@@ -771,6 +771,30 @@ def _harness_preset(
771771 params = {"max_new_tokens" : ("int:max_new_tokens" , "64" )},
772772 validate_reports = True , # §4 liveness gate on-device
773773 ),
774+ Preset (
775+ name = "mlx-kakeya-degen-probe" ,
776+ description = "DEBUG (Phase-1): full f_θ fused engine on a LONG prompt "
777+ "(--ignore-turn-stop, default 256 tokens) to characterize "
778+ "the long-decode degeneration onset. Emits KDBG NDJSON to "
779+ "stderr (captured in the bridge log) + transcript JSON. NOT "
780+ "gated — the degeneration is the thing being measured." ,
781+ command_templates = (
782+ (
783+ "python3" , "scripts/research/k3_integrated_niah_eval_mac.py" ,
784+ "--verifier-path" , "${ENV:KAKEYA_MAC_VERIFIER_PATH}" ,
785+ "--drafter-id" , "${ENV:KAKEYA_MAC_DRAFTER_ID}" ,
786+ "--f-theta-dir" , "${ENV:KAKEYA_MAC_FTHETA_DIR}" ,
787+ "--s5-exact-full-attn" , "--fused-specdecode" , "--force-f-theta" ,
788+ "--sink-size" , "4" , "--window-size" , "64" , "--block-size" , "4" ,
789+ "--max-new-tokens" , "{max_new_tokens}" , "--ignore-turn-stop" ,
790+ "--chat" , "--chat-scripted" , "请详细解释POW的工作原理" ,
791+ "--output" , "results/research/phase1_degeneration_chat.json" ,
792+ ),
793+ ),
794+ timeout_minutes = 90 ,
795+ params = {"max_new_tokens" : ("int:max_new_tokens" , "256" )},
796+ validate_reports = False ,
797+ ),
774798 )
775799}
776800
0 commit comments