The issue cannot be reproduced using the original repro code provided in the issue description.
However, the underlying issue mentioned by the maintainer (missing functions in `builder.py` and `trace_rules.py`) was never addressed and can still be reproduced with this test case:
```python
import torch
from torch.nn.attention import _cur_sdpa_kernel_backends
@torch.compile(fullgraph=True)
def test_function_that_triggers_error():
return _cur_sdpa_kernel_backends()
print("Calling torch.compile function...")
try:
result = test_function_that_triggers_error()
print(f"Success: {result}")
except Exception as e:
print(f"ERROR: {e}")
print(f"Error type: {type(e)}")
```
The original repro likely no longer triggers the issue due to code path changes in the SDPA implementation, while the direct call to `_cur_sdpa_kernel_backends()` exposes the underlying problem where certain torch._C functions returning non-Tensor values aren't properly handled by dynamo tracing.
I have implemented the changes by adding the missing functions to both `builder.py` and `trace_rules.py` to properly handle these cases during compilation.
@guilhermeleobas
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161169
Approved by: https://github.com/guilhermeleobas, https://github.com/StrongerXi
Initial prototype for dynamic int inputs, allows users to run with `torch.compile(f)(DynamicInt(4))`, compiling dynamically and using the underlying hint at runtime.
Current behavior:
- Also works in eager (mostly by subclassing int), as scalar input to torch functions, or numpy/math/etc. For example, `x = DynamicInt(3); torch.randn(x); torch.add(y, z, alpha=x); np.arange(x)` all act as if x = 3.
- Behavior for arithmetic ops is to return new DynamicInts rather than static ints; `DynamicInt(3) * 2 = DynamicInt(6)`. This is via SymNode magic methods, but coverage might not be 100% - for example, I had to explicitly override floordiv to avoid int casting. This is not necessarily the case for non-magic method ops (e.g. `math.cos(x)`). The alternative here is to int cast on all operations, but I opted for this for dynamism propagation in non-compiled regions.
- Doesn't ban fullgraph=False; DynamicInt objects might be leaked back to the user, but I guess this is fine, because they can be casted to ints when needed?
- Dynamo only allocates one symbol per DynamicInt; specifying the same DynamicInt for multiple inputs leads to input deduplication, and a guard installed.
- We don't raise on int specialization (in allowlist/maybe_mark_dynamic style) - but an easy change if needed.
- DynamicInts as nn.Module attributes are handled.
- We don't guard on the DynamicInt id, e.g. users can do the following without recompiling (maybe we should guard?)
```python
x = DynamicInt(4)
f(x)
f(1)
f(DynamicInt(3)) # same as f(3)
```
Follow-up work:
- Specifying shape constraints, either at the int-level, e.g.
```python
DynamicInt(64, name="s0", constraints=["s0 % 32 == 0", "s0 <= 1024"]
```
or at the compilation level, e.g. something like
```python
s0 = DynamicInt(64, name="s0")
s1 = DynamicInt(128, name="s1")
with some_compiler_config.dynamic_int_constraints(["s1 == 2*s0", "s0 % 32 == 0"]):
f(s0, s1)
```
This should subsume the need for specifying derived SymInts?
- SymFloat support - currently it seems backed floats are specialized by the tensorify float pass, and there's no handling in inductor.
- Propagating dynamism in tensor constructors, e.g. `x = DynamicInt(4); torch.randn(x)` could annotate `_dynamo_dynamic_indices`.
Differential Revision: D81698719
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162194
Approved by: https://github.com/bobrenjc93
I confirmed that the tracing was correct i.e. NamedTupleVariable had the correct dynamic attribute added to it.
The problem was that NamedTupleVariable was always marked as immutable. This does not reflect the behavior of namedtuple.
Subclasses of namedtuple may be mutable, so when a NamedTupleVariable is derived from a subclass that is mutable, I made NamedTupleVariable mutable as well. Then side_effects correctly updates the returned object.
Fixes#161610
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161645
Approved by: https://github.com/anijain2305, https://github.com/StrongerXi
Fixes#159590
This is similar to the reverted commit #156868, except it resolves an issue with two caches becoming misaligned, leading to incorrect objects for stateful placements (i.e. `_MaskPartial`) as in issue #159601. This adds little to no overhead in eager ([see past benchmarks](https://github.com/pytorch/pytorch/pull/156868#issuecomment-3047831149)).
This also handles cases such as #159590 where dynamo is disabled during tracing by entering the Python Dispatcher ahead of the sharding propogation during compile. Tests are added/modified to handle these, and the list/tuple inputs with the cat op.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160798
Approved by: https://github.com/bdhirsh
Renaming `set_fullgraph` to `error_on_graph_break` for now. There are no semantic differences yet. In a followup PR, we will introduce a new `torch.compile` option `error_on_graph_break` that has lower priority than `fullgraph` so that `fullgraph` really returns 1 graph.
I could keep `set_fullgraph` as a deprecated alias for `error_on_graph_break` for now, but I'm hoping that won't be necessary since it's still private API (there are no internal callsites yet, and there are no significant OSS callsites yet).
cc @albanD @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @Lucaskabela @mlazos @guilhermeleobas @xmfan as primary users for `set_fullgraph`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161739
Approved by: https://github.com/xmfan, https://github.com/Lucaskabela, https://github.com/anijain2305, https://github.com/mlazos
Before the change in this PR, we have an error for the following code
```python
import torch
torch._dynamo.config.capture_scalar_outputs = True
class M(torch.nn.Module):
def forward(self, idx, x):
u0 = idx.item()
x0 = x.select(0, u0)
def fn():
return x0.sin()
return torch.cond(x0.sum() > 0, fn, fn)
m = M()
out = torch.compile(m, fullgraph=True)(torch.tensor(0, dtype=torch.int64), torch.randn(3, 3))
```
The error is caused when speculate fn, and tries to lift symbol of x0.storage_offset() but found the symbols doesn't have a source associated with it.
What really happens is that, when input tensor is a scalar tensor of int type and resides on CPU, we have a short cut that creates a norm symint when .item() is called see https://github.com/pytorch/pytorch/pull/126245.
However, previously, we only track the unbacked symint output of an operation because we believe all the backed symint must have a source associated with it and has already bee lifted as input at the top-level. Now this invariant no longer holds, so we end up an error saying the symbol doesn't have source (because only input and symbols derided from inputs have source and result of .item() doesn't have a source).
In this PR, we start to also track the normal symint with the proxy that created it (i.e. in this case the proxy .item()).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161198
Approved by: https://github.com/zou3519
This adds a new function `bypass_package` and `CompilePackage.bypass_current_entry()`. This allows us to safely bypass if there are models with unserializable or incompatible parts. When we encounter something incompatible, we'll raise a bypass and ignore that particular code in DynamoCodeEntry.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160902
Approved by: https://github.com/zhxchen17
Summary: ONNX team and recent transformer upgrade ran into this error and we also ran into during our export benchmarking. This diff makes it possible to trace through vmap implementation in pre-dispatch IR. Note that we don't support serializing functorch ops in pre-dispatch IR and in the future, we should desugar them to post-grad ops.
The implementation strategy is:
1. We add python wrappers around vmap APIs so that we attach custom torch function handler that is only on during non-strict export. The reason is we don't want to add this to default torch_function handler because it will break BC.
2. Some dynamo changes to make sure it picks up new python wrapper APIs. The reason is when we do strict export, we need to re-materialize these APIs in pre-dispatch IR from torch IR. We can avoid this by special casing in dynamo for export to proxy different API calls but i feel that is too much chaos because you need to be able to proxy 2 different variants of same vmap API.
Test Plan: CI
Differential Revision: D75623875
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154650
Approved by: https://github.com/ezyang, https://github.com/zou3519
I feel uneasy about touching `__warningregistry__` since it is undocumented and private surface. The only public API hook that doesn't increment warnings version seems to be https://docs.python.org/3/library/warnings.html#warnings.showwarning.
So we could wack a mole all the warnings muters in compile to just not display warnings, and we wouldn't invalidate warnings cache. This PR adds it for torch/_dynamo, and I didn't find any warnings versioning mutation from torch/_inductor.
There is a behavior change if someone calls a compiled graph with simplefilter("error"):
```python
# e.g. test/dynamo_expected_failures/TestAutogradFallback.test_no_autograd_kernel_inplace_mode_nothing
with warnings.catch_warnings():
warnings.simplefilter("error") # turns all warnings into errors
compiled_fn() # will throw if any of the muted warnings fire
```
FIXES https://github.com/pytorch/pytorch/issues/128427
A note for the future: The warnings module doesn't offer a thread safe way of using it. Even regular filters have this problem, directly editing `__warningregistry__` would be very bad, and this PR would mute all threads. Someone will need to build a thread safe warnings interface.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158520
Approved by: https://github.com/anijain2305, https://github.com/zou3519
Before the PR, for code like this:
```
class Example2(torch.nn.Module):
def forward(self, x, trigger, target):
return torch.cond(
trigger == 1,
lambda: x + target,
lambda: x * target,
(),
)
m = Example2()
x = torch.randn(2)
trigger = 0
target = 2
args = (x, trigger, target)
ep = torch.export.export(
m, args, dynamic_shapes=(None, Dim.DYNAMIC, Dim.DYNAMIC)
)
```
dynamo will wrap "target" (i.e. a symInt) twice, once when we speculate the first lambda and find target is a symint and decides to wrap it up, creating a new SymNodeVariable and a placeholder input to the top-level graph.
The second time happens when we speculate the second lambda. Tensors are de-duplicated by checking tracked side effects to make sure object with the same id (though different sources) is mapped to the same TensorVaraible. For symints, two things are missing:
1. it's not in the _can_lift_attrs_to_input list (the change in builder.py)
2. it's not in the tracked by runahead_side_effects, so when speculate_subgraph finishes, they're discarded (the change in side_effects.py)
Note: the auto lifting mechanism for HOPs happens at proxy level when we trace the subgraph, which is after SymNodeVariable are created (they're created when realizing the args and bind them to subgraph). At that time, builder has created two unique SymNodeVariable for the same symint so the auto lifting in hops cannot de-dup them.
Differential Revision: [D78298163](https://our.internmc.facebook.com/intern/diff/D78298163)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158273
Approved by: https://github.com/avikchaudhuri, https://github.com/zou3519
@animesh pointed out using whitelist for strides can result in confusing graphs as follows
```
s60: "Sym(s60)", L_hidden_states_: "bf16[1, 4096, 3072][s60, 3072, 1]cuda:0"
```
We probably want to capture the relationship between sizes and strides anyways so let's make it so the whitelist only makes the sizes dynamic. That same graph now looks lik ethis
```
L_hidden_states_: "bf16[1, 4096, 64][262144, 64, 1]cuda:0"
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157960
Approved by: https://github.com/pianpwk
Dynamo was aggressively specializing on lazy VTs over `set_name_hint` in
`STORE_FAST`, etc., and `isinstance` in `LOAD_FAST_CHECK`. This causes
regional `torch.compile` from optimizing ComfyUI GGUF + LoRA to either
(1). exceed the recompialtion limit of 8, which results in suboptimal
performance, and (2). even if recompilation limit is increased, the
compilation time gets unnecessarily high (180s v.s. 20s for Flux).
This patch fixes the recompilation issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156891
Approved by: https://github.com/williamwen42, https://github.com/mlazos
vLLM profiler sets with_stack=True that shows the dict_getitem on the profiler, both inflating the numbers and confusing compile users. This PR keeps BINARY_SUBSCR for regular dicts, while using `dict.__getitem__` only for dict subclasses.
Using binary_subscr is little bit faster, but not enough to make any major latency improvements.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155727
Approved by: https://github.com/zou3519, https://github.com/StrongerXi, https://github.com/jansel
The goal of this multigraph work is to enable a compiled region that has a single dynamo trace but multiple backend specializations. This work was inspired by vLLM which does this in a somewhat hacky way where they use a custom backend to capture a dynamo graph and then manually invoke compile_fx multiple times to get specialized graphs.
There's really two parts of this work:
**The frontend changes:**
1) we introduce an optional kwarg `specialize_on` to mark_{dynamic,unbacked} that takes in a list of specializations. I debated other methods including specifying specializations via decorators, but ultimately decided this approach was more harmonious. The big issue with decorators is the difficulty of composing well with the rest of the torch.compile ecosystem including graph breaks, lazy initialization of variable trackers and symbolic variables, etc.
**The backend changes (this PR):**
1) We capture the backend_specialization specified in the mark_{dynamic,unbacked} API into a SymbolicContext. See changes in `/_dynamo/variables/builder.py`
2) After we are done dynamo tracing, we will lazily (more on this later) invoke `call_user_compiler` up to N + 1 times for N specializations and 1 generic graph. Under the hood this will call compile_fx, which composes nicely with both Async Compile and AOTAutogradCache. We do this by using a context manager to patch in specialization specific axioms into the ShapeEnv before invoking the user compiler.
3) When we have specializations, we install a lazy specialized dispatch function that checks each specialization and dispatches to the first one that matches. Instead of doing all of the specialization compiles up front, we do the compiles lazily. The first time a specialization is invoked, we will do the compilation and save it in a cache so subsequent invocations are fast. If none of the specializations match, we dispatch to the generic graph. I decided to do this over returning N different GuardedCodes since 1) it doesn't pollute the dynamo cache (eg. if you have 8 specializations, you would hit the cache limit) 2) it naturally incorporates the hierarchical lattice structure of the guards since the specializations are always necessarily stricter than the generic region's guards.
I benchmarked this PR stack with #152596 and found around a 50% reduction when dispatching to the specialized regions:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153449
Approved by: https://github.com/zou3519
ghstack dependencies: #153433
