spl16 commited on
Commit
126a524
Β·
verified Β·
1 Parent(s): 963d843

Upload index.html with huggingface_hub

Browse files
Files changed (1) hide show
  1. index.html +12 -12
index.html CHANGED
@@ -98,18 +98,18 @@
98
  <div class="field" data-show="gpu cpu"><label for="ram">System RAM</label><select id="ram"></select></div>
99
  <div class="field"><label for="ctx">Context length</label><select id="ctx"></select></div>
100
  <div class="field"><label for="kv">KV cache precision</label><select id="kv">
101
- <option value="f16/f16">Full (f16) β€” bench: reference</option>
102
- <option value="bf16/bf16">BF16 β€” bench: same bytes as f16, skip</option>
103
- <option value="q8_0/q8_0">q8_0 β€” bench: transparent</option>
104
- <option value="q8_0/q5_0">q8_0/q5_0 β€” bench: near-transparent</option>
105
- <option value="q5_0/q5_0">q5_0 β€” bench: tiny KL cost</option>
106
- <option value="q5_0/q4_1" selected>q5_0/q4_1 β€” bench knee, balanced default</option>
107
- <option value="q5_0/q4_0">q5_0/q4_0 β€” bench: compact, noisy on some models</option>
108
- <option value="q4_1/q4_1">q4_1 β€” bench: beats q4_0 at the same size</option>
109
- <option value="q4_1/q4_0">q4_1/q4_0 β€” bench: compact</option>
110
- <option value="q4_0/q4_0">q4_0 β€” bench: worst working pair</option>
111
- <option value="iq4_nl/iq4_nl">iq4_nl β€” bench: hangs at 64k, short ctx only</option>
112
- <option value="q5_1/q5_1">q5_1 β€” bench: same class as q5_0</option>
113
  </select></div>
114
  <div class="field"><label for="mmproj">Vision projector (vision models)</label><select id="mmproj">
115
  <option value="gpu">On the GPU, llama serve default</option>
 
98
  <div class="field" data-show="gpu cpu"><label for="ram">System RAM</label><select id="ram"></select></div>
99
  <div class="field"><label for="ctx">Context length</label><select id="ctx"></select></div>
100
  <div class="field"><label for="kv">KV cache precision</label><select id="kv">
101
+ <option value="f16/f16">Full (f16) β€” reference</option>
102
+ <option value="bf16/bf16">BF16 β€” not pareto optimal</option>
103
+ <option value="q8_0/q8_0">q8_0 β€” transparent</option>
104
+ <option value="q8_0/q5_0">q8_0/q5_0 β€” near-transparent</option>
105
+ <option value="q5_0/q5_0">q5_0 β€” not pareto optimal</option>
106
+ <option value="q5_0/q4_1" selected>q5_0/q4_1 β€” knee, default</option>
107
+ <option value="q5_0/q4_0">q5_0/q4_0 β€” not pareto optimal</option>
108
+ <option value="q4_1/q4_1">q4_1 β€” smallest frontier pair</option>
109
+ <option value="q4_1/q4_0">q4_1/q4_0 β€” not pareto optimal</option>
110
+ <option value="q4_0/q4_0">q4_0 β€” not pareto optimal</option>
111
+ <option value="iq4_nl/iq4_nl">iq4_nl β€” hangs at 64k</option>
112
+ <option value="q5_1/q5_1">q5_1 β€” not pareto optimal</option>
113
  </select></div>
114
  <div class="field"><label for="mmproj">Vision projector (vision models)</label><select id="mmproj">
115
  <option value="gpu">On the GPU, llama serve default</option>