ryankert01 opened a new pull request, #1366:
URL: https://github.com/apache/mahout/pull/1366
## Problem
`QdpBenchmark.run_latency()` and `run_throughput()` raised `TypeError` on
any call because `api.py` was unconditionally passing `dtype="f32"` to
`run_throughput_pipeline_py()` — but `QdpBenchmark` had no `_dtype` field and
no `.dtype()` builder method. The Rust binding already accepted `dtype`
(defaulting to `"f64"`) but the Python layer never wired it up.
## Changes
- Add `_dtype: str = "f64"` to `QdpBenchmark.__init__`
- Add `.dtype("f32" | "f64")` builder method with validation
- `"f32"` activates the zero-copy float32 batch path where the encoding
supports it; encodings without an f32 kernel fall back to f64 inside the Rust
pipeline automatically
- `"f64"` is the default, matching the Rust binding default and the
PyTorch reference path
- Pass `dtype=self._dtype` to both Rust call sites (`_run_throughput_rust`,
`_run_latency_rust`)
## Test
```python
from qumat_qdp import QdpBenchmark
for dt in ("f64", "f32"):
r = QdpBenchmark(device_id=0).qubits(4).encoding("amplitude").batches(4,
size=8).warmup(1).dtype(dt).run_latency()
print(f"dtype={dt}: {r.latency_ms_per_vector:.3f} ms/vector")
# dtype=f64: 0.019 ms/vector
# dtype=f32: 0.013 ms/vector
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
