# a0116b2 unchanged, STRATA_PREFILL_TIMING=1 STRATA_DECODE_TIMING=1 STRATA_VERIFY_PROFILE=1: three short and three long requests
strata serve: prompt chunk auto: 8192 tokens, a 384-slot ring
strata prefill timing: 559 tokens, GPU timeline 2469 ms, wall 2476 ms, host staging 11 ms: embed+steps 0 (0.0%) hc read 177 (7.2%) gdn 587 (23.8%) qsa proj 253 (10.3%) qsa indexer 0 (0.0%) qsa select 1 (0.0%) qsa attn 31 (1.3%) router+shared 82 (3.3%) host grouping 126 (5.1%) gather 8 (0.3%) wait copy 39 (1.6%) dequant 139 (5.6%) gemm gate/up 527 (21.3%) gemm down 160 (6.5%) combine 57 (2.3%) ple 7 (0.3%) kv stage 0 (0.0%) gdn conv+gates 9 (0.3%) gdn recurrence 34 (1.4%) gdn out proj 231 (9.4%)
strata prefill timing: host: chunk setup (PLE rows, the expert stream plan) 0 ms, waiting for each chunk 10 ms, after each chunk (the draft layer, progress) 7 ms, PLE 53 ms
strata decode timing: 116 windows, avg T 2.14, 1.72 tokens/window, 56.22 ms/window = verify 49.54 (GPU-reach wait 35.56 + per-layer host 8.48 [plan 0.18 actq 0.20 jobs 0.07 CPU 7.43] + stage 0.58) + commit/emit 0.36 + draft 4.04; per layer-window: CPU experts 1.80 (2.37 entries), VRAM hits 17.57, PCIe 0.00
strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 4.24 conv 0.20 ab 0.36 z 2.50 rec 0.91 out-proj 2.79 router+plan+bell 1.53 shared+quant 2.23 waitA 1.14 VRAM hits 8.49 waitB 0.31 PCIe grp 0.27 waitCPU 2.50 copy+combine 0.57 (gap) 0.34 head 2.87   hc0 norm 0.65   hc0 down 1.43   hc0 up 1.30   hc1 norm 0.33   hc1 down 1.40   hc1 up 1.38 | QSA layers: q8+kv-idx 0.13 k/v+norm-rope 0.35 kv+idx append 0.17 q+q-idx 1.83 scores+topk 0.11 kv-resolve 0.17 attention 0.57 gate 0.05 out-proj 0.98 router+plan+bell 0.51 shared+quant 0.73 waitA 0.11 VRAM hits 2.67 waitB 0.11 PCIe grp 0.09 waitCPU 0.54 copy+combine 0.19 (gap) 0.11   hc0 norm 0.10   hc0 down 0.48   hc0 up 0.43   hc1 norm 0.11   hc1 down 0.47   hc1 up 0.46 | total 49.21 ms/window over 117 windows
strata serve: prompt 566 tokens = 0 reused + 566 read in 2742 ms (206.4 tok/s), 200 generated in 6522 ms (30.7 tok/s), drafts accepted 85 of 132, 1 checkpoints
strata prefill timing: 557 tokens, GPU timeline 2372 ms, wall 2378 ms, host staging 20 ms: embed+steps 1 (0.0%) hc read 176 (7.4%) gdn 583 (24.6%) qsa proj 254 (10.7%) qsa indexer 0 (0.0%) qsa select 1 (0.0%) qsa attn 31 (1.3%) router+shared 80 (3.4%) host grouping 113 (4.8%) gather 8 (0.4%) wait copy 45 (1.9%) dequant 131 (5.5%) gemm gate/up 474 (20.0%) gemm down 156 (6.6%) combine 40 (1.7%) ple 6 (0.3%) kv stage 0 (0.0%) gdn conv+gates 9 (0.4%) gdn recurrence 34 (1.4%) gdn out proj 230 (9.7%)
strata prefill timing: host: chunk setup (PLE rows, the expert stream plan) 1 ms, waiting for each chunk 12 ms, after each chunk (the draft layer, progress) 5 ms, PLE 53 ms
strata decode timing: 113 windows, avg T 2.05, 1.77 tokens/window, 52.66 ms/window = verify 47.92 (GPU-reach wait 39.99 + per-layer host 3.31 [plan 0.16 actq 0.14 jobs 0.03 CPU 2.41] + stage 0.56) + commit/emit 0.27 + draft 3.86; per layer-window: CPU experts 0.47 (0.54 entries), VRAM hits 19.44, PCIe 0.00
strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 4.17 conv 0.20 ab 0.36 z 2.46 rec 0.89 out-proj 2.74 router+plan+bell 1.50 shared+quant 2.21 waitA 1.18 VRAM hits 9.59 waitB 0.31 PCIe grp 0.27 waitCPU 0.93 copy+combine 0.49 (gap) 0.34 head 2.83   hc0 norm 0.64   hc0 down 1.42   hc0 up 1.28   hc1 norm 0.33   hc1 down 1.39   hc1 up 1.36 | QSA layers: q8+kv-idx 0.13 k/v+norm-rope 0.34 kv+idx append 0.17 q+q-idx 1.80 scores+topk 0.10 kv-resolve 0.17 attention 0.57 gate 0.05 out-proj 0.97 router+plan+bell 0.50 shared+quant 0.73 waitA 0.12 VRAM hits 3.06 waitB 0.11 PCIe grp 0.09 waitCPU 0.17 copy+combine 0.16 (gap) 0.11   hc0 norm 0.10   hc0 down 0.47   hc0 up 0.43   hc1 norm 0.11   hc1 down 0.46   hc1 up 0.45 | total 48.27 ms/window over 114 windows
strata serve: prompt 564 tokens = 0 reused + 564 read in 2667 ms (211.5 tok/s), 200 generated in 5951 ms (33.6 tok/s), drafts accepted 87 of 119, 1 checkpoints
strata prefill timing: 560 tokens, GPU timeline 2369 ms, wall 2375 ms, host staging 29 ms: embed+steps 1 (0.0%) hc read 177 (7.5%) gdn 583 (24.6%) qsa proj 253 (10.7%) qsa indexer 0 (0.0%) qsa select 1 (0.0%) qsa attn 31 (1.3%) router+shared 80 (3.4%) host grouping 106 (4.5%) gather 8 (0.4%) wait copy 34 (1.4%) dequant 133 (5.6%) gemm gate/up 477 (20.1%) gemm down 157 (6.6%) combine 48 (2.0%) ple 6 (0.3%) kv stage 0 (0.0%) gdn conv+gates 9 (0.4%) gdn recurrence 34 (1.4%) gdn out proj 230 (9.7%)
strata prefill timing: host: chunk setup (PLE rows, the expert stream plan) 1 ms, waiting for each chunk 13 ms, after each chunk (the draft layer, progress) 5 ms, PLE 53 ms
strata decode timing: 112 windows, avg T 2.19, 1.79 tokens/window, 53.39 ms/window = verify 48.87 (GPU-reach wait 42.04 + per-layer host 2.23 [plan 0.16 actq 0.10 jobs 0.02 CPU 1.37] + stage 0.56) + commit/emit 0.22 + draft 4.11; per layer-window: CPU experts 0.25 (0.29 entries), VRAM hits 21.24, PCIe 0.00
strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 4.28 conv 0.21 ab 0.37 z 2.52 rec 0.91 out-proj 2.80 router+plan+bell 1.49 shared+quant 2.25 waitA 1.11 VRAM hits 10.16 waitB 0.32 PCIe grp 0.27 waitCPU 0.68 copy+combine 0.47 (gap) 0.34 head 2.90   hc0 norm 0.66   hc0 down 1.44   hc0 up 1.30   hc1 norm 0.33   hc1 down 1.42   hc1 up 1.39 | QSA layers: q8+kv-idx 0.13 k/v+norm-rope 0.35 kv+idx append 0.17 q+q-idx 1.85 scores+topk 0.11 kv-resolve 0.17 attention 0.58 gate 0.05 out-proj 0.99 router+plan+bell 0.50 shared+quant 0.74 waitA 0.12 VRAM hits 3.21 waitB 0.11 PCIe grp 0.09 waitCPU 0.06 copy+combine 0.15 (gap) 0.11   hc0 norm 0.10   hc0 down 0.48   hc0 up 0.44   hc1 norm 0.11   hc1 down 0.47   hc1 up 0.46 | total 49.13 ms/window over 113 windows
strata serve: prompt 567 tokens = 0 reused + 567 read in 2654 ms (213.7 tok/s), 200 generated in 5980 ms (33.4 tok/s), drafts accepted 88 of 133, 1 checkpoints
strata prefill timing: 7038 tokens, GPU timeline 16005 ms, wall 16070 ms, host staging 79 ms: embed+steps 8 (0.1%) hc read 2046 (12.8%) gdn 3129 (19.6%) qsa proj 1363 (8.5%) qsa indexer 0 (0.0%) qsa select 122 (0.8%) qsa attn 1858 (11.6%) router+shared 583 (3.6%) host grouping 39 (0.2%) gather 147 (0.9%) wait copy 28 (0.2%) dequant 305 (1.9%) gemm gate/up 2567 (16.0%) gemm down 1204 (7.5%) combine 665 (4.2%) ple 69 (0.4%) kv stage 0 (0.0%) gdn conv+gates 105 (0.7%) gdn recurrence 443 (2.8%) gdn out proj 1324 (8.3%)
strata prefill timing: host: chunk setup (PLE rows, the expert stream plan) 7 ms, waiting for each chunk 73 ms, after each chunk (the draft layer, progress) 63 ms, PLE 449 ms
strata decode timing: 37 windows, avg T 1.86, 1.62 tokens/window, 50.87 ms/window = verify 47.09 (GPU-reach wait 41.50 + per-layer host 1.20 [plan 0.13 actq 0.06 jobs 0.02 CPU 0.84] + stage 0.14) + commit/emit 0.16 + draft 3.50; per layer-window: CPU experts 0.14 (0.14 entries), VRAM hits 18.26, PCIe 0.00
strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 4.10 conv 0.20 ab 0.35 z 2.43 rec 0.88 out-proj 2.70 router+plan+bell 1.48 shared+quant 2.19 waitA 1.10 VRAM hits 9.55 waitB 0.32 PCIe grp 0.27 waitCPU 0.32 copy+combine 0.45 (gap) 0.34 head 2.79   hc0 norm 0.64   hc0 down 1.42   hc0 up 1.27   hc1 norm 0.33   hc1 down 1.39   hc1 up 1.35 | QSA layers: q8+kv-idx 0.13 k/v+norm-rope 0.33 kv+idx append 0.17 q+q-idx 1.79 scores+topk 0.62 kv-resolve 0.43 attention 0.97 gate 0.05 out-proj 0.95 router+plan+bell 0.49 shared+quant 0.72 waitA 0.12 VRAM hits 3.06 waitB 0.11 PCIe grp 0.09 waitCPU 0.06 copy+combine 0.15 (gap) 0.11   hc0 norm 0.10   hc0 down 0.47   hc0 up 0.42   hc1 norm 0.11   hc1 down 0.46   hc1 up 0.45 | total 48.21 ms/window over 38 windows
strata serve: prompt 7045 tokens = 0 reused + 7045 read in 16672 ms (422.6 tok/s), 60 generated in 1882 ms (31.9 tok/s), drafts accepted 25 of 32, 1 checkpoints
strata prefill timing: 7039 tokens, GPU timeline 15730 ms, wall 15794 ms, host staging 130 ms: embed+steps 6 (0.0%) hc read 1989 (12.6%) gdn 3136 (19.9%) qsa proj 1367 (8.7%) qsa indexer 0 (0.0%) qsa select 122 (0.8%) qsa attn 1860 (11.8%) router+shared 590 (3.8%) host grouping 40 (0.3%) gather 147 (0.9%) wait copy 22 (0.1%) dequant 307 (2.0%) gemm gate/up 2601 (16.5%) gemm down 1223 (7.8%) combine 363 (2.3%) ple 71 (0.5%) kv stage 0 (0.0%) gdn conv+gates 105 (0.7%) gdn recurrence 447 (2.8%) gdn out proj 1334 (8.5%)
strata prefill timing: host: chunk setup (PLE rows, the expert stream plan) 6 ms, waiting for each chunk 73 ms, after each chunk (the draft layer, progress) 63 ms, PLE 449 ms
strata decode timing: 38 windows, avg T 1.87, 1.58 tokens/window, 52.52 ms/window = verify 48.44 (GPU-reach wait 41.42 + per-layer host 2.52 [plan 0.16 actq 0.08 jobs 0.02 CPU 1.28] + stage 0.96) + commit/emit 0.28 + draft 3.70; per layer-window: CPU experts 0.19 (0.20 entries), VRAM hits 18.22, PCIe 0.00
strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 4.09 conv 0.20 ab 0.35 z 2.42 rec 0.88 out-proj 2.69 router+plan+bell 1.48 shared+quant 2.19 waitA 1.43 VRAM hits 9.56 waitB 0.32 PCIe grp 0.27 waitCPU 1.27 copy+combine 0.46 (gap) 0.34 head 2.78   hc0 norm 0.63   hc0 down 1.42   hc0 up 1.26   hc1 norm 0.33   hc1 down 1.39   hc1 up 1.34 | QSA layers: q8+kv-idx 0.13 k/v+norm-rope 0.33 kv+idx append 0.17 q+q-idx 1.78 scores+topk 0.62 kv-resolve 0.42 attention 0.98 gate 0.05 out-proj 0.95 router+plan+bell 0.49 shared+quant 0.72 waitA 0.18 VRAM hits 3.04 waitB 0.11 PCIe grp 0.09 waitCPU 0.06 copy+combine 0.15 (gap) 0.11   hc0 norm 0.10   hc0 down 0.47   hc0 up 0.42   hc1 norm 0.11   hc1 down 0.46   hc1 up 0.45 | total 49.49 ms/window over 39 windows
strata serve: prompt 7046 tokens = 0 reused + 7046 read in 16450 ms (428.3 tok/s), 60 generated in 1996 ms (30.1 tok/s), drafts accepted 22 of 33, 1 checkpoints
strata prefill timing: 7039 tokens, GPU timeline 15732 ms, wall 15796 ms, host staging 179 ms: embed+steps 3 (0.0%) hc read 1989 (12.6%) gdn 3141 (20.0%) qsa proj 1367 (8.7%) qsa indexer 0 (0.0%) qsa select 123 (0.8%) qsa attn 1860 (11.8%) router+shared 590 (3.8%) host grouping 39 (0.2%) gather 147 (0.9%) wait copy 24 (0.2%) dequant 307 (1.9%) gemm gate/up 2596 (16.5%) gemm down 1222 (7.8%) combine 363 (2.3%) ple 70 (0.4%) kv stage 0 (0.0%) gdn conv+gates 105 (0.7%) gdn recurrence 449 (2.9%) gdn out proj 1336 (8.5%)
strata prefill timing: host: chunk setup (PLE rows, the expert stream plan) 3 ms, waiting for each chunk 75 ms, after each chunk (the draft layer, progress) 63 ms, PLE 449 ms
strata decode timing: 39 windows, avg T 1.87, 1.54 tokens/window, 51.93 ms/window = verify 47.94 (GPU-reach wait 41.20 + per-layer host 2.13 [plan 0.17 actq 0.08 jobs 0.02 CPU 1.16] + stage 0.67) + commit/emit 0.19 + draft 3.67; per layer-window: CPU experts 0.17 (0.18 entries), VRAM hits 18.31, PCIe 0.00
strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 4.09 conv 0.20 ab 0.35 z 2.42 rec 0.88 out-proj 2.70 router+plan+bell 1.48 shared+quant 2.19 waitA 1.23 VRAM hits 9.60 waitB 0.32 PCIe grp 0.27 waitCPU 0.91 copy+combine 0.45 (gap) 0.34 head 2.78   hc0 norm 0.63   hc0 down 1.42   hc0 up 1.27   hc1 norm 0.33   hc1 down 1.38   hc1 up 1.34 | QSA layers: q8+kv-idx 0.13 k/v+norm-rope 0.33 kv+idx append 0.17 q+q-idx 1.78 scores+topk 0.63 kv-resolve 0.42 attention 0.97 gate 0.05 out-proj 0.95 router+plan+bell 0.49 shared+quant 0.72 waitA 0.12 VRAM hits 3.06 waitB 0.11 PCIe grp 0.09 waitCPU 0.13 copy+combine 0.15 (gap) 0.11   hc0 norm 0.10   hc0 down 0.47   hc0 up 0.42   hc1 norm 0.11   hc1 down 0.46   hc1 up 0.45 | total 48.99 ms/window over 40 windows
strata serve: prompt 7046 tokens = 0 reused + 7046 read in 16437 ms (428.7 tok/s), 60 generated in 2025 ms (29.6 tok/s), drafts accepted 21 of 34, 1 checkpoints
