30 Pull requests merged by 11 people

• [MoE training] torch.compile support for ScaledGroupedMMTensor

  #2509 merged Aug 2, 2025

• [ROCm CI] Migrate to MI325 Capacity

  #2662 merged Aug 2, 2025

• Make AWQ more general

  #2400 merged Aug 1, 2025

• backward pass for differentiable mxfp8 grouped gemm with dynamic quant

  #2639 merged Aug 1, 2025

• support for 2d-2d emulated mxfp8 grouped gemm

  #2632 merged Aug 1, 2025

• Add new ops to CMakeLists.txt

  #2647 merged Aug 1, 2025

• Add Aten operations

  #2664 merged Aug 1, 2025

• Add Float8BlockwiseLinear for training

  #2618 merged Aug 1, 2025

• [bc-breaking] Generalize FakeQuantizeConfig beyond intx

  #2628 merged Aug 1, 2025

• fix bc breakage flex path

  #2652 merged Aug 1, 2025

• Remove warnings in favor of skiptests for Moe code

  #2654 merged Aug 1, 2025

• skip rocm for moe training tests

  #2646 merged Jul 31, 2025

• Add op_exetorch

  #2645 merged Jul 31, 2025

• Clarifying the meaning of VERSION in AOBaseConfig

  #2635 merged Jul 31, 2025

• fix: improve formatting and resolve minor bug for better utility

  #2634 merged Jul 31, 2025

• Update test_dynamic_activation_lut.py

  #2637 merged Jul 30, 2025

• add differentiable mxfp8 grouped gemm with dynamic quant (forward pass)

  #2627 merged Jul 30, 2025

• Add tests

  #2624 merged Jul 30, 2025

• [Easy] Fix git repo url in citation

  #2599 merged Jul 30, 2025

• Add Triton kernels for fp8 blockwise quantization and GEMMs

  #2617 merged Jul 30, 2025

• mxfp8 emulated grouped gemm

  #2626 merged Jul 30, 2025

• update float8 readme with more recent performance numbers

  #2580 merged Jul 30, 2025

• Integrate PARQ with lowbit Arm CPU kernels

  #2622 merged Jul 29, 2025

• Update the op_-impl.h

  #2621 merged Jul 29, 2025

• Disable register_da8w4_concat_linear_cpu_pass

  #2623 merged Jul 29, 2025

• [OpenVINOQuantizer] Minor improvements

  #2581 merged Jul 29, 2025

• Update packed_weghts_header to include the new groupwise lowbit header.

  #2582 merged Jul 28, 2025

• reorganize MX inference code

  #2616 merged Jul 28, 2025

• delete outdated MX inference code

  #2615 merged Jul 28, 2025

• fix mx kernel tests

  #2614 merged Jul 28, 2025

23 Pull requests opened by 10 people

• mx: expose scaling calculation methods in training UX

  #2620 opened Jul 28, 2025

• New multi-step QAT API

  #2629 opened Jul 29, 2025

• Update test_bitpacking.cpp. Bug fix in test.

  #2633 opened Jul 30, 2025

• Deprecate old QAT APIs

  #2641 opened Jul 30, 2025

• Make scaling type configurable for MoE training

  #2642 opened Jul 30, 2025

• Convert model inference test from pytest to unittest

  #2644 opened Jul 31, 2025

• Update coreml codebook

  #2648 opened Jul 31, 2025

• Bump version for float8 dynamic quant and weight only quant configs

  #2650 opened Aug 1, 2025

• Check numerical equivalence / closeness between different kernel preferences

  #2651 opened Aug 1, 2025

• Migrate to unittest for files in test/float8

  #2655 opened Aug 1, 2025

• Add gpu_name as a parameter in roofline estimate utils

  #2657 opened Aug 1, 2025

• Migrate to unittest for test files in quantization, sparsity

  #2658 opened Aug 1, 2025

• Convert SmoothQuant test to unittest

  #2659 opened Aug 1, 2025

• Replace `torch.norm` with `torch.linalg.vector_norm` for PyTorch future update

  #2660 opened Aug 1, 2025

• mx: make CUDA kernel for dim1 cast in mxfp8_cublas recipe

  #2661 opened Aug 1, 2025

• [MoE training] Assert expert weights are column-major; preserve subclass with transpose

  #2663 opened Aug 1, 2025

• Update test scripts to include the new operation.

  #2665 opened Aug 1, 2025

• Add NVFP4 QAT

  #2666 opened Aug 1, 2025

• Add more metrics to dashboard

  #2667 opened Aug 1, 2025

• [moe training] use smaller block sizes for per group scaling kernels to improve perf

  #2668 opened Aug 2, 2025

• [moe training] add llama4 benchmarking script

  #2669 opened Aug 2, 2025

• [moe training] add benchmark script for moe layer

  #2671 opened Aug 4, 2025

• fix float8 rowwise inference perf with torch.compile

  #2672 opened Aug 4, 2025

1 Issue closed by 1 person

• What is the intention of "NF4WeightOnlyConfig" ?

  #2631 closed Jul 31, 2025

6 Issues opened by 4 people

• [moe training] fsdp2 bug for llama4 shared experts where num_experts=1

  #2673 opened Aug 4, 2025

• Deprecation for Float8DynamicActivationFloat8WeightConfig and Float8WeightOnlyConfig and the models

  #2649 opened Jul 31, 2025

• Refresh "Quantization Overview" docs page

  #2643 opened Jul 30, 2025

• Confusing behavior of `register_quantize_module_handler`

  #2640 opened Jul 30, 2025

• Deprecate and remove old QAT API

  #2630 opened Jul 30, 2025

• Int8QuantizedTrainingLinearWeight missing aten.cat support for distributed training

  #2619 opened Jul 28, 2025

20 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Add Float8Tensor

  #2463 commented on Aug 2, 2025 • 31 new comments

• Remove double baseline calculations for CI microbenchmarks

  #2613 commented on Aug 4, 2025 • 6 new comments

• Add vLLM x TorchAO integration workflow

  #2610 commented on Jul 29, 2025 • 2 new comments

• Migrate to unittest for files in test/dtypes

  #2605 commented on Jul 31, 2025 • 0 new comments

• wip MoE refactor

  #2600 commented on Aug 1, 2025 • 0 new comments

• Enable PyTorch/ao CI on Intel XPU with PyTorch Base Docker

  #2584 commented on Aug 4, 2025 • 0 new comments

• [Inductor][float8] Support qlinear for float8 in inductor

  #2565 commented on Aug 4, 2025 • 0 new comments

• Refactor Wanda for better readability

  #2538 commented on Aug 1, 2025 • 0 new comments

• feat: RGS for wanda++

  #2537 commented on Jul 29, 2025 • 0 new comments

• [CPU] Add support for dynamic float8 act float8 weight on CPU

  #2505 commented on Aug 1, 2025 • 0 new comments

• Add all fbgemm kernel Tensors into Int4WeightOnlyConfig and Float8DynamicActivationInt4WeightConfig

  #2474 commented on Aug 2, 2025 • 0 new comments

• [DRAFT] Enable CPU data layout convert to XPU

  #2441 commented on Jul 29, 2025 • 0 new comments

• [Inductor] Support scaled mm on inductor

  #2411 commented on Aug 1, 2025 • 0 new comments

• Enables the per_tensor lowering patterns for weight per_packing

  #2391 commented on Jul 31, 2025 • 0 new comments

• [roadmap/tracker] Low precision MoE training

  #2147 commented on Aug 4, 2025 • 0 new comments

• [ROCm] torchao.float8 should work properly on ROCm

  #1066 commented on Aug 4, 2025 • 0 new comments

• [feature request] np.packbits / np.unpackbits, general BitTensors (maybe can be just tensors with dtype torch.bits8 or have a new dtype torch.bits introduced) and bit packed tensors utilities for saving memory / accesses, support for BitTensors wherever BoolTensors are used

  #292 commented on Aug 3, 2025 • 0 new comments

• [Feature Req] Can you add *args and **kwargs to improve extensibility ?

  #2496 commented on Aug 1, 2025 • 0 new comments

• bug: int8 w8a8 doesn't work on 5090

  #2376 commented on Jul 28, 2025 • 0 new comments

• Missing benchmark for `sparse24_sm90_sparsify` overhead

  #2612 commented on Jul 28, 2025 • 0 new comments