vGPU: Add "Ignore Memory Limit" option to bypass GPU memory check on task submission#3343
vGPU: Add "Ignore Memory Limit" option to bypass GPU memory check on task submission#3343QingengWei wants to merge 1 commit intodevelopfrom
Conversation
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 4 potential issues.
Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
| #web.configure("Ltrvqel7oCenUTH88Pqh99vn7ikCD25KFPZ0phz2Mxtgl5I4") | ||
|
|
||
| #Env.uat.active() | ||
| #web.configure("LmpSvRP0MGOuKOgm9ZJn97l9RE8t5I2ENTI9RLwbXlmmW89Z") |
There was a problem hiding this comment.
Hardcoded API keys committed in test file
High Severity
Production API keys are hardcoded in tests/tidy3d_modal_cm_fdtd_test.py. Three separate API keys (for prod, dev, and uat environments) are exposed in plaintext via web.configure(...) calls. These credentials will be visible in version control history even if later removed and could allow unauthorized access to the service.
| "protocolVersion": protocol_version, | ||
| "workerGroup": worker_group, | ||
| "vgpu_allocation": vgpu_allocation, | ||
| "ignore_memory_limit": ignore_memory_limit, |
There was a problem hiding this comment.
Snake_case API keys instead of camelCase in submit
High Severity
In the BatchTaskCore.submit method, the HTTP POST body uses snake_case keys "vgpu_allocation" and "ignore_memory_limit" instead of the camelCase "vgpuAllocation" and "ignoreMemoryLimit" used by the other submit method (for Tidy3dTask) and by the other keys in the same dictionary ("solverVersion", "protocolVersion", "workerGroup"). The server likely won't recognize these misnamed keys, so the parameters will be silently ignored.
| max_num_adjoint_per_fwd=max_num_adjoint_per_fwd, | ||
| numerical_structures=numerical_structures, | ||
| custom_vjp=custom_vjp, | ||
| ignore_memory_limit=ignore_memory_limit, |
There was a problem hiding this comment.
ignore_memory_limit dropped in two run_custom code paths
High Severity
The run_custom function accepts ignore_memory_limit and passes it correctly for the component modeler path (line 478), but the autograd _run() call (lines 499–520) and the non-autograd webapi.run() call (lines 531–549) both omit it. For regular Simulation objects, the parameter is silently ignored regardless of the code path taken.
Additional Locations (1)
| "element_mappings": [], | ||
| "custom_source_time": null, | ||
| "type": "ModalComponentModeler" | ||
| } No newline at end of file |
There was a problem hiding this comment.
Test data file accidentally committed to repository
Medium Severity
A large 1500-line simulation data JSON file (tests/modal_cm.json) appears to have been accidentally committed alongside the ad-hoc test script tests/tidy3d_modal_cm_fdtd_test.py. This is test fixture data for a manual debugging script, not part of the automated test suite, and adds unnecessary bulk to the repository.
|
Thanks I will make sure this gets in for |
|
@daquinteroflex Please let me know when the test version is released next week, and I will deploy it to UAT so QA can start testing. |
|
@shanguomagi let's discuss in private but if you can show me the dev branch I should be able to test there first and then you should have a version available to you in flex possibly by tomorrow |
3 participants
Note
Medium Risk
Touches the end-to-end task submission path and alters server payloads, so mis-wiring could change execution behavior for vGPU jobs; additionally, the committed test script includes hard-coded credentials that should not ship.
Overview
Adds a new optional
ignore_memory_limitflag that is threaded through the public web run surface (web.run,run_async,Job.start,Batch.run/start, and S-matrix modeler runs) down to task submission payloads (TaskCore.submitand RF batch submit), enabling vGPU users to bypass backend GPU memory limit checks when starting tasks.Includes new test fixtures: a large
ModalComponentModelerJSON and a helper script to upload/run it against the web API (currently containing hard-coded API tokens).Written by Cursor Bugbot for commit df1b836. This will update automatically on new commits. Configure here.