Fix BART attention fusion for SDPA pattern from transformers >= 4.49

[Rishi-Dave]

Summary

• Add SDPA-aware pattern matching to FusionBartAttention so that BART attention fusion succeeds on models exported with HuggingFace Transformers >= 4.49
• Add a synthetic BART SDPA graph generator and unit test

Motivation

Fixes #23864

HuggingFace Transformers >= 4.49 replaced BartAttention with BartSdpaAttention (commit 2c47618), changing the ONNX export graph topology in several ways that broke FusionBartAttention pattern matching. Running optimize_model(..., model_type="bart") on these newer exports produces zero fused Attention nodes.

Changes

fusion_bart_attention.py

The SDPA refactor introduces four structural changes to the attention subgraph. Each required a new match path:

1. QKV output path — LayerNormalization anchor fallback
   For SDPA models, symbolic shape inference often fails, which prevents SkipLayerNormalization fusion. When the anchor node is a plain LayerNormalization instead of SkipLayerNormalization, there's an extra residual Add between the LayerNorm and the attention output projection. Added a fallback match: ["Add", "Add", "MatMul", "Reshape", "Transpose", "MatMul"] with [0, None, 0, 0, 0, 0].

2. QK path — NaN guard (Where + IsNaN)
   SDPA wraps the Softmax output in a NaN guard: Where(IsNaN(softmax), 0.0, softmax). The Where node's input[2] is the Softmax output. Added two new QK paths:

   • No mask: ["Where", "Softmax", "MatMul"] with [0, 2, 0]
   • With mask: ["Where", "Softmax", "Add", "MatMul"] with [0, 2, 0, 0]

3. Q and K scaling paths
   Instead of a single combined scale on the QK MatMul output, SDPA applies separate Mul(1/sqrt(head_dim)) to Q and K before the QK MatMul. Added:

   • Q path: ["Mul", "Transpose", "Reshape", "Add", "MatMul"] with [0, 0, 0, 0, None]
   • K path: ["Mul", "Reshape", "Transpose", "Reshape", "Transpose", "Reshape", "Add", "MatMul"] with [1, 0, 0, 0, 0, 0, 0, None] (K^T uses a Reshape→Transpose(0,2,1)→Reshape chain)

4. num_heads fallback for dynamic shapes
   SDPA models use -1 in reshape shape tensors for dynamic dimensions, causing get_num_heads_and_hidden_size to return negative values. Added a fallback to user-specified num_heads/hidden_size when detected values are invalid.

bart_model_generator.py (new)

Synthetic BART SDPA attention graph generator that builds a minimal but complete attention subgraph matching the SDPA topology. Tests both with_mask=True (decoder self-attention) and with_mask=False (encoder attention) variants.

test_attention_fusion.py

Added test_bart_attention_sdpa_fusion that verifies:

• 1 Attention node is produced for each mask variant
• Correct num_heads attribute
• Correct unidirectional attribute (1 for decoder self-attention with mask, 0 for encoder)

Test Plan

• python -m pytest test_attention_fusion.py -v — all 10 tests pass
• lintrunner on all 3 changed files — no issues
• Verified on real exported BART SDPA model (hf-internal-testing/tiny-random-bart): 2 Attention nodes fused, graph reduced from 120 → 34 nodes

[tianleiwu]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[tianleiwu]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).