[CUDNN] Update frontend to version 1.22 and add cuDNN 9.20 path for SM arch >100

[zmelumian972]

Summary

• Updates the cudnn-frontend submodule to version 1.22 (97f6cb3b)
• Adds a new cuDNN 9.20 path in nvte_get_fused_attn_backend for Blackwell (SM arch >= 100) that supports any head dimension, both forward and backward passes, non-paged layouts, and any sequence length

Changes

• 3rdparty/cudnn-frontend: Bump submodule from 7b9b711c to 97f6cb3b (cuDNN frontend v1.22)
• transformer_engine/common/fused_attn/fused_attn.cpp: Add cuDNN 9.20 backend selection condition:
  • Enables FusedAttn_F16_Arbitrary_Seqlen backend for SM >= 100 + cuDNN >= 9.20 + non-paged KV layouts
  • Fixes the logical operator joining the 9.11 condition from && to || to correctly OR the two Blackwell conditions

Test plan

• Verify FusedAttention with cuDNN 9.20+ on Blackwell (SM >= 100) hardware
• Confirm existing Hopper (SM 90) paths are unaffected
• Run fused attention unit tests for paged/non-paged layouts

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR updates the cudnn-frontend submodule to v1.22 and adds a cuDNN 9.20 backend-selection path in nvte_get_fused_attn_backend for Blackwell GPUs (SM≥100), enabling any head dimension for both forward and backward passes on non-paged KV layouts with any sequence length.

• 3rdparty/cudnn-frontend: Bumped from 7b9b711c to 97f6cb3b (v1.22) — straightforward dependency update.
• fused_attn.cpp: A new (sm_arch_ >= 100 && cudnn_runtime_version >= 92000 && layout_group != NVTE_Paged_KV_HD_HD_HD) branch is OR-ed with the existing 9.11 Blackwell-bprop condition. Prior to this PR, Blackwell bprop was restricted to head dims 192/128 only (9.11 path). The 9.20 path opens bprop to any head dimension (multiples of 8) on non-paged Blackwell.
• The && → || operator change on the 9.11 line is structurally necessary: it extends the OR-chain inside the head-dim group before the outer closing ) that separates it from the 9.11+ Hopper bug workaround &&. The parenthesisation is correct and the Hopper-specific bug workaround (sm_arch_ == 90) remains unaffected by the new Blackwell path.

Confidence Score: 4/5

PR is safe to merge; the logic change is structurally correct with one edge case worth verifying: sq=1+causal+fprop support in cuDNN 9.20 on Blackwell.

The parenthesis structure and operator fix are correct. The 9.20 condition is intentionally broader than predecessors (no seqlen or head-dim upper bound), consistent with the claimed cuDNN 9.20 capabilities for Blackwell. The only uncertainty is whether sq=1+causal/padding_causal fprop — previously excluded in the 9.10.2 path — is now supported by cuDNN 9.20 on Blackwell. All other downstream constraints (bias, mask type, qkv format, sliding window) still apply.

transformer_engine/common/fused_attn/fused_attn.cpp — verify cuDNN 9.20 supports sq=1+causal+non-paged+fprop on Blackwell before production use.

Important Files Changed

Filename	Overview
transformer_engine/common/fused_attn/fused_attn.cpp	Adds cuDNN 9.20 backend path for Blackwell (any head_dim, fprop+bprop, non-paged, any seqlen); fixes operator to correctly OR the 9.11 and 9.20 conditions within the head-dim group
3rdparty/cudnn-frontend	Submodule bump from 7b9b711c to 97f6cb3b (cuDNN frontend v1.22); straightforward dependency update

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[nvte_get_fused_attn_backend
F16/BF16 path] --> B{Architecture OK?
cudnn>=9.0.7 && sm>=100}
    B -- Yes --> C{head_dim % 8 == 0?}
    B -- No --> X[Other arch/version paths]
    C -- Yes --> D{head_dim <= 128?}
    D -- Yes --> PASS[Head-dim condition met]
    D -- No --> E{cudnn>=9.1/9.5 + Hopper
head_dim <= 256?}
    E -- Yes --> PASS
    E -- No --> F{cudnn>=9.9 + Blackwell
fprop + non-paged + sq>1?}
    F -- Yes --> PASS
    F -- No --> G{cudnn>=9.10.2
fprop + various layouts?}
    G -- Yes --> PASS
    G -- No --> H{cudnn>=9.11 + Blackwell
bprop + non-paged
d_qk=192 d_v=128?}
    H -- Yes --> PASS
    H -- No --> I{cudnn>=9.20 + Blackwell
non-paged + any sq
fprop or bprop NEW}
    I -- Yes --> PASS
    I -- No --> FAIL[No head-dim match]
    PASS --> J{Hopper 9.11+ bprop
bug workaround
sm==90 check}
    J -- Blocked --> FAIL
    J -- OK --> K[bias / mask / format /
sliding-window checks]
    K --> BACKEND[FusedAttn_F16_Arbitrary_Seqlen]

Loading

_{Reviews (1): Last reviewed commit: "FusedAttention: Add cudnn 9.20 path for ..." | Re-trigger Greptile}

[greptile-apps]

Verify sq=1 + causal/padding_causal fprop support in cuDNN 9.20

The 9.20 condition allows any max_seqlen_q (including sq = 1) with any mask type on non-paged Blackwell layouts. The preceding 9.10.2 fprop path explicitly excluded sq = 1 + causal and sq = 1 + padding_causal on non-paged layouts:

(max_seqlen_q == 1 && attn_mask_type != NVTE_Mask_Type::NVTE_CAUSAL_MASK &&
 attn_mask_type != NVTE_Mask_Type::NVTE_PADDING_CAUSAL_MASK)

With the 9.20 path (any sq, no mask-type restriction at the head-dim level), sq=1 + causal + non-paged + fprop on Blackwell/cuDNN≥9.20 will now pass this gate — where it was previously blocked. If cuDNN 9.20 lifts this restriction for SM≥100, this is correct. If not, passing this combination to the backend would produce a runtime error. Please confirm whether cuDNN 9.20 actually supports this combination on Blackwell.

[KshitijLakhani]

/te-ci jax L0