, list(reversed(stack_values)), root, output_replacements
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/output_graph.py", line 1382, in compile_and_call_fx_graph
compiled_fn = self.call_user_compiler(gm)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/output_graph.py", line 1432, in call_user_compiler
return self._call_user_compiler(gm)
~~~~~~~~~~~~~~~~~~~~~~~~^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/output_graph.py", line 1483, in _call_user_compiler
raise BackendCompilerFailed(self.compiler_fn, e).with_traceback(
e.__traceback__
) from None
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/output_graph.py", line 1462, in _call_user_compiler
compiled_fn = compiler_fn(gm, self.example_inputs())
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/repro/after_dynamo.py", line 130, in __call__
compiled_gm = compiler_fn(gm, example_inputs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/repro/after_dynamo.py", line 130, in __call__
compiled_gm = compiler_fn(gm, example_inputs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/__init__.py", line 2340, in __call__
return compile_fx(model_, inputs_, config_patches=self.config)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/compile_fx.py", line 1863, in compile_fx
return aot_autograd(
~~~~~~~~~~~~~
...<6 lines>...
cudagraphs=cudagraphs,
~~~~~~~~~~~~~~~~~~~~~~
)(model_, example_inputs_)
~^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/backends/common.py", line 83, in __call__
cg = aot_module_simplified(gm, example_inputs, **self.kwargs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_functorch/aot_autograd.py", line 1155, in aot_module_simplified
compiled_fn = dispatch_and_compile()
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_functorch/aot_autograd.py", line 1131, in dispatch_and_compile
compiled_fn, _ = create_aot_dispatcher_function(
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
functional_call,
^^^^^^^^^^^^^^^^
...<3 lines>...
shape_env,
^^^^^^^^^^
)
^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_functorch/aot_autograd.py", line 580, in create_aot_dispatcher_function
return _create_aot_dispatcher_function(
flat_fn, fake_flat_args, aot_config, fake_mode, shape_env
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_functorch/aot_autograd.py", line 830, in _create_aot_dispatcher_function
compiled_fn, fw_metadata = compiler_fn(
~~~~~~~~~~~^
flat_fn,
^^^^^^^^
...<2 lines>...
fw_metadata=fw_metadata,
^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 678, in aot_dispatch_autograd
compiled_fw_func = aot_config.fw_compiler(fw_module, adjusted_flat_args)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_functorch/aot_autograd.py", line 489, in __call__
return self.compiler_fn(gm, example_inputs)
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/compile_fx.py", line 1741, in fw_compiler_base
return inner_compile(
gm,
...<5 lines>...
boxed_forward_device_index=forward_device,
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/compile_fx.py", line 569, in compile_fx_inner
return wrap_compiler_debug(_compile_fx_inner, compiler_name="inductor")(
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
gm,
^^^
example_inputs,
^^^^^^^^^^^^^^^
**kwargs,
^^^^^^^^^
)
^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_dynamo/repro/after_aot.py", line 102, in debug_wrapper
inner_compiled_fn = compiler_fn(gm, example_inputs)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/compile_fx.py", line 685, in _compile_fx_inner
mb_compiled_graph = fx_codegen_and_compile(
gm, example_inputs, inputs_to_check, **graph_kwargs
)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/compile_fx.py", line 1129, in fx_codegen_and_compile
return scheme.codegen_and_compile(gm, example_inputs, inputs_to_check, graph_kwargs)
~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/compile_fx.py", line 1044, in codegen_and_compile
compiled_fn = graph.compile_to_module().call
~~~~~~~~~~~~~~~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/graph.py", line 2027, in compile_to_module
return self._compile_to_module()
~~~~~~~~~~~~~~~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/graph.py", line 2033, in _compile_to_module
self.codegen_with_cpp_wrapper() if self.cpp_wrapper else self.codegen()
~~~~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/graph.py", line 1964, in codegen
self.scheduler = Scheduler(self.operations)
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 1798, in __init__
self._init(nodes)
~~~~~~~~~~^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 1816, in _init
self.nodes = [self.create_scheduler_node(n) for n in nodes]
~~~~~~~~~~~~~~~~~~~~~~~~~~^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 1947, in create_scheduler_node
return SchedulerNode(self, node)
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 893, in __init__
self._compute_attrs()
~~~~~~~~~~~~~~~~~~~^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 907, in _compute_attrs
group_fn = self.scheduler.get_backend(device).group_fn
~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 3441, in get_backend
self.backends[device] = self.create_backend(device)
~~~~~~~~~~~~~~~~~~~^^^^^^^^
File "/home/crow/repos/praxis/.venv/lib/python3.13/site-packages/torch/_inductor/scheduler.py", line 3428, in create_backend
raise RuntimeError(
f"Found {device_props.name} which is too old to be supported by the triton GPU compiler, which is used as the backend. Triton only supports devices of CUDA Capability >= 7.0, but your device is of CUDA capability {device_props.major}.{device_props.minor}" # noqa: B950
)
torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
RuntimeError: Found NVIDIA GeForce GTX 1070 which is too old to be supported by the triton GPU compiler, which is used as the backend. Triton only supports devices of CUDA Capability >= 7.0, but your device is of CUDA capability 6.1
Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information