recipes/docker-images/mia-tp4/docs/results/tp4/validation-20260925-v2.txt

# Recorded in knapcio/DeepSeek-V4.1-Flash-4x-DGX-Spark-TP4 (fresh clone of that repository on all four nodes); commit ids refer to it.
health 200 after 10 s
log /home/knapcio/ds41-verify-0925/logs-tp4/dsv41.log
{"label": "final-1", "kind": "prose", "step_ms": 33.62, "tok_per_step": 2.274, "tps": 64.1, "n": 6}
{"label": "final-1", "kind": "code", "step_ms": 39.44, "tok_per_step": 3.865, "tps": 98.2, "n": 6}
{"label": "final-2", "kind": "prose", "step_ms": 32.41, "tok_per_step": 2.274, "tps": 65.0, "n": 6}
{"label": "final-2", "kind": "code", "step_ms": 38.92, "tok_per_step": 3.865, "tps": 98.1, "n": 6}
prose c1 per 86.39 agg 86.39
prose c1 per 85.82 agg 85.82
prose c1 per 87.16 agg 87.16
prose c1 per 87.17 agg 87.17
prose c1 per 86.54 agg 86.54
prose c1 per 86.7 agg 86.7
prose c1 per 85.96 agg 85.96
prose c2 per 61.85 agg 120.42
prose c4 per 41.38 agg 163.59
prose c8 per 30.81 agg 237.62
prose c16 per 22.24 agg 342.67
code c1 per 122.56 agg 122.56
code c2 per 88.37 agg 174.95
code c4 per 63.21 agg 246.8
code c8 per 41.43 agg 309.8
code c16 per 29.33 agg 438.25
structured c1 per 152.37 agg 152.37
structured c2 per 103.28 agg 177.59
structured c4 per 70.43 agg 240.4
structured c8 per 44.19 agg 295.51
structured c16 per 44.43 agg 572.17
json c1 per 118.9 agg 118.9
json c2 per 89.81 agg 174.15
json c4 per 76.2 agg 301.73
json c8 per 60.19 agg 471.48
json c16 per 42.7 agg 659.85
prefill [(4113, 4029.18), (16404, 5768.47), (32787, 5796.75), (65555, 5818.47), (131094, 5644.35), (262164, 5213.97)]
prefill [(4115, 4760.23), (16402, 5709.4), (32786, 5766.79), (65559, 5774.06), (131088, 5622.69), (262161, 5225.27)]
{"label": "final-1", "4k": [3767, 0.94, 4011], "16k": [13327, 2.81, 4740], "32k": [31079, 6.46, 4808], "64k": [57853, 11.91, 4856], "128k": [121918, 25.69, 4746]}
{"label": "final-2", "4k": [4009, 0.95, 4204], "16k": [15421, 3.2, 4820], "32k": [25715, 5.28, 4870], "64k": [57747, 11.94, 4835], "128k": [121267, 25.48, 4759]}
{"prose": 58.36, "structural": 94.25, "other": 71.77, "label": "final"}
26000 20001 expected 7 got '7' prompt 129025 25.5 s
26000 3333 expected 331 got '331' prompt 129025 1.0 s
52000 17777 expected 439 got '446' prompt 259025 31.7 s
52000 20001 expected 7 got '7' prompt 259025 0.9 s
197000 123457 expected 199 got '14' prompt 984025 281.1 s
  985k low-water: 05:54:22 head MemAvailable GiB 6
72/75 passed (96.0%), 0 hit the token cap, wall 1.8 min, median 83.52 tok/s
      1 DSV41 hc_fused: first call (1071 rows) bit-identical to stock
      2 DSV41 hc_fused: hc mix stats for >= 512 rows in one K walk
      1 [2026-09-25 03:33:38 TP0] DSV41 fast load: phase=target read 155.2 GB eagerly over 45 shards, 0 shards skipped (92160 expert tensors TP-sliced, 83.0 GB kept); host buffers: 330 requested (83.1 GB), 330 slabs of 256 MB, p
      1 [2026-09-25 03:33:52 TP0] DSV41 fast load: phase=draft read 4.3 GB eagerly over 3 shards, 45 shards skipped (2304 expert tensors TP-sliced, 2.5 GB kept); host buffers: 11 requested (2.5 GB), 11 slabs of 256 MB, peak 7 li
      1 [2026-09-25 03:34:18 TP0] max_total_num_tokens=6034944, chunked_prefill_size=4096, max_prefill_tokens=16384, max_running_requests=16, context_len=1048576, available_gpu_mem=16.34 GB
      1 [moe_b12x_next] FusedOpPool('none', 'flashinfer_mxfp4') -> b12x_next
      1 [moe_b12x_next] armed: routed MoE on b12x_next a7d7d29b (W4A8, in-place repack); graph bs [1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16], rows 5/6, ladder [128, 256, 512, 1024, 2048, 4096], tune=True
      1 [moe_b12x_next] geometry ready {'geometry': {'E': 128, 'K': 5120, 'N': 576, 'topk': 3, 'offset': 0}, 'capacities': [5, 6, 10, 12, 15, 18, 20, 24, 25, 30, 35, 36, 40, 42, 48, 50, 60, 70, 72, 80, 84, 96, 128, 256, 512, 102
      1 [moe_b12x_next] geometry ready {'geometry': {'E': 384, 'K': 5120, 'N': 576, 'topk': 6, 'offset': 0}, 'capacities': [5, 6, 10, 12, 15, 18, 20, 24, 25, 30, 35, 36, 40, 42, 48, 50, 60, 70, 72, 80, 84, 96, 128, 256, 512, 102
FINAL-DONE