評價此頁
★ ★ ★ ★ ★
fullgraph=False">
error_on_graph_break">

切換 error_on_graph_break#

建立日期：2025 年 9 月 3 日 | 最後更新日期：2025 年 9 月 3 日

摘要

  • 當 fullgraph=False 時，我們可以使用 torch._dynamo.error_on_graph_break() 來更靈活地處理圖中斷。

到目前為止，我們已經介紹了兩種處理 torch.compile 中圖中斷的方法。

  1. fullgraph=True 會在第一次圖中斷時報錯，並額外保證只從程式碼中跟蹤一個圖。

  2. fullgraph=False 在遇到圖中斷時也會繼續跟蹤。

如果我們希望大部分程式碼都不允許圖中斷，但有少數有問題的函式，這些函式的圖中斷很難消除，而我們對此是可以接受的，該怎麼辦？我們可以使用 torch._dynamo.error_on_graph_break() 來實現此目的。

torch.compile 有一個 error_on_graph_break 設定（初始值為 False）。如果在 error_on_graph_break 設定為 False 的情況下，在程式碼中發生圖中斷或編譯器錯誤，那麼 torch.compile 將嘗試在圖中斷/錯誤後繼續編譯。如果 error_on_graph_break 設定為 True，那麼 torch.compile 將中止編譯並將錯誤傳播給使用者程式碼。

error_on_graph_break=True 和 fullgraph=True 之間的顯著區別在於，前者**不保證捕獲單個圖**。error_on_graph_break**可以在編譯時任意切換**，透過使用 torch._dynamo.error_on_graph_break() 上下文管理器/裝飾器。相比之下，一旦 fullgraph 設定為 True，就無法將其重新設定為 False。最後，error_on_graph_break 的優先順序低於 fullgraph — error_on_graph_break 僅在 fullgraph=False 時生效。

error_on_graph_break(False) 示例#

@torch._dynamo.error_on_graph_break(False)
def code_with_a_difficult_graph_break(x):
    x = x + 1
    torch._dynamo.graph_break()
    return x + 2

def inner(x):
    return code_with_a_difficult_graph_break(x)

# NOTE: fullgraph=False
@torch._dynamo.error_on_graph_break(True)
@torch.compile
def fn(x):
    return inner(x)

# No error, but there is a graph break
fn(torch.randn(3))

Graph break in user code at /tmp/ipykernel_880/1452578661.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/1452578661.py", line 17, in <module>
    fn(torch.randn(3))
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/1452578661.py", line 14, in fn
    return inner(x)
  File "/tmp/ipykernel_880/1452578661.py", line 8, in inner
    return code_with_a_difficult_graph_break(x)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/1452578661.py", line 4, in code_with_a_difficult_graph_break
    torch._dynamo.graph_break()

Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1452578661.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1452578661.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1452578661.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

tensor([1.5304, 1.4924, 0.9838])

在 error_on_graph_break(True) 下使用 error_on_graph_break(False)，有助於在我們希望最小化圖中斷（即遵循 fullgraph=True 程式設計模型）但存在某些難以規避的、非效能關鍵的圖中斷程式碼段時。

error_on_graph_break() 也可以用作上下文管理器。

# NOTE: fullgraph=False
@torch._dynamo.error_on_graph_break(True)
@torch.compile
def fn(x):
    x = x + 1
    with torch._dynamo.error_on_graph_break(False):
        torch._dynamo.graph_break()  # no error
    return x + 2

# No error, but there is a graph break
fn(torch.randn(3))

Graph break in user code at /tmp/ipykernel_880/737485247.py:7
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/737485247.py", line 11, in <module>
    fn(torch.randn(3))
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/737485247.py", line 7, in fn
    torch._dynamo.graph_break()  # no error

tensor([4.5572, 3.4907, 1.6146])

您可以使用猴子補丁（monkey patching）來切換 error_on_graph_break，以處理您無法編輯原始碼的程式碼（例如，框架程式碼）。

class ThirdPartyModule(torch.nn.Module):
    def forward(self, x):
        x = x + 1
        torch._dynamo.graph_break()
        return x + 2

tp_mod = ThirdPartyModule()
tp_mod.forward = torch._dynamo.error_on_graph_break(False)(tp_mod.forward)

@torch._dynamo.error_on_graph_break(True)
@torch.compile
def fn(x):
    return tp_mod.forward(x)

# No error, but there is a graph break
fn(torch.randn(3))

Graph break in user code at /tmp/ipykernel_880/2112598647.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/2112598647.py", line 16, in <module>
    fn(torch.randn(3))
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/2112598647.py", line 13, in fn
    return tp_mod.forward(x)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/2112598647.py", line 4, in forward
    torch._dynamo.graph_break()

Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/2112598647.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/2112598647.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

tensor([1.7608, 2.9385, 2.1531])

error_on_graph_break(True) 示例#

@torch._dynamo.error_on_graph_break(True)
def inner2(x):
    x = x + 1
    torch._dynamo.graph_break()  # error
    return x + 2

def inner(x):
    return inner2(x)

# fullgraph=False, error_on_graph_break=False
@torch.compile
def fn(x):
    x = x + 4
    torch._dynamo.graph_break()  # no error
    return inner(x)

try:
    fn(torch.randn(3))
except Exception as e:
    print(e)

Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

from user code:
   File "/tmp/ipykernel_880/2379101916.py", line 15, in torch_dynamo_resume_in_fn_at_14
    return inner(x)
  File "/tmp/ipykernel_880/2379101916.py", line 8, in inner
    return inner2(x)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/2379101916.py", line 4, in inner2
    torch._dynamo.graph_break()  # error

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

Graph break in user code at /tmp/ipykernel_880/2379101916.py:14
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/2379101916.py", line 18, in <module>
    fn(torch.randn(3))
  File "/tmp/ipykernel_880/2379101916.py", line 14, in fn
    torch._dynamo.graph_break()  # no error

在 error_on_graph_break(False) 下使用 error_on_graph_break(True)，有助於在我們希望靈活使用 torch.compile（即遵循 fullgraph=False 程式設計模型）但某些程式碼段是效能關鍵的，並且我們希望確保這些程式碼段不包含圖中斷時。

error_on_graph_break 巢狀行為#

torch._dynamo.error_on_graph_break() 也會影響巢狀呼叫的 error_on_graph_break 設定。

def inner(x):
    x = x + 1
    torch._dynamo.graph_break()
    return x + 2

def inner2(x):
    with torch._dynamo.error_on_graph_break(False):
        return inner(x)

@torch._dynamo.error_on_graph_break(True)
@torch.compile
def fn(x):
    return inner2(x)

# no error
fn(torch.randn(3))

Graph break in user code at /tmp/ipykernel_880/1007149706.py:3
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/1007149706.py", line 16, in <module>
    fn(torch.randn(3))
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/1007149706.py", line 13, in fn
    return inner2(x)
  File "/tmp/ipykernel_880/1007149706.py", line 8, in inner2
    return inner(x)
  File "/tmp/ipykernel_880/1007149706.py", line 3, in inner
    torch._dynamo.graph_break()

Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1007149706.py:3
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1007149706.py:3
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

tensor([3.5942, 2.0674, 2.8445])

torch._dynamo.error_on_graph_break() 可以在另一個 torch._dynamo.error_on_graph_break() 區域下使用。

def inner(x):
    x = x + 1
    with torch._dynamo.error_on_graph_break(False):
        torch._dynamo.graph_break()
    return x + 2

def inner2(x):
    with torch._dynamo.error_on_graph_break(True):
        return inner(x)

@torch.compile
def fn(x):
    return inner2(x)

# no error
fn(torch.randn(3))

Graph break in user code at /tmp/ipykernel_880/1343774799.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/1343774799.py", line 16, in <module>
    fn(torch.randn(3))
  File "/tmp/ipykernel_880/1343774799.py", line 13, in fn
    return inner2(x)
  File "/tmp/ipykernel_880/1343774799.py", line 9, in inner2
    return inner(x)
  File "/tmp/ipykernel_880/1343774799.py", line 4, in inner
    torch._dynamo.graph_break()

Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1343774799.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
Graph break (user stack suppressed due to duplicate graph break) in user code at /tmp/ipykernel_880/1343774799.py:4
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

tensor([1.3759, 2.1866, 3.8692])

與 fullgraph 的互動#

fullgraph=True 的優先順序高於 error_on_graph_break。

@torch._dynamo.error_on_graph_break(False)
def inner(x):
    x = x + 1
    torch._dynamo.graph_break()
    return x + 2

@torch.compile(fullgraph=True)
def fn(x):
    return inner(x)

try:
    fn(torch.randn(3))
except Exception as e:
    print(e)

Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

from user code:
   File "/tmp/ipykernel_880/2331424258.py", line 9, in fn
    return inner(x)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/external_utils.py", line 209, in inner
    return func(*args, **kwargs)
  File "/tmp/ipykernel_880/2331424258.py", line 3, in inner
    x = x + 1

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

fullgraph=True 無法切換回 fullgraph=False。

@torch.compile(fullgraph=False)
def inner(x):
    x = x + 1
    torch._dynamo.graph_break()
    return x + 2

@torch.compile(fullgraph=True)
def fn(x):
    return inner(x)

try:
    fn(torch.randn(3))
except Exception as e:
    print(e)

Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

from user code:
   File "/tmp/ipykernel_880/262151723.py", line 9, in fn
    return inner(x)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/polyfills/__init__.py", line 264, in getattr_and_trace
    return fn(*args[2:], **kwargs)
  File "/tmp/ipykernel_880/262151723.py", line 3, in inner
    x = x + 1

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

@torch.compile(fullgraph=True)
def inner(x):
    x = x + 1
    torch._dynamo.graph_break()
    return x + 2

@torch.compile(fullgraph=False)
def fn(x):
    return inner(x)

try:
    fn(torch.randn(3))
except Exception as e:
    print(e)

Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html

from user code:
   File "/tmp/ipykernel_880/2119173801.py", line 3, in inner
    x = x + 1

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

Graph break in user code at /tmp/ipykernel_880/2119173801.py:3
Graph Break Reason: Call to `torch._dynamo.graph_break()`
  Explanation: User-inserted graph break. Message: None
  Hint: Remove the `torch._dynamo.graph_break()` call.

  Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`

 For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0025.html
User code traceback:
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/opt/conda/envs/py_3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel_launcher.py", line 18, in <module>
    app.launch_new_instance()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/traitlets/config/application.py", line 1075, in launch_instance
    app.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelapp.py", line 739, in start
    self.io_loop.start()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/tornado/platform/asyncio.py", line 211, in start
    self.asyncio_loop.run_forever()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 603, in run_forever
    self._run_once()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/base_events.py", line 1909, in _run_once
    handle._run()
  File "/opt/conda/envs/py_3.10/lib/python3.10/asyncio/events.py", line 80, in _run
    self._context.run(self._callback, *self._args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 519, in dispatch_queue
    await self.process_one()
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 508, in process_one
    await dispatch(*args)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 400, in dispatch_shell
    await result
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 368, in execute_request
    await super().execute_request(stream, ident, parent)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/kernelbase.py", line 767, in execute_request
    reply_content = await reply_content
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/ipkernel.py", line 455, in do_execute
    res = shell.run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/ipykernel/zmqshell.py", line 577, in run_cell
    return super().run_cell(*args, **kwargs)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3006, in run_cell
    result = self._run_cell(
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3061, in _run_cell
    result = runner(coro)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/async_helpers.py", line 129, in _pseudo_sync_runner
    coro.send(None)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3266, in run_cell_async
    has_raised = await self.run_ast_nodes(code_ast.body, cell_name,
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3445, in run_ast_nodes
    if await self.run_code(code, result, async_=asy):
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3505, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_880/2119173801.py", line 12, in <module>
    fn(torch.randn(3))
  File "/tmp/ipykernel_880/2119173801.py", line 9, in fn
    return inner(x)
  File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/polyfills/__init__.py", line 264, in getattr_and_trace
    return fn(*args[2:], **kwargs)
  File "/tmp/ipykernel_880/2119173801.py", line 3, in inner
    x = x + 1

fullgraph=True/False 與 error_on_graph_break 總結#

下表總結了 fullgraph=True/False 和 error_on_graph_break 之間的區別。

error_on_graph_break=True

error_on_graph_break=False (預設)

fullgraph=True

圖中斷會導致錯誤。只報告第一次圖中斷。保證一個圖。

無法將 fullgraph 切換到 False。error_on_graph_break 無效。

使用者程式碼必須完全相容 torch.compile。保證不會因圖中斷產生效能損失（因為沒有圖中斷）。

最適合對圖中斷敏感的程式碼：框架/庫程式碼或需要最大效能的場景。防止下游使用者程式碼無意中允許圖中斷。

與 fullgraph=True 和 error_on_graph_break=True 相同，因為當 fullgraph=True 時，error_on_graph_break 無效。

fullgraph=False (預設)

圖中斷會導致錯誤。只報告第一次圖中斷。不保證一個圖。

可以將 error_on_graph_break 切換到 False。

使用者程式碼必須完全相容 torch.compile。保證不會因圖中斷產生效能損失（因為沒有圖中斷）。

最適合對圖中斷敏感的使用者程式碼。error_on_graph_break 可以切換到 False 來處理那些難以規避的、有圖中斷的程式碼段。

遇到圖中斷後會繼續編譯。所有圖中斷都會被報告。

可以將 error_on_graph_break 切換到 True。

不需要大量使用者程式碼更改即可工作。效能可能會因圖中斷而受到負面影響。

最適合開箱即用的場景，“非特殊”程式碼，或不需要榨取最大效能的場景。
在此頁面上
在 GitHub 上編輯
顯示原始碼
PyTorch 庫