This partially solve the issue https://github.com/pytorch/pytorch/issues/163641. We do not need to ban unbacked to unbacked replacement if all rhs symbols are inputs since we know those symbols are seen by the whole program.
This issue was found as i was tracing some vllm models with unbacked, namely Qwen/Qwen2-1.5B-Instruct it makes reasoning logic easier to do those replacements.
as for data dependent similar pattern, I am thinking to create a set of replacements that we apply only during static eval
instead of none. to make reasoning better.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163652
Approved by: https://github.com/bobrenjc93
Summary:
I am really skeptical about inductor sizevars creating an empty shape env when not provided with one
i think we should fail there if the graph has dynamic shapes and no shape env is provided.
however i wonder if there are actually use cases that depends on the shape env not being there?
Reasoning APIs depends on facts in the shape env. and assumes some stuff exists for specific symbols.
Test Plan:
Fix the bug reported in creating simple e2e unit test is not trivial
https://www.internalfb.com/diff/D82337184
Rollback Plan:
Differential Revision: D82412384
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162927
Approved by: https://github.com/ezyang, https://github.com/eellison, https://github.com/jansel
Summary:
When rewriting sympy expressions in the compiler codebase we want to generate
FloorDiv(a, b) CleanDiv(a, b) directly and not a//b. since the later become floor(a*pow(b, -1))
For symnodes we automatically handle that conversions in the symnode op dispatch.
I will follow up with an issue to track all other usages of //.
Block internal Model.
Test Plan:
add test
run existing tests.
dakechen1993 testing on the model.
Rollback Plan:
Differential Revision: D82362241
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162869
Approved by: https://github.com/ezyang
Summary:
Sometimes `ShapeEnv.create_symbol` can return a `sympy.Integer`. This messes up our phantom symbol infra for derived dims.
Fixes#161902
Test Plan:
added test based on repro
Rollback Plan:
Differential Revision: D81960709
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162416
Approved by: https://github.com/tugsbayasgalan
Summary:
[reland]
Since `allow_complex_guards_as_runtime_asserts` is now sync'd with `prefer_deferred_runtime_asserts_over_guards`, we can kill the former (especially since it was a export-only concept).
Test Plan:
updated tests
Rollback Plan:
Differential Revision: D81334984
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161794
Approved by: https://github.com/zhxchen17
Summary: Since `allow_complex_guards_as_runtime_asserts` is now sync'd with `prefer_deferred_runtime_asserts_over_guards`, we can kill the former (especially since it was a export-only concept).
Test Plan:
updated tests
Rollback Plan:
Differential Revision: D79903317
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160198
Approved by: https://github.com/ezyang
Summary: Since `allow_complex_guards_as_runtime_asserts` is now sync'd with `prefer_deferred_runtime_asserts_over_guards`, we can kill the former (especially since it was a export-only concept).
Test Plan:
updated tests
Rollback Plan:
Differential Revision: D79903317
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160198
Approved by: https://github.com/ezyang
The motivation for this change can be seen through the following example:
```
import torch
GPU_TYPE = "cuda"
@torch.compile
def no_override(x):
return x.sum(dim=0)
@torch.compile
def override(x):
return x.sum(dim=0)
x_small = torch.randn(4096, 512, device=GPU_TYPE)
no_override(x_small)
torch._dynamo.decorators.mark_dynamic(x_small, 0, hint_override=4096 * 1000)
override(x_small)
```
Previously, when reductions were split, codegen relied only on the first observed shape. With a small input, this resulted in a small split size:
```
def triton_red_fused_sum_0(in_ptr0, out_ptr0, ks0, xnumel, r0_numel, XBLOCK : tl.constexpr, R0_BLOCK : tl.constexpr):
xnumel = 16384
rnumel = r0_numel
```
With the new scheme, inductor honors hint_override during codegen, producing larger and more appropriate split sizes:
```
def triton_red_fused_sum_0(in_ptr0, out_ptr0, ks0, xnumel, r0_numel, XBLOCK : tl.constexpr, R0_BLOCK : tl.constexpr):
xnumel = 1024000
rnumel = r0_numel
```
This addresses a broader problem with dynamism: performance and numerics previously depended on whichever shape was seen first. For example:
```
f(s0) -> f(s2)
f(s1) -> f(s2)
```
could generate different kernels. With the new approach, an explicit override pins the chosen configuration:
```
f(s0, hint_override=s0) -> f(s2)
f(s1, hint_override=s0) -> f(s2)
```
ensuring consistent kernel generation regardless of input order.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161007
Approved by: https://github.com/jansel
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
The motivation for this change can be seen through the following example:
```
import torch
GPU_TYPE = "cuda"
@torch.compile
def no_override(x):
return x.sum(dim=0)
@torch.compile
def override(x):
return x.sum(dim=0)
x_small = torch.randn(4096, 512, device=GPU_TYPE)
no_override(x_small)
torch._dynamo.decorators.mark_dynamic(x_small, 0, hint_override=4096 * 1000)
override(x_small)
```
Previously, when reductions were split, codegen relied only on the first observed shape. With a small input, this resulted in a small split size:
```
def triton_per_fused_sum_1(in_ptr0, out_ptr0, xnumel, r0_numel, XBLOCK : tl.constexpr):
xnumel = 512
r0_numel = 32
```
With the new scheme, inductor honors hint_override during codegen, producing larger and more appropriate split sizes:
```
def triton_red_fused_sum_0(in_ptr0, out_ptr0, xnumel, r0_numel, XBLOCK : tl.constexpr, R0_BLOCK : tl.constexpr):
xnumel = 16384
r0_numel = 128
```
This addresses a broader problem with dynamism: performance and numerics previously depended on whichever shape was seen first. For example:
```
f(s0) -> f(s2)
f(s1) -> f(s2)
```
could generate different kernels. With the new approach, an explicit override pins the chosen configuration:
```
f(s0, hint_override=s0) -> f(s2)
f(s1, hint_override=s0) -> f(s2)
```
ensuring consistent kernel generation regardless of input order.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161007
Approved by: https://github.com/jansel
When select has data dependent input, we cant tell if the actual index shall be index+size or index.
to avoid throwing dde, we allocate a new unbacked symbol to represent the storage offset of the
output view and we compute its value dynamically at runtime when inductor is lowered.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157605
Approved by: https://github.com/ColinPeppler
As part of better engineering week, we would like to improve out type support to improve dev experience in dynamo
This PR adds strict typing support to a critical set of files for dynamo, `source.py` and the base `_guards.py`
Running
```
mypy torch/_dynamo/source.py torch/_guards.py --linecount-report /tmp/coverage_log
```
| -------- | Lines Unannotated | Lines Total | % lines covered | Funcs Unannotated | Funcs Total | % funcs covered |
| -------- | ------- | -------- | ------- | ------- | ------- | ------- |
| Main | 1227 | 2208 | 55.57% | 207 | 362 | 57.18% |
| This PR | 2217 | 2217 | 100.00% | 362 | 362 | 100.00% |
| Delta | +990 | +9 | +44.43% | +155 | 0 | +42.82% |
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158397
Approved by: https://github.com/anijain2305
When select has data dependent input, we cant tell if the actual index shall be index+size or index.
to avoid throwing dde, we allocate a new unbacked symbol to represent the storage offset of the
output view and we compute its value dynamically at runtime when inductor is lowered.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157605
Approved by: https://github.com/ColinPeppler
Creating contiguous strides creates an expression max(1, x). Often we know that x >= 1, in
which case we should simplify max(1, x) to x.
This appeared in two situations:
1) An internal user complained about statically_known_true(x == max(1, x)) failing (internal link: https://fb.workplace.com/groups/1028545332188949/permalink/1232958568414290).
This https://github.com/pytorch/pytorch/pull/155938 won't be needed with this.
3) Not simplifying the above could result in wrong ConstraintViolationErrors.
Because we assume non-trival single arg guards shall evaporate see the logic in the function
issue_guard in symbolic_shapes.py
with this change we longer throw ConstraintViolationErrors with the program bellow
this is blocking landing this [PR](https://github.com/pytorch/pytorch/pull/155590) from landing
internally. Due to internal export tests throwing ConstraintViolationErrors.
like
```
Constraints violated (width)!
- Not all values of width = L['x'].size()[3] in the specified range 224 <= width <= 455 satisfy the generated guard max(1, 1 + (((-1) + L['x'].size()[3]) // 2)) == (1 + (((-1) + L['x'].size()[3]) // 2)).
````
```
x = torch.rand(10)
torch._dynamo.mark_dynamic(x, 0, max=20, min=5)
@torch.compile(fullgraph=True, dynamic=True)
def func(x):
if max(1, (-1 + x.size()[0]//2)) == (-1+x.size()[0]//2):
return x*400
else:
return (x*10)*100
func(x)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157189
Approved by: https://github.com/pianpwk
Differential Revision: D77249427
Due to memoization and graph order update, it can happen that a backed symbol is passed into compute_unbacked_bindings and lead to failure. An example as follow:
- There are 2 boolean indexing operators (e.g. op1 and op2) with the same mask.
- A unbacked symint is generated from op1, and then op2 reuses the unbacked symint due to a nonzero_memo in nonzero's fake implementation and no rebinding is needed for op2.
- Since op1 generated the unbacked symint, its meta has "unbacked_bindings" field filled and op2's meta doesn't have it.
- Output from op1 and op2 are later concated with others with backed symint, so that the unbacked symint can be replaced by a backed symint.
- In Inductor, during fake tensor prop, there is no memoi because new fake tensor is always generated (for the same node). op1 generates an unbacked symint and the unbacked can be rebound successfully to the backed symint. Since there is no memoi, op2 also generates a new unbacked symint, but no rebinding can happen because op2's meta doesn't have "unbacked_bindings". And "compute_unbacked_bindings/_rename_unbacked_to" fails to assert op2's old symbol to be unbacked.
From discussion with [@ezyang](https://www.internalfb.com/intern/profile/?id=503862770), there is no easy way to fix this issue.
- We can try to enable memoization for fake tensor prop in Inductor, however, we need to ensure that op1 is visited before op2 during Inductor fake tensor prop for this to work (op2's meta doesn't have "unbacked_bindings" so no rebinding can happen and we need to do rebinding from op1. But there are passes such as reorder_for_locality that can change the graph order so this doesn't work.
- A simple hack is to just replace the unbacked symbol in op2 by the backed symbol.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156911
Approved by: https://github.com/ezyang
