Repository navigation
Conversation
BDonnot
commented
Sep 11, 2026
BDonnot
commented
Contributor
- a diffirentiable powerflow there
Assisted-by: Claude Code (Sonnet 5) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>
Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>
Assisted-by: Claude Code (Fable 5.1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>
…ernel Assisted-by: Claude Code (Opus 5) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>
Three bugs in the persistent-driver path of PR #14, each with a regression test: - ScenarioSweepSession::run(): a hot run (only injections / gen_v changed) keeps the live batch source but used to clear every row's disconnected / masked_buses flags, which only a new source recomputes. Islanded rows still came back NaN while get_disconnected() and n_disconnected reported nothing. The path is now chosen first and the flags are reset only when a new source is built. - ScenarioSweepSession::set_branch_data(): the live driver uploads the branch admittances / flow buffers once, so data set after the first run never reached the GPU (wrong flows, and an out-of-bounds write in zero_branch_flows_kernel if the branch count grew). It now invalidates the driver's cached branch data so the next run / compute_flows re-uploads. - BatchPowerFlow: the topology and generator masks were cached before the session received them; a call raising in between (e.g. a bad gen_status shape) left the cache ahead of the session, and the next call with the same line_status was treated as unchanged and solved on the old trips. The masks are now committed only once the session accepted them. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01FZwHuL9k42bdoU93GqaPQW Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>
Found by a second review of PR #14 (including cc28dd5); each fix comes with a regression test. - BatchPowerFlow: the "session still holds a trip list / generator mask for another row count" check compared against _last_n_scen, which is updated before the op runs. After a call that failed inside the op, the next call with the same row count sent nothing and every run() was then refused ("row count no longer matches"). The row count of what the session holds is now recorded when the session accepts it. - ScenarioSweepSession::run(): a warm run kept the chunk capacity the cold run sized for its own active rows (1 when all but one row was islanded), so a later warm run with every row active was solved in one-row chunks. The driver is now rebuilt when the live capacity would need more than twice a fresh driver's chunk count; the old driver is freed before the new one is allocated. - compute_limit_violations = False had no effect on a reused driver: the fused check stayed armed with the old limits. run() now disarms it. - set_branch_data(): refuses data that drops a branch the current topology trips (before replacing anything), rebuilds the per-row Ybus patches from the new admittances, and drops per-branch limits of a different length so compute_limit_violations asks for set_limits() again instead of reading past their end. - differentiable compute_flows(): a half-open branch's open end (bus -1) read V[-1] and vn_kv[-1], i.e. the last bus. It now follows compute_branch_flows_kernel: V = 0 and no terminal current on that side, base current from the live endpoint. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01FZwHuL9k42bdoU93GqaPQW Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>
Each fix comes with a regression test. - BatchPfDriver::solve_JT_batch: refuse a forward that ran in several chunks. Every adjoint buffer holds one chunk; with a caller-supplied J snapshot (which passes the shape check) the rhs gather wrote n_active rows into it, out of bounds, and the gen_v kernel read past d_V_batch / d_Ybus_values_batch. - ScenarioSweepGPU: send gen_v to the session with the generators that set_contingency_gens() takes out of a row NaN'd there (re-sent when the mask changes). The conflict check already ignored them, but the session still applied their set-point, so an off generator could impose its voltage on a bus another connected generator keeps PV. - ScenarioSweepSession::run(): with keep_final_jacobian, also rebuild the driver when the live capacity needs more than one chunk and a fresh driver would not; the forward otherwise refused a batch_size that was already large enough, on every topology change. - ScenarioSweepSession::run(): count run_counter on entry and clear last_run_kept_jacobian_. A run() that threw after replacing the source or the driver left the counter unchanged, so a pending alias-mode BatchPowerFlow backward read the new buffers as its own. - acpf_nr.cu: the single-system ||F||_inf reductions (NR loop and the presolved check) used thrust::maximum from 0, which drops NaN: an all-NaN F read as 0 and was reported converged. They now propagate NaN like the batched compute_residuals_kernel. test_matching_alg's NaN case now also asserts converged is False; docstrings updated. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01FZwHuL9k42bdoU93GqaPQW Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>
Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>
Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>
BDonnot
added a commit
that referenced
this pull request
Sep 30, 2026
Brings in the build CI (#16), the Modal GPU test job (#17) and the scenario-sweep / adjoint fixes of #14. Conflict resolutions: - scenario_sweep_session.cu: keep main's warm path (rebuild the driver when the live chunk capacity would split the active rows into too many chunks, or break keep_final_jacobian's single chunk) and its reset_rows lambda; drop main's per-row slack weights from the ScenarioSweepBatch constructor, since this branch hands them over after the build through set_slack_redistribution. - _batch_pf.py: take main's rule that _topology_mask only changes once the session accepted the topology; _row_masks (redistribute_slack's pre-pass, which runs before the op hands the topology over) now reads the queued topology and keys its cache on that tensor. - _flows.py: per-terminal ampere bases (this branch) with main's half-open handling; a -1 side falls back to the live end's vn_kv, as set_branch_data does. - scenario_sweep/gpu_facade.py: main's _push_gen_v helper, with this branch's gen_v_bus mapping. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01Afi372v35Q3tbaDtypd9Fa Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.