Replace the C physics port with a Numba accelerator - #37

Merged

swissskimmilk merged 20 commits into

mainfrom

engine-design/numba-accelerator

Sep 5, 2026

Commits on Sep 4, 2026

Add numba physics kernels and 3-way Layer-1 benchmark, declare numba dep
swissskimmilk
committed
db6ef59
Free numba kernels from the C EdEngineState with pure-Python param extraction
swissskimmilk
committed
c308962
Port ablative cooling to numba, closing the last coverage gap vs the C kernel
swissskimmilk
committed
235cd2b
Add live A/B parity suite for the numba accelerator vs C and Python
swissskimmilk
committed
f3ee62d
Move accelerator into engine/accel and drop the C injector solve from evaluate
swissskimmilk
committed
aea1d6a
Route production through engine.accel: solve, chamber_solve, dispatcher, call sites
swissskimmilk
committed
53a610b
CI: replace native-parity with a matrixed accel-parity job covering numba and C
swissskimmilk
committed
c3296fd
Vectorise the chug sweep and add a numba kernel for it, removing the last C dependency
swissskimmilk
committed
c4e56d6
Retire the C oracle from the parity suite; guard numba against Python at 1e-6
swissskimmilk
committed
80d023d
Drop the ED_ACCEL=c dispatch branch now that the parity suite no longer needs it
swissskimmilk
committed
4561b11
Delete the C physics port; numba is the accelerator
swissskimmilk
committed
f76e667
Port the pintle injector to numba, extending accelerated coverage beyond impinging
swissskimmilk
committed
1a13957
Discriminate unsupported-config from non-convergence in the accelerator API
swissskimmilk
committed
fb84c7d
Fix parity suite under strict mode: forcing the Python reference is not a failure
swissskimmilk
committed
9f8588c
Add a guard that production actually calls the accelerator, not just that it works
swissskimmilk
committed
761fee3
Strengthen the bypass guard: per-candidate ratio, chug seam, call-site notes
swissskimmilk
committed
2388118
Fix Layer 1 convergence plot x-axis, evaluation counter, and chart legend
swissskimmilk
committed
4418536
Memoize decompressed CEA tables so a Layer 1 run reads the .npz once, not 73 times
swissskimmilk
committed
e347ccc
Floor CMA iter_budget at 1 so max_iterations below the restart count cannot divide by zero
swissskimmilk
committed
b4c9f1b
Drop two redundant deepcopies of the worker evaluate() payload
swissskimmilk
committed
ce92dd5

Navigation Menu

Replace the C physics port with a Numba accelerator - #37

Replace the C physics port with a Numba accelerator #37
swissskimmilk merged 20 commits into
mainfrom
engine-design/numba-accelerator

Commits on Sep 4, 2026

Add numba physics kernels and 3-way Layer-1 benchmark, declare numba dep

Free numba kernels from the C EdEngineState with pure-Python param extraction

Port ablative cooling to numba, closing the last coverage gap vs the C kernel

Add live A/B parity suite for the numba accelerator vs C and Python

Move accelerator into engine/accel and drop the C injector solve from evaluate

Route production through engine.accel: solve, chamber_solve, dispatcher, call sites

CI: replace native-parity with a matrixed accel-parity job covering numba and C

Vectorise the chug sweep and add a numba kernel for it, removing the last C dependency

Retire the C oracle from the parity suite; guard numba against Python at 1e-6

Drop the ED_ACCEL=c dispatch branch now that the parity suite no longer needs it

Delete the C physics port; numba is the accelerator

Port the pintle injector to numba, extending accelerated coverage beyond impinging

Discriminate unsupported-config from non-convergence in the accelerator API

Fix parity suite under strict mode: forcing the Python reference is not a failure

Add a guard that production actually calls the accelerator, not just that it works

Strengthen the bypass guard: per-candidate ratio, chug seam, call-site notes

Fix Layer 1 convergence plot x-axis, evaluation counter, and chart legend

Memoize decompressed CEA tables so a Layer 1 run reads the .npz once, not 73 times

Floor CMA iter_budget at 1 so max_iterations below the restart count cannot divide by zero

Drop two redundant deepcopies of the worker evaluate() payload