AGILLM3.5 is the AGILLM3 checkpoint/tokenizer contract running on the AGILLM4 runtime and DiffusionBlock training path.

The runnable artifact is agillm35.py. The helper modules are folded into that one file so the runtime can be cloned, inspected, and launched without restoring the whole AGILLM4 source tree.

public_join/agillm35_network_host.py starts a signed-lease HTTPS coordinator for people who want to run their own network.

public_join/agillm35_join_worker.py is an outbound-only worker for untrusted joiners. It requests short-lived leases, verifies package hashes, runs a local worker command, and submits results to quarantine rather than exposing SSH or writing directly into the master merge path.

distributed_infer/agillm35_distributed_infer.py is a single-file distributed AR inference harness for the real AGILLM3.5 transformer. It splits contiguous transformer/DiffusionBlock layer ranges across local or HTTP worker stages, using the actual Block implementation and MoE FFNs from the checkpoint config.

Plan layer ranges:

Start a worker for one layer range:

Run the coordinator:

Network tensor payloads use a small raw tensor wire format rather than unpickling remote worker responses. Use TLS plus a bearer token for workers exposed beyond localhost. --cache-mode kv is the default and keeps per-session KV state on each worker after the prompt prefill, so decode steps send only the new hidden token through the pipeline. --cache-mode full is kept for comparison/debugging. SAT/NAT distributed decoding is a later phase.

For inference against the live round-299 checkpoint, prefer the HF inference-slim artifact distributed/inference/master_r299_20260602-205914_ar_infer_slim.pt; it drops optimizer/SAT/disaggregated training state while preserving AR transformer inference.

This repository contains code only, not AGILLM3 checkpoint weights.

DiffusionBlock logs report raw CE-style loss plus the actual EDM-weighted training objective as weighted. The weighted value is the optimization target; the raw value is the sanity-check number to compare with ordinary AR/SAT loss.

The Linux smoke test compiles the single file and completes a one-step synthetic training save. The full AGILLM3.5 continuation run is managed separately by the disaggregated Hetzner worker setup.