Add torch.uint16, torch.uint32

[Freed-Wu]

• [TP] reorder MXFP8 wrapper over DTensor ([TP] reorder MXFP8 wrapper over DTensor #4010)
• Patch nightly Linux AArch64-specific wheel build (Patch nightly Linux AArch64-specific wheel build #3954)
• [ROCm] Fix ROCm CI failures ([ROCm] Fix ROCm CI failures #4061)
• Add CuTeDSL kernel for 3D tensor quantization to MXFP8 (Add CuTeDSL kernel for 3D tensor quantization to MXFP8 #4090)
• [mxfp8] fix exp2f_rcp to handle special case of unbiased exponent of 127 ([mxfp8] fix exp2f_rcp to handle special case of unbiased exponent of 127 #4117)
• add bandwidth benchmarking script for fp8 quant blockwise (add bandwidth benchmarking script for fp8 quant blockwise #4128)
• Delete autoquant_v2 and subgraph_utils (Delete autoquant_v2 and subgraph_utils #4121)
• Delete deprecated autoquant v1 and all references (Delete deprecated autoquant v1 and all references #4122)
• Delete deprecated CutlassSemiSparseLayout and related code (Delete deprecated CutlassSemiSparseLayout and related code #4126)
• Upgrade torchao build configs to C++20 (Upgrade torchao build configs to C++20 (#4025) #4025)
• [mxfp8 moe training] remove unnecessary memory fences in cutedsl kernel; unify redundant conditionals ([mxfp8 moe training] remove unnecessary memory fences in cutedsl kernel; unify redundant conditionals #4129)
• Delete deprecated Float8Layout and Float8AQTTensorImpl (Delete deprecated Float8Layout and Float8AQTTensorImpl #4127)
• [moe training] Optimize FP8 MoE backward pass: fused colwise kernel + AMD tuning ([moe training] Optimize FP8 MoE backward pass: fused colwise kernel + AMD tuning #4069)
• [claude] Make MXFP8 cuda kernels ABI stable ([claude] Make MXFP8 cuda kernels ABI stable #3610)
• Add cross attention support for low precision attention API (Add cross attention support for low precision attention API #4142)
• [mxfp8 training] require torch nightly and cuda 12.8+ ([mxfp8 training] require torch nightly and cuda 12.8+ #4141)
• Delete deprecated BlockSparseLayout and related code (Delete deprecated BlockSparseLayout and related code #4143)
• Delete deprecated GemlitePackedLayout and GemliteUIntXWeightOnlyConfig (Delete deprecated GemlitePackedLayout and GemliteUIntXWeightOnlyConfig #4144)
• [X86] Re-enable DA8W4 path on X86 CPU ([X86] Re-enable DA8W4 path on X86 CPU #4033)
• Remove version compatibility table (Remove version compatibility table #4154)
• Suppress deprecation warnings from Deprecate AQT and related classes #4074 and Deprecate TorchAODType #4100 (Suppress deprecation warnings from #4074 and #4100 #4155)
• Delete deprecated UintxLayout and related code (Delete deprecated UintxLayout and related code #4145)
• Delete deprecated QDQLayout and related code (Delete deprecated QDQLayout and related code #4146)
• Delete deprecated Int4XPULayout and related code (Delete deprecated Int4XPULayout and related code #4147)
• Delete deprecated SemiSparseLayout and related code (Delete deprecated SemiSparseLayout and related code #4148)
• Delete deprecated Int4CPULayout and related code (Delete deprecated Int4CPULayout and related code #4149)
• Delete deprecated PackedLinearInt8DynamicActivationIntxWeightLayout and related code (Delete deprecated PackedLinearInt8DynamicActivationIntxWeightLayout and related code #4150)
• Fix test_sparse_api failures for builds without hipSPARSELt (Fix test_sparse_api failures for builds without hipSPARSELt (#4125) #4125) (Fix test_sparse_api failures for builds without hipSPARSELt (#4125) #4125)
• [reland][xpu] INT8 quantization on Intel XPU ([reland][xpu] INT8 quantization on Intel XPU #3782)
• Skip test_fsdp2 if PyTorch version is 2.11.0 or higher (Skip test_fsdp2 if PyTorch version is 2.11.0 or higher #4168)
• Move bitpacking.py to prototype and add uintx_utils.py (Move bitpacking.py to prototype and add uintx_utils.py #4152)
• Fix rocm CI (Fix rocm CI #4167)
• [xpu][test] Skip WIP config for Intel GPU in test_safetensors_support.py and test_x86inductor_fusion.py ([xpu][test] Skip WIP config for Intel GPU in test_safetensors_support.py and test_x86inductor_fusion.py #4049)
• Remove tensor parallel test for v1 of Int8DynamicActivationInt8WeightConfig (Remove tensor parallel test for v1 of Int8DynamicActivationInt8WeightConfig #4169)
• clean up unused rocm references in test_training.py (clean up unused rocm references in test_training.py #4170)
• [mxfp8 training] add cutedsl kernel for mxfp8 quantation along dim0 ([mxfp8 training] add cutedsl kernel for mxfp8 quantation along dim0 #4156)
• Add PerGroup granularity for Float8WeightOnly config (Add PerGroup granularity for Float8WeightOnly config #4174)
• Add embedding op support for Float8Tensor with PerGroup quantization (Add embedding op support for Float8Tensor with PerGroup quantization #4175)
• remove per kernel fp8 triton quantized benchmarks (remove per kernel fp8 triton quantized benchmarks #4166)
• Add WAN style rope detection pattern for low precision attention API (Add WAN style rope detection pattern for low precision attention API #4173)
• [mxfp8 moe training] refactor autograd func ([mxfp8 moe training] refactor autograd func #4176)
• [mxfp8 moe training] remove unused block_size arg ([mxfp8 moe training] remove unused block_size arg #4177)
• Add better debug print for failed prepare (Add better debug print for failed prepare #3950)
• fix undefined values for tail elements in act quant kernels (fix undefined values for tail elements in act quant kernels #4186)
• [mxfp8 moe training] register sharding rules for cutedsl 2d quant kernel ([mxfp8 moe training] register sharding rules for cutedsl 2d quant kernel #4178)
• [CPU] Add software prefetch to overlap bandwidth for scaled_embedding_bag ([CPU] Add software prefetch to overlap bandwidth for scaled_embedding_bag #4171)
• [X86] intmm: Use u8s8 when only support avx512-vnni ([X86] intmm: Use u8s8 when only support avx512-vnni #4103)
• Add CodeML'25 BibTeX (Add CodeML'25 BibTeX #3978)
• add rope fusion detection tests (add rope fusion detection tests #4183)
• Bump version from 0.17.0 to 0.18.0 (Bump version from 0.17.0 to 0.18.0 #4202)
• [xpu][test] Add weekly test for Intel GPU ([xpu][test] Add weekly test for Intel GPU #4190)
• add torch.compile test for Float8BlockwiseLinear (add torch.compile test for Float8BlockwiseLinear #4187)
• install torchvision to resolve ci issue (install torchvision to resolve ci issue #4203)
• Create recipe for flux2pro running on AMD (Create recipe for flux2pro running on AMD (#4200) #4200)
• fix ruff on main branch (fix ruff on main branch #4207)
• Add CLAUDE.md with AI agent instructions and quick reference (Add CLAUDE.md with AI agent instructions and quick reference #4195)
• Add llms.txt to docs site for LLM crawler discoverability (Add llms.txt to docs site for LLM crawler discoverability #4196)
• Skip test_comm for nvfp4 (Skip test_comm for nf4 #4210)
• [CPU] Update qsdpa lowering to return only the output node ([CPU] Update qsdpa lowering to return only the output node #4162)
• Bump version from 0.17.0 to 0.18.0 (Bump version from 0.17.0 to 0.18.0 #4172)
• Fix NF4Tensor narrow() crash with recent PyTorch nightly (Fix NF4Tensor narrow() crash with recent PyTorch nightly #4208)
• Delete deprecated TensorCoreTiledLayout and related code (Delete deprecated TensorCoreTiledLayout and related code #4153)
• [CPU] Re-enable the previously failing test_fp8_q_attention_block UT ([CPU] Re-enable the previously failing test_fp8_q_attention_block UT #4212)
• [CPU] Rename Int4WeightOnlyOpaqueTensorConfig to PrototypeInt4WeightOnlyConfig ([CPU] Rename Int4WeightOnlyOpaqueTensorConfig to PrototypeInt4WeightOnlyConfig #4205)
• add hadamard option to low precision attention api (add hadamard option to low precision attention api #4194)
• [ROCm] Enable FP8 QAT test on MI300/MI350 ([ROCm] Enable FP8 QAT test on MI300/MI350 #4132)
• Delete benchmark_e2e_fp8_sparse_linear.py (Delete benchmark_e2e_fp8_sparse_linear.py #4217)
• Add RunningAbsMaxSmoothQuantObserver for memory-efficient calibration (Add RunningAbsMaxSmoothQuantObserver for memory-efficient calibration (#3946) #3946)
• Add zero_point and act_mapping_type checks in int8 quantization (Add zero_point and act_mapping_type checks in int8 quantization #4160)
• Fix ROCM CI: skipped some tests, added checks, install torchvision (Fix ROCM CI: skipped some tests, added checks, install torchvision #4218)
• support pinning for mx and nvfp4 tensors (support pinning for mx and nvfp4 tensors #4192)
• Expand CODEOWNERS and add GitHub issue templates (Expand CODEOWNERS and add GitHub issue templates #4197)
• Delete deprecated PlainLayout, PlainAQTTensorImpl and related v1 code paths (Delete deprecated PlainLayout, PlainAQTTensorImpl and related v1 code paths #4151)
• [nvfp4] Make per_tensor_scale optional for triton kernel path ([nvfp4] Make per_tensor_scale optional for triton kernel path #4188)
• unbreak gptq int4 test and add to CI (unbreak gptq int4 test and add to CI #4223)
• add sharding rules for fp8 blockwise kernels (add sharding rules for fp8 blockwise kernels #4220)
• Replace mock-based MI350 tests with functional PerRow FP8 test (Replace mock-based MI350 tests with functional PerRow FP8 test (#4228) #4228)
• gptq: rename W to W_t (gptq: rename W to W_t #4224)
• gptq: more obvious block var names (gptq: more obvious block var names #4225)
• gptq: more obvious group var names (gptq: more obvious group var names #4226)
• gptq: more obvious inner loop var names (gptq: more obvious inner loop var names #4227)
• fix ruff (fix ruff #4235)
• [moe training] fix fp8 grouped mm compile issue ([moe training] fix fp8 grouped mm compile issue #4233)
• [mxfp8 training] make alignment size configurable in benchmark script ([mxfp8 training] make alignment size configurable in benchmark script #4246)
• [mxfp8 training] add cutedsl kernel for 32x1 mxfp8 quantization on 2d tensors ([mxfp8 training] add cutedsl kernel for 32x1 mxfp8 quantization on 2d tensors #4239)
• [pt2e] Skip linear+bn fusion when input is higher than 2-D ([pt2e] Skip linear+bn fusion when input is higher than 2-D #4242)
• Add alg_id argument to SemiSparseWeightConfig (Add alg_id argument to SemiSparseWeightConfig (#4238) #4238) (Add alg_id argument to SemiSparseWeightConfig (#4238) #4238)
• Update clean_release_notes.py to use per-module highlights (Update clean_release_notes.py to use per-module highlights #4250)
• [nvfp4 training] emulated nvfp4 grouped gemm ([nvfp4 training] emulated nvfp4 grouped gemm #4240)
• [mxfp8 training] update triton_to_mxfp8_dim0 nan handling; fix offset int32 overflow issue ([mxfp8 training] update triton_to_mxfp8_dim0 nan handling; fix offset int32 overflow issue #4201)
• Delete AffineQuantizedTensor, AQTTensorImpl, and Layout (Delete AffineQuantizedTensor, AQTTensorImpl, and Layout #4245)
• Adding V-only Hadamard support to low precision attention API (Adding V-only Hadamard support to low precision attention API #4249)
• Remove tutorials/calibration_flow/ and update stale references (Remove tutorials/calibration_flow/ and update stale references #4257)
• Remove LinearActivationQuantizedTensor and all related code (Remove LinearActivationQuantizedTensor and all related code #4258)
• Prevent folding of mutable input to copy_ and put_ operators (Prevent folding of mutable input to copy_ and put_ operators #4181)
• [xpu][test] Ignore test_x86inductor_fusion.py on XPU CI as it is not target ([xpu][test] Ignore test_x86inductor_fusion.py on XPU CI as it is not target #4265)
• Remove LinearActivationWeightObservedTensor and insert_observers_ (Remove LinearActivationWeightObservedTensor and insert_observers_ #4261)
• Remove WeightTensorWithLinearActivationScaleMetadata and related code (Remove WeightTensorWithLinearActivationScaleMetadata and related code #4262)
• Remove WeightTensorWithLinearActivationQuantizationMetadata (Remove WeightTensorWithLinearActivationQuantizationMetadata #4263)
• Use fixed scale for Float8 softmax quantization instead of observer (Use fixed scale for Float8 softmax quantization instead of observer #4260)
• Add torch.uint16, torch.uint32

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4269

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Rolling out OSDC (ARC) runners on pull workflow for PyTorch trunk commits

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