GPUEngine.Reshape silently ignores dst argument (compute.Engine contract violation)

[dndungu]

Problem

compute/gpu_engine_memory.go:614 GPUEngine.Reshape takes a variadic dst ...*tensor.TensorNumeric[T] parameter, matching the compute.Engine contract shared with CPUEngine.Reshape. However the zero-copy GPUStorage fast-path at line 644 ignores dst entirely:

// GPUStorage[T]: zero-copy reshape.
if gs, ok := a.GetStorage().(*tensor.GPUStorage[T]); ok && isFloat32[T]() && newSize == currentSize {
    return tensor.NewWithStorage[T](inferredShape, gs.View(gs.Len()))
}

It returns a brand-new tensor aliasing a's storage and never touches the caller-provided dst. Callers that (like other ops in the engine) discard the return value and rely on dst being the updated tensor silently receive stale pre-allocated storage.

Impact

This is what caused the PatchTST GPU training convergence regression tracked in #79 — see the postmortem on that issue. A single-line zerfoo workaround (use the return value) restored convergence, but any future caller that follows the same pattern as other engine ops (pass a dst, discard the return) will re-hit the silent-zero trap.

Expected behavior

When dst is provided, Reshape should either:

• (a) SetStorage on dst[0] to alias the view, matching the zero-copy semantics and leaving the GPUStorage shared with a; or
• (b) copy the source data into dst[0]'s existing storage, matching the caller's intent to use its pre-allocated buffer.

Option (a) preserves the zero-copy win and is consistent with how other engine ops mutate dst. Either choice must be a deliberate documented contract — currently the behavior is neither.

Repro

See .claude/scratch/wave-7-probe-logs.txt in zerfoo main (commit 1ebdb787) for first-batch GPU probe evidence showing fc.dFlat populated by MatMul and fc.dX (dst of Reshape) left all-zero with a different storage pointer.

Suggested fix

In gpu_engine_memory.go:644, after constructing the view, if len(dst) > 0 && dst[0] != nil, update dst[0]'s storage and shape (SetStorage + shape update) and return dst[0]. Add a test in compute/ that passes a pre-allocated dst with poisoned data and asserts the returned tensor AND dst both reflect the reshaped source.

Related

• GPU engine dst-output routing: tensors written by kernels return zero via .Data() — freezes PatchTST training on GB10 #79 (root cause investigation — closed)
• fix(timeseries): PatchTST GPU convergence — use Reshape return value zerfoo#371 (zerfoo-side workaround — merged)