📢 cuDNN Frontend v1.14 — Release Notes

Preferred version for: cuDNN 9.12.0 and above
Minimum Python version: 3.9 (previously 3.8, now obsolete)
Updated pip wheels: Available for Python 3.13

🚀 Improvements

🔹 SDPA

• Introduced a unified SDPA node → reduces graph creation latency & simplifies the SDPA graph creation.
  (No API changes required from users.)
• Improved support checks for SDPA FP8 forward propagation (fprop).
• Improved support checks for Hopper backward propagation (bprop) to fix a bug (introduced in 9.11) affecting certain large head-dimension combinations of d_qk and d_v.
• Added new SDPA samples with:
  • Paged prefill
  • Ragged Q tensor decode

🔹 Normalizations

• Added support for fused LayerNorm with ReLU.
• Included sample: LayerNorm with ReLU bitmask dump.

🔹 Matmul

• Added Python sample for low-precision FP8/FP4 matrix multiplications.

🔹 Other Updates

• Added Python bindings for deviceless graph compilation.
  → Sample: test_deviceless_aot_compilation.py
• Addressed GitHub issue: #151.

✅ Recommended Action: Upgrade to cuDNN Frontend v1.14,0 for full compatibility with cuDNN 9.12.0+, improved SDPA support, additional normalization support, and deviceless graph compilation features.