Ifu dev 20260706 v2.17 by AllenFarcas · Pull Request #660 · ROCm/TransformerEngine

AllenFarcas · 2026-07-07T16:25:50Z

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

Type of change

Documentation change (change only to the documentation, either a fix or a new content)
Bug fix (non-breaking change which fixes an issue)
New feature (non-breaking change which adds functionality)
Breaking change (fix or feature that would cause existing functionality to not work as expected)
Infra/Build change
Code refactoring

Changes

Please list the changes introduced in this PR:

Change A
Change B

Checklist:

I have read and followed the contributing guidelines
The functionality is complete
I have commented my code, particularly in hard-to-understand areas
I have made corresponding changes to the documentation
My changes generate no new warnings
I have added tests that prove my fix is effective or that my feature works
New and existing unit tests pass locally with my changes

Signed-off-by: Przemek Tredak <[email protected]>

* Reduced padding kernel compilation time Signed-off-by: Oleg Goncharov <[email protected]> * Completely removed unroll for better performance Signed-off-by: Oleg Goncharov <[email protected]> --------- Signed-off-by: Oleg Goncharov <[email protected]>

…(#2905) Zero out padded region when swizzling via group quantize Signed-off-by: Kirthi Shankar Sivamani <[email protected]>

…2879) * fix broken links in README Signed-off-by: Santosh Bhavani <[email protected]> * update README to modernize description and standardize terminology Signed-off-by: Santosh Bhavani <[email protected]> * Update README.rst Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> Signed-off-by: Santosh Bhavani <[email protected]> * Removed the duplicate line Signed-off-by: Przemek Tredak <[email protected]> --------- Signed-off-by: Santosh Bhavani <[email protected]> Signed-off-by: Przemek Tredak <[email protected]> Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> Co-authored-by: Przemek Tredak <[email protected]>

* initial implementation for mxfp8 Signed-off-by: Charlene Yang <[email protected]> * semi-working FP8; broken F16 Signed-off-by: Charlene Yang <[email protected]> * clean up last commit Signed-off-by: Charlene Yang <[email protected]> * comment out F16 pass Signed-off-by: Charlene Yang <[email protected]> * pull in grouped_quantize for MXFP8 Signed-off-by: Charlene Yang <[email protected]> * grouped tensor - pytorch Signed-off-by: Charlene Yang <[email protected]> * quantize mxfp8 Signed-off-by: Charlene Yang <[email protected]> * fix shapes/strides Signed-off-by: Charlene Yang <[email protected]> * fix unfused; clean up Signed-off-by: Charlene Yang <[email protected]> * split d to d_qk/d_v; attempt at bwd Signed-off-by: Charlene Yang <[email protected]> * fix last merge Signed-off-by: Charlene Yang <[email protected]> * update FE Signed-off-by: Charlene Yang <[email protected]> * attempt at SWA/MLA Signed-off-by: Charlene Yang <[email protected]> * remove prints Signed-off-by: Charlene Yang <[email protected]> * remove leftover prints Signed-off-by: Charlene Yang <[email protected]> * Revert "update FE" This reverts commit d9ff5662aa4b4b6267c77baf614aada6602fa133. Signed-off-by: Charlene Yang <[email protected]> * update FE Signed-off-by: Charlene Yang <[email protected]> * fix MLA O strides; add bottom_right_diagonal Signed-off-by: Charlene Yang <[email protected]> * attempt at bwd Signed-off-by: Charlene Yang <[email protected]> * fix get_quantizers; attempt at bwd Signed-off-by: Charlene Yang <[email protected]> * fix fprop; add o_format Signed-off-by: Charlene Yang <[email protected]> * attempt at bwd with o_format/d_out_format/dqkv_layout Signed-off-by: Charlene Yang <[email protected]> * fix dtype/o_format/etc in bwd calls Signed-off-by: Charlene Yang <[email protected]> * fix generateMatrixStridesWithFormats and _v1; fix padding for mxfp8 Signed-off-by: Charlene Yang <[email protected]> * fix upon last commit for paddedsizes Signed-off-by: Charlene Yang <[email protected]> * add mxfp8 env var Signed-off-by: Charlene Yang <[email protected]> * disable FA for mxfp8 Signed-off-by: Charlene Yang <[email protected]> * add mha test Signed-off-by: Charlene Yang <[email protected]> * attempt at bwd; force determinism; fix shapes Signed-off-by: Charlene Yang <[email protected]> * remove prints Signed-off-by: Charlene Yang <[email protected]> * update FE Signed-off-by: Charlene Yang <[email protected]> * update FE from pre-merge branch to post-merge develop Signed-off-by: Charlene Yang <[email protected]> * allow MXFP8 linear + f16 attn Signed-off-by: Charlene Yang <[email protected]> * test cp a2a Signed-off-by: Charlene Yang <[email protected]> * remove prints temporarily Signed-off-by: Charlene Yang <[email protected]> * test cp p2p Signed-off-by: Charlene Yang <[email protected]> * minor fixes for mla Signed-off-by: Charlene Yang <[email protected]> * open up a2a for mla Signed-off-by: Charlene Yang <[email protected]> * test ag Signed-off-by: Charlene Yang <[email protected]> * tweaks for last commit Signed-off-by: Charlene Yang <[email protected]> * enable mla ag Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix merge Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix merge Signed-off-by: Charlene Yang <[email protected]> * revert to main grouped tensor impl Signed-off-by: Charlene Yang <[email protected]> * minor tweaks to return to main Signed-off-by: Charlene Yang <[email protected]> * remove prints Signed-off-by: Charlene Yang <[email protected]> * fix combine_and_quantize for f16 Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * minor tweaks Signed-off-by: Charlene Yang <[email protected]> * tweak tests Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix ds descale_o Signed-off-by: Charlene Yang <[email protected]> * Revert "fix ds descale_o" This reverts commit cd0bd82e239ff01210338b4e34cb8784109d22ec. Signed-off-by: Charlene Yang <[email protected]> * minor fixes for p2p and ag Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * tweak cp test skips Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * update FE Signed-off-by: Charlene Yang <[email protected]> * fix bwd KV tensors Signed-off-by: Charlene Yang <[email protected]> * tweak recipe control and backend selection Signed-off-by: Charlene Yang <[email protected]> * tweak quantizer logic Signed-off-by: Charlene Yang <[email protected]> * minor fixes after last two commits Signed-off-by: Charlene Yang <[email protected]> * improve generate strides Signed-off-by: Charlene Yang <[email protected]> * minor fixes for previous commit Signed-off-by: Charlene Yang <[email protected]> * fix bwd for current/delayed Signed-off-by: Charlene Yang <[email protected]> * tweak test configs Signed-off-by: Charlene Yang <[email protected]> * fix dO/dO_f16 strides Signed-off-by: Charlene Yang <[email protected]> * fix tests: SWA logic/test configs Signed-off-by: Charlene Yang <[email protected]> * fix ag Signed-off-by: Charlene Yang <[email protected]> * add fp8 sink attn Signed-off-by: Charlene Yang <[email protected]> * fix a2a comm for F16 Signed-off-by: Charlene Yang <[email protected]> * remove nan/inf print in test Signed-off-by: Charlene Yang <[email protected]> * fix fa a2a Signed-off-by: Charlene Yang <[email protected]> * fix fa a2a+p2p f16 Signed-off-by: Charlene Yang <[email protected]> * update FE to include new fixes Signed-off-by: Charlene Yang <[email protected]> * fix thd for bwd Signed-off-by: Charlene Yang <[email protected]> * refactor a2a for fu/fa Signed-off-by: Charlene Yang <[email protected]> * update FE to fix d64 Signed-off-by: Charlene Yang <[email protected]> * refactor ag Signed-off-by: Charlene Yang <[email protected]> * refactor p2p/a2a+p2p; mostly regarding shapes Signed-off-by: Charlene Yang <[email protected]> * add shadow f16 fwd Signed-off-by: Charlene Yang <[email protected]> * update FE to fix SWA/BRCM Signed-off-by: Charlene Yang <[email protected]> * switch to GH FE temporarily Signed-off-by: Charlene Yang <[email protected]> * switch back to GL FE Signed-off-by: Charlene Yang <[email protected]> * update FE to latest commit Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * update group tensor usage after merge main Signed-off-by: Charlene Yang <[email protected]> * env vars for qdq(q,k), o_f16 tests Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * allow other recipes than mxfp8 Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix grouped tensor for MLA Signed-off-by: Charlene Yang <[email protected]> * change cp test configs Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add shadow f16 bwd Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix a2a+p2p for sbhd Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix last commit and causal flag for fa Signed-off-by: Charlene Yang <[email protected]> * enable fp8 sink and disable fp8_mha Signed-off-by: Charlene Yang <[email protected]> * minor cleanup for cp/non-cp Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * update FE for FP8 sink Signed-off-by: Charlene Yang <[email protected]> * fix TE for FP8 sink Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * temporary: random sink/print sink Signed-off-by: Charlene Yang <[email protected]> * Revert "temporary: random sink/print sink" This reverts commit 706095f802e04cbdd5d88ee53849cc5ec938203f. Signed-off-by: Charlene Yang <[email protected]> * replace d_out_format with do_format Signed-off-by: Charlene Yang <[email protected]> * fix compare_and_assert for None cases Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * remove logic for b and simplify logic for dqkv types Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * minor fix for ndim_q/kv Signed-off-by: Charlene Yang <[email protected]> * add explanation of fp8_output/grad in MHA Signed-off-by: Charlene Yang <[email protected]> * tidy up FP8 checks for bhsd/learnable Signed-off-by: Charlene Yang <[email protected]> * remove leading underscores in nvte_convert_qkv_format Signed-off-by: Charlene Yang <[email protected]> * simplify logic in generateMatrixStridesWithLayout Signed-off-by: Charlene Yang <[email protected]> * clean up strides/ifelse-recipe logic Signed-off-by: Charlene Yang <[email protected]> * tweak checks in utils.py Signed-off-by: Charlene Yang <[email protected]> * tweak UnfusedDPA Signed-off-by: Charlene Yang <[email protected]> * enable testing for ag+swa and disable fp8_mha Signed-off-by: Charlene Yang <[email protected]> * tweak FusedAttn, fp8/f16 tensor naming/docstring Signed-off-by: Charlene Yang <[email protected]> * replace d_out_format with do_format Signed-off-by: Charlene Yang <[email protected]> * fix lint Signed-off-by: Charlene Yang <[email protected]> * clean up a2a Signed-off-by: Charlene Yang <[email protected]> * clean up ag Signed-off-by: Charlene Yang <[email protected]> * clean up p2p/a2a+p2p Signed-off-by: Charlene Yang <[email protected]> * tweak test configs Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * qdq dO in bwd shadow f16 path Signed-off-by: Charlene Yang <[email protected]> * tweak qdq dO logic Signed-off-by: Charlene Yang <[email protected]> * remove prints in shadow paths Signed-off-by: Charlene Yang <[email protected]> * update FE to allow non-determinism Signed-off-by: Charlene Yang <[email protected]> * fuse qkv transposes; first pass Signed-off-by: Charlene Yang <[email protected]> * remap parallelism to grid(bh, splits, 3) block(s/splits x d); use nvec = 128 bits Signed-off-by: Charlene Yang <[email protected]> * allocate contiguous block for qkv Signed-off-by: Charlene Yang <[email protected]> * fix grouped tensor row/col scale_inv offsets Signed-off-by: Charlene Yang <[email protected]> * use fused permute kernels Signed-off-by: Charlene Yang <[email protected]> * quantize row/col as needed in fwd/bwd, non-cp/cp Signed-off-by: Charlene Yang <[email protected]> * Revert "quantize row/col as needed in fwd/bwd, non-cp/cp" This reverts commit ca5376956e8b8f662c7fa88661695b3e9eda4f8f. Signed-off-by: Charlene Yang <[email protected]> * Reapply "quantize row/col as needed in fwd/bwd, non-cp/cp" This reverts commit f19e852be3463210f2b3be5839ae8931e5ad92d0. Signed-off-by: Charlene Yang <[email protected]> * fix v_col format when row is quantized Signed-off-by: Charlene Yang <[email protected]> * add back necessary bwd quants for shadow paths/cp a2a Signed-off-by: Charlene Yang <[email protected]> * remove ZInv for all layouts except T3HD Signed-off-by: Charlene Yang <[email protected]> * fix cp p2p with zinv Signed-off-by: Charlene Yang <[email protected]> * temporarily switch to GH FE main Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * switch back to GL FE Signed-off-by: Charlene Yang <[email protected]> * fix ag after merge main Signed-off-by: Charlene Yang <[email protected]> * add condition for qdq(do) to not affect other tests Signed-off-by: Charlene Yang <[email protected]> * fix custom_mha_fp8 test Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix amax dqkv Signed-off-by: Charlene Yang <[email protected]> * fix fp8_recipe in DPA utils Signed-off-by: Charlene Yang <[email protected]> * remove use of amax for mxfp8 Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add o_format/do_format/dqkv_layout to cache indicators for fp8 and f16 Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * enable sink attn + FP8 in CP Signed-off-by: Charlene Yang <[email protected]> * update FE to GH v1.22.0 Signed-off-by: Charlene Yang <[email protected]> * fix for inconsistent kwarg name in permute to grouped tensor Signed-off-by: Charlene Yang <[email protected]> * add TMA permute Signed-off-by: Charlene Yang <[email protected]> * Revert "add TMA permute" This reverts commit 2532a50e829144bee290fc94acb8f3f154a62ea9. Signed-off-by: Charlene Yang <[email protected]> * TMA load for bhsd transposes Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix some lint Signed-off-by: Charlene Yang <[email protected]> * temp: quant+perm+swizzle, rope, perm_fused Signed-off-by: Charlene Yang <[email protected]> * remove mla_rope for now; clean up quant+permute+pad_swizzle; create multi_tensor_swizzle Signed-off-by: Charlene Yang <[email protected]> * fix last commit Signed-off-by: Charlene Yang <[email protected]> * implement narrow-m for col swizzle; reorder to pad+perm+swizzle Signed-off-by: Charlene Yang <[email protected]> * fused pad into perm; remove at::zeros as zeros done in perm kernels Signed-off-by: Charlene Yang <[email protected]> * remove shadow code Signed-off-by: Charlene Yang <[email protected]> * minor fix for permute shapes Signed-off-by: Charlene Yang <[email protected]> * check smem size before entering narrow-k/m kernels Signed-off-by: Charlene Yang <[email protected]> * expand permute to multi_tensor_ Signed-off-by: Charlene Yang <[email protected]> * refactor qkv/do quant; create a fast_path call Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix lint Signed-off-by: Charlene Yang <[email protected]> * cleanup grouped tensor fix Signed-off-by: Charlene Yang <[email protected]> * remove _with_amax for create_unquantized_tensor Signed-off-by: Charlene Yang <[email protected]> * fix last commit Signed-off-by: Charlene Yang <[email protected]> * reimplement inplace_multi_tensor_swizzle Signed-off-by: Charlene Yang <[email protected]> * fix last commit; set swizzled flag in python Signed-off-by: Charlene Yang <[email protected]> * remove permute_to_grouped_tensor_bwd; clean up fwd Signed-off-by: Charlene Yang <[email protected]> * add doxygen for multi_tensor_swizzle Signed-off-by: Charlene Yang <[email protected]> * clean up nvte_convert_qkv_format Signed-off-by: Charlene Yang <[email protected]> * fixes based on code review Signed-off-by: Charlene Yang <[email protected]> * group layouts/formats in APIs Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * rename nvte_convert_qkv_format to nvte_convert_qkv_shape Signed-off-by: Charlene Yang <[email protected]> * remove MXFP8 create_unquantized_tensor Signed-off-by: Charlene Yang <[email protected]> * rename permute_to_grouped_tensor to transpose_to_bhsd Signed-off-by: Charlene Yang <[email protected]> * add multi_tensor_swizzle_xx_unchecked and split the calls/paths Signed-off-by: Charlene Yang <[email protected]> * straighten up indexing for multi_tensor_pad Signed-off-by: Charlene Yang <[email protected]> * batch up kernel calls per-16-tensors for pad and permute Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * remove nvec128; rename nvec64 back to nvec Signed-off-by: Charlene Yang <[email protected]> * add Macros/arch specifics for compilation Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix lint Signed-off-by: Charlene Yang <[email protected]> * attempt 1: MLA RoPE Signed-off-by: Charlene Yang <[email protected]> * Revert "attempt 1: MLA RoPE" This reverts commit 79229248718d26a0ae7029206adc26c687bb42a7. Signed-off-by: Charlene Yang <[email protected]> * fix kv_cache tests for Fused, is_page=True Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * attempt 2: MLA RoPE Signed-off-by: Charlene Yang <[email protected]> * use DIVUP/_TO_MULTIPLE for pad_s_d_for_mxfp8 Signed-off-by: Charlene Yang <[email protected]> * remove CUDNN_VERSION 8900 macros Signed-off-by: Charlene Yang <[email protected]> * add narrow-k/m swizzle tests Signed-off-by: Charlene Yang <[email protected]> * compile flash_attn.cu with special archs Signed-off-by: Charlene Yang <[email protected]> * Revert "attempt 2: MLA RoPE" This reverts commit 3b854b29a3677de2005fecff821d801ccd9bf5d4. Signed-off-by: Charlene Yang <[email protected]> * make contiguous instead of check is_contiguous Signed-off-by: Charlene Yang <[email protected]> * remove unused s_q/s_kv Signed-off-by: Charlene Yang <[email protected]> * remove unused issue_tma_store_strided Signed-off-by: Charlene Yang <[email protected]> * add version gate for mxfp8 for CPP users Signed-off-by: Charlene Yang <[email protected]> * replace nvte_get_qkv_shape with AttentionShape Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * populate nvte_ changes to Jax Signed-off-by: Charlene Yang <[email protected]> * update FE to 1.22.1 Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix minor merge issue Signed-off-by: Charlene Yang <[email protected]> * revert to FE 1.21 since it's what mxfp8 needs Signed-off-by: Charlene Yang <[email protected]> * udpate jax attention shapes Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Revert "revert to FE 1.21 since it's what mxfp8 needs" This reverts commit f09961a03bd7f5a316474b5d77b8292c7a49c1a6. Signed-off-by: Charlene Yang <[email protected]> * pick FE 1.22 to support mxfp8 and avoid rng issue in 1.22.1 Signed-off-by: Charlene Yang <[email protected]> * fix CP AG test on Hopper Signed-off-by: Charlene Yang <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Charlene Yang <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* starting grouped linear integration, not tested, grouped_bias_add optimized and uses scales now Signed-off-by: Varun Thumbe <[email protected]> * all changes Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * optimize grouped bias add kernel to 4TB/s handling load imbalance Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * kernel optimized + review comments Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Remove uneecessary load from the kernel Remove unnecessary pointer reinterpretation for bias input. Signed-off-by: vthumbe1503 <[email protected]> * No need for text for grouped bias add Signed-off-by: vthumbe1503 <[email protected]> * Remove lru_cache from _is_deterministic_mode Removed lru_cache decorator from the _is_deterministic_mode function. Signed-off-by: vthumbe1503 <[email protected]> * address more review comments Signed-off-by: Varun Thumbe <[email protected]> * for better perf Signed-off-by: Varun Thumbe <[email protected]> * remove unecessary comments Signed-off-by: vthumbe1503 <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * rename a bit Signed-off-by: Varun Thumbe <[email protected]> --------- Signed-off-by: Varun Thumbe <[email protected]> Signed-off-by: vthumbe1503 <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add better ordering enforcment to split_overlap_rs gemms. This adds a short delay kernel to the split_overlap_rs function, which ensures that the gemms are properly ordered when run with cuda graphs. Signed-off-by: Chase Block <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Chase Block <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Przemyslaw Tredak <[email protected]>

Signed-off-by: Björn Buschkämper <[email protected]>

…caches (#2805) * fixing mem leaks Signed-off-by: Peter St. John <[email protected]> * update xfail message Signed-off-by: Peter St. John <[email protected]> * addressing greptile comments Signed-off-by: Peter St. John <[email protected]> * remove fsdp_safe Signed-off-by: Peter St. John <[email protected]> * fix Float8BlockScaling backward override, unused imports, xfail MXFP8 HSDP Fix three issues: 1. LayerNormLinear weight quantizer was requesting columnwise usage even when backward_override is set. The FSDP2 refactor on this branch changed the columnwise condition to `is_grad_enabled and not is_fsdp2` but omitted the `backward_override is None` guard. This caused the weight to carry unnecessary columnwise data, which the dequantized backward path rejects. Also harden Float8BlockQuantizer::create_tensor to pass py::none() based on usage flags rather than relying on at::Tensor::defined() which is unreliable for default-constructed tensors in some PyTorch builds. 2. Remove unused imports: Float8Tensor from linear.py, and the entire float8_tensor import line from layernorm_linear.py (Float8Quantizer, Float8CurrentScalingQuantizer, Float8Tensor were all unused). 3. xfail MXFP8BlockScaling + fp8_init + HSDP in FSDP2 model tests. Pre-existing bug (confirmed on main) where fsdp_post_all_gather receives fewer output tensors than fsdp_pre_all_gather sent when the HSDP shard dimension is trivial (size 1). Signed-off-by: Peter St. John <[email protected]> * address review comments Signed-off-by: Peter St. John <[email protected]> * address greptile review Signed-off-by: Peter St. John <[email protected]> * no need Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * revert grouped linear changes Signed-off-by: Varun Thumbe <[email protected]> * fix all tests Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix cudnn commit Signed-off-by: Varun Thumbe <[email protected]> * fix lint Signed-off-by: Varun Thumbe <[email protected]> * revert Signed-off-by: Varun Thumbe <[email protected]> * fix lint Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Peter St. John <[email protected]> Signed-off-by: Varun Thumbe <[email protected]> Signed-off-by: vthumbe1503 <[email protected]> Co-authored-by: Varun Thumbe <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Signed-off-by: Vladimir Cherepanov <[email protected]>

…rentScalingQuantizer (#2816) * Make Float8CurrentScalingQuantizer stateless (no amax/scale members) Remove amax, scale, and use_existing_amax from Float8CurrentScalingQuantizer on both C++ and Python sides. All amax/scale allocations are now ad-hoc at quantization time: - quantize() allocates a combined 2-element tensor for amax+scale - quantize_with_amax() accepts amax as a parameter - create_unquantized_tensor_with_amax() returns amax in a tuple - set_quantization_params() is now a no-op Update all call sites in activation.cpp, bias.cpp, normalization.cpp, attention.cpp, and cast.cpp to propagate the amax buffer. For FusedAdam FP8 kernel: when scale_ptr is null (current scaling), derive scale from scale_inv and skip writing amax/scale_inv metadata. Python side passes empty(0) tensors for scale/amax to signal this. Signed-off-by: Pawel Gadzinski <[email protected]> Made-with: Cursor * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Guard quantizer.amax/scale access with recipe.delayed() checks Add delayed() guards in context_parallel.py for amax writeback and scale cloning, since Float8CurrentScalingQuantizer no longer has these attributes. Allocate scratch scale buffer in _cast_master_weights_to_fp8_current_scaling instead of reading quantizer.scale. Signed-off-by: Pawel Gadzinski <[email protected]> Made-with: Cursor * Fix pylint unused-argument warnings in Float8CurrentScalingQuantizer The device, use_existing_amax, scale, and amax parameters are kept for backward compatibility but not used internally. Signed-off-by: Pawel Gadzinski <[email protected]> Made-with: Cursor * Clear dangling scale/amax pointers in Float8CurrentScalingQuantizer::quantize_impl amax_buf and scale_buf are caller-owned local tensors whose storage is released as soon as quantize() returns, leaving raw pointers stored in `out` dangling. No current caller reads out.scale()/out.amax() after quantize_impl returns, so this is currently safe, but it is a silent invariant that could turn into a use-after-free if new callers are added. Defensively clear both pointers at the end of quantize_impl (and in the empty-input early return), mirroring the existing set_amax(nullptr, ...) call already present before nvte_quantize_v2. Signed-off-by: Pawel Gadzinski <[email protected]> Made-with: Cursor * Route current-scaling Float8Tensor through FP32 master path in FusedAdam Float8Tensor with Float8CurrentScalingQuantizer now goes through the FP32 master + requantize path (same as MXFP8/NVFP4/blockwise) instead of the fused FP8 Adam kernel. The fused FP8 kernel stays for delayed- scaling Float8Tensor only. Also revert adam.cu to upstream — current scaling no longer needs the scale_inv-derived path in the kernel. Signed-off-by: Pawel Gadzinski <[email protected]> Made-with: Cursor * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Restore intent/perf comments in FusedAdam quantized dispatch - Note about a possible fused Adam+requantize kernel removing the FP32 round-trip for the QuantizedTensor path. - Justification for casting BF16/FP16 grads to FP32 before the FP32 Adam kernel. Also drop a stray f-string prefix on a literal-only RuntimeError part. Signed-off-by: Pawel Gadzinski <[email protected]> Made-with: Cursor * tests skip Signed-off-by: Pawel Gadzinski <[email protected]> * tests skip Signed-off-by: Pawel Gadzinski <[email protected]> * Warn on deprecated kwargs in Float8CurrentScalingQuantizer Signed-off-by: Pawel Gadzinski <[email protected]> * Guard remaining quantizer.scale/amax accesses with recipe.delayed() Followup to e06253d7. Three more spots accessed Float8CurrentScaling- Quantizer.scale/.amax (which no longer exist after the stateless refactor): - AttnFuncWithCPAndKVAllGather.forward used `not mxfp8()` instead of `delayed()` when cloning .scale, causing AttributeError under Float8CurrentScaling. - AttnFuncWithCPAndKVP2P fwd/bwd assigned .amax on per-step quantizer copies under `not mxfp8()`; harmless for current scaling (Python attaches a dynamic attribute that nobody reads) but inconsistent with the delayed()-guarded amax aggregation that follows. - print_quantizers (debug-only) read .scale/.amax for both DS and CS; restrict to DS only. Signed-off-by: Pawel Gadzinski <[email protected]> --------- Signed-off-by: Pawel Gadzinski <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

fix fp8 and is_bwd_fp8 relationship Signed-off-by: Charlene Yang <[email protected]>

… (#2728) * fix: scope get_full_cu_seqlens cache key by device and inference mode Signed-off-by: Dongmin Ra <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Dongmin Ra <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Charlene Yang <[email protected]>

Fix FA4 selection when FA3 is unavailable. Signed-off-by: Björn Buschkämper <[email protected]>

* Add AIR TopK support to TE JAX extension Adds a custom AIR TopK implementation (header-only, vendored into transformer_engine/common/util/) exposed as a JAX FFI custom call via the TE JAX extension. Key changes: - transformer_engine/common/util/air_topk.cu: AIR TopK CUDA kernel - transformer_engine/common/util/standalone_air_topk.cuh: vendored header - transformer_engine/common/include/transformer_engine/air_topk.h: C API - transformer_engine/jax/csrc/extensions/air_topk.cpp: JAX FFI binding - transformer_engine/jax/cpp_extensions/air_topk.py: Python wrapper - CMakeLists.txt: compile new kernel; use CCCL from CUDA toolkit - CMakeLists.txt: fix SM100 arch handling when all arches are special-cased Co-Authored-By: Claude Sonnet 4.6 <[email protected]> Signed-off-by: dcampora <[email protected]> Signed-off-by: Daniel Campora <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Signed-off-by: Daniel Campora <[email protected]> * Address PR review comments: fix namespace pollution, unused var, missing export, cache sm_cnt - Move WARP_SIZE/WARP_BITS/FULL_WARP_MASK/VECTORIZED_READ_SIZE into namespace nv - Remove unused keys_element_bytes variable in AirTopkFFI; collapse switch to dtype validation - Add missing `from .air_topk import *` export in jax/cpp_extensions/__init__.py - Cache sm_cnt per device with static vars to avoid repeated cudaGetDevice/cudaDeviceGetAttribute calls - Add CMAKE_BUILD_WITH_INSTALL_RPATH=ON to build_ext.py Signed-off-by: dcampora <[email protected]> Co-Authored-By: Claude Sonnet 4.6 <[email protected]> Signed-off-by: Daniel Campora <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Signed-off-by: Daniel Campora <[email protected]> * Rename air_topk -> topk throughout JAX extension Remove the `air_` prefix from all TopK-related identifiers: file names, C API functions (nvte_air_topk -> nvte_topk), FFI handler/primitive names (te_air_topk_ffi -> te_topk_ffi), Python symbols, and the internal `air_topk` namespace in standalone_topk.cuh. No functional changes. Signed-off-by: Diego Campora <[email protected]> Signed-off-by: dcampora <[email protected]> Signed-off-by: Daniel Campora <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Signed-off-by: Daniel Campora <[email protected]> * Address ptrendx review comments and fix CI lint issues - Follow TE workspace convention: remove nvte_get_topk_workspace_bytes() and implement empty-workspace size-query pattern in nvte_topk() instead - Remove unnecessary nv_detail::float_to_T helper; replace usages with static_cast<T>() directly - Remove unrelated CMAKE_CUDA_ARCHITECTURES OFF block from CMakeLists.txt - Fix cpplint errors in standalone_topk.cuh: replace unsigned long long int with uint64_t, add NOLINT for constexpr-sized arrays and else-with-comment Co-Authored-By: Claude Sonnet 4.6 <[email protected]> Signed-off-by: Daniel Campora <[email protected]> * Add assertion for 2D input in topk Co-Authored-By: Claude Sonnet 4.6 <[email protected]> Signed-off-by: Daniel Campora <[email protected]> --------- Signed-off-by: dcampora <[email protected]> Signed-off-by: Daniel Campora <[email protected]> Signed-off-by: Diego Campora <[email protected]> Co-authored-by: Claude Sonnet 4.6 <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

…ed quantization kernel (#2921) Fix the race in the dbias computation Signed-off-by: Przemek Tredak <[email protected]>

* remove ctype to eliminate memory usage from the cudnn kernel Signed-off-by: Varun Thumbe <[email protected]> * Remove c_dtype from fusible ops test Signed-off-by: Kirthi Shankar Sivamani <[email protected]> --------- Signed-off-by: Varun Thumbe <[email protected]> Signed-off-by: Kirthi Shankar Sivamani <[email protected]> Co-authored-by: Kirthi Shankar Sivamani <[email protected]>

Signed-off-by: Muu <[email protected]>

* [Common, PyTorch] Add triton mHC kernels & pytorch operators Signed-off-by: Kaining Zhong <[email protected]> * fix Signed-off-by: Kaining Zhong <[email protected]> * nit Signed-off-by: Kaining Zhong <[email protected]> * make linter happy Signed-off-by: Kaining Zhong <[email protected]> * nit Signed-off-by: Kaining Zhong <[email protected]> * ah OK Signed-off-by: Kaining Zhong <[email protected]> * new configs to improve perf Signed-off-by: Kaining Zhong <[email protected]> * add APIs to docs Signed-off-by: Kaining Zhong <[email protected]> * fix typos, check deterministic, refactor Signed-off-by: Kaining Zhong <[email protected]> * fix Signed-off-by: Kaining Zhong <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * reset rng for all tests Signed-off-by: Kaining Zhong <[email protected]> * add docstring Signed-off-by: Kaining Zhong <[email protected]> * fix api doc Signed-off-by: Kaining Zhong <[email protected]> * whoops Signed-off-by: Kaining Zhong <[email protected]> * grad_x doesn't have to zero Signed-off-by: Kaining Zhong <[email protected]> * nit Signed-off-by: Kaining Zhong <[email protected]> * nit Signed-off-by: Kaining Zhong <[email protected]> * force pytorch to not use bf16 for reduction Signed-off-by: Kaining Zhong <[email protected]> * use TE's general_gemm instead Signed-off-by: Kaining Zhong <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Looks like this is how to make TE use fp32 acc Signed-off-by: Kaining Zhong <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Kaining Zhong <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

…True (#2936) * fix Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add test Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * cleanup Signed-off-by: Varun Thumbe <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * zero_out should also be tested Signed-off-by: Varun Thumbe <[email protected]> --------- Signed-off-by: Varun Thumbe <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: root <[email protected]>

@Oleg-Goncharov

… (#2924) * Fix contiguous path for k=2880 Signed-off-by: Kirthi Shankar Sivamani <[email protected]> * format Signed-off-by: Kirthi Shankar Sivamani <[email protected]> * Review suggestion from @Oleg-Goncharov Signed-off-by: Kirthi Shankar Sivamani <[email protected]> * Add test for swizzle + padding fusion Signed-off-by: Kirthi Shankar Sivamani <[email protected]> * Address review comments Signed-off-by: Kirthi Shankar Sivamani <[email protected]> --------- Signed-off-by: Kirthi Shankar Sivamani <[email protected]>

…942) * accumulate bias in fp32 instead of bf16 in ref impl dbias to avoid accumulated numerical error Signed-off-by: tdophung <[email protected]>

* loosen up thresholds. Also only check min loss of last 10% of steps to avoid failing by noise near convergence Signed-off-by: tdophung <[email protected]> * add deterministic flag for mnist run Signed-off-by: tdophung <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* feat: add support for grouped GEMM swizzling with variable shapes and update C++ operator interface Signed-off-by: Abhishek <[email protected]> * Added confirmation with uniformity in one of the dimensions Signed-off-by: Abhishek <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Using single kernel for variable m and k Signed-off-by: Abhishek <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Cached blocks per sm for device and removed redundant checks Signed-off-by: Abhishek <[email protected]> * Updated the code with newer changes in main Signed-off-by: Abhishek <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Abhishek <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: vthumbe1503 <[email protected]>

…929) * Avoid removing usages from quantized weight in linear op Quantized weight tensor may be used across steps, so removing a usage is not safe. Signed-off-by: Tim Moon <[email protected]> * Tweak test to catch bug when alternating train and infer steps Signed-off-by: Tim Moon <[email protected]> * Avoid removing usages from quantized weights in grouped linear op Signed-off-by: Tim Moon <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Restore pre-forward quantizer config in ops Turns out we still need this in case the quantizer is used before the forward, e.g. in previous ops or CPU offloading. Signed-off-by: Tim Moon <[email protected]> * Blindly preserve quantizer usages in quantized weight params. Signed-off-by: Tim Moon <[email protected]> --------- Signed-off-by: Tim Moon <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

… (#2943) * Patch for NVFP4 test suite Signed-off-by: Oleg Goncharov <[email protected]> * C++ tests fix Signed-off-by: Oleg Goncharov <[email protected]> * Cleanup Signed-off-by: Oleg Goncharov <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Removed dead code Signed-off-by: Oleg Goncharov <[email protected]> * Set the golden value for amax in tests Signed-off-by: Oleg Goncharov <[email protected]> * Fixed memory leakage Signed-off-by: Oleg Goncharov <[email protected]> --------- Signed-off-by: Oleg Goncharov <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

… (#2948) * Switch to cuDNN-FE min version 1.23.0 to enable fused grouped MLP Signed-off-by: Kirthi Shankar Sivamani <[email protected]> * Fix tests Signed-off-by: Kirthi Shankar Sivamani <[email protected]> --------- Signed-off-by: Kirthi Shankar Sivamani <[email protected]>

…n expert (#2947) Add workaround for cuteDSL stride requirement for zero token expert Signed-off-by: Kirthi Shankar Sivamani <[email protected]>

Remove unused NVFP4 quantize kernel Signed-off-by: Tim Moon <[email protected]> Co-authored-by: Oleg Goncharov <[email protected]>

Follow-up to bffa8f0, trimmed to the necessary changes: - quantization_ref_nvfp4.py: keep only the be3feea revert (global scale); restore the incidental is_fp8_fnuz cleanup (no-op on gfx950, unreached on gfx942) to keep the diff minimal. - test_dequantize_nvfp4.cu: skip on gfx942 (te_fp8_fnuz) — the dequant kernel now uses OCP 448 while its reference is fnuz-240, so it fails there. - test_nvfp4_rht_quantize_exact.py: gate the ROCm-only test_hadamard_transform_amax on recipe_available (was IS_HIP_EXTENSION-only; uses the reverted reference). Drops the unnecessary cast/grouped-gemm cpp skips (those tests pass on gfx942). Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

68bc6e0 added the te_fp8_fnuz() skip to the CUDA-branch dequant bodies (#else of __HIP_PLATFORM_AMD__), which don't run on ROCm — so gfx942 still failed — and would break the CUDA build (te_fp8_fnuz is ROCm-only). Move the skip to the ROCm-branch DequantizeNVFP4 test body (the one that actually runs on gfx942) and drop it from the CUDA bodies. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

AllenFarcas · 2026-07-22T17:08:13Z

@ipanfilo @wangye805 related to your comment on tests/pytorch/distributed/test_fusible_ops.py line 1049 "TODO probably is not applicable - we know why the code is different. So it should rather be other comment, similar to test_cast_master_weights_to_fp8() which later solved the same problem"

I cannot comment there but ill add here my reply: I replaced the TODO with an explanatory comment in the # ROCm: style used in test_cast_master_weights_to_fp8.py: # ROCm: fnuz FP8 (gfx942) caps E4M3 at 240; OCP (gfx950+/CUDA) caps at 448., so the reference max must branch on is_fp8_fnuz(). Also removed an unused transformer_engine_torch import left behind.

ipanfilo

Number of files have AMD copyright header incorrect format

ipanfilo · 2026-07-23T14:09:11Z

 import transformer_engine_torch as tex

 from transformer_engine.common.recipe import Recipe
-from transformer_engine.pytorch.tensor._quantization_helpers import (


So on ROCm _quantized_helpers import quntized_tensor? Why?

AllenFarcas · 2026-07-23T14:40:06Z

@ipanfilo In response to your question about _quantized_helpers: the import was removed because it was the load-time dependency creating the quantized_tensor -> tensor/init -> quantized_tensor cycle on ROCm.

The cycle is ROCm-specific because init_mxfp4_extension() (guarded by #ifdef USE_ROCM in pybind) imports mxfp4_tensor -> triton_kernels.cast -> quantized_tensor at extension load, entering it before the tensor package; on CUDA that block is compiled out.

AllenFarcas · 2026-07-23T15:22:00Z

@ipanfilo I fixed all remaining AMD copyright and added any missing ones.

new_empty([]) crashed staging a rank-0 NVFP4 placeholder; shape it like the source instead. The per-tensor amax rides fsdp_pre_all_gather metadata, which async DCP bf16-casts, so async cannot be bit-exact -- use the documented loose tolerance (as with Float8CurrentScaling). Sync stays bit-exact. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

TE caches the FP8 weight columnwise transpose during forward instead of recreating it in backward, accumulating ~0.68 MiB/layer under FSDP2. The modules already skip caching when use_fsdp2 is set, which the test never did; enable use_fsdp2 on submodules in _shard_model after sharding. Also trims the verbose new_empty comment added in the prior NVFP4 async-DCP fix. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

# Conflicts: # transformer_engine/common/fused_attn_rocm/fused_attn.cpp

ipanfilo · 2026-07-24T02:08:01Z

@ipanfilo I fixed all remaining AMD copyright and added any missing ones.
It doesn't seem so. There are still files with wrong copyright header format

ipanfilo · 2026-07-24T01:15:41Z

 __device__ __forceinline__ float compute_global_encode_scaling_factor_FP4(const float global_amax) {
  using namespace detail;
-  constexpr float fp8_max = TypeExtrema<fp8e4m3>::max;  // 448.0f;
+  static_assert(E4M3_MAX == 448 || E4M3_MAX == 256, "Unsupported NVFP4 E4M3 max.");
+  constexpr float fp8_max = static_cast<float>(E4M3_MAX);


Are any of those changes ROCm specific or all come from upstream?

I reverted the copyright since this file is all from upstream.

AllenFarcas · 2026-07-24T14:34:08Z

@ipanfilo During my check of copyright I found a couple of files not touched by this IFU but which had AMD content but the copyright was missing: test_cast_transpose_current_scaling.cu, tests/jax/test_layer.py, jax/csrc/extensions/quantization.cpp, pytorch/csrc/pybind.h

I checked again and now everything else should be clean with the copyrights.

ptrendx and others added 30 commits April 20, 2026 11:45

Changed version to 2.16.0.dev0

fff2245

Signed-off-by: Przemek Tredak <[email protected]>

[PyTorch] Fix cuteDSL kernel incorrect numerics when K is 64 aligned …

2d92aa6

…(#2905) Zero out padded region when swizzling via group quantize Signed-off-by: Kirthi Shankar Sivamani <[email protected]>

Fix flash attention version check. (#2910)

4014f7f

Signed-off-by: Björn Buschkämper <[email protected]>

Make NS coefficients parameter 2D in Python API (#2904)

3c62f42

Signed-off-by: Vladimir Cherepanov <[email protected]>

[PyTorch] Fix CP A2A F16 when NVTE_FP8_DPA_BWD=1 (#2917)

424b031

fix fp8 and is_bwd_fp8 relationship Signed-off-by: Charlene Yang <[email protected]>

[PyTorch] Fix FA4 selection when FA3 is unavailable. (#2909)

9e55a25

Fix FA4 selection when FA3 is unavailable. Signed-off-by: Björn Buschkämper <[email protected]>

Fix the race in the dbias computation in MXFP8 quantization and group…

5d947a0

…ed quantization kernel (#2921) Fix the race in the dbias computation Signed-off-by: Przemek Tredak <[email protected]>

fix: TransformerEngineBaseModule quantizers init values type (#2927)

f2e31db

Signed-off-by: Muu <[email protected]>

[Common] Fix "0" literal for compilation (#2934)

82ace62

[JAX] Fix bf16 precision loss in TestGroupedDense reference dbias (#2…

cc05742

…942) * accumulate bias in fp32 instead of bf16 in ref impl dbias to avoid accumulated numerical error Signed-off-by: tdophung <[email protected]>

[PyTorch] Add workaround for cuteDSL stride requirement for zero-toke…

36fc336

…n expert (#2947) Add workaround for cuteDSL stride requirement for zero token expert Signed-off-by: Kirthi Shankar Sivamani <[email protected]>

[Core] Remove unused NVFP4 quantize kernel (#2946)

7e8bc98

Remove unused NVFP4 quantize kernel Signed-off-by: Tim Moon <[email protected]> Co-authored-by: Oleg Goncharov <[email protected]>

AllenFarcas and others added 2 commits July 22, 2026 14:07

Addressed PR reviews

94896e0

AllenFarcas requested review from ipanfilo, matthiasdiener, sudhu2k and wangye805 July 22, 2026 19:18

wangye805 approved these changes Jul 23, 2026

View reviewed changes

ipanfilo reviewed Jul 23, 2026

View reviewed changes

AllenFarcas added 2 commits July 23, 2026 16:02

Addressed PR comments

42ef2c4

Fixing GroupedLinear CI test

d1ea5ed

AllenFarcas requested review from ipanfilo July 23, 2026 16:13

AllenFarcas marked this pull request as ready for review July 23, 2026 16:13

Merge remote-tracking branch 'origin/dev' into IFU-dev-20260706-v2.17

97b5305

AllenFarcas added ci-level 1 CI test level 1 and removed ci-level 1 CI test level 1 labels Jul 23, 2026

AllenFarcas and others added 3 commits July 23, 2026 20:31

Merge remote-tracking branch 'origin/dev' into IFU-dev-20260706-v2.17

14113a3

# Conflicts: # transformer_engine/common/fused_attn_rocm/fused_attn.cpp

AllenFarcas added the ci-level 3 CI test level 3 label Jul 23, 2026

ipanfilo reviewed Jul 24, 2026

View reviewed changes

AllenFarcas added 2 commits July 24, 2026 14:01

Removed copyright

594d8fe

Added copyrights to files that were missing it

95d6cdf

Uh oh!

Conversation

AllenFarcas commented Jul 7, 2026

Description

Type of change

Changes

Checklist:

Uh oh!

AllenFarcas commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ipanfilo left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ipanfilo Jul 23, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

AllenFarcas commented Jul 23, 2026

Uh oh!

AllenFarcas commented Jul 23, 2026

Uh oh!

ipanfilo commented Jul 24, 2026

Uh oh!

ipanfilo Jul 24, 2026

Choose a reason for hiding this comment

Uh oh!

AllenFarcas Jul 24, 2026

Choose a reason for hiding this comment

Uh oh!

AllenFarcas commented Jul 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

20 participants

AllenFarcas commented Jul 22, 2026 •

edited

Loading