PyTorch Tensor Shape Tracer & AOT Compilation Visualizer

CONTRACT STATUS:

● SHAPE TYPING VALID

TARGET: AOTInductor (Triton/C++ CUDA)

Symbolic Tensor Inputs

ARCHITECTURE PRESET

BATCH (B)

SEQUENCE (S)

EMBEDDING (D)

ATTN HEADS (H)

TENSOR DTYPE

ENABLE GPU KV-CACHE (All-GPU Hit Scenario)

Static Shape Constraints:

• Linear layers require in_features == prev.shape[-1]

• Multi-Head Attention requires D % H == 0 (Head Dim = D / H)

2 Fusion Groups

AOTI INFERENCE SPEEDUP Triton Server

2.20x - 2.38x

AOTI + KV Cache GPU Residency (Ideal Cache Hit)

Standard Eager (Python Overhead) 1.00x (Baseline)

100% Latency

AOTInductor (No KV-Cache) 1.14x - 1.28x

78% Latency

AOTInductor + KV-Cache Hit 2.20x - 2.38x

42% Latency

GENERATED PYTORCH 2.X CODE