Will perform strided perplexity calculation -> adjusting context size from 3072 to 3264
llama_model_loader: loaded meta data with 39 key-value pairs and 363 tensors from granite-4.1-8b-Q4_K.gguf (version GGUF V3 (latest))
llama_model_loader: - type  f32:   81 tensors
llama_model_loader: - type q4_K:   90 tensors
llama_model_loader: - type q5_K:   79 tensors
llama_model_loader: - type iq3_xxs:    2 tensors
llama_model_loader: - type iq3_s:    1 tensors
llama_model_loader: - type iq4_xs:  110 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 4.61 GiB (4.50 BPW) 

====== Perplexity statistics ======
Mean PPL(Q)                   :   8.867438 ±   0.067303
Mean PPL(base)                :   8.691178 ±   0.065443
Cor(ln(PPL(Q)), ln(PPL(base))):  98.88%
Mean ln(PPL(Q)/PPL(base))     :   0.020077 ±   0.001134
Mean PPL(Q)/PPL(base)         :   1.020280 ±   0.001157
Mean PPL(Q)-PPL(base)         :   0.176260 ±   0.010114

====== KL divergence statistics ======
Mean    KLD:   0.047917 ±   0.000392
Maximum KLD:   8.513770
99.9%   KLD:   1.939411
99.0%   KLD:   0.563371
95.0%   KLD:   0.167563
90.0%   KLD:   0.095562
Median  KLD:   0.017061
10.0%   KLD:   0.000210
 5.0%   KLD:   0.000043
 1.0%   KLD:   0.000003
 0.1%   KLD:  -0.000001
Minimum KLD:  -0.000004

====== Token probability statistics ======
Mean    Δp: -0.365 ± 0.017 %
Maximum Δp: 95.922%
99.9%   Δp: 43.653%
99.0%   Δp: 16.925%
95.0%   Δp:  6.946%
90.0%   Δp:  3.742%
75.0%   Δp:  0.557%
Median  Δp: -0.002%
25.0%   Δp: -0.901%
10.0%   Δp: -4.570%
 5.0%   Δp: -8.319%
 1.0%   Δp: -22.260%
 0.1%   Δp: -58.978%
Minimum Δp: -98.363%
RMS Δp    :  6.446 ± 0.053 %
Same top p: 90.937 ± 0.076 %