{"model":"deepseek-4.1-flash","title":"Schaka publishes first working recipe - pivoted","status":"positive","summary":"Schaka 170hx-journey (TP4xPP2) is the first working V4.1 on SM80: 43.8/72.9/104.2 tok/s @ c1/c4/c8, prefill 1,584/1,650/1,389 tok/s @ 39k/119k/319k, KV pool 14,058,003 tokens. His kit matches our bug: wo_a wrong-shaped under Marlin, and V4.0's 9-arg qnorm op dropping V4.1's apply_q_norm…","occurred_at":"2026-09-11T00:15:00Z","config":{"serving":"TP4xPP2 (Schaka 170hx-journey)","engine":"Schaka/170hx-journey patch kit"},"benchmarks":[{"metric":"c1_decode_tok_s","value":43.8,"unit":"tok/s","context":"Schaka box, TP4xPP2 (author-reported)"},{"metric":"agg_decode_c4_tok_s","value":72.9,"unit":"tok/s","context":"Schaka box (author-reported)"},{"metric":"agg_decode_c8_tok_s","value":104.2,"unit":"tok/s","context":"Schaka box (author-reported)"},{"metric":"prefill_tok_s","value":1584,"unit":"tok/s","context":"Schaka box @39k (1,650 @119k, 1,389 @319k)"},{"metric":"kv_pool_tokens","value":14058003,"unit":"tokens","context":"Schaka TP4xPP2 pool"}],"fixed":["wo_a wrong-shaped under Marlin -\u003e emulation kernel","V4.0 9-arg qnorm op dropped V4.1 apply_q_norm -\u003e Q normalized twice -\u003e wrong output, no error","pp_kv_group_relay.py unlocks PP (previously thought impossible)"]}