| # ARCHON Inference Turbo |
|
|
| 12 modules pour faire jouer ARCHON 282M comme un 7B+ à l'inférence. |
|
|
| Plan complet: `../docs/ARCHON_INFERENCE_TURBO_PLAN.md` |
|
|
| ## Modules Phase A (drop-in, déployables immédiat post-SFT v2) |
|
|
| - `m7_dola.py` — Contrastive decoding entre layer L6 et L18 |
| - `m9_xgrammar.py` — Wrapper logit processor structured output |
| - `m11_react_tools.py` — ReAct loop + 70 tools registry (port NEXUS) |
| - `m6_snapkv.py` — KV cache compression prefill phase |
|
|
| ## Modules Phase B (engine v1, instance GPU requise) |
|
|
| - `m1_mtp_self_spec.py` — MTP-5 self-speculative decoding |
| - `m10_nvfp4_quant.py` — NVFP4 PTQ Blackwell |
| - `m8_graphrag.py` — HippoRAG2 federated retrieval |
| - `m5_prm_bestof_n.py` — PRM 50M + best-of-N |
|
|
| ## Modules Phase C (training requis) |
|
|
| - `m2_extended_thinking.py` — Thinking budget loop |
| - `m3_coconut.py` — Latent continuous thought |
| - `m12_ttt_e2e.py` — Test-time training LoRA |
|
|
| ## Module Phase D (architecture mod) |
|
|
| - `m4_mor.py` — Mixture of Recursions router |
|
|
| ## Engine |
|
|
| - `engine.py` — Chaîne complète orchestrée |
| - `bench.py` — Benchmark vs Qwen2.5-7B, Phi-4, Llama 3.1 8B |
|
|
| ## Status code |
|
|
| Phase A skeletons codés 2026-06-05. Reste à brancher sur ARCHON loaded model post-SFT v2 finish. |
|
|
