dynamo: fix the issue of aten.expand when the source and expaned size are all symbolic size #101173

XiaobingSuper · 2023-05-11T09:09:38Z

Stack from ghstack (oldest at bottom):

-> dynamo: fix the issue of aten.expand when the source and expaned size are all symbolic size #101173

For the dynamic shape of TIMM eca_halonext26ts model(python -m torch.backends.xeon.run_cpu --node_id 0 benchmarks/dynamo/timm_models.py --performance --float32 -dcpu --inference -n5 --inductor --dynamic-shapes --only eca_halonext26ts), there meets an error when calling aten.expand when src and the expanded size are all symbolic size:

Traceback (most recent call last):
  File "/home/xiaobing/pytorch-offical/torch/_dynamo/utils.py", line 1309, in run_node
    return node.target(*args, **kwargs)
  File "/home/xiaobing/pytorch-offical/torch/_decomp/decompositions.py", line 3270, in matmul
    tensor2_expanded = tensor2.expand(tensor2_expand_size).reshape(
  File "/home/xiaobing/pytorch-offical/torch/utils/_stats.py", line 20, in wrapper
    return fn(*args, **kwargs)
  File "/home/xiaobing/pytorch-offical/torch/_subclasses/fake_tensor.py", line 1105, in __torch_dispatch__
    return self.dispatch(func, types, args, kwargs)
  File "/home/xiaobing/pytorch-offical/torch/_subclasses/fake_tensor.py", line 1269, in dispatch
    return decomposition_table[func](*args, **kwargs)
  File "/home/xiaobing/pytorch-offical/torch/_refs/__init__.py", line 2799, in expand
    check(
  File "/home/xiaobing/pytorch-offical/torch/_prims_common/__init__.py", line 1648, in check
    raise exc_type(s())
RuntimeError: expand: attempting to expand a dimension of length 8*s0!

the src size is 8*s0, the expanded size is ((s0*(((s2 - 1)//16))**2 + 2*s0*(((s2 - 1)//16)) + s0)//(8*(((((s2 - 1)//16) + 1)//8))**2)).

This PR will try to fix it.

cc @ezyang

…ze are all symbolic size [ghstack-poisoned]

pytorch-bot · 2023-05-11T09:09:41Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/101173

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 14 New Failures

As of commit 7b95b3f:

NEW FAILURES - The following jobs have failed:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

…ze are all symbolic size ghstack-source-id: 848eb52 Pull Request resolved: #101173

XiaobingSuper · 2023-05-11T09:24:15Z

Note: a mini reproduced code didn't create even following the steps at https://github.com/pytorch/pytorch/blob/main/docs/source/compile/troubleshooting.rst.

XiaobingSuper · 2023-05-11T10:17:20Z

torch/_refs/__init__.py

@@ -2791,11 +2793,12 @@ def expand(a: Tensor, *shape) -> Tensor:
    for idx, x in enumerate(a.shape):
        offset_idx = idx + offset
        requested_length = shape[offset_idx]
+        requested_length = guard_int(requested_length)
+        x = guard_int(x)


I don't know if it is ok to add such a guard, for timm model, the root error is caused by https://github.com/pytorch/pytorch/blob/main/torch/_decomp/decompositions.py#L3266 which matmul calling expand.

No it's not ok unfortunately

Do you have any suggestions?

So, the error message looks legit: you can't expand a non-one dimension. But somehow, the dimension is 8*s0. You must figure out what caused the dimension to be this size; that is where the bug truly is.

XiaobingSuper · 2023-06-08T04:53:11Z

@ezyang , there has a smaller case to reproduce this issue:

import torch
import torch._dynamo
import torch.nn.functional as F

try:
    from torch import _assert
except ImportError:
    def _assert(condition: bool, message: str):
        assert condition, message

torch._dynamo.config.dynamic_shapes = True
torch._dynamo.config.assume_static_by_default = False

class HaloAttn(torch.nn.Module):
    def __init__(
            self, dim=256, qkv_bias=False):
        super().__init__()
        self.num_heads = 8
        self.dim_head_qk = 16
        self.dim_head_v = 32
        self.dim_out_qk = 128
        self.dim_out_v = 256
        self.scale = 0.25
        self.scale_pos_embed = False
        self.block_size = self.block_size_ds = 8
        self.halo_size = 2
        self.win_size = 12  # neighbourhood window size
        self.block_stride = 1

        self.q = torch.nn.Conv2d(dim, self.dim_out_qk, 1, stride=self.block_stride, bias=qkv_bias)
        self.kv = torch.nn.Conv2d(dim, self.dim_out_qk + self.dim_out_v, 1, bias=qkv_bias)

    def forward(self, x):
        B, C, H, W = x.shape
        _assert(H % self.block_size == 0, '')
        _assert(W % self.block_size == 0, '')

        num_h_blocks = H // self.block_size
        num_w_blocks = W // self.block_size
        num_blocks = num_h_blocks * num_w_blocks

        q = self.q(x)
        # unfold
        q = q.reshape(
            -1, self.dim_head_qk,
            num_h_blocks, self.block_size_ds, num_w_blocks, self.block_size_ds).permute(0, 1, 3, 5, 2, 4)
        # B, num_heads * dim_head * block_size ** 2, num_blocks
        q = q.reshape(B * self.num_heads, self.dim_head_qk, -1, num_blocks).transpose(1, 3)
        # B * num_heads, num_blocks, block_size ** 2, dim_head

        kv = self.kv(x)
        # Generate overlapping windows for kv. This approach is good for GPU and CPU. However, unfold() is not
        # lowered for PyTorch XLA so it will be very slow. See code at bottom of file for XLA friendly approach.
        # FIXME figure out how to switch impl between this and conv2d if XLA being used.
        kv = F.pad(kv, [self.halo_size, self.halo_size, self.halo_size, self.halo_size])
        kv = kv.unfold(2, self.win_size, self.block_size).unfold(3, self.win_size, self.block_size).reshape(
            B * self.num_heads, self.dim_head_qk + self.dim_head_v, num_blocks, -1).permute(0, 2, 3, 1)
        k, v = torch.split(kv, [self.dim_head_qk, self.dim_head_v], dim=-1)
        # B * num_heads, num_blocks, win_size ** 2, dim_head_qk or dim_head_v

        return q @ k.transpose(-1, -2)

x = torch.rand(128, 256, 16, 16)

model = HaloAttn().eval()
model = model.eval()
print(model)
opt_model = torch.compile(model, dynamic=True)
with torch.no_grad():
    for i in range(3):
        out = opt_model(x)

For this case, if I comment _assert(H % self.block_size == 0, '') and _assert(W % self.block_size == 0, ''), it can works.

XiaobingSuper · 2023-06-08T07:28:00Z

I checked that expand input sizes are : src[8*s0, ((s2//8))**2, 16, 144], dst[((s0*s2**2)//(8*((s2//8))**2)), ((s2//8))**2, 16, 144], the expand check it is true for 8*s0 == ((s0*s2**2)//(8*((s2//8))**2)) when removing assert code, but it is false when has assert code, it seems the guard has some issue.

ngimel · 2023-06-08T18:20:05Z

Likely we'd have to create a guard that s2 % 8 == 0 and, if we don't have the necessary optimization for sympy expr, an optimization pass that'll simplify ((s0*s2**2)//(8*((s2//8))**2)) to 8 * s0. The latter part probably works if there's a guard on s2

XiaobingSuper · 2023-06-09T03:12:50Z

Likely we'd have to create a guard that s2 % 8 == 0 and, if we don't have the necessary optimization for sympy expr, an optimization pass that'll simplify ((s0*s2**2)//(8*((s2//8))**2)) to 8 * s0. The latter part probably works if there's a guard on s2

Where should I add a guard for s2 % 8 == 0?

ezyang · 2023-06-10T19:43:50Z

@ysiraichi wondering if you could run TV on this and see if it's a sympy simplification bug

ysiraichi · 2023-06-13T19:29:41Z

I tried running it with TV, but it failed before actually the validation check.
Investigating it a bit further, I think the problem is somewhat related to a SymPy issue sympy/sympy#25146. Here's what I'm seeing:

With the asserts:

symbolic_shapes: [DEBUG] 0.0: eval Eq(((s0*s2**2)//(8*((s2//8))**2)), 8*s0) == False [statically known]

Without the asserts:

symbolic_shapes: [INFO] 0.0: eval Eq(8*s0, ((s0*s2**2)//(8*((s2//8))**2))) [guard added]

The Actual Problem

That expression goes through ShapeEnv.simplify and ends up being transformed into, since it was marked as "divisible" due to the assertion:

Eq(8*s0, ((s0*s2**2)//(8*((s2/8))**2)))

From there, SymPy kicks in and apply the following transformations:

>>> Eq(8*s0, ((s0*s2**2)//(8*((s2/8))**2)))
>>> Eq(8*s0, s0//Rational(1, 8))  # Correct.
>>> False  # Incorrect.

A Possible Solution

The easiest way I see in order to solve this problem is to special case it (Rational divisor) in FloorDiv.eval.
That said, maybe we should also check if one of the sides of the equality is real. If so, we multiply the other side by 1.0.

ezyang · 2023-06-13T21:04:03Z

Hmm, failing before the validation check sucks. Maybe we should have a way to trigger validation early... we ought to be able to attempt to produce_guards at any point during compilation.

ngimel · 2023-06-13T21:23:42Z

Special-casing Rational in FloorDiv (especially when Rational numerator divides FloorDiv numerator) sgtm, I'd be wary of actually multiplying by 1.0, there are very few cases we encounter non-integer expressions, and imo we should first try staying in the integer land.

ysiraichi · 2023-08-01T19:46:13Z

@XiaobingSuper may I take over this one?

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. [ghstack-poisoned]

…ng error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. ghstack-source-id: 5aa7711 Pull Request resolved: #106645

…tal.py. on "Handle `Rational` divisors in `FloorDiv`." Follow-up: #101173 This PR fixes the bug presented in #101173 by creating a special case for `sympy.Rational` divisors, inside `FloorDiv` evaluation. In summary: ```python FloorDiv(a, Rational(1, b)) a * b ``` Besides that, this PR also does 2 other things: - Replaces the use of the old `sympy.Mod` by the internal `Mod` (there were a few places that were still looking for the SymPy one) - Introduces debugging logs to the translation validator. These can be seen by setting the environment variable: `TORCH_LOGS=+torch.fx.experimental.validator` [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. ghstack-source-id: f0175c7 Pull Request resolved: #106645

…on on tracing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…ror." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…acing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. ghstack-source-id: b307a71 Pull Request resolved: #106645

ysiraichi · 2023-08-09T21:21:29Z

Solved at #106644

…validation on tracing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…acing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…ror." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. ghstack-source-id: 76f312c Pull Request resolved: #106645

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…cing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. ghstack-source-id: 2cf7caa Pull Request resolved: #106645

…on "Run translation validation on tracing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…ion validation on tracing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

… translation validation on tracing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…idation on tracing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

…acing error." This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy chenyang78 aakhundov [ghstack-poisoned]

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. ghstack-source-id: 5d0116b Pull Request resolved: #106645

This PR wraps `InstructionTranslator` run with a try-catch block so as to run the translation validation (TV) if it ends up raising an error. In this context, we run TV so as to catch simplification errors. These may turn `ShapeEnv.divisible` and `ShapeEnv.replacements` incorrect. For example: #101173 describes a SymPy simplification bug that doesn't reach TV, since it's run only in the end of the tracing. Pull Request resolved: #106645 Approved by: https://github.com/ezyang

Follow-up: pytorch#101173 This PR fixes the bug presented in pytorch#101173 by creating a special case for `sympy.Rational` divisors, inside `FloorDiv` evaluation. In summary: ```python FloorDiv(a, Rational(1, b)) a * b ``` Besides that, this PR also does 2 other things: - Replaces the use of the old `sympy.Mod` by the internal `Mod` (there were a few places that were still looking for the SymPy one) - Introduces debugging logs to the translation validator. These can be seen by setting the environment variable: `TORCH_LOGS=+torch.fx.experimental.validator` Pull Request resolved: pytorch#106644 Approved by: https://github.com/ezyang ghstack dependencies: pytorch#106643

inductor: fix the issue of aten.expand when the source and expaned si…

7b95b3f

…ze are all symbolic size [ghstack-poisoned]

XiaobingSuper added a commit that referenced this pull request May 11, 2023

inductor: fix the issue of aten.expand when the source and expaned si…

23666a5

…ze are all symbolic size ghstack-source-id: 848eb52 Pull Request resolved: #101173

pytorchbot added the open source label May 11, 2023

XiaobingSuper changed the title ~~inductor: fix the issue of aten.expand when the source and expaned size are all symbolic size~~ dynamo: fix the issue of aten.expand when the source and expaned size are all symbolic size May 11, 2023

XiaobingSuper added module: dynamic shapes release notes: dynamo labels May 11, 2023

XiaobingSuper requested review from ngimel and lezcano May 11, 2023 09:46

XiaobingSuper commented May 11, 2023

View reviewed changes

lezcano requested review from ezyang and removed request for lezcano May 11, 2023 10:20

XiaobingSuper marked this pull request as draft May 11, 2023 13:26

This was referenced Aug 4, 2023

Handle Rational divisors in FloorDiv. #106644

Closed

Run translation validation on tracing error. #106645

Closed

ysiraichi closed this Aug 9, 2023

facebook-github-bot deleted the gh/XiaobingSuper/114/head branch September 9, 2023 14:21

dynamo: fix the issue of aten.expand when the source and expaned size are all symbolic size #101173

dynamo: fix the issue of aten.expand when the source and expaned size are all symbolic size #101173

Uh oh!

Conversation

XiaobingSuper commented May 11, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented May 11, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/101173

❌ 14 New Failures

Uh oh!

XiaobingSuper commented May 11, 2023

Uh oh!

XiaobingSuper May 11, 2023

Choose a reason for hiding this comment

Uh oh!

ezyang May 11, 2023

Choose a reason for hiding this comment

Uh oh!

XiaobingSuper May 11, 2023

Choose a reason for hiding this comment

Uh oh!

ezyang May 11, 2023

Choose a reason for hiding this comment

Uh oh!

XiaobingSuper commented Jun 8, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

XiaobingSuper commented Jun 8, 2023

Uh oh!

ngimel commented Jun 8, 2023

Uh oh!

XiaobingSuper commented Jun 9, 2023

Uh oh!

ezyang commented Jun 10, 2023

Uh oh!

ysiraichi commented Jun 13, 2023

The Actual Problem

A Possible Solution

Uh oh!

ezyang commented Jun 13, 2023

Uh oh!

ngimel commented Jun 13, 2023

Uh oh!

ysiraichi commented Aug 1, 2023

Uh oh!

ysiraichi commented Aug 9, 2023

Uh oh!

Uh oh!

XiaobingSuper commented May 11, 2023 •

edited

Loading

pytorch-bot bot commented May 11, 2023 •

edited

Loading

XiaobingSuper commented Jun 8, 2023 •

edited

Loading