What
The QDP documentation has stale CUDA tensor dtype guidance:
qdp/qdp-python/README.md says the pipeline float32 batch path is amplitude-only.docs/qdp/getting-started.md says batched angle CUDA tensor inputs require float64.
The implementation now supports float32 CUDA tensor inputs for angle and basis single-sample and batched encoding paths.
Why
Users reading the docs may unnecessarily cast supported CUDA tensors to float64, or assume that angle and basis float32 batched tensor paths are not available.
That makes the docs disagree with the current QDP CUDA implementation and tests.
How
Update the QDP Python README and getting-started guide to document the current CUDA tensor dtype matrix:
amplitude: float32 or float64 CUDA tensors, single-sample or batchedangle: float32 or float64 CUDA tensors, single-sample or batchedbasis: int64 or float32 CUDA index tensors, single-sample or batchediqp / iqp-z: float64 CUDA tensorsphase: host input only for now
Also clarify that streaming basis files still read indices as float64 because float32 cannot exactly represent large integer basis indices.
What
The QDP documentation has stale CUDA tensor dtype guidance:
qdp/qdp-python/README.mdsays the pipeline float32 batch path is amplitude-only.docs/qdp/getting-started.mdsays batched angle CUDA tensor inputs require float64.The implementation now supports float32 CUDA tensor inputs for angle and basis single-sample and batched encoding paths.
Why
Users reading the docs may unnecessarily cast supported CUDA tensors to float64, or assume that angle and basis float32 batched tensor paths are not available.
That makes the docs disagree with the current QDP CUDA implementation and tests.
How
Update the QDP Python README and getting-started guide to document the current CUDA tensor dtype matrix:
amplitude: float32 or float64 CUDA tensors, single-sample or batchedangle: float32 or float64 CUDA tensors, single-sample or batchedbasis: int64 or float32 CUDA index tensors, single-sample or batchediqp/iqp-z: float64 CUDA tensorsphase: host input only for nowAlso clarify that streaming basis files still read indices as float64 because float32 cannot exactly represent large integer basis indices.