Default tolerant zephyr models#33
Conversation
|
Proof of concept that this works is already finished, IP approved and discussed publicly 2026-06-24-AQC-Jack Raymond.pptx . Codebase is already in public pull request dwavesystems/dwave-experimental#54. Just some differences in the training and archiecture conventions to understand (and test). |
|
Since this pull request was made there have been some method developments relative to fully_yielded_zephyr_subgraph. That need to be incorporated |
thisac
left a comment
thisac
left a comment
There was a problem hiding this comment.
Apart for some minor comments, LGTM. Would be good to have someone more familiar with the defect zephyr subgraph embedding implementation to review and approve before merging though.
| subgraph = greedy_get_subgraph(n_nodes=n_latents, random_seed=random_seed, graph=qpu_graph) | ||
| _, mapping = get_graph_mapping(subgraph) | ||
| latent_mapping = [mapping[node] for node in subgraph.nodes()] | ||
| except Exception: |
There was a problem hiding this comment.
Specify exception caught, or at least print error message? I assume you're trying to catch a specific error here for when greedy_get_subgraph fails for find a subgraph?
| raise | ||
| fallback = _try_zephyr_sublattice_fallback( | ||
| n_nodes=n_latents, qpu_graph=qpu_graph, random_seed=random_seed | ||
| ) | ||
| if fallback is None: | ||
| raise |
There was a problem hiding this comment.
Missing exceptions and messages.
| try: | ||
| samples = self._grbm.sample( # type: ignore | ||
| sampler=self.sampler, | ||
| prefactor=self.PREFACTOR, | ||
| linear_range=self.linear_range, | ||
| quadratic_range=self.quadratic_range, | ||
| device=spins.device, | ||
| sample_params=self.sampler_kwargs, | ||
| ) | ||
| except BinaryQuadraticModelStructureError: | ||
| self._rebuild_sampler() | ||
| samples = self._grbm.sample( # type: ignore | ||
| sampler=self.sampler, | ||
| prefactor=self.PREFACTOR, | ||
| linear_range=self.linear_range, | ||
| quadratic_range=self.quadratic_range, | ||
| device=spins.device, | ||
| sample_params=self.sampler_kwargs, | ||
| ) |
There was a problem hiding this comment.
For the sake of not copying this twice, you could:
| try: | |
| samples = self._grbm.sample( # type: ignore | |
| sampler=self.sampler, | |
| prefactor=self.PREFACTOR, | |
| linear_range=self.linear_range, | |
| quadratic_range=self.quadratic_range, | |
| device=spins.device, | |
| sample_params=self.sampler_kwargs, | |
| ) | |
| except BinaryQuadraticModelStructureError: | |
| self._rebuild_sampler() | |
| samples = self._grbm.sample( # type: ignore | |
| sampler=self.sampler, | |
| prefactor=self.PREFACTOR, | |
| linear_range=self.linear_range, | |
| quadratic_range=self.quadratic_range, | |
| device=spins.device, | |
| sample_params=self.sampler_kwargs, | |
| ) | |
| for _ in range(2): | |
| try: | |
| samples = self._grbm.sample( # type: ignore | |
| sampler=self.sampler, | |
| prefactor=self.PREFACTOR, | |
| linear_range=self.linear_range, | |
| quadratic_range=self.quadratic_range, | |
| device=spins.device, | |
| sample_params=self.sampler_kwargs, | |
| ) | |
| except BinaryQuadraticModelStructureError: | |
| self._rebuild_sampler() |
| try: | ||
| subgraph = greedy_get_subgraph(n_nodes=n_latents, random_seed=random_seed, graph=qpu_graph) | ||
| mapped_graph, mapping = get_graph_mapping(subgraph) | ||
| sampler = FixedEmbeddingComposite(qpu, {l_: [p] for p, l_ in mapping.items()}) | ||
| except Exception: | ||
| qpu_topology = qpu.properties["topology"]["type"] | ||
| if qpu_topology != "zephyr": | ||
| raise | ||
| fallback = _try_zephyr_sublattice_fallback( | ||
| n_nodes=n_latents, qpu_graph=qpu_graph, random_seed=random_seed | ||
| ) | ||
| if fallback is None: | ||
| raise | ||
| mapped_graph, chain_embedding = fallback | ||
| sampler = FixedEmbeddingComposite(qpu, chain_embedding) |
There was a problem hiding this comment.
Should this just be put into a helper function and imported into callback_helpers.py?
| import random | ||
| from typing import Callable, Literal, Optional | ||
|
|
||
| import dwave_networkx as dnx |
There was a problem hiding this comment.
Better use dwave-graphs here instead, since dnx is deprecated.
There was a problem hiding this comment.
Yeah I have this update locally along with the new implementation of Jacks changes I was just waiting for his PR to be merged so I can test before making a commit
There was a problem hiding this comment.
Although maybe it just worth merging these changes and making a followup PR once his PR is merged? Let me know your thoughts
3 participants
Feature Implemented
Losing a qubit from the Advantage2 typically renders the models (trained on the Advantage2) unusable. The changes in this PR allow for these models to continue to be used even after a qubit or two has been lost.
Note: The Advantage2 models in this branch are old and have not been updated to the new models. These models were trained prior to the last qubit loss and therefore should be helpful for testing.
AI Generation Disclosure
Copilot was given the fully_yielded_zephyr_subgraph file and implemented this guided but mostly unaided. I have reviewed it and tested it. Please review carefully.