Repository navigation
SEP 44: read the transverse edge where the state lives, and what else a native run pays the host for - #520
Conversation
A plane the step that reached it already read crossed the state, reduced it and only then found nothing to record. A replay decided its reads from dz_init, which it never updates: a short one read a step late, past the spacing, and a long one at every step. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
An atomic per element of the state took most of a monitor reading on a device: 184 ms on an NVIDIA H200 and 1.24 s on an AMD Instinct MI325X at 768 x 768 x 600, against steps of 239 and 224 ms. A runner's points now share the walk-off table too, which is a sixteenth of a state to upload. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
pack_energy summed every mode on one GPU thread, and took 53 of the 55 ms of an energy call on an NVIDIA H200 at 768 x 768 x 600. statistics_peak has the same shape. A low-storage step copied its whole table of partials to read one row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
A first saved plane took the state and the whole history on the host to compare their size with the limit, then dropped them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
The method behind the device measurements of the boundary check, in the tree so that they can be repeated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
A copy to fresh pageable memory takes its first-touch page faults inside the driver, one thread at a time: 1.8 GiB/s on an NVIDIA H200 and 7.9 GiB/s on an AMD Instinct MI325X, for a 5.27 GiB state, against about 52 GiB/s pinned. A bare vendor test of this scheme reached 33 and 50 GiB/s into a fresh huge-page buffer. Field, stage, held-plane and Adams-row reads and field uploads now take it above 256 MiB on a device build. Only the first such crossing allocates the ring, so a field-free run holds none. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
A mask of the whole state, on one thread, took twice the copy of it on a CPU and 0.73 s at 768 x 768 x 600 on an NVIDIA H200's host. A sum holding a NaN or an infinity isn't finite, so only one that overflows needs a second look. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
spectrum_to_optics allocated two state-sized arrays and took most of a saved plane once its crossing was fast. An in-memory store now converts into its own row, and its copy goes. Stored bits are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
The adapter always takes the general engine, so these eight tests held only code nothing can reach. They keep their assertions and tolerances. Refs #88. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
BoundState built the other engine only for a problem the adapter never sends. Its refusals and fallbacks go with the dispatch. Two members that never checked `require_open` keep their test of the pointer, which is null after `close`. Refs #88. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Nothing chooses an engine any more, so the adapter's constant and the query it fed have no reader. `_core` is private: the kernel schema does not change. Refs #88. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
SEP 44 phase 1. A backend now hands the edge an EdgeReading, numbers and never the field, through StepDriver._edge_reading. HostEdgeReduction is the reference, on a host copy, and native takes it until phase 2, so nothing moves yet. READ_RESOLUTION goes, so a read now synthesises every plane. The launch's safety-ramp judgement reads clean columns formed at the load, which needs no state on the host. A resume now restores a collar's current, baselines and peaks (#512). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Nothing can reach it: the general engine runs every problem, and makes every check this engine's validator made. `fft`, `multiply_phase` and `compact_problem` stay in core.cpp, and `_core.State` keeps its name. OccupancyScatter and its ScatterView include had no other user. GeneralState's monitor comments deferred to the deleted class's, so they take over its text. Refs #88. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
A load zeroes outside the active band, so a random field never came back bit for bit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
SEP 44 phase 2, for a run without a requested collar. The edge hands native its weight tables, four regions on the nodes, and native returns the region sums and each mode's energy. Native holds no face, mask or judgement, and a field-free default run brings no state to the host, which a test now holds it to. A collar still reduces a host copy. An inverse transform into the free scratch, then one sum: the same state read on both sides agrees within the transform's rounding on the radial, Cartesian, line and planar routes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
A reduction's value has to fit the device's shared memory per block. HIP's smallest block is 64 threads, and 128 doubles asked for 67,584 bytes of its 65,536, so every reading with buckets raised on an AMD Instinct MI325X. CUDA shrinks its block instead, so the NVIDIA H200 didn't show it, and an offline HIP compile can't. On HIP a reading now takes two passes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
`_device_build` stays private, as this work's new names do, and a test that unwraps the run to reach its refusal unwraps the body now. On a real device the first scan of the kept test kept a backend too, so its reference is now held to one a point. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
[skip ci] Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7c11633bb0
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| ): | ||
| simulation.run(capture, progress=progress, stop=stop) | ||
| return |
There was a problem hiding this comment.
Close the kept backend before running a non-device point
When a local scan runs a device-backed point followed by a CPU-backed or TDSE point, this early return bypasses the compatibility check and leaves the previous runner in self._kept. The new point therefore constructs its backend while the old backend still owns its device allocations and pinned host buffers (potentially 1.5 GiB for the crossing ring alone), and those resources remain held through subsequent non-device points until the scan finishes. Call self.close() before this fallback run so that switching to a point that cannot reuse the runner releases the previous backend first.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Right, thanks. Fixed in 3eec501d: a point off the device, or one with a TDSE, now closes the kept backend before it runs. test_a_point_off_the_device_closes_the_kept_backend_first (043de3ba) runs a device point, then one off the device, then a device point again. It checks that each builds its own backend and that every backend is closed. It fails without the fix.
A point that can't share the kept backend, because it is on the CPU or has a TDSE, returned before the check and left the last point's backend open. It then built its own beside it, holding the device's memory and the crossing ring through every such point until the scan ended. The method's docstring already promised two are never held. Found by the Codex review of this pull request. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
043de3b's message describes this line, but that commit holds only its test, since the edit hadn't applied when it was made. The test fails without this and passes with it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
It adds the rows the history's figures left out, and kerr_beam_384 at one thread narrowly misses the plan's aim, at x1.0803. It replaces the non-interleaved figures in the ledger and the SEP's report. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
The maintainer chose to set Final without an AMD pass at the head: a rented AMD Instinct MI325X had no capacity for a day, and an AMD Instinct MI300X failed twice before it started. The NVIDIA H200 and NVIDIA A40 ran the later commits, and the HIP build with tests compiles. What hasn't run on an AMD card is in #541, named in the completion report beside the other follow-ups. The SEP moves to the archive with its links, its roadmap entry goes, and the two manual pages it changed link to it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
…ve-edge The API index keeps both rows' additions, `CallbackOutput` from here and the field plane's file from main, and the layers figure is redrawn from the merged modules. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
GitHub retired its macOS 14 image: `minimum (macos-14)` was cancelled before a step ran, with "The configuration 'macos-14-arm64-us-default' is not supported". Publishing waits for that job, so a release would have waited on it too. The oldest macOS a hosted runner offers is now 15. The wheels still target 14.0, which no hosted machine now runs, and the workflow says so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
SEP 44's Final and today's main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ
One pull request for this work, at the maintainer's request, since each PR and each push costs a full CI run. It folds in #516, #517, #518, #522, #523 and #524, now closed, whose descriptions keep the full tables and per-section verification. The ledger entries in
docs/dev/benchmarks.mdhold the measurements. SEP 44's text, accepted in #515, is here too.The aim: a native run's field stays in its memory space from load to close, and only numbers cross, unless a caller asks to keep a field.
What's here
energyand the peak are real reductions;benchmarks.device_splitKERNEL_SCHEMA_VERSION; the deadnative::Stategoes; the run layer's duck-typed hooks are declaredstop=at each call's return, sosolprop runstops on SIGTERM rather than reporting a stop that didn't happenbeam_radius,fwhm_r,on_axis_intensityand theon_axis_*row reduce on the stateCallbackOutput, anOutputbuilt from a callable, so a reducing scan is a few lines; withfields="none"it holds no field volumeexpis skipped, input goes through tiles, and the accept is one reductionobserve_members(field=): a member's field crosses only at a plane the scan reads, under the rule a lone point keepsLocalscan on CUDA or HIP keeps its backend across points that differ only in inputsolprop runexits 0 and says so when a signal arrives too late to stop a runCloses #510, closes #511, closes #503, closes #502, closes #88, closes #385, closes #512, closes #519, closes #493, closes #499, closes #521, closes #299, closes #359, closes #362, closes #298, closes #514. #346 is mostly done and stays open for the plane's record; #385's Metal part isn't touched. Relates to #347, #396, #365, #361, #103, #357, #504 and #388. #151 (#525) is separate. What hasn't run on an AMD card is #541.
What it does for a run
On the data-centre cards, an NVIDIA H200 at
8cd170e2and an AMD Instinct MI325X at9a5d2346, against the same probe before this work. Field-free,python -m benchmarks.device_split. The MI325X's commit still took the collar a whole state at a time, now dropped, which adds 2.2 % to a step, and it hasn't the threaded finiteness pass. No AMD card has run this head (#541).field_to_host, 5.27 GiB, msenergy, msbeam_radius, msSEP 44's gates are met on both cards, and so is every target in the plan but one:
kerr_beam_384at one thread on an AMD EPYC 7763, below.python -m benchmarks ab --deviceagainstmainat8ed26dd7, 3 rounds, the same work:kerr_beam_384x0.051 on the H200 and x0.117 on the MI325X,kerr_beam_192x0.141 and x0.138, fields within 6.2e-16.A requested collar's read, at 384 x 384 x 600, native against the host's reduction of a host copy, with a step of 72 ms: the current 4.5 ms against 2.7 s on the H200 and 3.9 ms against 1.2 s on the MI325X, a read with its strip 18.7 ms against 12.9 s and 14.1 ms against 6.2 s.
A scan. A new backend costs 2.10 s of a 4.69 s point at 768 x 768 x 600 on the H200, and 0.98 s of 4.24 s at 384 x 384 x 600 over 51 steps, so a
Localscan of inputs on a device now keeps one (#298). A lockstep batch's plane no one reads falls from 5 to 7 ms of observing to 0.23 ms, of a plane interval of 23 to 28 ms (#299).A find, not fixed here (#529). FFT plans are the largest row of a device's memory, 21 of 39 GiB at 768 x 768 x 600, because each holds its own work area. At some transverse sizes a pair of them needs none: 900 x 900 holds less than 768 x 768 with 37 % more points, and 750 x 750 is 11 % faster with 5 % fewer. A follow-up shares one area across a device's plans: on the A40 at 768 x 768 x 600 the plans fall from 21.09 to 5.28 GiB and the run from 38.46 to 22.64 GiB, with the same fields and time.
Against 0.1.0, on an AMD EPYC 7763 at
ccd701d5, interleaved over 3 rounds, the same work:kerr_beam_192x1.059 at one thread and x0.902 at 16,kerr_beam_384x1.0803 and x0.941. The plan's target was under x1.08, sokerr_beam_384at one thread narrowly misses it.kerr_beam_192's final field differs from 0.1.0's by 1.8e-7, which is SEP 40's safety ramp: with the ramp off it's within 5.4e-16.On CPU,
python -m benchmarks abagainste49925d1, 3 rounds, the same work, fields within 4.2e-15: geometric mean 0.85, withkerr_beam_192at 16 threads x0.48. SEP 40's four-way control, default check against no ramp and no reads:line+15.8 %,planar+7.6 %,radial_filament_256+4.1 %,kerr_beam_192+8.3 %.Setup, at 768 x 768 x 600: on an Intel Xeon Gold 6240
from_free169–179 s → 14.9 s andtranslate28–30 s → 7.2–7.8 s. On the H200's host 5.31 s and 3.65 s, on the MI325X's 3.41 s and 1.63 s.The full tables, with what a run holds and the Python share, are in
docs/dev/benchmarks.md, "The native read on the data-centre cards".How it was verified
9a5d2346:pytest -m gpuovertests130 passed, 1 skipped, and overnative/tests31 passed; the edge's, collar's, widths' and statistics' modules with the transverse backends 269 passed, 3 skipped;test_transverse_reads.py,test_transverse_monitor.pyandtest_stop.py66 passed; the targeted device set 210 passed, 7 skipped. The MI325X is the first AMD card to run the read.8cd170e2, the node-chunk collar, the batch's field and the replaced input included: the edge's, collar's, split rows', batch's, widths', statistics' and transverse modules 326 passed, 3 skipped; the scan batch, runner, stop and transverse modules 135 passed;pytest -m gpuon the far field, widths and statistics 177 passed and 3 failed. All three are one seed of the new fuzz that drew only masked bins, a bug in the test, since fixed and passing on CPU in every seed.ccd701d5:pytest -m gpu tests130 passed, 2 skipped; the split rows', edge's and collar's read modules 172 passed; the transverse backends, widths, statistics and march lifecycle 118 passed, 3 skipped. Atec7b5bab, with the kept backend: the kept-backend, scan and batch modules 149 passed;pytest -m gpuon the far field, with the fuzz in all 120 cases, 134 passed.9fd4ceab: all green, 60 checks: every test lane, 100 % coverage, the docs, clang-tidy, the sanitizers, the slow lane and the wheels on every platform.f4fd0913, the head: all green, 63 checks, with today'smainmerged in, so Compile the CUDA and HIP targets in CI #525's twodevice-compilelanes, HIP under ROCm 7.2 and CUDA 12.8, run on this branch for the first time and pass. The wheels' minimum-systems lane failed at8c30cfdcbecause GitHub has retired its macOS 14 image, so that lane now checks macOS 15, the oldest hosted macOS; the wheels still target 14.0, which no hosted machine runs now.3eec501d: HIP under ROCm 7.2.4 and CUDA 12.8.1, tests included and warnings as errors, both compile. Nothing undernative/changes after that commit except whatmainbrought.AGENTS.mdsays. One full gate was run here by mistake, early on.Three separate agents, given the code and not my conclusions, read SEP 44's collar, the statistics and the device step for faults. They found no wrong number. They did find a race on FFTW's planner, a resume test that couldn't fail, a device test that ran one block three times, a device-only attribute missing from the seam's tuple and several weak tests, all fixed. Seven deliberate breaks of the collar and 18 of the device step were each caught, the last two of those only after tests were added.
Not run
9a5d2346(#541). A rented MI325X had no capacity for a day, and an MI300X failed twice before it started. The HIP build with tests compiles at the head, and the only C++ changed since that commit keeps the collar at the node chunk, which the MI325X measured, and answers the lint and tidy lanes. SEP 44's later commits haven't run on an AMD card #541 lists what to run there:pytest -m gpu, a whole scan on a kept backend, the fuzz on Far-field aperture shapes, a custom absorbed_power channel, and the boundary items #497's base and rocFFT's work areas.Decided by the maintainer
A rental for phase 4 was approved, with a cap of US$15, later US$20, or US$25 for an AMD Instinct MI325X, for SEP 44's later commits haven't run on an AMD card #541's pass. It has spent US$12.22, none of it left running. US$5.92 of that went on eight machines that never came up: four NVIDIA H200s that never answered, two AMD Instinct MI300Xs, and two H200 creates that failed on the provider's side.
check_buildloses itsproblemargument, a public signature, with a fragment. An edit undernative/refuses an editable build until rebuilt, even a comment.This work's new names are private, since
solpropandsolprop-nativeare released together:_EdgeReading,_EdgeTables,_EdgeLine,_HostEdgeReduction,_edge_lines,_transverse_edge_bytesand_STRIP_BATCH_BYTESinsolprop.transverse_monitor, and_FluenceRings,_fluence_ringsand_ring_fluence_widthsinsolprop.spatial.The widths' refusal band stays. Native's budget uses a Cauchy-Schwarz bound on the reference's magnitude, so a negative value between the two budgets, about 4e-11 of the peak, is refused by
numpy_refand read as rounding by native. The oracle is left alone.The native Adams call's bound stays 0.1 s, as SEP 39 recorded. A stop waits for at most that and the attempt in flight, however long an attempt takes, as on every other march.
kerr_beam_384at one thread stays at x1.0803 against 0.1.0. The other rows meet the target and both 16-thread rows are faster, and getting it back would probably mean restoring plane skipping on CPU.Keeping a scan's backend is the default on a device, with no setting. It's limited to
Localscans on CUDA and HIP, and a point's record differs only increatedandplanning_seconds.SEP 44 is
Finalin this PR's last commit, without the AMD pass, which SEP 44's later commits haven't run on an AMD card #541 carries. The SEP is in the archive, and the two manual pages it changed link to it.CallbackOutput's design stands. It implementsOutputdirectly, notPlaneStore, whose save converts into a new volume and which counts as aCheckpointOutput, so a stop would log thatresume=Truecontinues. So it has nooutcome, and TDSE is refused.fieldsstays a run setting and not a handler argument. An emptystat_namesmeans optmat runs lose the peak-field report andMaterialRangeWarning, which is documented, with an info log at run time.check_cache,resumeandresumabilitywiden their annotations toOutputin order to refuse it by name.The low-storage march on a GPU is 1.35 to 1.67 times a single-transform engine #365 stays open. A dedicated engine takes 159.4 ms a step at 21.19 GiB on the H200 at 768 x 768 x 600, against 220 ms and 38.96 GiB here. The axis order on the volume route and A device's FFT plans each hold a work area, four states' worth at 768 x 768 #529's shared work area go in one follow-up PR, not here.
Backendstays a protocol for now. MakingStepDriverthe contract is noted on Write a SEP for the kernel seam: the integrator and its tolerance, and the seam as a union of routes #110, as a simplification of its own.The whole-block collar stays dropped here,
ae82a2a9. It missed The node chunk is sized at translation, whatever the device, and each chunk is a kernel launch #361's 10 % on both cards, x0.996 and x0.978, so a device keeps the node chunk andTransform.chunk_budgetedgoes. Share one work area across a device's FFT plans of the scratch #542 brings it back on a device without a seam field, and keeps it only if an interleaved AMD run gains in every round; otherwise it goes for good, and The node chunk is sized at translation, whatever the device, and each chunk is a kernel launch #361 says so.Known and left: a brief 128 MiB reallocation of far-field partials at 2048 x 2048 x 2049 with two polarisations, by arithmetic;
test_the_native_pass_does_not_depend_on_the_thread_countasserts the host's threads and fails on a CUDA build, from before this work.🤖 Generated with Claude Code
https://claude.ai/code/session_01XGENPtq9Gp7VCh5QmcDdzQ