talon-backend 0.0.3

TALON Backend: Config generation, CPU simulation, and FPGA deployment for Talon IR graphs

TALON Backend

Config generation and CPU simulation for TALON IR graphs. Compiles partitioned graphs into JSON/binary descriptors for hardware runtime and provides a CPU simulator for correctness validation and profiling.

Installation

uv add talon-backend

Features

• Config Generator: Produces JSON/binary hardware descriptors (not bitstream)
• CPU Simulator: Validates graph execution correctness
• From-Scratch Profiler: Latency, energy, and memory estimation without external deps
• Backend Registry: Plugin system for custom hardware backends
• DAG + Cyclic Support: Auto-detects topology and selects execution mode

Quick Start

from talon import ir, backend

# Load a partitioned graph
graph = ir.read("model.t1c")

# Compile to hardware descriptor
cpu = backend.get_backend("cpu")
descriptor = cpu.compile(graph)
descriptor.to_json("model_config.json")

# Simulate for correctness
result = cpu.simulate(graph, input_data, n_steps=100)
print(f"Valid: {result.outputs_valid}")

# Profile (from scratch, no external deps)
profile = cpu.profile(graph, input_data, n_steps=100)
print(f"Latency: {profile.total_latency_us} us")
print(f"Energy: {profile.energy_estimate_uj} uJ")

License

MIT