Add BF16 kernels in several ops for Gemma-3 #26102
kunal-vaishnavi
merged
15 commits into
microsoft:main
from
kunal-vaishnavi:kvaishnavi/gemma3-bf16
Remove duplicate op counting
9916a246
Register SkipLayerNorm for bfloat16
8c3f997a
Fix warnings during build
980f9d8c
Add Conv-22 with bfloat16
5c7151c3
Add Pow-15 with bfloat16
b4d030a9
Add AveragePool-22 with bfloat16
27242b7c
Add weight-only quantization for MatMulNBits with bfloat16
92d3524d
nenad1002
commented
on 2025-09-23
tianleiwu
commented
on 2025-09-23
Add end versions to op registrations
daa3fc77
Add unit tests
69cb2153
Return early if BF16 is not supported
7d6119a0
Remove BF16 CUDA NHWC kernels
82349667
Merge branch 'main' into kvaishnavi/gemma3-bf16
7d6cee3d
Fix build break with BF16 Conv test
fc9b62b4
Increase threshold for BF16 CUDA SkipLayerNorm test
cc483622
Add failing ONNX tests to list
3d1d2df7
tianleiwu
commented
on 2025-09-25
tianleiwu
approved these changes
on 2025-09-25
Login to write a write a comment.
Login via GitHub