AMD Life Science 26.10 release notes#
3 min read time
AMD Life Science 26.10 adds geometric-nn-ops 1.0.0. The release adds fused triangle attention, the triangle multiplicative update, attention with pair bias, and a hipequivariance-ops-torch compatibility package for protein structure prediction models such as OpenFold3.
Note
hipCIM, MONAI on ROCm, MONAI Model Zoo, and MONAILabel are available in the AMD Life Science 26.08 release.
AMD Life Science release highlights#
The following are notable new features and improvements in AMD Life Science 26.10 since the release of 26.08. For detailed changes to individual components, see detailed component changes.
geometric-nn-ops 1.0.0: This is the first release of geometric-nn-ops. geometric-nn-ops is a library of Triton GPU kernels for protein structure prediction models such as OpenFold3. The same kernel source runs on AMD GPUs with ROCm. The release adds fused triangle attention, the triangle multiplicative update, attention with pair bias, fp32 precision control, and a
hipequivariance-ops-torchcompatibility package.
AMD Life Science components#
The following table lists the versions of AMD Life Science components for AMD Life Science 26.10. Click the GitHub icon to go to the component’s source code.
| Category | Component name | Version | Source code |
|---|---|---|---|
| AI/ML | geometric-nn-ops | 1.0.0 |
Note
hipCIM, MONAI on ROCm, MONAI Model Zoo, and MONAILabel are available in the AMD Life Science 26.08 release.
Detailed component changes#
The following sections describe changes to the AMD Life Science components.
geometric-nn-ops (1.0.0)#
This release is the first release of geometric-nn-ops and includes the following enhancements:
Added#
Fused triangle attention:
triangle_attentionhas native Triton forward and backward passes. It supports an optional pair bias and a boolean key mask. It runs bf16, fp16, and fp32.dynamic=Trueavoids recompiling for each sequence length.use_exp2=Trueselects a base-2 exponential softmax.Triangle multiplicative update:
triangle_multiplicative_updatehas a fused Triton forward pass and acuequivariance-compatible signature. When gradients are needed, it switches to a differentiable PyTorch implementation.Attention with pair bias:
attention_pair_biasimplements Algorithm 24 from the AlphaFold 3 paper in PyTorch. The interface matchescuequivariance0.9, and it can cache the projected pair bias.fp32 precision control: The Triton kernels follow
torch.backends.cuda.matmul.fp32_precision. With"tf32", they use a matmul that is at least TF32-accurate, for example XF32 MFMA on AMD CDNA3.Removed bias arguments:
attention_pair_biasno longer acceptsln_z_biasorproj_z_bias. The softmax cancels these parameters, so they can’t affect the output.Drop-in replacement for
cuequivariance_ops_torch: Thehipequivariance-ops-torchpackage lets unmodifiedcuequivariance_torch0.9.1 code run on AMD GPUs. The shim translates tensor layouts and output ranks to matchcuequivariance.Installation guards: Importing the shim raises an error if
cuequivariance-ops-torchwheels are also installed. It warns if the installedcuequivariance-torchversion isn’t 0.9.1.
Resolved issues#
These changes are in the triangle attention kernel originally contributed to OpenFold3.
Fixed NaN output in fp16 when a whole tile of keys was masked.
Fixed incorrect gradients for fully masked rows in the backward pass.