Using the same repro from the issue (but with BatchNorm2D)
Rectifies native_batch_norm schema by splitting the schema into 2:
1. one will have NON-optional alias-able running_mean and running_var inputs
2. the other will just not have those parameters at all (no_stats variation)
**Calling for name suggestions!**
## test plan
I've added tests in test_functionalization.py as well as an entry in common_method_invocations.py for `native_batch_norm_legit`
CI should pass.
## next steps
Because of bc/fc reasons, we reroute native_batch_norm to call our new schemas ONLY through the python dispatcher, but in 2 weeks or so, we should make `native_batch_norm_legit` the official batch_norm.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88697
Approved by: https://github.com/albanD
This is an interesting one
Since this is an operation that's intrinsically defined on the reals,
we should perform the ops on that dtype always, and just cast to
the desired dtype at the end. This simplifies the decomposition.
Now, I started looking at this one when I started seeing failures on a
test that's added in a later PR. What's going on here is that, by doing
an upcast to a higher dtype and then cast down to integers, sometimes
there's an off-by-one error. I think this is fine, as the decomposition
is more accurate than the original function, which goes in line with
the whole PrimTorch effort.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87203
Approved by: https://github.com/mruberry
Fixes: https://github.com/pytorch/pytorch/issues/88010
This PR does a couple things to stop slow gradcheck from timing out:
- Splits out test_ops_fwd_gradients from test_ops_gradients, and factors out TestFwdGradients and TestBwdGradients which both inherit from TestGradients, now situated in common_utils (maybe there is a better place?)
- Skips CompositeCompliance (and several other test files) for slow gradcheck CI since they do not use gradcheck
- because test times for test_ops_fwd_gradients and test_ops_gradients are either unknown or wrong, we hardcode them for now to prevent them from being put together. We can undo the hack after we see actual test times are updated. ("def calculate_shards" randomly divides tests with unknown test times in a round-robin fashion.)
- Updates references to test_ops_gradients and TestGradients
- Test files that are skipped for slow gradcheck CI are now centrally located in in run_tests.py, this reduces how fine-grained we can be with the skips, so for some skips (one so far) we still use the old skipping mechanism, e.g. for test_mps
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88216
Approved by: https://github.com/albanD
`diag` was unnecessarily implemented as a kernel rather than as a composite
function, which made it unnecessarily difficult (explicit backward + all it entails).
We also change a few uses of `diag` on 2D tensors for `diagonal()`. The
latter returns a view rather than creating a new tensor.
We also upgrade its meta implementation to a fully-fledged
decomposition
I tried implementing the backwards of `diagonal()` via `diag_scatter` (or better `diag_scatter_` to keep the perf) but functionalisation was failing and I was not sure how to fix this, so I moved on. It may be possible to simplify that one as well if @soulitzer or someone knows how to do this.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87180
Approved by: https://github.com/ngimel, https://github.com/albanD, https://github.com/mruberry
`embedding_backward`'s decomposition is less accurate for bf16.
Currently bfloat16 is skipped in both forward and backward, but the
forward decomposition matches 1-1 with the ATen implementation so this
re-enables the test for the forwards decomposition.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84554
Approved by: https://github.com/albanD
Based on @ezyang's suggestion, mode stack now has "one true mode" which is the _only_ mode that can ever be active at the C++ level. That mode's torch dispatch is just to take the top mode in the stack, reenable itself (if we aren't at the end of the mode stack), and run the top mode's torch_{dispatch|function}
This maintains that in the middle of a mode's torch dispatch, the mode itself will not be active. It changes the function the user has to call to see what the current mode is (no longer queries the C++, it's python only) but allows the user to also see the entire mode stack easily
Removes `enable_torch_dispatch_mode` and `.restore()` since neither makes sense in this new setup
### Background
Why do we want this? Well, a pretty common pattern that was coming up was that users had to do something like
```python
## PRE-PR UX
def f(mode):
with mode.restore(): # user needs to understand this restore thing?
...
with Mode() as m:
pass
f(m)
```
Many users were getting error from forgetting to call `.restore` or from forgetting to add the (tbh weird) "mode instantiation" step where they use the mode as a context manager with an empty body. Really, they wanted to treat modes like context managers and just write
```python
## FROM FEEDBACK, USER DESIRED CODE. POSSIBLE POST-PR
def f(mode):
with mode:
...
f(Mode())
```
** Technical Details **
With the old mode stack, we basically had a linked list so the mode itself could only be used once and had a fixed parent. In this new design, the mode stack is just a python list that we're pushing to and popping from. There's only one mode that's ever active at the C++ level and it runs the next mode in the Python list. The modes don't have state on them anymore
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84774
Approved by: https://github.com/ezyang, https://github.com/zou3519
FYI, this decomposition seems to be significantly slower than the lowering in torchinductor:
```
------------------------------------- upsample_bicubic2d -------------------------------------]
| lowering | Inductor | Eager
32 threads: ------------------------------------------------------------------------------------
(torch.Size([16, 4, 128, 256]),), ((512, 1024), True) | 1.8 | 3.880 | 1.4
(torch.Size([16, 4, 128, 256]),), ((512, 1024), False) | 1.9 | 3.887 | 1.4
```
This seems related to the fact that in the lowering we can use int32s as the indices and in the decomp we can only use int64s (see https://github.com/pytorch/torchdynamo/issues/1293).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85403
Approved by: https://github.com/ngimel
Not sure why this check was necessary? Tests seem to run fine without
it.
There were definitely tests this was skipping before that it shouldn't,
e.g., pretty much all of the tests for `torch.nn.functional.interpolate`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85402
Approved by: https://github.com/ezyang
`torch.norm` is very odd. Some notable issues are:
- The default value of `"fro"` in `torch.norm` has an odd behaviour when `dim=None`. This is handled in the new dispatch
- The treatment of the `dtype` argument in `torch.norm` was completely wrong. This should fix it
- Some `out=` variants in the previous implementation were also wrong. This should fix those.
- This new dispatch should make some paths much faster. For example, `torch.norm(x)` where `x` is complex.
I'll try to make the changes in these PRs as incremental as possible as this is a tricky one.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81761
Approved by: https://github.com/ngimel
Also Back out "Revert D39075159: [acc_tensor] Use SymIntArrayRef for overloaded empty.memory_format's signature"
Original commit changeset: dab4a9dba4fa
Original commit changeset: dcaf16c037a9
Original Phabricator Diff: D38984222
Original Phabricator Diff: D39075159
Also update Metal registrations for C++ registration changes.
Also update NNPI registration to account for tightened schema checking
Differential Revision: [D39084762](https://our.internmc.facebook.com/intern/diff/D39084762/)
**NOTE FOR REVIEWERS**: This PR has internal Facebook specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D39084762/)!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84173
Approved by: https://github.com/Krovatkin
This PR adds nvfuser-specific primitive - `var_mean`.
Interpretation `torch.var_mean` -> `torch.ops.nvprims.var_mean` is handled by `TorchRefsNvfuserCapabilityMode` context manager.
I moved some helper code from `_prims/__init__.py` to `_prims_common`. Correctness is tested with OpInfo tests (see `PythonRefInfo("ops.nvprims.var_mean"`).
Layer norm reference now uses `torch.var_mean` instead of `torch._refs.var_mean` to allow interception. Here's a simple comparison of performance with this PR and master (on 3080ti):
```py
import torch
from torch._prims.context import TorchRefsNvfuserCapabilityMode
from torch.fx.experimental.proxy_tensor import make_fx
from torch._prims.executor import execute
def func(a):
return torch.native_layer_norm(a, (1024,), None, None, 1e-6)
a = torch.randn(10, 512, 1024, dtype=torch.float16, device="cuda")
with TorchRefsNvfuserCapabilityMode():
gm = make_fx(func)(a)
for _ in range(10):
execute(gm, a, executor="strictly_nvfuser");
```
run with `PYTORCH_NVFUSER_DUMP=dump_eff_bandwidth python script.py`
```py
# WITH THIS PR
# kernel1 run in 0.032768 ms, achieved: 641.25 GB/s
# kernel1 run in 0.033792 ms, achieved: 621.818 GB/s
# kernel1 run in 0.032768 ms, achieved: 641.25 GB/s
# kernel1 run in 0.032608 ms, achieved: 644.396 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# kernel1 run in 0.032768 ms, achieved: 641.25 GB/s
# kernel1 run in 0.03072 ms, achieved: 684 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# ON MASTER
# kernel1 run in 0.05632 ms, achieved: 373.091 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.043808 ms, achieved: 479.649 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
```
So this PR gives about 35% improvement in performance using nvfuser executor with this specific normalized shape.
Also this PR fixes https://github.com/pytorch/pytorch/issues/83506 (see the change in `torch/csrc/jit/python/pybind_utils.cpp`).
Ref. https://github.com/pytorch/pytorch/issues/80187
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83508
Approved by: https://github.com/ngimel
This PR adds nvfuser-specific primitive - `var_mean`.
Interpretation `torch.var_mean` -> `torch.ops.nvprims.var_mean` is handled by `TorchRefsNvfuserCapabilityMode` context manager.
I moved some helper code from `_prims/__init__.py` to `_prims_common`. Correctness is tested with OpInfo tests (see `PythonRefInfo("ops.nvprims.var_mean"`).
Layer norm reference now uses `torch.var_mean` instead of `torch._refs.var_mean` to allow interception. Here's a simple comparison of performance with this PR and master (on 3080ti):
```py
import torch
from torch._prims.context import TorchRefsNvfuserCapabilityMode
from torch.fx.experimental.proxy_tensor import make_fx
from torch._prims.executor import execute
def func(a):
return torch.native_layer_norm(a, (1024,), None, None, 1e-6)
a = torch.randn(10, 512, 1024, dtype=torch.float16, device="cuda")
with TorchRefsNvfuserCapabilityMode():
gm = make_fx(func)(a)
for _ in range(10):
execute(gm, a, executor="strictly_nvfuser");
```
run with `PYTORCH_NVFUSER_DUMP=dump_eff_bandwidth python script.py`
```py
# WITH THIS PR
# kernel1 run in 0.032768 ms, achieved: 641.25 GB/s
# kernel1 run in 0.033792 ms, achieved: 621.818 GB/s
# kernel1 run in 0.032768 ms, achieved: 641.25 GB/s
# kernel1 run in 0.032608 ms, achieved: 644.396 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# kernel1 run in 0.032768 ms, achieved: 641.25 GB/s
# kernel1 run in 0.03072 ms, achieved: 684 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# kernel1 run in 0.031744 ms, achieved: 661.935 GB/s
# ON MASTER
# kernel1 run in 0.05632 ms, achieved: 373.091 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.043808 ms, achieved: 479.649 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
# kernel1 run in 0.044032 ms, achieved: 477.209 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
# kernel1 run in 0.043008 ms, achieved: 488.571 GB/s
```
So this PR gives about 35% improvement in performance using nvfuser executor with this specific normalized shape.
Also this PR fixes https://github.com/pytorch/pytorch/issues/83506 (see the change in `torch/csrc/jit/python/pybind_utils.cpp`).
Ref. https://github.com/pytorch/pytorch/issues/80187
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83508
Approved by: https://github.com/ngimel
Previously, we introduced new SymInt overloads for every function we wanted. This led to a lot of boilerplate, and also a lot of confusion about how the overloads needed to be implemented.
This PR takes a simpler but more risky approach: just take the original function and changes its ints to SymInts.
This is BC-breaking in the following ways:
* The C++ API for registering implementations for aten operators will change from int64_t to SymInt whenever you make this change. Code generated registrations in PyTorch do not change as codegen handles the translation automatically, but manual registrations will need to follow the change. Typically, if you now accept a SymInt where you previously only took int64_t, you have to convert it back manually. This will definitely break XLA, see companion PR https://github.com/pytorch/xla/pull/3914 Note that not all dispatch keys get the automatic translation; all the composite keys and Meta keys are modified to take SymInt directly (because they should handle them directly), and so there are adjustments for this.
This is not BC-breaking in the following ways:
* The user facing C++ API remains compatible. Even if a function changes from int to SymInt, the default C++ binding still takes only ints. (e.g., at::empty(IntArrayRef, ...). To call with SymInts, you must call at::empty_symint instead. This involved adding two more signatures to CppSignatureGroup; in many cases I refactored code to iterate over all signatures in the group instead of hard-coding the two that previously existed.
* This is TorchScript compatible; internally we treat SymInts as ints so there is no change to what happens at runtime in TorchScript. In particular, it's OK to reference an empty schema by its old type (using int types), as long as you're not doing string equality (which you shouldn't be), these parse to the same underyling type.
Structure of the PR:
* The general strategy of this PR is that, even when you write `SymInt` inside `native_functions.yaml`, sometimes, we will treat it *as if* it were an `int`. This idea pervades the codegen changes, where we have a translation from SymInt to c10::SymInt or int64_t, and this is controlled by a symint kwarg which I added and then audited all call sites to decide which I wanted. Here are some of the major places where we pick one or the other:
* The C++ FunctionSchema representation represents `SymInt` as `int`. There are a few places we do need to know that we actually have a SymInt and we consult `real_type()` to get the real type in this case. In particular:
* When we do schema validation of C++ operator registration, we must compare against true schema (as the C++ API will provide `c10::SymInt`, and this will only be accepted if the schema is `SymInt`. This is handled with cloneWithRealTypes before we check for schema differences.
* In `toIValue` argument parsing, we parse against the true schema value. For backwards compatibility reasons, I do still accept ints in many places where Layout/SymInt/etc were expected. (Well, accepting int where SymInt is expected is not BC, it's just the right logic!)
* In particular, because SymInt never shows up as type() in FunctionSchema, this means that we no longer need a dedicated Tag::SymInt. This is good, because SymInts never show up in mobile anyway.
* Changes to functorch/aten are mostly about tracking changes to the C++ API registration convention. Additionally, since SymInt overloads no longer exist, registrations for SymInt implementations are deleted. In many cases, the old implementations did not properly support SymInts; I did not add any new functionality with this PR, but I did try to annotate with TODOs where this is work to do. Finally, because the signature of `native::` API changed from int to SymInt, I need to find alternative APIs for people who were directly calling these functions to call. Typically, I insert a new dispatch call when perf doesn't matter, or use `at::compositeexplicitautograd` namespace to handle other caes.
* The change to `make_boxed_from_unboxed_functor.h` is so that we accept a plain IntList IValue anywhere a SymIntList is expected; these are read-only arguments so covariant typing is OK.
* I change how unboxing logic works slightly. Previously, we interpret the C++ type for Layout/etc directly as IntType JIT type, which works well because the incoming IValue is tagged as an integer. Now, we interpret the C++ type for Layout as its true type, e.g., LayoutType (change to `jit_type.h`), but then we accept an int IValue for it anyway. This makes it symmetric with SymInt, where we interpret the C++ type as SymIntType, and then accept SymInt and int IValues for it.
* I renamed the `empty.names` overload to `empty_names` to make it less confusing (I kept mixing it up with the real empty overload)
* I deleted the `empty.SymInt` overload, which ended up killing a pile of functions. (This was originally a separate PR but the profiler expect test was giving me grief so I folded it in.)
* I deleted the LazyDynamicOpsTest tests. These were failing after these changes, and I couldn't figure out why they used to be passing: they make use of `narrow_copy` which didn't actually support SymInts; they were immediately converted to ints.
* I bashed LTC into working. The patches made here are not the end of the story. The big problem is that SymInt translates into Value, but what if you have a list of SymInt? This cannot be conveniently represented in the IR today, since variadic Values are not supported. To work around this, I translate SymInt[] into plain int[] (this is fine for tests because LTC dynamic shapes never actually worked); but this will need to be fixed for proper LTC SymInt support. The LTC codegen also looked somewhat questionable; I added comments based on my code reading.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83628
Approved by: https://github.com/albanD, https://github.com/bdhirsh
