26 Pull requests merged by 12 people

• Update README.md to include Flux-Fast

  #2457 merged Jun 28, 2025

• Graduate debug handle in torchao

  #2452 merged Jun 27, 2025

• [float8 moe training] fix bug affecting mixed precision training

  #2451 merged Jun 27, 2025

• Add exportable coreml codebook quantization op

  #2443 merged Jun 27, 2025

• Store NVFP4 block scales in swwizzled layout on tensor

  #2438 merged Jun 26, 2025

• Fixes issue #156414: Fixes bug in implementation of _combine_histogram (Follow up)

  #2418 merged Jun 26, 2025

• Improve tiling params to speed up prefill

  #2406 merged Jun 26, 2025

• float8 readme: add key features section

  #2448 merged Jun 26, 2025

• float8 readme: remove duplication

  #2447 merged Jun 26, 2025

• [CPU] Fix ref path of DA8W4 cpp kernel

  #2444 merged Jun 26, 2025

• Call out axolotl + QAT integration on README

  #2442 merged Jun 25, 2025

• [CPU] Enable DA8W4 on CPU

  #2128 merged Jun 25, 2025

• [float8] Prevent quantize_affine_float8/dequantize_affine_float8 decomposed on inductor

  #2379 merged Jun 25, 2025

• solve the test issue

  #2432 merged Jun 24, 2025

• add-to-benchmarks

  #2427 merged Jun 24, 2025

• Gemlite generate.py fix

  #2372 merged Jun 24, 2025

• enable tensor parallelism for MXLinear

  #2434 merged Jun 24, 2025

• NVfp4

  #2408 merged Jun 24, 2025

• enable to_mxfp8 cast for DTensor

  #2420 merged Jun 24, 2025

• Update github links in torchao pt2e tutorial

  #2435 merged Jun 24, 2025

• [float8] add _auto_filter_for_recipe to float8

  #2410 merged Jun 24, 2025

• make dtensor shared test util more generic

  #2416 merged Jun 24, 2025

• rename torchao.testing.float8 to torchao.testing.training

  #2415 merged Jun 24, 2025

• fix float8 training TP+SP integration tests

  #2414 merged Jun 24, 2025

• Revert "Build mxfp4 kernel for sm120a"

  #2428 merged Jun 24, 2025

• mitigate the numeric test issue

  #2426 merged Jun 23, 2025

14 Pull requests opened by 10 people

• Add support for Int4GroupwisePreshuffleTensor for fbgemm

  #2421 opened Jun 22, 2025

• Remove `transpose_input` from fbgemm configs

  #2422 opened Jun 22, 2025

• enabling xpu in UT test

  #2424 opened Jun 23, 2025

• [float8 moe training] Add TP support

  #2425 opened Jun 23, 2025

• [DOC] Set strict export explicitly for API change

  #2430 opened Jun 24, 2025

• test rowwise fp32

  #2431 opened Jun 24, 2025

• mxfp8 training: add TP sharding strategy for dim1 kernel

  #2436 opened Jun 24, 2025

• Add support for Float8ActivationInt4GroupwisePreshuffleTensor for fbgemm

  #2437 opened Jun 24, 2025

• Add kernel

  #2439 opened Jun 25, 2025

• [DRAFT] Enable CPU data layout convert to XPU

  #2441 opened Jun 25, 2025

• fix scale shape in mx_format conversion

  #2446 opened Jun 26, 2025

• [NF4] Support nf4 tensor shard and gather

  #2449 opened Jun 26, 2025

• [float8] add tests for float8 _auto_filter_for_recipe

  #2450 opened Jun 26, 2025

• [WIP] [moe training] ScaledGroupedMMTensor - set dtype

  #2455 opened Jun 27, 2025

3 Issues closed by 3 people

• Benefits of Using QAT Before GGUF Quantization?

  #2419 closed Jun 25, 2025

• [Quant] Can quant not be decomposed on inductor?

  #2228 closed Jun 25, 2025

• Question about the choice of use_fast_accum in FP8 Training

  #2377 closed Jun 22, 2025

7 Issues opened by 7 people

• How to not decompose the choose_qparams_affine call_func

  #2456 opened Jun 27, 2025

• `fill_()` not implemented for `torchao.quantization.linear_activation_quantized_tensor.LinearActivationQuantizedTensor`

  #2454 opened Jun 27, 2025

• [MoE training] crash with FSDP if shared_expert uses float8 in torchtitan llama4

  #2453 opened Jun 27, 2025

• Gradient Checkpoint makes FP8 Training Slow

  #2445 opened Jun 26, 2025

• [BUG] TorchDynamo frame conversion fails for autoquant because __tensor_flatten__ is not implemented for Float32Tensor

  #2440 opened Jun 25, 2025

• [BUG]: Gemlite integration not working

  #2429 opened Jun 24, 2025

• Training hangs after 1 epoch when using QAT

  #2423 opened Jun 22, 2025

17 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.

• Inference tutorial - Part 3 of e2e series [WIP]

  #2343 commented on Jun 26, 2025 • 14 new comments

• [Inductor] Support scaled mm on inductor

  #2411 commented on Jun 25, 2025 • 4 new comments

• Add benchmark numbers to dashboard

  #2260 commented on Jun 28, 2025 • 2 new comments

• Align scale dtype with model precision in GPTQ

  #2403 commented on Jun 26, 2025 • 1 new comment

• [WIP] Make AWQ more general

  #2400 commented on Jun 25, 2025 • 0 new comments

• Enables the per_tensor lowering patterns for weight per_packing

  #2391 commented on Jun 26, 2025 • 0 new comments

• moe quant with dedicated kernels [wip]

  #2325 commented on Jun 26, 2025 • 0 new comments

• Eval hf models using lm_eval

  #2179 commented on Jun 23, 2025 • 0 new comments

• [low-bit optim] Add coat for float8 optimizer

  #1231 commented on Jun 26, 2025 • 0 new comments

• [roadmap/tracker] Low precision MoE training

  #2147 commented on Jun 28, 2025 • 0 new comments

• Support `torch.int4` `target_dtype` for ops `choose_qparams_affine`, `quantize_affine`, `dequantize_affine`

  #2354 commented on Jun 27, 2025 • 0 new comments

• MX single node performance tracker

  #1768 commented on Jun 26, 2025 • 0 new comments

• low precision training upcoming feature tracker

  #556 commented on Jun 26, 2025 • 0 new comments

• int4_weight_only get plain weight are padded

  #2249 commented on Jun 24, 2025 • 0 new comments

• TP + FSDP + MXFP8 fails during compile

  #2393 commented on Jun 24, 2025 • 0 new comments

• BF16 stochastic rounding does not work distributed (FSDP)

  #2296 commented on Jun 23, 2025 • 0 new comments

• TorchAO Paper

  #2412 commented on Jun 23, 2025 • 0 new comments