[NEP50 NEP52 NEP56] NumPy 2.0 Breaking Changes

Overview

NumPy 2.0 introduced significant breaking changes affecting type promotion, API surface, and array standard compliance.

---

NEP 50: Promotion Rules for Python Scalars

Status: Final | Full Text

The Problem Solved

NumPy 1.x used value-based promotion - inspecting scalar values to determine output types:

np.result_type(np.int8, 1) == np.int8      # 1 fits in int8
np.result_type(np.int8, 255) == np.int16   # 255 doesn't fit - UPCASTED

NumPy 2.x Behavior

Weak scalar promotion - Python scalars defer to array dtype:

uint8(1) + 2 → uint8(3)           # Python int defers to uint8
uint8(1) + 255 → uint8(0)         # Overflow with warning
int16(2) + int64(3) → int64(5)    # NumPy scalar is "strong"

Kind Hierarchy

• boolean < integral < inexact
• Cross-kind uses default precision (int64, float64, complex128)

---

NEP 52: Python API Cleanup for NumPy 2.0

Status: Final | Full Text

Removed Aliases

Removed	Canonical
np.round_	np.round
np.product	np.prod
np.sometrue	np.any
np.alltrue	np.all

Removed Functions

• byte_bounds, disp, safe_eval, who
• maximum_sctype and related sctype functions

Namespace Changes

• numpy.core → numpy._core (private)
• np.compat removed

---

NEP 56: Array API Standard Support

Status: Final | Full Text

New Strict Behaviors

• .T errors for ndim > 2
• cross() errors on size-2 vectors
• outer() raises on >1-D inputs

DType Changes

• ceil/floor/trunc return integer dtype (was float)

New Functions

• isdtype(dtype, kind) - dtype introspection
• unique_values(), unique_counts(), unique_inverse(), unique_all()
• matrix_transpose(), vecdot(), matrix_norm(), vector_norm()

New Aliases

• Trig: acos, asin, atan, atan2 (for arc* versions)
• Other: concat, permute_dims, pow, bitwise_*

New Properties

• ndarray.mT - matrix transpose (last 2 axes)
• ndarray.device - returns CPU device

copy= Semantics

np.asarray(x, copy=True)   # Always copy
np.asarray(x, copy=False)  # Never copy (raise if needed)
np.asarray(x, copy=None)   # Copy if necessary

---

Suggested Implementation for NumSharp

Phase 1: Type Promotion (NEP 50)

• Audit np._FindCommonType for value-based logic
• Implement weak scalar semantics in arithmetic operators
• Add overflow warnings

Phase 2: API Cleanup (NEP 52)

• Audit for deprecated function names
• Ensure only canonical names exposed
• Address dead code per CLAUDE.md

Phase 3: Array API (NEP 56)

• Add function aliases (acos, concat, etc.)
• Implement isdtype() function
• Add unique_* family
• Implement .mT property
• Update copy parameter semantics
• Add .T validation for ndim > 2

---

Documentation

See docs/neps/NEP50.md, docs/neps/NEP52.md, docs/neps/NEP56.md