Skip to content

Repository files navigation

RecursiveCompute

Use AI to build optimal Inference Compute systems

Plans

Initial Smoke Commands

Run the standard-library tests locally:

python3 -m unittest discover -s dv/tests -v

This includes the Phase 8 firmware/driver suite: freestanding C compilation, MMIO boot simulation, secure-boot negative paths, graph submission, KV fault replay, and telemetry validation.

Run the first event-model simulation:

python3 -m sim.event_model.rcif_sim sim/workloads/sample_agentic_coding_trace.json --pretty

Generate a synthetic agentic coding trace:

python3 -m sim.tracegen.generate_agentic_trace \
  --requests 4 \
  --prefix-tokens 131072 \
  --reuse-ratio 0.9 \
  --output /tmp/generated_agentic_trace.json

RTL Simulation

Verilator lint/build/simulation runs on Modal CPU workers by default. Local Verilator is not required.

make lint
make verilate
make sim
make regress
make phase9
make phase10

If modal is not on PATH:

MODAL=/Users/kb/Library/Python/3.9/bin/modal make regress

The local-* targets are only for debugging on machines that already have Verilator installed:

make local-lint
make local-sim
make local-regress
make local-phase9

phase9 adds assertion-enabled full-chip directed and randomized simulation, security/memory-safety invariant tests, real descriptor-trace replay, exhaustive cross coverage, reset/backpressure stress, and an enforced Verilator coverage report. See the full-chip verification plan for bounded scope and physical-signoff gates.

phase10 adds the reduced FPGA shell, an AXI external-DDR proxy, real runtime token-graph submission, a bit-exact toy attention/GEMV block, synthetic agentic trace replay, KV page-fault migration, and a 5% cycle-model agreement gate. See the FPGA prototype guide for the reproducible emulation scope and the physical-board closure requirements.

The current cloud verification record is Modal RTL Verification Record — 2026-07-21. It records passing subsystem regression, Phase 9 full-chip coverage, and Phase 10 FPGA-shell/DDR simulation runs, including the exact tool revision and Modal run IDs.

About

Use AI to build optimal Inference Compute systems

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages