PyTorch Static Tensor Shape Type Checker & Contract Lab

Model Tensor Shapes

ARCHITECTURE PRESET

Batch Size B 4

Sequence Length S 128

Model Dim D 512

Heads H 8

INJECT STATIC SHAPE MISMATCHES

Static Shape Constraint Verification CHECKING

Static Verification Status: Sound (0 Type Errors)
Active Contract: Tensor[B=4, S=128, D=512] -> MultiHeadAttention -> Tensor[B=4, S=128, D=512]

PEP 695 / PyreFly Type Stub

Static Assertion Contract

# Generated PyreFly Tensor Contract Stub
from typing import TypeVar
import torch
from torch import Tensor

B = TypeVar("B")  # Batch
S = TypeVar("S")  # Sequence
D = TypeVar("D")  # Dimension
H = TypeVar("H")  # Heads

Tensor dimensions must align before reshape and matrix multiplication

Your browser does not support video.

Read the explanation

The educational shape checker divides model dimension by head count and flags a remainder. Hypothetical dimension five hundred twelve with eight and sixteen heads gives sixty-four and thirty-two elements per head. Bars share four pixels per elements-per-head count. This illustrates static shape arithmetic, not executing a neural model. The source query-key-value projection maps D features to three times D. Hypothetical D five hundred twelve and two hundred fifty-six yield one thousand five hundred thirty-six and seven hundred sixty-eight projected features. Bars share point two pixels per projected feature. A later split requires compatible head dimensions; projection alone does not validate all tensor operations. Attention scores have two sequence dimensions S by S per head. Hypothetical sequence lengths one hundred twenty-eight and sixty-four give sixteen thousand three hundred eighty-four and four thousand ninety-six entries. Bars share point zero two pixels per entry. Doubling sequence length quadruples that matrix size. Full actual Python-stub download, native reset, responsive and first-footer checks preserve the source educational stages. Generated typing/runtime framework claims and untested fault buttons remain separate; no backend model request is made.