{"model":"deepseek-4.1-flash","title":"Correctness root cause localized","status":"positive","summary":"Per-branch probe: attention 'o' absmax ~2.0 (sparse MLA works) but _o_proj output absmax ~0.013 - a ~150x shrink; residual goes FFN-only and explodes (0.6 -\u003e 3.5e29 at KV-source layers). One bug: fused inverse-RoPE + wo_a einsum + wo_b.","occurred_at":"2026-09-10T23:47:00Z","config":{"serving":"TP8xPP1, DSpark k=5, fp8_ds_mla, 1M ctx","engine":"local/vllm-dsv41:sm80 @ dsv41-feat@e47aa780"},"found":["o absmax ~2.0, _o_proj absmax ~0.013 -\u003e ~150x shrink of the attention contribution"],"failed":["Ruled out: tokenization/template, Engram (DSV41_ZERO_ENGRAM), wo_a layout, backend selection, NaN"]}