cuda.core: public API class naming audit (PEP 8 and consistency)

[mdboom]

@mdboom's commentary: This all seems pretty low stakes, but now is the time to address any inconsistencies before 1.0.0.

Summary

An audit of all public class names in cuda.core found three categories of naming inconsistency: a terminology mismatch ("Config" vs "Options"), one unusual abbreviation, and several PEP 8 acronym-capitalization violations in the cuda.core.system submodule.

Original prompt

Can you please look at all classes exposed in the public API of cuda_core. I want you to look just at the names of the classes, not any members or methods. Please make sure they are all PEP8 compliant and that the same words are used for the same concepts.

---

Findings

1. "Config" vs "Options" inconsistency

LaunchConfig is the only class that uses "Config" for a configuration/options object. Every other such class uses "Options":

Pattern-breaker	Consistent peers
LaunchConfig	EventOptions, LinkerOptions, StreamOptions, ProgramOptions, DeviceMemoryResourceOptions, ManagedMemoryResourceOptions, PinnedMemoryResourceOptions, VirtualMemoryResourceOptions, TensorMapDescriptorOptions, GraphAllocOptions, GraphCompleteOptions, GraphDebugPrintOptions

Suggestion: LaunchConfig → LaunchOptions

2. GraphDef abbreviation

GraphDef abbreviates "Definition" while the rest of the API spells words out (e.g. TensorMapDescriptor, not TensorMapDesc; ConditionalNode, not CondNode). The AllocNode / GraphAllocOptions abbreviations are defensible since "alloc/free" are standard CUDA/C terms, but GraphDef is less conventional.

Suggestion: GraphDef → GraphDefinition

3. Acronym capitalization (PEP 8 violation in cuda.core.system)

PEP 8 says: "When using abbreviations in CapWords, capitalize all the letters of the abbreviation. Thus HTTPServerError is better than HttpServerError."

The system module is internally inconsistent — P2P and BAR1 are correctly all-caps, while other acronyms are treated as regular CamelCase words:

Current name	PEP 8 name	Acronym
GpuDynamicPstatesInfo	GPUDynamicPStatesInfo	GPU, PStates
GpuDynamicPstatesUtilization	GPUDynamicPStatesUtilization	GPU, PStates
GpuP2PCapsIndex	GPUP2PCapsIndex	GPU
GpuP2PStatus	GPUP2PStatus	GPU
GpuTopologyLevel	GPUTopologyLevel	GPU
Pstates	PStates	PStates
PcieUtilCounter	PCIeUtilCounter	PCIe
PciInfo	PCIInfo	PCI
GpuIsLostError	GPUIsLostError	GPU
GpuNotFoundError	GPUNotFoundError	GPU
VgpuEccNotSupportedError	VGPUECCNotSupportedError	VGPU, ECC
IrqIssueError	IRQIssueError	IRQ
LibRmVersionMismatchError	LibRMVersionMismatchError	RM

Note: BAR1MemoryInfo, GpuP2PCapsIndex (the P2P part), and GpuP2PStatus (the P2P part) already correctly capitalize their acronyms, making the inconsistency with Gpu, Pci, etc. more visible.

---

What already looks good

• All class names use CamelCase (PEP 8 compliant).
• The *MemoryResource / *MemoryResourceOptions family is consistently named.
• The *Node subclasses in cuda.core.graph are consistently named.
• The *Error exception classes consistently use the Error suffix.
• Event/EventOptions, Stream/StreamOptions, Program/ProgramOptions, Linker/LinkerOptions all follow the same pairing pattern.

---

Recommended changes by priority

1. LaunchConfig → LaunchOptions — terminology inconsistency with the rest of the API
2. GraphDef → GraphDefinition — unusual abbreviation
3. Acronym capitalization in cuda.core.system — 13 classes violate PEP 8's explicit acronym rule and are internally inconsistent with P2P/BAR1 in the same module