================================================================
Krasis Benchmark — 2026-08-18 18:55:39
================================================================
Model:            DeepSeek-V4-Flash-0731
Architecture:     deepseek_v4, 43 layers, 256 experts, top-6
PP Partition:     [43] (1 GPUs)

Hardware:
  CPU:            AMD EPYC 7742 64-Core Processor (64 cores)
  RAM:            995 GB total, 152.2 GB used by process
  GPU 0:          NVIDIA RTX A6000 (49140 MB), 8300 MB allocated

Quantization:
  GPU experts:    INT4 (Marlin)
  CPU experts:    INT4
  Attention:      HQQ6
  Shared expert:  INT8
  Dense MLP:      INT8
  LM head:        INT8
  KV cache:       native

Strategy:
  Layer group size: 1 (layer_grouped(1))
  Prefill threshold: 1
  Mode:           gpu_decode (layer_grouped(1))

Prefill (internal) — 6 runs at different lengths:
  Run 1:  112.6 tok/s (1,000 tokens in 8878.1ms)
  Run 2:  484.1 tok/s (5,000 tokens in 10327.4ms)
  Run 3:  1,020.5 tok/s (10,000 tokens in 9799.6ms)
  Run 4:  1,162.0 tok/s (20,000 tokens in 17211.8ms)
  Run 5:  919.0 tok/s (35,000 tokens in 38083.0ms)
  Run 6:  983.7 tok/s (39,920 tokens in 40582.7ms)
  Best:   1,162.0 tok/s (20,000 tokens)

Decode (internal) — 50/100/250 tokens, 3 separate prompts:
   50 tokens: 13.36 tok/s (74.8ms/tok)
  100 tokens: 14.26 tok/s (70.1ms/tok)
  250 tokens: 13.51 tok/s (74.0ms/tok)
  Best:    14.26 tok/s
  HCS:     2775/11008 experts (25.2%)
  Min free VRAM: 1042 MB

Round trip (network) — 50/100/250 tokens via HTTP:
   50 tokens: 24.37 tok/s (10.52s total)
  100 tokens: 18.64 tok/s (13.78s total)
  250 tokens: 14.67 tok/s (25.35s total)
  Best:    24.37 tok/s
================================================================