Implements https://github.com/pytorch/pytorch/issues/144908.
Implementation notes:
- `set_fullgraph` is implemented using `patch_config`, which changes config correctly during runtime and tracing.
- Moved setting `config.error_on_graph_break` from convert_frame.py to eval_frame.py. This is because this should only be done at the top-level decorated function. If we kept this in convert_frame.py, we would be changing `config.error_on_graph_break` on every top-level frame, which causes confusing behavior (see added test for example).
- InstructionTranslator reads from `config.error_on_graph_break` every `step()`. This is to determine the value of `config.error_on_graph_break` at the time of the graph break, because tracer cleanup will restore the value of `config.error_on_graph_break` .
- `convert_frame.py` determines whether we should abort tracing (fullgraph=True) or continue (fullgraph=False) by reading the value of the tracer's `error_on_graph_break`. If there is no tracer (failed to initialize), then default to reading `config.error_on_graph_break`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154289
Approved by: https://github.com/jansel, https://github.com/zou3519
ghstack dependencies: #154283
`torch.compile` now always goes through `torch._dynamo._optimize`. fullgraph is now implemented in `torch.compile` by looking at `config.error_on_graph_break`. Export still goes through `torch._dynamo._optimize_assert`, which uses `tx.one_graph` instead of `config.error_on_graph_break`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154283
Approved by: https://github.com/jansel, https://github.com/anijain2305
I had to create a new PR for this because of @atalman request of temporary reverting the previous PR to restore diff train sync. Nothing has changed from this PR and the original one.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156639
Approved by: https://github.com/atalman
In practice `bool(...)` is either constant folded by Dynamo or used for
branching (so most of its emulation logic lived in
`InstructionTranslator.generic_jump`.
This patch adds a dedicated `bool` hanlder (only for symbolic
bool/int/float for now), and fixes#136075.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155863
Approved by: https://github.com/williamwen42
This PR refers to the issue: https://github.com/pytorch/pytorch/issues/155352
This PR uses torch._dynamo.utils.warn_once so that this warning only emits once, clarifies in the warning that silent incorrectness is potential, not observed, Doesn't warn for functions that come from torch.*
As of right now with this code change the terminal outputs:
if the code came from torch.* :
Nothing, as we shouldn't warn for functions that come from torch.*
else:
/data/users/ssubbarao8/pytorch/torch/_dynamo/variables/functions.py:1565: UserWarning: Dynamo detected a call to a `functools.lru_cache`-wrapped function. Dynamo ignores the cache wrapper and directly traces the wrapped function. Silent incorrectness is only a *potential* risk, not something we have observed. Enable TORCH_LOGS="+dynamo" for a DEBUG stack trace.
torch._dynamo.utils.warn_once(msg)
If the user runs the command 'TORCH_LOGS="+dynamo" python foo4.py', in the debug logs it shows(this log below is based on chillee's repro:
/data/users/ssubbarao8/pytorch/torch/_dynamo/variables/functions.py:1565: UserWarning: Dynamo detected a call to a `functools.lru_cache`-wrapped function. Dynamo ignores the cache wrapper and directly traces the wrapped function. Silent incorrectness is only a *potential* risk, not something we have observed. Enable TORCH_LOGS="+dynamo" for a DEBUG stack trace.
torch._dynamo.utils.warn_once(msg)
V0619 21:00:16.504000 956424 torch/_dynamo/variables/functions.py:1575] [0/0] call to a lru_cache` wrapped function from user code at: /data/users/ssubbarao8/pytorch/foo4.py:9
V0619 21:00:16.504000 956424 torch/_dynamo/variables/functions.py:1575] [0/0] File "/data/users/ssubbarao8/pytorch/foo4.py", line 9, in <module>
V0619 21:00:16.504000 956424 torch/_dynamo/variables/functions.py:1575] [0/0] torch.compile(foo, backend="eager")(torch.randn(4))
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156463
Approved by: https://github.com/williamwen42
dynamo: Don't crash when someone tries to access a non existent list member
Test added which reproduces the failure. Note that I'm using the new
unimplemented_v2 API. Let me know if people have a strong preference that I use
something else.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156335
Approved by: https://github.com/jansel
Implements https://github.com/pytorch/pytorch/issues/144908.
Implementation notes:
- `set_fullgraph` is implemented using `patch_config`, which changes config correctly during runtime and tracing.
- Moved setting `config.error_on_graph_break` from convert_frame.py to eval_frame.py. This is because this should only be done at the top-level decorated function. If we kept this in convert_frame.py, we would be changing `config.error_on_graph_break` on every top-level frame, which causes confusing behavior (see added test for example).
- InstructionTranslator reads from `config.error_on_graph_break` every `step()`. This is to determine the value of `config.error_on_graph_break` at the time of the graph break, because tracer cleanup will restore the value of `config.error_on_graph_break` .
- `convert_frame.py` determines whether we should abort tracing (fullgraph=True) or continue (fullgraph=False) by reading the value of the tracer's `error_on_graph_break`. If there is no tracer (failed to initialize), then default to reading `config.error_on_graph_break`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154289
Approved by: https://github.com/jansel, https://github.com/zou3519
ghstack dependencies: #154283
`torch.compile` now always goes through `torch._dynamo._optimize`. fullgraph is now implemented in `torch.compile` by looking at `config.error_on_graph_break`. Export still goes through `torch._dynamo._optimize_assert`, which uses `tx.one_graph` instead of `config.error_on_graph_break`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154283
Approved by: https://github.com/jansel, https://github.com/anijain2305
Recursive function collect_temp_source with closure in PyCodegen caused cycle reference issue when torch.compile is used.
This issue may cause major tensors will not freed timely even there are no user references to these tensors.
We saw OOM issues because of this problem in many cases including training and inference using torch.compile.
The fix is to use iterative function implementation to replace the recursive function implementation.
Fixes#155778
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155791
Approved by: https://github.com/ezyang
Implements https://github.com/pytorch/pytorch/issues/144908.
Implementation notes:
- `set_fullgraph` is implemented using `patch_config`, which changes config correctly during runtime and tracing.
- Moved setting `config.error_on_graph_break` from convert_frame.py to eval_frame.py. This is because this should only be done at the top-level decorated function. If we kept this in convert_frame.py, we would be changing `config.error_on_graph_break` on every top-level frame, which causes confusing behavior (see added test for example).
- InstructionTranslator reads from `config.error_on_graph_break` every `step()`. This is to determine the value of `config.error_on_graph_break` at the time of the graph break, because tracer cleanup will restore the value of `config.error_on_graph_break` .
- `convert_frame.py` determines whether we should abort tracing (fullgraph=True) or continue (fullgraph=False) by reading the value of the tracer's `error_on_graph_break`. If there is no tracer (failed to initialize), then default to reading `config.error_on_graph_break`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154289
Approved by: https://github.com/jansel, https://github.com/zou3519
ghstack dependencies: #154283
`torch.compile` now always goes through `torch._dynamo._optimize`. fullgraph is now implemented in `torch.compile` by looking at `config.error_on_graph_break`. Export still goes through `torch._dynamo._optimize_assert`, which uses `tx.one_graph` instead of `config.error_on_graph_break`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154283
Approved by: https://github.com/jansel, https://github.com/anijain2305
This is enough to get @XilunWu 's stack in a state where his flex_attention DTensor implementations worked E2E for me. It also required these changes on the DTensor side, to properly add a DTensor rule for flex backward: P1789852198
There are two problems:
(1) in the normal dispatcher, we have a precedence ordering between modes and subclasses. Modes are dispatched to first, but modes are allowed to return NotImplemented, giving subclasses a chance to run.
This normally happens automatically in `FakeTensorMode.__torch_dispatch__` and `FunctionalTensorMode.__torch_dispatch__`. However, since HOPs implement these two modes themselves, HOPs do not get this benefit. For now, I ended up hardcoding this `NotImplemented` logic directly into the functional/fake rules for flex attention.
Having to do this for every HOP seems a bit painful. If we could plumb every HOP through `Fake[|Functional]TensorMode.__torch_dispatch__` then we would get this support. Another option could be to just assume that most HOP <> mode implementations want the same treatment by default, and hardcode this `NotImplemented` logic into `torch/_ops.py`. I'm not sure if we'd need a way for the HOP to opt out of this though.
(2) We were hardcoding a call to flex attention's fake implementation in dynamo to run fake prop. This is technically wrong for subclasses, because it doesn't give subclasses the chance to interpose on the op and desugar it before fake prop runs. I tweaked dynamo's logic to call the op, and let the dispatcher handle invoking the fake implementation.
**Testing** Xilun is adding some DTensor tests in his PR that will end up testing this logic. If folks would prefer, though, I can try to add a test that uses another subclass instead that is maybe more basic.
This is the tlparse that his DTensor test gnerated for me: https://manifold.edge.x2p.facebook.net/v0/read/tree/logs/hirsheybar/0196c1d3-a9a2-46ea-a46d-aa21618aa060/custom/rank_0/index.html?bucketName=tlparse_reports&apiKey=tlparse_reports-key&withPayload=1&timeoutMsec=10000
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151719
Approved by: https://github.com/ydwu4
Co-authored-by: drisspg <drisspguessous@gmail.com>
Summary:
This PR adds support for torch.cuda.FloatTensor and friends in Dynamo.
These are indeed legacy APIs, but that doesn't stop us from adding
support for them in torch.compile.
I add support for these in the same way that we support torch.Tensor:
these APIs can be safely put into the Dynamo graph.
Fixes#130722
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156107
Approved by: https://github.com/williamwen42
This adds support for user-defined triton kernels using TensorDescriptor.from_tensor into triton_kernel_wrap: i.e. storing metadata about the TMA descriptors and doing mutation analysis.
Major changes:
* TMADescriptorMetadata has changed: previously it was a dict[str, tuple[list[int], list[int], int]]. But now there are two metadata formats: one for experimental API and one for stable API. Now the metadata format is dict[str, tuple[str, tuple[...]]], where tuple[...] is tuple[list[int], list[int], int] for experimental and tuple[list[int],] for stable API. And then most handling of the metadata has to be branched based on whether the metadata represents a stable or experimental TMA descriptor
* mutation analysis: unlike experimental TMA (where the mutation analysis / ttir analysis pretends that the TMA descriptor is actually just a tensor), we need to construct an actual TMA descriptor before getting the Triton frontend to create the TTIR (otherwise assertions fail). A TensorDescriptor (i.e. stable TMA API descriptor) passed into a python triton kernel actually turns into 1 + 2*N parameters in the TTIR (for a rank-N tensor), so the arg list also needs to be patched for this reason (in generate_ttir)
* mutation analysis: now we also need to pass tma_descriptor_metadata into the mutation analysis, in order to create the TMA descriptors that are passed into the frontend code (ie. the previous point). This is why all the mutation tests are modified with an extra return value (the tma_descriptor_metadata)
Inductor is not modified (Inductor just errors out if you use a stable API tma descriptor). This will be the next PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155777
Approved by: https://github.com/aakhundov
