Update conv1d meta kernel to match eager

[jansel]

Stack from ghstack (oldest at bottom):

• -> Update conv1d meta kernel to match eager #163584
• Fix prims.broadcast_in_dim functionalization #163377
• [inductor] Fix divmod error in decomp #163482
• [inductor] Fix bugs in emulate_precision_casts #163520
• [inductor] Fix issue with scalar arg handling #163481
• [inductor] Fix error from custom CUDA allocators #163422
• [dynamo] Fix TorchFunctionMode handling with get_rng_state #163412
• [inductor] Support out_dtype arg to matmul #163393
• [inductor] Freeze layouts in FlexAttention #163434
• [inductor] libdevice.sqrt => tl.sqrt_rn #163419

Fixes #163569

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/163584

• 📄 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

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

✅ No Failures

As of commit 5c8e8dd with merge base 51152ef ():
💚 Looks good so far! There are no failures yet. 💚

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

[thenumberouscode]

I actually have a previous PR (https://github.com/pytorch/pytorch/pull/160408/files) that implements a similar approach, which could both resolve #163569 and #159462.

No offense intended - this is purely a technical discussion. I believe we should align the conv1d meta kernel with the actual convolution kernel implementation. Specifically, we could utilize _select_conv_backend and _conv_determine_backend_memory_format within the conv function in fake_impls.py to determine the correct memory format. This would eliminate the need for the current workaround in pick_memory_format and provide a unified method for memory format selection.

convolution kernel implementation:

pytorch/aten/src/ATen/native/Convolution.cpp

Lines 1488 to 1503 in 8f30a8d

	// Expand 1d -> 2d.
	// This is only done for backends that don't natively support 1d spatial input.
	if (k == 3 && !input.is_mkldnn() && !input.is_xpu()) {
	// avoid accidentally going through NHWC for permuted 3d input.
	input = input.contiguous();
	params.view1d_as_2d();
	input = view4d(input);
	weight = view4d(weight);
	}
	// Select appropriate backend to use.
	auto bias_sizes_opt = bias.defined() ? std::optional<IntArrayRef>(bias.sizes()) : std::nullopt;
	bool need_backward = GradMode::is_enabled() &&
	(input.requires_grad() || weight.requires_grad() || (bias.defined() && bias.requires_grad()));
	ConvBackend backend = _select_conv_backend(input, weight, bias, c10::OptionalIntArrayRef(bias_sizes_opt), need_backward, params);
	at::MemoryFormat backend_memory_format = determine_backend_memory_format(input, weight, backend);

cc @eellison

[thenumberouscode]

Using only the weight memory format isn't enough to determine the correct format. Issue #159462 provides a test case that demonstrates this limitation.

[jansel]

@thenumberouscode let's go with your implementation then, thanks for catching this I wasn't aware of your PR