IFU dev v2.6

[wangye805]

Description

targeted NV upstream commit: ca7407e on 2025/07/18 based on our rocm dev commit 6bbd03c
Fixes https://github.com/ROCm/frameworks-internal/issues/13729

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

See NV upstream release doc for upstream changes.
Our IFU conflict resolving are listed in the following commits:
1). common: 4d3ca4d
2). jax extension: 5ce0afd
3). pytorch extension: c9c9126
4). build/installation: 9730903
5). cpp gtests: 51bdbb8
6). pytorch pytests: ba59f81
7). jax pytests: 5842c24

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

[alextmagro]

LGTM -- only covered common dir and cpp tests.

[Micky774]

I mainly focused on the attention sections, and the install/build. Overall looks good, just some minor nits/questions.

[Micky774]

Why do we not use is_training = config.head_dim_qk <= 192 and config.head_dim_v <= 128 as in line 400 later in this function for determining the available backend? Won't this potentially cause issues if the later is_training=False, where we could have had a certain backend enabled at this step but didn't because we assumed is_training=True? Or is that not a problem?

[wangye805]

Missed this conflict with git merge. Fixed. Thanks

[Micky774]

Indent to keep aligned

[wangye805]

Done. Thanks

[Micky774]

To clarify, this means that when building TE one must have both JAX and PyTorch installed in order to build even just for a single framework right?

[wangye805]

I'm not quite familiar with this new pyproject.toml. Based on my experience with pip install --no-build-isolation, the source build does not interact with this pyproject.toml in my mind.

Please correct me if I was wrong

[AllenFarcas]

For lines 369-371

# Deallocate GEMM input tensor if no longer needed
if not weight.requires_grad and not return_layernorm_output:
    ln_out = ln_out_total = None
    clear_tensor_data(ln_out, ln_out_total)

we shouldn't make ln_out and ln_out_total None first, we should clear tensor data first and then make them none

Also my commit ([Feat] Add transpose cache to LayerNorm kernel (#279) ) had instead:

if not weight.requires_grad:
    if not return_layernorm_output:
        clear_tensor_data(ln_out, ln_out_total)
    ln_out = None

[wangye805]

Done. Thanks

[ipanfilo]

I think test_checkpoint does not involve calling different fused_attn backends so whole this addition should be under if [ $_fus_attn = "$_DEFAULT_FUSED_ATTN" ]

[wangye805]

In fact, I did find this test involves attention:

TransformerEngine/tests/pytorch/test_checkpoint.py

Line 64 in a00236c

return te.TransformerLayer(1, 1, 1)

[ipanfilo]

It indeed creates TransformerLayer but it neither calls fwd nor bwd so FA backend is called. It tests saving of Torch.nn.module derivative classes state save/loading. In fact, even Flash vs Fused attn does not make difference in the state

[wangye805]

I see. Done. Thanks