Fix CPU LSTM input validation and refactor ValidateInputs to frontends

[Youngsik Yang (vacu9708)]

Summary

Follow-up to #27737, which fixed out-of-bounds shape reads in the CUDA LSTM path but explicitly deferred the CPU counterpart. This PR addresses CPU LSTM validation gaps.

---

Problems fixed

1. Structural: validation in the backend instead of the frontend

W's third dimension input_size must equal X's third dimension input_size at runtime,
Validating W therefore depends on X. Beyond this data dependency,
input validation is not the responsibility of the computation. ComputeImpl()
should receive already-validated inputs, not validate them itself.
In addition, ValidateInputs() no longer receives six individual tensors from the caller.

2. Bug: out-of-bounds dimension reads before rank check

In ValidateInputs(), X.Shape()[0] and X.Shape()[1] were read before the NumDimensions() != 3 guard:

// Before — UB on rank < 2
auto& X_shape = X.Shape();
int64_t seq_length = X_shape[0];   // ← read before rank check
int64_t batch_size = X_shape[1];   // ← read before rank check
if (X_shape.NumDimensions() != 3)
  return error;

// After — guard first
if (X_shape.NumDimensions() != 3)
  return error;
int64_t seq_length = X_shape[0];
int64_t batch_size = X_shape[1];

This mirrors the fix applied to the CUDA path in #27737.

3. Missing: CPU LSTM never validated W or R shapes

Neither DeepCpuLstmOp::Compute() nor DynamicQuantizeLSTM::Compute() validated W or R shapes.

DynamicQuantizeLSTM uses transposed W and R layouts.
[num_directions, input_size, 4*hidden_size] and [num_directions, hidden_size, 4*hidden_size] respectively
That's why W/R validation cannot be integrated into ValidateInputs()

---

Changes

File	Change
lstm_base.h	ValidateInputs() declaration: (OpKernelContext& context)
lstm_base.cc	1. ValidateInputs() fetches inputs internally. 2. Rank check reordered before dimension reads. 3. call removed from ComputeImpl
deep_cpu_lstm.cc	ValidateInputs(*context) + W/R validation called at top of Compute()
dynamic_quantize_lstm.cc	Same; W/R validation added for the first time (transposed W layout: [num_directions, input_size, 4*hidden_size])
deep_cpu_lstm_op_test.cc	LSTMTest.InvalidInputShapes: wrong-rank W, wrong-dim W, wrong-rank R, wrong-dim R; plus X-rank regression test against ONNX shape inference
quantize_lstm_op_test.cc	DynamicQuantLSTMTest.InvalidInputShapes: same four cases for the quantized path

[Youngsik Yang (vacu9708)]

@microsoft-github-policy-service agree

[Copilot]

Pull request overview

This PR closes CPU-side validation gaps for LSTM kernels by moving common input validation to the frontend (Compute()), fixing an out-of-bounds dimension-read hazard in ValidateInputs(), and adding missing runtime shape validation for W and R (including the transposed layout used by DynamicQuantizeLSTM).

Changes:

• Refactor LSTMBase::ValidateInputs to take OpKernelContext&, fetch inputs internally, and guard X rank before reading dimensions.
• Move ValidateInputs(*context) calls into CPU LSTM frontends and add explicit W/R shape checks in DeepCpuLstmOp and DynamicQuantizeLSTM.
• Add negative tests covering invalid W/R ranks and dimensions for both float and dynamic-quantized CPU paths.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
onnxruntime/core/providers/cpu/rnn/lstm_base.h	Changes ValidateInputs API to accept OpKernelContext& for frontend-driven validation.
onnxruntime/core/providers/cpu/rnn/lstm_base.cc	Implements the refactor; fixes rank-check ordering to avoid OOB reads; removes validation from ComputeImpl.
onnxruntime/core/providers/cpu/rnn/deep_cpu_lstm.cc	Calls ValidateInputs up-front and adds missing W/R runtime shape validation for CPU LSTM.
onnxruntime/contrib_ops/cpu/quantization/dynamic_quantize_lstm.cc	Calls ValidateInputs up-front and adds W/R runtime shape validation for the transposed quantized layout.
onnxruntime/test/providers/cpu/rnn/deep_cpu_lstm_op_test.cc	Adds CPU LSTM invalid-shape tests (including an X-rank regression case).
onnxruntime/test/contrib_ops/quantize_lstm_op_test.cc	Adds DynamicQuantizeLSTM invalid-shape tests for W/R.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.