From fe88faaf9cd35cb1f41024a7903e0cf3ae37bce0 Mon Sep 17 00:00:00 2001 From: "Xia, Weiwen" Date: Tue, 8 Jul 2025 16:29:56 +0000 Subject: [PATCH 1/2] Update 2.8.0/todo/result_quantization.md --- 2.8.0/todo/result_quantization.md | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/2.8.0/todo/result_quantization.md b/2.8.0/todo/result_quantization.md index cdba137..a66484e 100644 --- a/2.8.0/todo/result_quantization.md +++ b/2.8.0/todo/result_quantization.md @@ -27,6 +27,11 @@ The categories below are as follows: ### bc breaking ### deprecation ### new features +- [Quant][PT2E] add a lowering pass for x86 backend ([#149708](https://github.com/pytorch/pytorch/pull/149708)) +- [Quant][PT2E][X86] enable qconv1d-relu fusion ([#150751](https://github.com/pytorch/pytorch/pull/150751)) +- [Quant][X86] add an op to compute uint8 pointwise mul ([#151112](https://github.com/pytorch/pytorch/pull/151112)) +- [Quant][X86] add ops to compute uint8 pointwise add/add_relu ([#152411](https://github.com/pytorch/pytorch/pull/152411)) +- [Quant][X86] add an op to compute uint8 batch norm 2d ([#152811](https://github.com/pytorch/pytorch/pull/152811)) ### improvements ### bug fixes ### performance @@ -41,24 +46,19 @@ The categories below are as follows: - add `torch.float4_e2m1fn_x2` to PyTorch ([#148791](https://github.com/pytorch/pytorch/pull/148791)) - Improve attr mismatch msg ([#149576](https://github.com/pytorch/pytorch/pull/149576)) - [Intel GPU] Allow XPU backend in Quantize operators ([#150288](https://github.com/pytorch/pytorch/pull/150288)) -- [Quant][PT2E] add a lowering pass for x86 backend ([#149708](https://github.com/pytorch/pytorch/pull/149708)) - [AO] update port_metadata_pass to support quant_affine ops ([#150642](https://github.com/pytorch/pytorch/pull/150642)) - [AO] Add Moving Average Affine Observer ([#150643](https://github.com/pytorch/pytorch/pull/150643)) - [AO] Refactor convert and add QuantAffinePlaceholderObserver ([#150644](https://github.com/pytorch/pytorch/pull/150644)) -- [Quant][PT2E][X86] enable qconv1d-relu fusion ([#150751](https://github.com/pytorch/pytorch/pull/150751)) - [AO] fix per token block size calculation ([#150890](https://github.com/pytorch/pytorch/pull/150890)) - Fix torchscript issues with reference quantized modules ([#150870](https://github.com/pytorch/pytorch/pull/150870)) - [BE][1/2] Move original_weights_lookup attribute to constant ([#151241](https://github.com/pytorch/pytorch/pull/151241)) - [Sana][HybridCache] Fix bug in detect_attr_assignment ([#151824](https://github.com/pytorch/pytorch/pull/151824)) - [standalone_compile] Dynamic shape handling ([#151788](https://github.com/pytorch/pytorch/pull/151788)) -- [Quant][X86] add an op to compute uint8 pointwise mul ([#151112](https://github.com/pytorch/pytorch/pull/151112)) - [1/N] Deprecate c10::string_view and at::string ([#151972](https://github.com/pytorch/pytorch/pull/151972)) - [fbgemm] Implement __obj_flatten__ for LinearPackedParamsBase ([#152619](https://github.com/pytorch/pytorch/pull/152619)) - Avoid std::chrono::system_clock ([#153135](https://github.com/pytorch/pytorch/pull/153135)) - [torch][ao] Properly strip tracking stats in _fold_conv_bn_qat for 1D ([#152982](https://github.com/pytorch/pytorch/pull/152982)) - [Refactor] Explicilty spell out the namespace for device() function ([#153248](https://github.com/pytorch/pytorch/pull/153248)) -- [Quant][X86] add ops to compute uint8 pointwise add/add_relu ([#152411](https://github.com/pytorch/pytorch/pull/152411)) -- [Quant][X86] add an op to compute uint8 batch norm 2d ([#152811](https://github.com/pytorch/pytorch/pull/152811)) - Add deprecation warning for `torch.ao.quantization` ([#153892](https://github.com/pytorch/pytorch/pull/153892)) - BE: Type previously untyped decorators ([#154515](https://github.com/pytorch/pytorch/pull/154515)) - [Reland][pytorch] Patch the _is_conv_node function ([#154473](https://github.com/pytorch/pytorch/pull/154473)) From 2782a38704ba9edfcb372dd2ef25c9993120f859 Mon Sep 17 00:00:00 2001 From: "Xia, Weiwen" Date: Thu, 10 Jul 2025 09:19:37 +0000 Subject: [PATCH 2/2] Update 2.8.0/todo/result_quantization.md --- 2.8.0/todo/result_quantization.md | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/2.8.0/todo/result_quantization.md b/2.8.0/todo/result_quantization.md index a66484e..c4c5f92 100644 --- a/2.8.0/todo/result_quantization.md +++ b/2.8.0/todo/result_quantization.md @@ -27,11 +27,11 @@ The categories below are as follows: ### bc breaking ### deprecation ### new features -- [Quant][PT2E] add a lowering pass for x86 backend ([#149708](https://github.com/pytorch/pytorch/pull/149708)) -- [Quant][PT2E][X86] enable qconv1d-relu fusion ([#150751](https://github.com/pytorch/pytorch/pull/150751)) -- [Quant][X86] add an op to compute uint8 pointwise mul ([#151112](https://github.com/pytorch/pytorch/pull/151112)) -- [Quant][X86] add ops to compute uint8 pointwise add/add_relu ([#152411](https://github.com/pytorch/pytorch/pull/152411)) -- [Quant][X86] add an op to compute uint8 batch norm 2d ([#152811](https://github.com/pytorch/pytorch/pull/152811)) +- Add a lowering pass for x86 backend of PT2E quantization ([#149708](https://github.com/pytorch/pytorch/pull/149708)) +- Enable qconv1d-relu fusion for PT2E quantization on X86 CPU ([#150751](https://github.com/pytorch/pytorch/pull/150751)) +- Add an op to compute uint8 pointwise mul for PT2E quantization on X86 CPU ([#151112](https://github.com/pytorch/pytorch/pull/151112)) +- Add ops to compute uint8 pointwise add/add_relu for PT2E quantization on X86 CPU ([#152411](https://github.com/pytorch/pytorch/pull/152411)) +- Add an op to compute uint8 batch norm 2d for PT2E quantization on X86 CPU ([#152811](https://github.com/pytorch/pytorch/pull/152811)) ### improvements ### bug fixes ### performance