[Copilot]

Description

• Parsing: Read total_sequence_length using the supplied TensorProto’s type, accepting propagated int64 values and int32 constants while preserving single-element validation.
• Coverage: Add scalar and one-element regressions for constants and static-shape expressions, with and without past caches; verify inferred output dimensions.

Motivation and Context

Shape → Cast(INT32) → Squeeze produces a valid int32 sequence length, but shape propagation exposes its value as an int64 TensorProto. GQA’s unconditional int32 parsing rejects it during session creation with ParseData type mismatch, even when graph optimizations are disabled.

• Fixes GroupQueryAttention fails to load when total_sequence_length is computed from static shapes (ParseData type mismatch) #33865

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟢 Approval recommended

No active review issues are identified for the GroupQueryAttention sequence-length shape-inference change.

No open issues 🎉

What changed in this PR

Fixes GroupQueryAttention shape inference for int32 sequence lengths derived from static shapes, addressing the session-creation failure reported in #33865.

Changes:

• Parses sequence-length values using the supplied TensorProto type, supporting propagated int64 values and int32 constants while retaining single-element validation.
• Adds regression coverage for scalar and one-element inputs, with and without past caches, and checks inferred output dimensions.

File	Description
onnxruntime/​test/​contrib_ops/​group_query_attention_shape_inference_test.cc	Adds constant and shape-derived sequence-length regression cases.
onnxruntime/​core/​graph/​contrib_ops/​bert_defs.cc	Selects sequence-length parsing based on the TensorProto data type.

_{Review effort: Balanced}

[Tianlei Wu (tianleiwu)]

Maintenance update at 85f9bf12132941743ed1a75e6d5fb83bd31d95e9: the sequence-length changes, main refresh and reviewer fixes remain intact.

The separately reviewable native/navigation repairs are retained. e27bfcf3e9 completes AsyncWorker before resolve/reject dispatch can release its last reference, preserving once-only completion if dispatch throws. The original header's post-dispatch UAF was reproduced with real pinned Hermes/JSI; the fixed header passed 320 synchronized resolve/reject/cancel/Env-invalidation/dispatch-exception scenarios in each of two independent ASan+LSan processes, 640 total executions. f3b45608a9 gives the actual native Button a unique testID and uses by.id on both platforms without changing accessibility or the full14 assertion. All eight component/navigation cases passed, with all scoped linters passing and no introduced App type diagnostics; the pre-existing MNIST diagnostics were explicitly unchanged.

The old f3 iOS attempt 3 ran zero device cases because cold boot exceeded the original setup deadline, then cleanup/hanging Jest reached the job limit. 85f9bf1213 backports only the verified selected-simulator preparation workflow/docs repair: exact configured simulator selection, targeted reset, boot and app-service readiness. The 240-second Detox deadline, uninstall/reinstall policy, assertions and cleanup policy remain unchanged. All nine protocol cases executed against this branch's actual YAML passed.

The genuine current-head React Native workflow 42604665463 completed successfully. Fresh iOS artifact 13670441653 and Android artifact 13668343373 each contain exactly three real passing MNIST, Basic Types and EPContext cases, no failures/errors/skips, and the unchanged 14/14 checks passed assertion in device logs/traces. These are current85f9 executed simulator/device cases, not superseded f3 retries or zero-case setup reports.

The fresh complete all-20 audit at 2026-11-09T13:20:03Z verified this exact current85f9 head open, non-draft, mergeable, with zero unresolved threads and all current-head checks green. The current device and retained640-execution native proof remain valid. This PR's CI preparation is complete; normal human approval remains external. No tests, tolerances, runtime checks or hardware gates were weakened.