[All] Bump minimum supported cuDNN version to 9.11

[cyanguwa]

Description

This PR raises the minimum supported cuDNN version from 9.3 to 9.11 in Transformer Engine. 9.11 is about a year old and we encourage users to update their cuDNN version if they haven't.

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

• README.rst: bump documented minimum cuDNN under System Requirements from 9.3+ to 9.11+.
• docs/installation.rst: bump the cuDNN prerequisite from cuDNN 9.3 or later to cuDNN 9.11 or later.

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

[greptile-apps]

Greptile Summary

This PR updates the documented minimum cuDNN version from 9.3 to 9.11 across the two user-facing documentation files. No source code, build system, or test logic is modified.

• README.rst: System Requirements table entry for cuDNN updated from 9.3+ to 9.11+.
• docs/installation.rst: Prerequisites list item updated from cuDNN 9.3 or later to cuDNN 9.11 or later.

Confidence Score: 5/5

Safe to merge — changes are confined to two documentation files with no impact on build logic or runtime behavior.

Both changes are single-line documentation edits that align the stated prerequisite with the team's new minimum. No code, build scripts, or tests are touched, so there is no risk of regression.

Files Needing Attention: No files require special attention. The runtime source code still contains version-guard branches for cuDNN < 9.3 that are now unreachable given the new minimum, but cleaning those up is a separate refactoring task.

Important Files Changed

Filename	Overview
README.rst	Single-line documentation update: cuDNN minimum version in System Requirements bumped from 9.3+ to 9.11+.
docs/installation.rst	Single-line documentation update: cuDNN prerequisite bumped from "cuDNN 9.3 or later" to "cuDNN 9.11 or later".

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User reads docs] --> B{cuDNN version?}
    B -- "< 9.11 (old: < 9.3)" --> C[Unsupported — upgrade required]
    B -- ">= 9.11" --> D[Supported — proceed with install]
    D --> E[Build Transformer Engine]

Loading

_{Reviews (3): Last reviewed commit: "Merge branch 'main' into update-min-cudn..." | Re-trigger Greptile}

[KshitijLakhani]

IIUC, the minimum version check does not mean that we cannot build with <9.11 right ? It just means that older versions (features) are just unsupported, right ? So the 9.11 is merely a lowest tested/supported version and not a hard runtime/build floor right ?
If my understand is right we do not need to worry about compilation checks during build and these:

TransformerEngine/transformer_engine/jax/csrc/extensions/attention.cpp

Line 186 in 65bdbe3

if (is_ragged && cudnn_runtime_version >= 90300) {

?

TransformerEngine/transformer_engine/jax/cpp_extensions/attention.py

Line 602 in 65bdbe3

# cuDNN version < 9.3.0:

TransformerEngine/transformer_engine/common/fused_attn/fused_attn.cpp

Line 364 in 65bdbe3

(cudnn_runtime_version >= 90300 &&

TransformerEngine/tests/pytorch/attention/test_attention.py

Line 1035 in 65bdbe3

# cuDNN 9.3.0+ is required to run pad_between_seqs = False/True in the same run

---

However, I think these below two might needs anotehr look right ?

• Wouldn't the notebook here need a refresh : docs/examples/attention/attention.ipynb ?
• Also, would we want to refresh the assert in here: tutorial_generation_gemma_with_te.ipynb ?

[sbhavani]

For min version, maybe we could match with a previous PyTorch NGC container (e.g. cuDNN 9.12.0.46 for 25.08) from frameworks support matrix.

[cyanguwa]

I'll do a bit more cleanup throughout the code to support this update, and will update to cuDNN 9.12.