[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

Bumps cudnn-frontend to v1.22 and adds a cuDNN 9.20 head-dimension OR branch for Blackwell (SM≥100, non-paged, fprop+bprop, any sq) inside the existing flag_arb check in nvte_get_fused_attn_backend. The parenthesis accounting is correct: the trailing && on the 9.11 clause is correctly changed to || to insert the new sub-condition inside the head-dim OR group, and all outer gates (architecture, mask type, QKV format, sliding window, determinism) continue to apply to the 9.20 path unchanged.

Confidence Score: 5/5

Safe to merge — the logic change is structurally correct and all pre-existing runtime gates still apply to the new 9.20 path.

No new P0/P1 findings. The &&→|| restructuring is correct: it moves the closing ) of the 9.11 sub-condition one level inward and appends the 9.20 clause as a sibling OR branch before the outer group closes, keeping the bug-exclusion && in the right place. The 9.20 condition inherits all surrounding guards (SM≥100 arch check, mask-type block, QKV-format block, Blackwell determinism check), so no previously-gated combinations are accidentally opened beyond what cuDNN 9.20 is expected to support. The one open question about sq=1 + causal on Blackwell/9.20 was already raised in a prior review thread.

No files require special attention.

Important Files Changed

Filename	Overview
transformer_engine/common/fused_attn/fused_attn.cpp	Adds cuDNN 9.20 head-dim OR branch for Blackwell (SM≥100, fprop+bprop, non-paged); changes trailing && to `
3rdparty/cudnn-frontend	Submodule bump from 7b9b711c to 97f6cb3b (cudnn-frontend v1.22); no source changes in this repo.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["F16/BF16 path entered"] --> B{"arch check\nSM≥100 requires\ncuDNN≥9.7"}
    B -->|pass| C{"head_dim % 8 == 0\nfor qk and v?"}
    B -->|fail| Z["flag_arb = false"]
    C -->|yes| D{"Which head_dim\ncondition matches?"}
    C -->|no| Z
    D --> D1["≤128 (any version)"]
    D --> D2["≤256 + Hopper\n(9.1 fprop / 9.5 bprop)"]
    D --> D3["any dim + Blackwell\n+ fprop + non-paged\n+ sq>1  (9.9)"]
    D --> D4["any dim + fprop\n+ any arch (9.10.2)"]
    D --> D5["dqk=192/dv=128\n+ Blackwell + bprop\n+ non-paged (9.11)"]
    D --> D6["any dim + Blackwell\n+ fprop/bprop\n+ non-paged (9.20 NEW)"]
    D1 & D2 & D3 & D4 & D5 & D6 --> E{"Hopper bprop\nbug exclusion\n(sm==90 only)"}
    E -->|not blocked| F{"mask type,\nformat, SWA,\ndeterminism\ngates"}
    E -->|blocked| Z
    F -->|all pass| G["flag_arb = true\n→ NVTE_F16_arbitrary_seqlen"]
    F -->|any fail| Z

Loading

_{Reviews (2): 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