Unified · Split
| @@ -41,8 +41,8 @@ pipeline on real shards. Single‑stream, greedy, CUDA graphs (2026‑07‑09/10 · 40 unmodified lines | ||
| 41 | 41 | a small gate‑filled FP4 pool for precision. With MTP k=2 (acceptance 3.0) and miss‑tolerance |
| 42 | 42 | 8: **45–52 tok/s** (strict zero‑tolerance: 21–24). Needle 4/4 PASS across the window. |
| 43 | 43 | Bare‑2‑bit quality artifacts ("capital of Poland: Krakow", garbled Polish) are corrected by |
| 44 | − the FP4 tier. The window is KV‑budget‑bound, not architectural — the record run gave KV | |
| 45 | − 8 GiB; the same box boots 262K windows on the nvfp4 KV cache (retrieval validation pending). | |
| 44 | + the FP4 tier. The window is KV‑budget‑bound, not architectural — the record config simply | |
| 45 | + allotted KV 8 GiB/rank; a larger window trades against expert‑cache coverage. | |
| 46 | 46 | - **NVFP4 KV cache** (`--kv-cache-dtype nvfp4`): packed 352 B/token vs 656 B `fp8_ds_mla` — |
| 47 | 47 | **+38% KV pool** (415K → 571K tokens at equal settings) at decode parity, or the freed VRAM |
| 48 | 48 | goes to the FP4 pool (the standing 4‑card config runs a 19.6 GiB/GPU pool + 175K‑token KV). |