[Bug][QDP] Align Iris GPU training path with sibling pipelines (real dtype + leaf weights)

[ryankert01]

Related Issues

Closes #1160

Changes

• Bug fix

Why

The Iris amplitude-encoding GPU training path (_run_training_gpu) had
drifted from its sibling pipelines (mnist_amplitude.py, svhn_iqp.py)
in two ways, both of which break it:

1. Complex dtype leaks into trainable params/labels (issue [need triage] we cast from complex64 to real value everytime #1160).
   dtype = encoded_train.dtype is complex64/complex128 from the QDP
   amplitude encoding. That complex dtype was inherited by the weights
   (used as qml.Rot angles), the bias, and the ±1 labels — all of
   which must be real.

2. Non-leaf weight tensor. weights = 0.01 * torch.randn(..., requires_grad=True)
   is a non-leaf tensor (the multiply has a grad_fn), so
   torch.optim.SGD rejects it with can't optimize a non-leaf Tensor,
   crashing the path before training starts.

Both are already handled correctly in the sibling pipelines; Iris was
simply not aligned.

How

Mirror the siblings:

• Derive a real dtype: real_dtype = torch.float64 if encoded_train.dtype == torch.complex128 else torch.float32, used for weights, bias and labels (encoded state vector stays complex for StatePrep).
• Initialise weights as a leaf: (0.01 * torch.randn(...)).detach().requires_grad_(True).

Split into two commits so each fix can be reviewed independently.

Verification

qumat_qdp / lightning.gpu / CUDA were not available in my environment,
so I exercised the real _run_training_gpu directly on CPU with
default.qubit and a synthetic complex64 state-vector batch (stubbing
the unused qumat_qdp import). Before/after:

	weights dtype	optimizer	training
before	complex64	rejects (non-leaf)	crashes
after	float32	accepts (leaf)	runs 10 epochs, returns results

Note: this confirms the code path executes with correct real dtypes and
gradients flow; it does not validate model accuracy (synthetic input, not
real QDP-encoded Iris). End-to-end accuracy validation still needs the
qumat_qdp extension + a CUDA/lightning.gpu stack.

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

_{No automated tests exist for these benchmark pipelines (the GPU path needs CUDA + lightning.gpu); verified manually as described above.}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.