Summary:
The basic concept behind this diff is to modify Dynamo's tracing behavior when it encounters a KeyedJaggedTensor that is synced (aka has `_length_per_key` and `_offset_per_key` populated). These fields are lists of integers; ordinarily, Dynamo will optimistically try to specialize on integers, however, for KJTs, we know that these integers will definitely vary from run-to-run. Furthermore, ordinarily, we would also specialize these integers if they are 0/1, but we will frequently expect features in KJTs to be 0/1.
The fix is to detect KJTs and treat these integers as *unbacked integers*. This is NOT a universally sound optimization: when treating these integers as unbacked, we never report them as equal to zero or one. In return, we always generate graphs that generalize no matter the length of values on features. This is enough to trace through APS sparse arch, torchrec_dlrm and some small split-cat examples.
The special integer behavior is triggered by a dynamically scoped `force_unspec_int_unbacked_size_like` variable on TracingContext, which we trigger when we wrap a KJT. There probably are other ways to do this, but this was simple and worked.
Test Plan:
```
buck2 test mode/dev-nosan //pytorch/benchmark/fb/test_gpu:run_test_gpu
```
from aakhundov
1. first build feed_lower_benchmark:
```
buck2 build --show-output mode/opt -c python.package_style=inplace -c fbcode.enable_gpu_sections=true -c fbcode.platform=platform010 -c fbcode.split-dwarf=true hpc/new/models/feed/benchmark:feed_lower_benchmark
```
2. then run the lowering of the model with it:
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCHINDUCTOR_UNIQUE_KERNEL_NAMES=1 TORCH_LOGS="output_code,graph_code" TORCH_COMPILE_DEBUG=1 ../buck-out/v2/gen/fbcode/79c6b019ee0f9469/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --load=manifold://ig_inference_model/tree/user/facebook/fblearner/predictor/960999465/60/gpu_lowering/input.predictor --skip-trt --skip-ait --sync-mode=0 --enable-aot-inductor --lower-presets="ig_stories" --gpu-trace
```
cf https://docs.google.com/document/d/1yD30xYrdmM8r2HTdmXnZTg0-MHVexfVrAa0294m1AUE/edit?pli=1#heading=h.qiv3fp7e6zg0
From torchrec: https://www.internalfb.com/intern/wiki/Torchrec/Development/Testing_production_models/
From ge0405
baseline (without your diff): f477293168
your diff: f477292363
```
buck2 test //caffe2/test/dynamo:test_dynamo_torchrec
buck2 run 'fbcode//mode/opt' fbcode//pytorch/benchmark/fb/test_gpu:run_test_gpu -- 'pytorch.benchmark.fb.test_gpu.test_gpu.TestBenchmarkFbGpu.test_train_blue_reels_vdd_v3_inductor_speedup'
```
Differential Revision: D49236757
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109216
Approved by: https://github.com/voznesenskym
before the PR, for HF's ModelOutput class, we use dicts.py/DataClassVariable with our own implementation on __getItem__, __setAttr__, __setItem__. There is a risk that ModelOutput logic may change since it is a user code
after the PR, we inline __getItem__, __setAttr__, __setItem__ using dicts.py/CustomizedDictVariable so the logic always keep AA
unit test
* python test/dynamo/test_model_output.py -k test_HF_bert_model_output
test on HF benchmark
* python benchmarks/dynamo/huggingface.py -d cuda --inference --accuracy --progress --inductor --print-dataframe-summary 2>&1
* all metric are the same before/after the PR, including pass rate, unique_graphs, graph_breaks, unique_graph_breaks
* before the PR: P790393916
* after the PR: P790368991
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105044
Approved by: https://github.com/jansel
The strategy in this PR is pretty straightforward.
There are 2 kinds of hooks:
1) Hooks on objects with sources (inputs, params)
2) Hooks on objects w/o sources (intermediaries, and outputs).
Note: As outputs can be made simple by how dynamo handles residuals, they could actually be handled as if they were inputs, but, for the sake of this PR, we will refer to hooks as either hooks on inputs (sourced), or hooks on intermediaries (not sourced).
The plan:
**For tensors w/ a source:**
We record registered hooks, store them as a global, and associate them with the tensor in residuals. This means that when dynamo goes to create the frame, where we produce bytecode to stitch together our PT2 modified bytecode with the original eager code, we call `register_hook`. This registration of hooks in residuals is sound because (a) it happens right after a Pt2 frame region ends and (b) we know that the tensor is alive in f_locals, f_globals, or a module in the users invoking frame. This means we can soundly know it will be around to invoke `register_hook` on. As long as we guard on the identity of the lifted function, this is sound to do.
**For tensors w/o a source:**
Graph break - we will support this in a subsequent PR
**Handles:**
An interesting new component here is the creation of a `STORE_FAST `->`LOAD_FAST` associated with the handle, the return result of `register_hook`. If the user code stored the result of `register_hook` in a handle, we need to honor that. We do so by interceding into `STORE_FAST`, and recording the name of the local variable as directed by user code. We then honor that same name in the reconstructed bytecode. If the user did not store a hook, we merely pop the produced value to preserve the stack.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108903
Approved by: https://github.com/ezyang
ghstack dependencies: #108846, #109092
Summary:
The basic concept behind this diff is to modify Dynamo's tracing behavior when it encounters a KeyedJaggedTensor that is synced (aka has `_length_per_key` and `_offset_per_key` populated). These fields are lists of integers; ordinarily, Dynamo will optimistically try to specialize on integers, however, for KJTs, we know that these integers will definitely vary from run-to-run. Furthermore, ordinarily, we would also specialize these integers if they are 0/1, but we will frequently expect features in KJTs to be 0/1.
The fix is to detect KJTs and treat these integers as *unbacked integers*. This is NOT a universally sound optimization: when treating these integers as unbacked, we never report them as equal to zero or one. In return, we always generate graphs that generalize no matter the length of values on features. This is enough to trace through APS sparse arch, torchrec_dlrm and some small split-cat examples.
The special integer behavior is triggered by a dynamically scoped `force_unspec_int_unbacked_size_like` variable on TracingContext, which we trigger when we wrap a KJT. There probably are other ways to do this, but this was simple and worked.
Test Plan:
```
buck2 test mode/dev-nosan //pytorch/benchmark/fb/test_gpu:run_test_gpu
```
from aakhundov
1. first build feed_lower_benchmark:
```
buck2 build --show-output mode/opt -c python.package_style=inplace -c fbcode.enable_gpu_sections=true -c fbcode.platform=platform010 -c fbcode.split-dwarf=true hpc/new/models/feed/benchmark:feed_lower_benchmark
```
2. then run the lowering of the model with it:
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCHINDUCTOR_UNIQUE_KERNEL_NAMES=1 TORCH_LOGS="output_code,graph_code" TORCH_COMPILE_DEBUG=1 ../buck-out/v2/gen/fbcode/79c6b019ee0f9469/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --load=manifold://ig_inference_model/tree/user/facebook/fblearner/predictor/960999465/60/gpu_lowering/input.predictor --skip-trt --skip-ait --sync-mode=0 --enable-aot-inductor --lower-presets="ig_stories" --gpu-trace
```
cf https://docs.google.com/document/d/1yD30xYrdmM8r2HTdmXnZTg0-MHVexfVrAa0294m1AUE/edit?pli=1#heading=h.qiv3fp7e6zg0
From torchrec: https://www.internalfb.com/intern/wiki/Torchrec/Development/Testing_production_models/
From ge0405
baseline (without your diff): f477293168
your diff: f477292363
Differential Revision: D49019987
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108960
Approved by: https://github.com/voznesenskym
In some cases, a UserFunctionVariable can be constructed when the underlying function is actually a TorchVariable. One example is when an attribute on a UnspecializedNNModuleVariable is a torch function. In those cases, we should treat the UserFunctionVariable as a TorchVariable.
This adds a check in UserDefinedObjectVariable.var_getattr() to try to create a TorchVariable instead of a UserFunctionVariable.
Fixes#104172
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104231
Approved by: https://github.com/williamwen42, https://github.com/jansel
Fixes: #101979
This PR adds support for dictionaries with torch object as keys in dynamo.
The main problem was that, for example, the source built for `d[torch.float]` (`d` being a
dictionary) was `ODictGetItemSource(GlobalSource('d'), index=torch.float)`. When
`Source.name` method was called, we got `odict_getitem(G['d'], torch.float)`. Evaluating
that string raised an error, since `torch` was only available in the global dictionary `G`
as `G["torch"]`.
Instead, this PR builds the source:
`ODictGetItemSource(GlobalSource('d'), index=AttrSource(GlobalSource('torch'), 'float'))`.
The to-be-evaluated string is correctly generated as:
`odict_getitem(G['d'], G['torch'].float)`.
Here's a minimal example that reproduces the error, before this PR:
```python
import torch
d = {
torch.float16: torch.float32,
}
@torch.compile
def f():
return torch.randn(3, dtype=d[torch.float16])
f()
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103158
Approved by: https://github.com/mlazos
Fixes#102315
The root cause is for ```UnspecializedNNModuleVariable``` which extends from ```UserDefinedObjectVariable```, if ```__bool__``` is missing, we should use ```__len__``` to infer a truth value.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/102583
Approved by: https://github.com/jansel
Stable Diffusion has a pattern like this:
```
def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, **cross_attention_kwargs):
# The `Attention` class can call different attention processors / attention functions
# here we simply pass along all tensors to the selected processor class
# For standard processors that are defined here, `**cross_attention_kwargs` is empty
return self.processor(
self,
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
```
Wherein processor is something like `AttnProcessor2_0`, which is callable but not an NNModule.
This allows for a significant speedup in stable diffusion.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100444
Approved by: https://github.com/anijain2305
This PR fixes capturing static methods for FSDP-managed modules. Previously, if a static method was invoked using `self.<staticmethod>`, then Dynamo would pass `self` twice to the method, causing a graph break due to the method being "unsupported". This PR achieves this by checking for `staticmethod` and using `UserFunctionVariable` instead of `UserMethodVariable`, which handles the correct calling convention.
This fixes FSDP + PT2 on HuggingFace's `T5ForConditionalGeneration`, which otherwise reports an error like the following based on the most recent trunk:
```
Output 0 of AsStridedBackward0 is a view of a view which was created in no_grad mode and is being modified inplace with grad mode enabled.
```
This is in reference to the `scores` tensor in `scores += position_bias_masked` ([code](a0ae2310ec/src/transformers/models/t5/modeling_t5.py (L559))).
I am not clear if this PR's fix is actually masking a different problem though. I wonder if there are edge cases with respect to Dynamo resuming execution and input mutations. Possibly, this PR only side steps the problem because there is no more recompilation at the static method `_relative_position_bucket()` ([code](a0ae2310ec/src/transformers/models/t5/modeling_t5.py (L443))).
In `UserDefinedObjectVariable.var_getattr()`, there is an existing branch:
e5291e633f/torch/_dynamo/variables/user_defined.py (L395-L398)
I am not clear on when this branch can be triggered since if `subobj` is a static method, it still takes the `FunctionTypes` branch:
e5291e633f/torch/_dynamo/variables/user_defined.py (L403-L404)
To preserve backward compatibility, the current version of this PR only modifies this `FunctionTypes` branch to differentiate between `staticmethod` and not `staticmethod`.
The PR that added this `FunctionTypes` branch is https://github.com/pytorch/pytorch/pull/92050/, and I checked that the added test `test_torch_distributions_functions()` only exercises the non-`staticmethod` case (since `Independent.log_prob` is not a `staticmethod`).
The last commit in `pytorch` that touched the `staticmethod` branch before https://github.com/pytorch/pytorch/pull/92050/ was the move from the `torchdynamo` repo into `pytorch`, so I cannot easily tell which test cases it corresponds to.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100117
Approved by: https://github.com/anijain2305
This is a draft version of generic context manager, I believe there are some scenarios that I didn't anticipate. I posted this draft for discussion and check if this is the right direction.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98725
Approved by: https://github.com/jansel
**Summary**: torch.nn.Module implementations previously did not support custom implementations of `__getattr__`; if a torch.nn.Module subclass implemented `__getattr__` and we tried to access an attribute that was expected to be present in `__getattr__`, dynamo would not check `__getattr__` and would error out with an AttributeError. This PR copies the functionality from UserDefinedObjectVariable into torch.nn.Module so that it also supports `__getattr__`
Example of a module which previously would fail:
```python
class MyMod(torch.nn.Module):
def __init__(self):
super().__init__()
self.custom_dict = {"queue": [torch.rand((2, 2)) for _ in range(3)]}
self.other_attr = torch.rand((2, 2))
def __getattr__(self, name):
custom_dict = self.custom_dict
if name in custom_dict:
return custom_dict[name]
return super().__getattr__(name)
def forward(self, x):
return x @ self.other_attr + self.queue[-1]
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94658
Approved by: https://github.com/yanboliang, https://github.com/jansel
**Motivation**
When adding support for default args (#90575), a lot of VariableTrackers missing sources were encountered. Currently, in a lot of cases it seems OK to skip the source for VariableTrackers created (especially during inlining), but that assumption breaks down when inlining functions with default arguments.
**Summary** of changes
- propagate the self.source of the VariableBuilder to the new variables being built, which seems like it was an omission previously
- Add SuperSource to track usages of super(), so that SuperVariables can support function calls with default args
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91729
Approved by: https://github.com/ezyang
