Linting

Author: kavanase

CHANGELOG.md

@@ -10,7 +10,7 @@ Most recent change on the top.
 
 ### Changed
 - `nequip-compile --device cpu` now defaults to scalar C++ codegen (Inductor's `cpp.simdlen=0`) instead of vectorized codegen, which is 1.5-2x faster for NequIP and 6-11x faster for smaller Allegro models in CPU inference, and avoids the CPU AOTInductor compilation failure on PyTorch 2.10/2.11 entirely. Controllable with `ATEN_CPU_CAPABILITY=<your ISA>` or `--inductor-configs cpp.simdlen=<width>`.
-- PyTorch compatibility docs: the CPU + AOTInductor compilation failure affects both PyTorch 2.10.0 and 2.11.0 (fixed upstream in 2.12), now avoided with default NequIP scalar codegen compilation 
+- PyTorch compatibility docs: the CPU + AOTInductor compilation failure affects both PyTorch 2.10.0 and 2.11.0 (fixed upstream in 2.12), now avoided with default NequIP scalar codegen compilation.
 
 ### Fixed
 - `--inductor-configs` values were passed to `torch` as strings, so numeric options mis-compared. Values are now YAML-parsed, as for `nequip-package modify` modifier kwargs.

docs/guide/getting-started/workflow.md

@@ -168,7 +168,7 @@ For further speed-ups where appropriate (e.g. often permissible for MD simulatio
 
 ```{important}
 When compiling for **CPU** (`--device cpu`), `nequip-compile` turns off Inductor's vectorized C++ codegen and emits scalar code instead, by default.
-This is typically significantly faster for model inference. To compile with vectorization instead, give your CPU's vector width with `--inductor-configs cpp.simdlen=512` (`512` for AVX-512, `256` for AVX2, `128` for ARM NEON), or set torch's own `ATEN_CPU_CAPABILITY` to an x86 ISA name (e.g. `ATEN_CPU_CAPABILITY=avx512 nequip-compile ...`). 
+This is typically significantly faster for model inference. To compile with vectorization instead, give your CPU's vector width with `--inductor-configs cpp.simdlen=512` (`512` for AVX-512, `256` for AVX2, `128` for ARM NEON), or set torch's own `ATEN_CPU_CAPABILITY` to an x86 ISA name (e.g. `ATEN_CPU_CAPABILITY=avx512 nequip-compile ...`).
 `Allegro-L` is the one NequIP/Allegro foundation model we have measured to be faster with vectorized codegen (by 1.34x), so can be worth opting in there.
 ```
 

tests/unit/scripts/test_compile_configs.py

@@ -13,7 +13,10 @@
     "arg,expected",
     [
         ("cpp.simdlen=256", {"cpp.simdlen": 256}),
-        ("cpp.simdlen=0", {"cpp.simdlen": 0}),   # a bare `false` left as a string would be non-empty, i.e. truthy
+        (
+            "cpp.simdlen=0",
+            {"cpp.simdlen": 0},
+        ),  # a bare `false` left as a string would be non-empty, i.e. truthy
         ("cpp.cpp_wrapper=false", {"cpp.cpp_wrapper": False}),
         ("cpp.cpp_wrapper=true", {"cpp.cpp_wrapper": True}),
         # bare strings must survive YAML parsing unchanged: