Dont batch unless separate samplers are used for each subcircuit #333
Conversation
caleb-johnson
added
quantum performance
Pull Request Test Coverage Report for Build 5623635187
💛 - Coveralls |
| num_qpd_bits = np.reshape( | ||
| num_qpd_bits_flat, (len(subexperiments), len(subexperiments[0])) | ||
| ) | ||
| quasi_dists_reshaped: list[list[QuasiDistribution]] = [[] for _ in subexperiments] |
There was a problem hiding this comment.
Can't rely on numpy here anymore because there are now potentially different subcircuits coming in on the same batch, and those can have different numbers of subexperiments associated with them; therefore, we can't trivially reshape to the input size anymore.
We now have to "stream" the quasi-dists into the output data structure, using the input as a guide for how to re-build the structure
|
If we get this one merged soon, it's probably worth a patch release to 0.2.1 due to the large amount of time saved in cases where the circuit is separated and the user runs on quantum hardware with no reservation |
That reminds me: this definitely deserves a release note. |
This thought occurred to me before my coffee. We already have new features in main, it's too late to consider a bugfix release probably :) |
Now that CKT supports Python 3.11, we might as well use the latest version in the Dockerfile.
* Add support for `SwapGate` * Reorder terms [ci skip] * Add missing terms * DRY the coefficients * Fix coverage * Add support for `iSwapGate` * Fix black * Add to release note * Fix type hint * Gates without parameters are nicer to work with and can be singletons, one day! * Remove a line * Add comments describing channels * `_copy_unique_sublists` * Add `DCXGate` * Tweak * Implement cutting of general 2-qubit unitaries Builds on #294. Closes #186. * Add tests of additional gates * Fix type annotation * Add explanatory comments * `supported_gates()` -> `explicitly_supported_gates()` * Add to references * Improved error message and test about `to_matrix` conversion failing * Add xref to `QPDBasis` in docstrings * Add `qpdbasis_from_gate` to Sphinx build * Make `explicitly_supported_gates` private and remove its release note It's not clear that this function remains useful now that we support essentially all 2-qubit gates. If we find a use for it in the future, we can re-introduce it (or something like it) as a public interface. * Fix intersphinx link * Release note * Update qpd.py: remove extraneous `from None`
I've also tried to make the developer documentation easier to discover
* Make the repository link more obvious from the Sphinx build * Enable "edit" link in the header * Add comment
* Add docs badge to link to stable docs * Add ruff badge * Remove ruff
Making directions on opening the docs a little more succinct.
* Implement wire cutting as a two-qubit instruction * Update type annotation * s/gate/instruction/ * Add overhead test for `Move` instruction * Add wire cutting tutorial * Add `Move` to Sphinx build * Doc updates suggested by Caleb * Add release note and link to new tutorial * Clarify wording following #174 (comment) * Improvements to `Move` docstring * Use svg as the plot format This avoids pixelation on high-dpi displays * Remove unnecessary uses of `CircuitInstruction` https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/pull/174/files#r1278067109 * The notebook tests should ignore any files that crop up in `docs/_build` `matplotlib.sphinxext.plot_directive` likes to leave python files there
|
@garrison the change log should now be correct in this PR now, so it is actually reviewable :) |
There was a problem hiding this comment.
Thanks, this looks reasonable, and it appears basically correct as far as I can tell without working out every last detail explicitly about the orderings in these reshaped lists. In order to raise my confidence to a pretty high bar, I tested this with a modified roundtrip test, and everything works out. I will submit my modified test in a PR soon, as I think it is worth doing on an ongoing basis. Thanks!
This is my PR promised at #333 (review). Essentially, I really want the roundtrip tests to regularly test both experiment workflows: the one that uses a single sampler, and the one that uses multiple samplers. Of course, all sampler(s) used should be `ExactSampler`s, and the expectation values are required to work out exactly, up to floating point error. This PR achieves this by flipping a coin each time a roundtrip test is called.
This is my PR promised at #333 (review). Essentially, I really want the roundtrip tests to regularly test both experiment workflows: the one that uses a single sampler, and the one that uses multiple samplers. Of course, all sampler(s) used should be `ExactSampler`s, and the expectation values are required to work out exactly, up to floating point error. This PR achieves this by flipping a coin each time a roundtrip test is called.
Resolves #330