19 Pull requests merged by 13 people

• [float8 moe training] FSDP support

  #2413 merged Jun 21, 2025

• Build mxfp4 kernel for sm120a

  #2285 merged Jun 21, 2025

• Remove more Galore bits

  #2417 merged Jun 21, 2025

• Enable cpp kernel building

  #2402 merged Jun 19, 2025

• Replace debug handle with from_node to trace operator transformation

  #2339 merged Jun 18, 2025

• [float8 moe training] make using triton kernels for per-group scaling configurable

  #2405 merged Jun 18, 2025

• Add part 2 of end-to-end tutorial: fine-tuning

  #2394 merged Jun 18, 2025

• Fix ruff broken on main

  #2404 merged Jun 18, 2025

• fix torchao quantized model in fbcode

  #2396 merged Jun 18, 2025

• [BE] Convert quant_primitives methods private

  #2350 merged Jun 18, 2025

• Delete Galore

  #2397 merged Jun 18, 2025

• Add inplace quantizer examples

  #2345 merged Jun 18, 2025

• Update index.rst

  #2395 merged Jun 17, 2025

• Add pt2e tutorials to torchao doc page

  #2384 merged Jun 17, 2025

• deduplicate torch ao debugger tests between pytorch/ao and ExecuTorch

  #2390 merged Jun 17, 2025

• [float8 training] update torchtitan benchmark script args

  #2392 merged Jun 17, 2025

• turn off building tests with cpuinfo

  #2324 merged Jun 17, 2025

• remove torchao dependency from torchao build script

  #2383 merged Jun 17, 2025

• remove rocm source files when not building for rocm

  #2385 merged Jun 16, 2025

21 Pull requests opened by 13 people

• Add support for resharding and int4 preshuffle kernel

  #2387 opened Jun 16, 2025

• Enables the per_tensor lowering patterns for weight per_packing

  #2391 opened Jun 17, 2025

• Unskip tests

  #2398 opened Jun 18, 2025

• [WIP] Add AWQ quantization with QDQLayout support for ExecuTorch

  #2399 opened Jun 18, 2025

• [WIP] Make AWQ more general

  #2400 opened Jun 18, 2025

• Align scale dtype with model precision in GPTQ

  #2403 opened Jun 18, 2025

• Improve tiling params to speed up prefill

  #2406 opened Jun 18, 2025

• Groupwise low bit LUT based model quantization.

  #2407 opened Jun 18, 2025

• NVfp4

  #2408 opened Jun 18, 2025

• [float8] add auto_filter_for_recipe to float8

  #2410 opened Jun 18, 2025

• [Inductor] Support scaled mm on inductor

  #2411 opened Jun 19, 2025

• fix float8 training TP+SP integration tests

  #2414 opened Jun 20, 2025

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

  #2415 opened Jun 20, 2025

• make dtensor shared test util more generic

  #2416 opened Jun 20, 2025

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

  #2418 opened Jun 21, 2025

• enable to_mxfp8 cast for DTensor

  #2420 opened Jun 21, 2025

• 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

• mitigate the numeric test issue

  #2426 opened Jun 23, 2025

3 Issues closed by 3 people

• Question about the choice of use_fast_accum in FP8 Training

  #2377 closed Jun 22, 2025

• [Question] Combining QAT and Sparsity Training

  #2310 closed Jun 20, 2025

• [Windows][build]two Build failure on Windows on latest main branch

  #2297 closed Jun 16, 2025

5 Issues opened by 4 people

• Training hangs after 1 epoch when using QAT

  #2423 opened Jun 22, 2025

• Benefits of Using QAT Before GGUF Quantization?

  #2419 opened Jun 21, 2025

• TorchAO Paper

  #2412 opened Jun 19, 2025

• TP + FSDP + MXFP8 fails during compile

  #2393 opened Jun 17, 2025

• Implement an AWQ algorithm with dynamic activation quantization for ExecuTorch

  #2388 opened Jun 16, 2025

14 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 23, 2025 • 21 new comments

• Add Claude MD file

  #2311 commented on Jun 18, 2025 • 5 new comments

• skip quant/dequant decomposed

  #2299 commented on Jun 19, 2025 • 3 new comments

• [float8] Prevent quantize_affine_float8/dequantize_affine_float8 decomposed on inductor

  #2379 commented on Jun 20, 2025 • 3 new comments

• [CPU] Enable DA8W4 on CPU

  #2128 commented on Jun 20, 2025 • 1 new comment

• [roadmap/tracker] Low precision training for MoEs

  #2147 commented on Jun 16, 2025 • 0 new comments

• BF16 stochastic rounding does not work distributed (FSDP)

  #2296 commented on Jun 18, 2025 • 0 new comments

• [QAT] Low-bit FSDP all-gather for QAT

  #1224 commented on Jun 19, 2025 • 0 new comments

• Eval hf models using lm_eval

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

• Add round_scales_to_power_of_2 option for float quantization

  #2323 commented on Jun 18, 2025 • 0 new comments

• moe quant with dedicated kernels [wip]

  #2325 commented on Jun 20, 2025 • 0 new comments

• Update to new PT Theme

  #2361 commented on Jun 17, 2025 • 0 new comments

• [CPU INT8 SDPA] use manual transpose and pack

  #2380 commented on Jun 20, 2025 • 0 new comments

• [not for land] float8 blockwise scaling training prototype using deep_gemm

  #2386 commented on Jun 18, 2025 • 0 new comments