Nekochu commited on
Commit
9021aa9
·
verified ·
1 Parent(s): 4bc36ac

Add README, KLD plots, measurement

Browse files
Files changed (6) hide show
  1. .gitattributes +3 -0
  2. README.md +186 -0
  3. example_dashboard.jpg +3 -0
  4. kld_plot.png +3 -0
  5. kld_spread_plot.png +3 -0
  6. measurement.json +2388 -0
.gitattributes CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ kld_plot.png filter=lfs diff=lfs merge=lfs -text
37
+ kld_spread_plot.png filter=lfs diff=lfs merge=lfs -text
38
+ example_dashboard.jpg filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,186 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: microsoft/AesCode-32B
4
+ base_model_relation: quantized
5
+ quantized_by: WeReCooking
6
+ pipeline_tag: image-text-to-text
7
+ tags:
8
+ - exl3
9
+ - qwen3_vl
10
+ - html
11
+ ---
12
+
13
+ <style>
14
+ .container-dark { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Arial, sans-serif; line-height: 1.6; color: #d4d4d4; }
15
+ a { color: #569cd6; text-decoration: none; font-weight: 600; }
16
+ a:hover { text-decoration: underline; }
17
+ .card-dark { background-color: #252526; border-radius: 12px; padding: 24px; margin-bottom: 20px; box-shadow: 0 4px 12px rgba(0,0,0,0.3); border: 1px solid #3c3c3c; }
18
+ .card-dark h1 { font-size: 2.2em; color: #ffffff; text-align: center; margin-bottom: 10px; }
19
+ .card-dark .subtitle { text-align: center; font-size: 1.1em; color: #a0a0a0; }
20
+ .card-dark h2 { font-size: 1.5em; margin-top: 0; padding-bottom: 10px; border-bottom: 1px solid #3c3c3c; color: #c586c0; }
21
+ .styled-table { display: table; border: none; width: 100%; font-size: 0.95em; }
22
+ .styled-table thead th { background-color: #333333; color: #c586c0; text-align: left; padding: 12px 15px; }
23
+ .styled-table td { padding: 0; border-bottom: 1px solid #3c3c3c; }
24
+ .styled-table tbody tr { transition: background-color 0.1s ease; }
25
+ .styled-table tbody tr:hover { background-color: #3a3a3a; }
26
+ .styled-table tr:last-child td { border-bottom: none; }
27
+ .styled-table td a { display: block; padding: 12px 15px; }
28
+ .styled-table td a.fake-link { text-decoration:none; color:inherit; }
29
+ details { margin-top: 20px; border: 1px solid #3c3c3c; border-radius: 8px; overflow: hidden; }
30
+ summary { cursor: pointer; padding: 12px 18px; background-color: #6A5ACD; font-weight: 600; display: flex; align-items: center; gap: 10px; justify-content: space-between; list-style: none; }
31
+ summary::-webkit-details-marker { display: none; }
32
+ summary:hover { filter: brightness(1.1); }
33
+ summary::after { content: ''; display: inline-block; width: 8px; height: 8px; border-bottom: 2px solid white; border-right: 2px solid white; transform: rotate(45deg); transition: transform 0.3s ease; }
34
+ details[open] > summary::after { transform: rotate(225deg); }
35
+ .details-content { padding: 18px; }
36
+ </style>
37
+
38
+ <div class="container-dark">
39
+
40
+ <div class="card-dark">
41
+ <h1>AesCode-32B EXL3</h1>
42
+ <p class="subtitle">
43
+ EXL3 quants of <a href="https://huggingface.co/microsoft/AesCode-32B">microsoft/AesCode-32B</a>
44
+ using <a href="https://github.com/turboderp-org/exllamav3/">exllamav3</a> v1.6.1
45
+ </p>
46
+ </div>
47
+
48
+ <div class="card-dark">
49
+ <h2>KL Divergence vs VRAM</h2>
50
+ <img src="kld_plot.png" alt="KLD plot" style="width:100%; border-radius: 8px;" />
51
+ <p class="subtitle">Reference: BF16 original (PPL 6.843). Noise floor KLD 0.000151. Measured with exllamav3 <code>eval/qbench.py</code> on wikitext-2 (20 rows, 2048 ctx). VRAM = decoder + lm_head weights.</p>
52
+ <details>
53
+ <summary>KLD spread (median, IQR, noise floor)</summary>
54
+ <div class="details-content">
55
+ <img src="kld_spread_plot.png" alt="KLD spread" style="width:100%; border-radius: 8px;" />
56
+ </div>
57
+ </details>
58
+ </div>
59
+
60
+ <div class="card-dark">
61
+ <h2>Quants</h2>
62
+ <table class="styled-table">
63
+ <thead>
64
+ <tr><th>Branch</th><th>BPW</th><th>Head</th><th>VRAM (GiB)</th><th>KLD mean</th><th>KLD median</th><th>PPL</th><th>Type</th></tr>
65
+ </thead>
66
+ <tbody>
67
+ <tr>
68
+ <td><a href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">3.0bpw_H6</a></td>
69
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">3.0</a></td>
70
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">6</a></td>
71
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">11.5</a></td>
72
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">0.0360</a></td>
73
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">0.01522</a></td>
74
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">6.987</a></td>
75
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.0bpw_H6">base</a></td>
76
+ </tr>
77
+ <tr>
78
+ <td><a href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">3.73bpw_H6</a></td>
79
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">3.73</a></td>
80
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">6</a></td>
81
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">14.1</a></td>
82
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">0.0157</a></td>
83
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">0.00616</a></td>
84
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">6.902</a></td>
85
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/3.73bpw_H6">optimized</a></td>
86
+ </tr>
87
+ <tr>
88
+ <td><a href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">4.0bpw_H6</a></td>
89
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">4.0</a></td>
90
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">6</a></td>
91
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">15.1</a></td>
92
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">0.0093</a></td>
93
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">0.00392</a></td>
94
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">6.876</a></td>
95
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.0bpw_H6">base</a></td>
96
+ </tr>
97
+ <tr>
98
+ <td><a href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">4.54bpw_H6</a></td>
99
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">4.54</a></td>
100
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">6</a></td>
101
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">17.0</a></td>
102
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">0.0065</a></td>
103
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">0.00260</a></td>
104
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">6.863</a></td>
105
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/4.54bpw_H6">optimized</a></td>
106
+ </tr>
107
+ <tr>
108
+ <td><a href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">5.0bpw_H6</a></td>
109
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">5.0</a></td>
110
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">6</a></td>
111
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">18.7</a></td>
112
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">0.0027</a></td>
113
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">0.00122</a></td>
114
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">6.871</a></td>
115
+ <td><a class="fake-link" href="https://huggingface.co/WeReCooking/AesCode-32B-EXL3/tree/5.0bpw_H6">base</a></td>
116
+ </tr>
117
+ </tbody>
118
+ </table>
119
+ <p class="subtitle">Bases are direct converts. Optimized = KLD-guided tensor mix of the bases + attention recompiled at 5bpw; on this model they land on the base size/KLD curve, so treat them as in-between sizes and pick by VRAM. Vision tower 6bpw in every branch.</p>
120
+ </div>
121
+
122
+ <div class="card-dark">
123
+ <h2>Download and run</h2>
124
+ <details>
125
+ <summary>Download</summary>
126
+ <div class="details-content">
127
+ <pre><code>pip install -U "huggingface_hub"
128
+ hf download WeReCooking/AesCode-32B-EXL3 --revision 4.0bpw_H6 --local-dir ./AesCode-32B-EXL3</code></pre>
129
+ </div>
130
+ </details>
131
+ <details>
132
+ <summary>Run with exllamav3</summary>
133
+ <div class="details-content">
134
+ <b>Text chat, official exllamav3 CLI</b> (model card sampling). Install exllamav3 first (<a href="https://github.com/turboderp-org/exllamav3/releases">prebuilt wheels</a>):
135
+ <pre><code>git clone https://github.com/turboderp-org/exllamav3 && cd exllamav3
136
+ pip install pyperclip prompt_toolkit
137
+ python examples/chat.py -m ../AesCode-32B-EXL3 -mode chatml -cs 24576 -maxr 12000 -temp 0.8 -topp 0.95 -topk 20 -minp 0</code></pre>
138
+ <b>One shot to file:</b>
139
+ <pre><code>python examples/chat.py -m ../AesCode-32B-EXL3 -mode chatml -cs 24576 -maxr 12000 -temp 0.8 -topp 0.95 -topk 20 -minp 0 -prompt "Create a portrait poster infographic about coffee origins as one complete HTML page." -save page.md</code></pre>
140
+ <b>With a reference image</b> (how AesCode is meant to be used, based on <code>examples/multimodal.py</code>):
141
+ <pre><code>from exllamav3 import Cache, ComboSampler, Config, Generator, Model, Tokenizer
142
+ from PIL import Image
143
+
144
+ config = Config.from_directory("AesCode-32B-EXL3")
145
+ vision = Model.from_config(config, component="vision"); vision.load()
146
+ model = Model.from_config(config); cache = Cache(model, max_num_tokens=24576); model.load()
147
+ tok = Tokenizer.from_config(config)
148
+ emb = vision.get_image_embeddings(tokenizer=tok, image=Image.open("reference.png"))
149
+ prompt = ("&lt;|im_start|&gt;user\n" + emb.text_alias + "Create a landscape dashboard ... as one complete HTML page."
150
+ "&lt;|im_end|&gt;\n&lt;|im_start|&gt;assistant\n")
151
+ gen = Generator(model=model, cache=cache, tokenizer=tok)
152
+ print(gen.generate(prompt=prompt, embeddings=[emb], encode_special_tokens=True, completion_only=True,
153
+ max_new_tokens=12000, stop_conditions=["&lt;|im_end|&gt;"],
154
+ sampler=ComboSampler(temperature=0.8, top_p=0.95, top_k=20)))</code></pre>
155
+ <p>The repo's training system prompt and spec template (<a href="https://github.com/microsoft/AesCode/blob/main/inference/prompts.py">inference/prompts.py</a>) give the best results.</p>
156
+ </div>
157
+ </details>
158
+ </div>
159
+
160
+ <div class="card-dark">
161
+ <h2>Example output (4.0bpw_H6)</h2>
162
+ <img src="example_dashboard.jpg" alt="Reference image vs EXL3 output" style="width:100%; border-radius: 8px;" />
163
+ <p class="subtitle">Repo example <code>hospital_dashboard</code> with the training prompt template and reference image. Exact text 11/11, chart data 5/5, table 0/1, 1 element outside the 1280x720 canvas. 5097 tokens at 58.0 tok/s on an RTX 5090.</p>
164
+ </div>
165
+
166
+ <div class="card-dark">
167
+ <h2>Build Details</h2>
168
+ <details>
169
+ <summary>How these were made</summary>
170
+ <div class="details-content">
171
+ <p><b>Base quants:</b> <code>convert.py -b &lt;bpw&gt;</code> (3.0, 4.0, 5.0), head 6bpw, vision 6bpw. Calibration: 250 rows x 2048 (default) for 3.0, 128 rows x 2048 for 4.0/5.0 (32 GB RAM build box)</p>
172
+ <p><b>KLD measurement:</b> <code>util/measure.py -r &lt;bf16&gt; -i &lt;3.0&gt; &lt;4.0&gt; &lt;5.0&gt;</code> (<code>measurement.json</code> on main)</p>
173
+ <p><b>Optimized:</b> <code>util/optimize.py -i &lt;3.0&gt; &lt;4.0&gt; &lt;5.0&gt; -m measurement.json -b &lt;target&gt;</code> then <code>util/recompile.py -or override.yaml</code> with <code>*.self_attn.*</code> from 5.0bpw</p>
174
+ <p><b>Eval:</b> <code>eval/qbench.py</code>, BF16 streamed as reference</p>
175
+ <p>Docs: <a href="https://github.com/turboderp-org/exllamav3/blob/master/doc/convert.md">exllamav3 convert.md</a></p>
176
+ </div>
177
+ </details>
178
+ </div>
179
+
180
+ <div class="card-dark">
181
+ <h2>Files</h2>
182
+ <p><code>main</code> branch: <code>measurement.json</code>, KLD plots, example image, README</p>
183
+ <p>Each bpw branch: quantized shards (text + vision) + config + tokenizer + processor configs</p>
184
+ </div>
185
+
186
+ </div>
example_dashboard.jpg ADDED

Git LFS Details

  • SHA256: 42b0f7fc793bd550318618136e772f5f0b123e09900b6b08f3f32a04e9ea89c4
  • Pointer size: 131 Bytes
  • Size of remote file: 406 kB
kld_plot.png ADDED

Git LFS Details

  • SHA256: 9f7252c1107e4b7fb1d3eaeefcc00f631d83d56b17f81605c7dce0b7d9dcbe50
  • Pointer size: 131 Bytes
  • Size of remote file: 159 kB
kld_spread_plot.png ADDED

Git LFS Details

  • SHA256: 0f6be599fa819ddcd59f24809aeb22d027299d97ff0f39f0a6f7af3b2a1c0540
  • Pointer size: 131 Bytes
  • Size of remote file: 280 kB
measurement.json ADDED
@@ -0,0 +1,2388 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base": {
3
+ "dir": "D:\\main\\program\\AI\\tmp\\exl3_aescode\\quants\\3.0bpw_H6",
4
+ "bpw": 3.004301534673219
5
+ },
6
+ "alts": [
7
+ {
8
+ "dir": "D:\\main\\program\\AI\\tmp\\exl3_aescode\\quants\\4.0bpw_H6",
9
+ "bpw": 4.004301534673219
10
+ },
11
+ {
12
+ "dir": "D:\\main\\program\\AI\\tmp\\exl3_aescode\\quants\\5.0bpw_H6",
13
+ "bpw": 5.004301534673219
14
+ }
15
+ ],
16
+ "groups": [
17
+ {
18
+ "idx": 0,
19
+ "layers": [
20
+ "model.language_model.layers.0.self_attn.q_proj",
21
+ "model.language_model.layers.0.self_attn.k_proj",
22
+ "model.language_model.layers.0.self_attn.v_proj",
23
+ "model.language_model.layers.0.self_attn.o_proj"
24
+ ],
25
+ "candidates": [
26
+ {
27
+ "dkld": 0.005342455487698294,
28
+ "dbits": 94371840
29
+ },
30
+ {
31
+ "dkld": 0.002878085989505036,
32
+ "dbits": 188743680
33
+ }
34
+ ]
35
+ },
36
+ {
37
+ "idx": 1,
38
+ "layers": [
39
+ "model.language_model.layers.0.mlp.gate_proj",
40
+ "model.language_model.layers.0.mlp.up_proj",
41
+ "model.language_model.layers.0.mlp.down_proj"
42
+ ],
43
+ "candidates": [
44
+ {
45
+ "dkld": 0.0010677056852728128,
46
+ "dbits": 393216000
47
+ },
48
+ {
49
+ "dkld": -0.050666560931131255,
50
+ "dbits": 786432000
51
+ }
52
+ ]
53
+ },
54
+ {
55
+ "idx": 2,
56
+ "layers": [
57
+ "model.language_model.layers.1.self_attn.q_proj",
58
+ "model.language_model.layers.1.self_attn.k_proj",
59
+ "model.language_model.layers.1.self_attn.v_proj",
60
+ "model.language_model.layers.1.self_attn.o_proj"
61
+ ],
62
+ "candidates": [
63
+ {
64
+ "dkld": -0.1397632632870227,
65
+ "dbits": 94371840
66
+ },
67
+ {
68
+ "dkld": -0.15157142621465028,
69
+ "dbits": 188743680
70
+ }
71
+ ]
72
+ },
73
+ {
74
+ "idx": 3,
75
+ "layers": [
76
+ "model.language_model.layers.1.mlp.gate_proj",
77
+ "model.language_model.layers.1.mlp.up_proj",
78
+ "model.language_model.layers.1.mlp.down_proj"
79
+ ],
80
+ "candidates": [
81
+ {
82
+ "dkld": 0.0028491628821939174,
83
+ "dbits": 393216000
84
+ },
85
+ {
86
+ "dkld": 0.0034188396297395174,
87
+ "dbits": 786432000
88
+ }
89
+ ]
90
+ },
91
+ {
92
+ "idx": 4,
93
+ "layers": [
94
+ "model.language_model.layers.2.self_attn.q_proj",
95
+ "model.language_model.layers.2.self_attn.k_proj",
96
+ "model.language_model.layers.2.self_attn.v_proj",
97
+ "model.language_model.layers.2.self_attn.o_proj"
98
+ ],
99
+ "candidates": [
100
+ {
101
+ "dkld": -0.1418308971915394,
102
+ "dbits": 94371840
103
+ },
104
+ {
105
+ "dkld": -0.011705041117966192,
106
+ "dbits": 188743680
107
+ }
108
+ ]
109
+ },
110
+ {
111
+ "idx": 5,
112
+ "layers": [
113
+ "model.language_model.layers.2.mlp.gate_proj",
114
+ "model.language_model.layers.2.mlp.up_proj",
115
+ "model.language_model.layers.2.mlp.down_proj"
116
+ ],
117
+ "candidates": [
118
+ {
119
+ "dkld": -0.0067629903089255206,
120
+ "dbits": 393216000
121
+ },
122
+ {
123
+ "dkld": -0.00024901125580073513,
124
+ "dbits": 786432000
125
+ }
126
+ ]
127
+ },
128
+ {
129
+ "idx": 6,
130
+ "layers": [
131
+ "model.language_model.layers.3.self_attn.q_proj",
132
+ "model.language_model.layers.3.self_attn.k_proj",
133
+ "model.language_model.layers.3.self_attn.v_proj",
134
+ "model.language_model.layers.3.self_attn.o_proj"
135
+ ],
136
+ "candidates": [
137
+ {
138
+ "dkld": -0.0072793466039002,
139
+ "dbits": 94371840
140
+ },
141
+ {
142
+ "dkld": -0.1536645043641329,
143
+ "dbits": 188743680
144
+ }
145
+ ]
146
+ },
147
+ {
148
+ "idx": 7,
149
+ "layers": [
150
+ "model.language_model.layers.3.mlp.gate_proj",
151
+ "model.language_model.layers.3.mlp.up_proj",
152
+ "model.language_model.layers.3.mlp.down_proj"
153
+ ],
154
+ "candidates": [
155
+ {
156
+ "dkld": -0.006463171308860205,
157
+ "dbits": 393216000
158
+ },
159
+ {
160
+ "dkld": 0.008568887179717405,
161
+ "dbits": 786432000
162
+ }
163
+ ]
164
+ },
165
+ {
166
+ "idx": 8,
167
+ "layers": [
168
+ "model.language_model.layers.4.self_attn.q_proj",
169
+ "model.language_model.layers.4.self_attn.k_proj",
170
+ "model.language_model.layers.4.self_attn.v_proj",
171
+ "model.language_model.layers.4.self_attn.o_proj"
172
+ ],
173
+ "candidates": [
174
+ {
175
+ "dkld": 0.004537224862724537,
176
+ "dbits": 94371840
177
+ },
178
+ {
179
+ "dkld": 0.0033987248782068324,
180
+ "dbits": 188743680
181
+ }
182
+ ]
183
+ },
184
+ {
185
+ "idx": 9,
186
+ "layers": [
187
+ "model.language_model.layers.4.mlp.gate_proj",
188
+ "model.language_model.layers.4.mlp.up_proj",
189
+ "model.language_model.layers.4.mlp.down_proj"
190
+ ],
191
+ "candidates": [
192
+ {
193
+ "dkld": 0.00795677597634492,
194
+ "dbits": 393216000
195
+ },
196
+ {
197
+ "dkld": -0.1482118721585721,
198
+ "dbits": 786432000
199
+ }
200
+ ]
201
+ },
202
+ {
203
+ "idx": 10,
204
+ "layers": [
205
+ "model.language_model.layers.5.self_attn.q_proj",
206
+ "model.language_model.layers.5.self_attn.k_proj",
207
+ "model.language_model.layers.5.self_attn.v_proj",
208
+ "model.language_model.layers.5.self_attn.o_proj"
209
+ ],
210
+ "candidates": [
211
+ {
212
+ "dkld": -0.009429611032828689,
213
+ "dbits": 94371840
214
+ },
215
+ {
216
+ "dkld": -0.014525620499625802,
217
+ "dbits": 188743680
218
+ }
219
+ ]
220
+ },
221
+ {
222
+ "idx": 11,
223
+ "layers": [
224
+ "model.language_model.layers.5.mlp.gate_proj",
225
+ "model.language_model.layers.5.mlp.up_proj",
226
+ "model.language_model.layers.5.mlp.down_proj"
227
+ ],
228
+ "candidates": [
229
+ {
230
+ "dkld": 0.002557756146416068,
231
+ "dbits": 393216000
232
+ },
233
+ {
234
+ "dkld": 0.004029345652088512,
235
+ "dbits": 786432000
236
+ }
237
+ ]
238
+ },
239
+ {
240
+ "idx": 12,
241
+ "layers": [
242
+ "model.language_model.layers.6.self_attn.q_proj",
243
+ "model.language_model.layers.6.self_attn.k_proj",
244
+ "model.language_model.layers.6.self_attn.v_proj",
245
+ "model.language_model.layers.6.self_attn.o_proj"
246
+ ],
247
+ "candidates": [
248
+ {
249
+ "dkld": -0.05095413839444518,
250
+ "dbits": 94371840
251
+ },
252
+ {
253
+ "dkld": -0.060551986517384665,
254
+ "dbits": 188743680
255
+ }
256
+ ]
257
+ },
258
+ {
259
+ "idx": 13,
260
+ "layers": [
261
+ "model.language_model.layers.6.mlp.gate_proj",
262
+ "model.language_model.layers.6.mlp.up_proj",
263
+ "model.language_model.layers.6.mlp.down_proj"
264
+ ],
265
+ "candidates": [
266
+ {
267
+ "dkld": 0.0034258951433002727,
268
+ "dbits": 393216000
269
+ },
270
+ {
271
+ "dkld": 0.004030669433996076,
272
+ "dbits": 786432000
273
+ }
274
+ ]
275
+ },
276
+ {
277
+ "idx": 14,
278
+ "layers": [
279
+ "model.language_model.layers.7.self_attn.q_proj",
280
+ "model.language_model.layers.7.self_attn.k_proj",
281
+ "model.language_model.layers.7.self_attn.v_proj",
282
+ "model.language_model.layers.7.self_attn.o_proj"
283
+ ],
284
+ "candidates": [
285
+ {
286
+ "dkld": 0.00012257243506608173,
287
+ "dbits": 94371840
288
+ },
289
+ {
290
+ "dkld": -2.733203582466759e-05,
291
+ "dbits": 188743680
292
+ }
293
+ ]
294
+ },
295
+ {
296
+ "idx": 15,
297
+ "layers": [
298
+ "model.language_model.layers.7.mlp.gate_proj",
299
+ "model.language_model.layers.7.mlp.up_proj",
300
+ "model.language_model.layers.7.mlp.down_proj"
301
+ ],
302
+ "candidates": [
303
+ {
304
+ "dkld": -0.00040053990669550177,
305
+ "dbits": 393216000
306
+ },
307
+ {
308
+ "dkld": -0.00045577390119433403,
309
+ "dbits": 786432000
310
+ }
311
+ ]
312
+ },
313
+ {
314
+ "idx": 16,
315
+ "layers": [
316
+ "model.language_model.layers.8.self_attn.q_proj",
317
+ "model.language_model.layers.8.self_attn.k_proj",
318
+ "model.language_model.layers.8.self_attn.v_proj",
319
+ "model.language_model.layers.8.self_attn.o_proj"
320
+ ],
321
+ "candidates": [
322
+ {
323
+ "dkld": 3.334633074700277e-05,
324
+ "dbits": 94371840
325
+ },
326
+ {
327
+ "dkld": -0.00012314599007368088,
328
+ "dbits": 188743680
329
+ }
330
+ ]
331
+ },
332
+ {
333
+ "idx": 17,
334
+ "layers": [
335
+ "model.language_model.layers.8.mlp.gate_proj",
336
+ "model.language_model.layers.8.mlp.up_proj",
337
+ "model.language_model.layers.8.mlp.down_proj"
338
+ ],
339
+ "candidates": [
340
+ {
341
+ "dkld": -0.00010292534716427326,
342
+ "dbits": 393216000
343
+ },
344
+ {
345
+ "dkld": -7.305359467865702e-05,
346
+ "dbits": 786432000
347
+ }
348
+ ]
349
+ },
350
+ {
351
+ "idx": 18,
352
+ "layers": [
353
+ "model.language_model.layers.9.self_attn.q_proj",
354
+ "model.language_model.layers.9.self_attn.k_proj",
355
+ "model.language_model.layers.9.self_attn.v_proj",
356
+ "model.language_model.layers.9.self_attn.o_proj"
357
+ ],
358
+ "candidates": [
359
+ {
360
+ "dkld": 2.7341675013281552e-05,
361
+ "dbits": 94371840
362
+ },
363
+ {
364
+ "dkld": 9.749447926876709e-05,
365
+ "dbits": 188743680
366
+ }
367
+ ]
368
+ },
369
+ {
370
+ "idx": 19,
371
+ "layers": [
372
+ "model.language_model.layers.9.mlp.gate_proj",
373
+ "model.language_model.layers.9.mlp.up_proj",
374
+ "model.language_model.layers.9.mlp.down_proj"
375
+ ],
376
+ "candidates": [
377
+ {
378
+ "dkld": 0.0003252701833844074,
379
+ "dbits": 393216000
380
+ },
381
+ {
382
+ "dkld": 0.00020022029057142898,
383
+ "dbits": 786432000
384
+ }
385
+ ]
386
+ },
387
+ {
388
+ "idx": 20,
389
+ "layers": [
390
+ "model.language_model.layers.10.self_attn.q_proj",
391
+ "model.language_model.layers.10.self_attn.k_proj",
392
+ "model.language_model.layers.10.self_attn.v_proj",
393
+ "model.language_model.layers.10.self_attn.o_proj"
394
+ ],
395
+ "candidates": [
396
+ {
397
+ "dkld": 0.00011542919091878101,
398
+ "dbits": 94371840
399
+ },
400
+ {
401
+ "dkld": 7.19995703548082e-05,
402
+ "dbits": 188743680
403
+ }
404
+ ]
405
+ },
406
+ {
407
+ "idx": 21,
408
+ "layers": [
409
+ "model.language_model.layers.10.mlp.gate_proj",
410
+ "model.language_model.layers.10.mlp.up_proj",
411
+ "model.language_model.layers.10.mlp.down_proj"
412
+ ],
413
+ "candidates": [
414
+ {
415
+ "dkld": -0.0002729880623519587,
416
+ "dbits": 393216000
417
+ },
418
+ {
419
+ "dkld": -0.00010036383755507305,
420
+ "dbits": 786432000
421
+ }
422
+ ]
423
+ },
424
+ {
425
+ "idx": 22,
426
+ "layers": [
427
+ "model.language_model.layers.11.self_attn.q_proj",
428
+ "model.language_model.layers.11.self_attn.k_proj",
429
+ "model.language_model.layers.11.self_attn.v_proj",
430
+ "model.language_model.layers.11.self_attn.o_proj"
431
+ ],
432
+ "candidates": [
433
+ {
434
+ "dkld": 4.468513652680639e-05,
435
+ "dbits": 94371840
436
+ },
437
+ {
438
+ "dkld": 5.694916471837841e-05,
439
+ "dbits": 188743680
440
+ }
441
+ ]
442
+ },
443
+ {
444
+ "idx": 23,
445
+ "layers": [
446
+ "model.language_model.layers.11.mlp.gate_proj",
447
+ "model.language_model.layers.11.mlp.up_proj",
448
+ "model.language_model.layers.11.mlp.down_proj"
449
+ ],
450
+ "candidates": [
451
+ {
452
+ "dkld": -0.000621724314987665,
453
+ "dbits": 393216000
454
+ },
455
+ {
456
+ "dkld": -0.0003705095499754174,
457
+ "dbits": 786432000
458
+ }
459
+ ]
460
+ },
461
+ {
462
+ "idx": 24,
463
+ "layers": [
464
+ "model.language_model.layers.12.self_attn.q_proj",
465
+ "model.language_model.layers.12.self_attn.k_proj",
466
+ "model.language_model.layers.12.self_attn.v_proj",
467
+ "model.language_model.layers.12.self_attn.o_proj"
468
+ ],
469
+ "candidates": [
470
+ {
471
+ "dkld": 2.4668266996735744e-05,
472
+ "dbits": 94371840
473
+ },
474
+ {
475
+ "dkld": -9.445403702557087e-05,
476
+ "dbits": 188743680
477
+ }
478
+ ]
479
+ },
480
+ {
481
+ "idx": 25,
482
+ "layers": [
483
+ "model.language_model.layers.12.mlp.gate_proj",
484
+ "model.language_model.layers.12.mlp.up_proj",
485
+ "model.language_model.layers.12.mlp.down_proj"
486
+ ],
487
+ "candidates": [
488
+ {
489
+ "dkld": -0.0003460992127657103,
490
+ "dbits": 393216000
491
+ },
492
+ {
493
+ "dkld": -0.00034626852720975876,
494
+ "dbits": 786432000
495
+ }
496
+ ]
497
+ },
498
+ {
499
+ "idx": 26,
500
+ "layers": [
501
+ "model.language_model.layers.13.self_attn.q_proj",
502
+ "model.language_model.layers.13.self_attn.k_proj",
503
+ "model.language_model.layers.13.self_attn.v_proj",
504
+ "model.language_model.layers.13.self_attn.o_proj"
505
+ ],
506
+ "candidates": [
507
+ {
508
+ "dkld": -0.000152776110917352,
509
+ "dbits": 94371840
510
+ },
511
+ {
512
+ "dkld": -0.00025113103911281187,
513
+ "dbits": 188743680
514
+ }
515
+ ]
516
+ },
517
+ {
518
+ "idx": 27,
519
+ "layers": [
520
+ "model.language_model.layers.13.mlp.gate_proj",
521
+ "model.language_model.layers.13.mlp.up_proj",
522
+ "model.language_model.layers.13.mlp.down_proj"
523
+ ],
524
+ "candidates": [
525
+ {
526
+ "dkld": -0.0007007658015936724,
527
+ "dbits": 393216000
528
+ },
529
+ {
530
+ "dkld": -0.00030329353176059004,
531
+ "dbits": 786432000
532
+ }
533
+ ]
534
+ },
535
+ {
536
+ "idx": 28,
537
+ "layers": [
538
+ "model.language_model.layers.14.self_attn.q_proj",
539
+ "model.language_model.layers.14.self_attn.k_proj",
540
+ "model.language_model.layers.14.self_attn.v_proj",
541
+ "model.language_model.layers.14.self_attn.o_proj"
542
+ ],
543
+ "candidates": [
544
+ {
545
+ "dkld": 4.437807947396677e-05,
546
+ "dbits": 94371840
547
+ },
548
+ {
549
+ "dkld": -2.7742749080073015e-05,
550
+ "dbits": 188743680
551
+ }
552
+ ]
553
+ },
554
+ {
555
+ "idx": 29,
556
+ "layers": [
557
+ "model.language_model.layers.14.mlp.gate_proj",
558
+ "model.language_model.layers.14.mlp.up_proj",
559
+ "model.language_model.layers.14.mlp.down_proj"
560
+ ],
561
+ "candidates": [
562
+ {
563
+ "dkld": -0.00022678198292852558,
564
+ "dbits": 393216000
565
+ },
566
+ {
567
+ "dkld": -0.0002609751652926373,
568
+ "dbits": 786432000
569
+ }
570
+ ]
571
+ },
572
+ {
573
+ "idx": 30,
574
+ "layers": [
575
+ "model.language_model.layers.15.self_attn.q_proj",
576
+ "model.language_model.layers.15.self_attn.k_proj",
577
+ "model.language_model.layers.15.self_attn.v_proj",
578
+ "model.language_model.layers.15.self_attn.o_proj"
579
+ ],
580
+ "candidates": [
581
+ {
582
+ "dkld": -0.00016537196934224285,
583
+ "dbits": 94371840
584
+ },
585
+ {
586
+ "dkld": -0.00027730665169656277,
587
+ "dbits": 188743680
588
+ }
589
+ ]
590
+ },
591
+ {
592
+ "idx": 31,
593
+ "layers": [
594
+ "model.language_model.layers.15.mlp.gate_proj",
595
+ "model.language_model.layers.15.mlp.up_proj",
596
+ "model.language_model.layers.15.mlp.down_proj"
597
+ ],
598
+ "candidates": [
599
+ {
600
+ "dkld": -0.0002389888744801505,
601
+ "dbits": 393216000
602
+ },
603
+ {
604
+ "dkld": -0.00022584153339266777,
605
+ "dbits": 786432000
606
+ }
607
+ ]
608
+ },
609
+ {
610
+ "idx": 32,
611
+ "layers": [
612
+ "model.language_model.layers.16.self_attn.q_proj",
613
+ "model.language_model.layers.16.self_attn.k_proj",
614
+ "model.language_model.layers.16.self_attn.v_proj",
615
+ "model.language_model.layers.16.self_attn.o_proj"
616
+ ],
617
+ "candidates": [
618
+ {
619
+ "dkld": 8.467026054859161e-05,
620
+ "dbits": 94371840
621
+ },
622
+ {
623
+ "dkld": 6.083534099160515e-05,
624
+ "dbits": 188743680
625
+ }
626
+ ]
627
+ },
628
+ {
629
+ "idx": 33,
630
+ "layers": [
631
+ "model.language_model.layers.16.mlp.gate_proj",
632
+ "model.language_model.layers.16.mlp.up_proj",
633
+ "model.language_model.layers.16.mlp.down_proj"
634
+ ],
635
+ "candidates": [
636
+ {
637
+ "dkld": -0.00029788915999234833,
638
+ "dbits": 393216000
639
+ },
640
+ {
641
+ "dkld": -0.00023291488178076025,
642
+ "dbits": 786432000
643
+ }
644
+ ]
645
+ },
646
+ {
647
+ "idx": 34,
648
+ "layers": [
649
+ "model.language_model.layers.17.self_attn.q_proj",
650
+ "model.language_model.layers.17.self_attn.k_proj",
651
+ "model.language_model.layers.17.self_attn.v_proj",
652
+ "model.language_model.layers.17.self_attn.o_proj"
653
+ ],
654
+ "candidates": [
655
+ {
656
+ "dkld": -0.00013766633346679047,
657
+ "dbits": 94371840
658
+ },
659
+ {
660
+ "dkld": -0.0002778722904622777,
661
+ "dbits": 188743680
662
+ }
663
+ ]
664
+ },
665
+ {
666
+ "idx": 35,
667
+ "layers": [
668
+ "model.language_model.layers.17.mlp.gate_proj",
669
+ "model.language_model.layers.17.mlp.up_proj",
670
+ "model.language_model.layers.17.mlp.down_proj"
671
+ ],
672
+ "candidates": [
673
+ {
674
+ "dkld": 3.100791946053505e-05,
675
+ "dbits": 393216000
676
+ },
677
+ {
678
+ "dkld": -1.545818522574738e-05,
679
+ "dbits": 786432000
680
+ }
681
+ ]
682
+ },
683
+ {
684
+ "idx": 36,
685
+ "layers": [
686
+ "model.language_model.layers.18.self_attn.q_proj",
687
+ "model.language_model.layers.18.self_attn.k_proj",
688
+ "model.language_model.layers.18.self_attn.v_proj",
689
+ "model.language_model.layers.18.self_attn.o_proj"
690
+ ],
691
+ "candidates": [
692
+ {
693
+ "dkld": 2.829506993293207e-05,
694
+ "dbits": 94371840
695
+ },
696
+ {
697
+ "dkld": -4.0968367829935515e-05,
698
+ "dbits": 188743680
699
+ }
700
+ ]
701
+ },
702
+ {
703
+ "idx": 37,
704
+ "layers": [
705
+ "model.language_model.layers.18.mlp.gate_proj",
706
+ "model.language_model.layers.18.mlp.up_proj",
707
+ "model.language_model.layers.18.mlp.down_proj"
708
+ ],
709
+ "candidates": [
710
+ {
711
+ "dkld": -8.97957943379879e-06,
712
+ "dbits": 393216000
713
+ },
714
+ {
715
+ "dkld": -0.00019588181748986244,
716
+ "dbits": 786432000
717
+ }
718
+ ]
719
+ },
720
+ {
721
+ "idx": 38,
722
+ "layers": [
723
+ "model.language_model.layers.19.self_attn.q_proj",
724
+ "model.language_model.layers.19.self_attn.k_proj",
725
+ "model.language_model.layers.19.self_attn.v_proj",
726
+ "model.language_model.layers.19.self_attn.o_proj"
727
+ ],
728
+ "candidates": [
729
+ {
730
+ "dkld": 0.0001437588594853656,
731
+ "dbits": 94371840
732
+ },
733
+ {
734
+ "dkld": 0.0001899583265185245,
735
+ "dbits": 188743680
736
+ }
737
+ ]
738
+ },
739
+ {
740
+ "idx": 39,
741
+ "layers": [
742
+ "model.language_model.layers.19.mlp.gate_proj",
743
+ "model.language_model.layers.19.mlp.up_proj",
744
+ "model.language_model.layers.19.mlp.down_proj"
745
+ ],
746
+ "candidates": [
747
+ {
748
+ "dkld": -0.0001980282366275954,
749
+ "dbits": 393216000
750
+ },
751
+ {
752
+ "dkld": -0.00021657398901880898,
753
+ "dbits": 786432000
754
+ }
755
+ ]
756
+ },
757
+ {
758
+ "idx": 40,
759
+ "layers": [
760
+ "model.language_model.layers.20.self_attn.q_proj",
761
+ "model.language_model.layers.20.self_attn.k_proj",
762
+ "model.language_model.layers.20.self_attn.v_proj",
763
+ "model.language_model.layers.20.self_attn.o_proj"
764
+ ],
765
+ "candidates": [
766
+ {
767
+ "dkld": -0.00013915137387812693,
768
+ "dbits": 94371840
769
+ },
770
+ {
771
+ "dkld": -7.083872333168983e-05,
772
+ "dbits": 188743680
773
+ }
774
+ ]
775
+ },
776
+ {
777
+ "idx": 41,
778
+ "layers": [
779
+ "model.language_model.layers.20.mlp.gate_proj",
780
+ "model.language_model.layers.20.mlp.up_proj",
781
+ "model.language_model.layers.20.mlp.down_proj"
782
+ ],
783
+ "candidates": [
784
+ {
785
+ "dkld": -1.2000277638452017e-05,
786
+ "dbits": 393216000
787
+ },
788
+ {
789
+ "dkld": -6.155669689178467e-05,
790
+ "dbits": 786432000
791
+ }
792
+ ]
793
+ },
794
+ {
795
+ "idx": 42,
796
+ "layers": [
797
+ "model.language_model.layers.21.self_attn.q_proj",
798
+ "model.language_model.layers.21.self_attn.k_proj",
799
+ "model.language_model.layers.21.self_attn.v_proj",
800
+ "model.language_model.layers.21.self_attn.o_proj"
801
+ ],
802
+ "candidates": [
803
+ {
804
+ "dkld": 0.00012218807823954903,
805
+ "dbits": 94371840
806
+ },
807
+ {
808
+ "dkld": 0.00018510879017410997,
809
+ "dbits": 188743680
810
+ }
811
+ ]
812
+ },
813
+ {
814
+ "idx": 43,
815
+ "layers": [
816
+ "model.language_model.layers.21.mlp.gate_proj",
817
+ "model.language_model.layers.21.mlp.up_proj",
818
+ "model.language_model.layers.21.mlp.down_proj"
819
+ ],
820
+ "candidates": [
821
+ {
822
+ "dkld": -0.00025624348782005146,
823
+ "dbits": 393216000
824
+ },
825
+ {
826
+ "dkld": -0.0002975346054882033,
827
+ "dbits": 786432000
828
+ }
829
+ ]
830
+ },
831
+ {
832
+ "idx": 44,
833
+ "layers": [
834
+ "model.language_model.layers.22.self_attn.q_proj",
835
+ "model.language_model.layers.22.self_attn.k_proj",
836
+ "model.language_model.layers.22.self_attn.v_proj",
837
+ "model.language_model.layers.22.self_attn.o_proj"
838
+ ],
839
+ "candidates": [
840
+ {
841
+ "dkld": -0.00014645280316472054,
842
+ "dbits": 94371840
843
+ },
844
+ {
845
+ "dkld": -0.0002074604388326573,
846
+ "dbits": 188743680
847
+ }
848
+ ]
849
+ },
850
+ {
851
+ "idx": 45,
852
+ "layers": [
853
+ "model.language_model.layers.22.mlp.gate_proj",
854
+ "model.language_model.layers.22.mlp.up_proj",
855
+ "model.language_model.layers.22.mlp.down_proj"
856
+ ],
857
+ "candidates": [
858
+ {
859
+ "dkld": -2.5052111596035598e-06,
860
+ "dbits": 393216000
861
+ },
862
+ {
863
+ "dkld": 8.436250500379328e-05,
864
+ "dbits": 786432000
865
+ }
866
+ ]
867
+ },
868
+ {
869
+ "idx": 46,
870
+ "layers": [
871
+ "model.language_model.layers.23.self_attn.q_proj",
872
+ "model.language_model.layers.23.self_attn.k_proj",
873
+ "model.language_model.layers.23.self_attn.v_proj",
874
+ "model.language_model.layers.23.self_attn.o_proj"
875
+ ],
876
+ "candidates": [
877
+ {
878
+ "dkld": -7.000300101936419e-05,
879
+ "dbits": 94371840
880
+ },
881
+ {
882
+ "dkld": -5.9204548597341367e-05,
883
+ "dbits": 188743680
884
+ }
885
+ ]
886
+ },
887
+ {
888
+ "idx": 47,
889
+ "layers": [
890
+ "model.language_model.layers.23.mlp.gate_proj",
891
+ "model.language_model.layers.23.mlp.up_proj",
892
+ "model.language_model.layers.23.mlp.down_proj"
893
+ ],
894
+ "candidates": [
895
+ {
896
+ "dkld": -0.0003282364923506975,
897
+ "dbits": 393216000
898
+ },
899
+ {
900
+ "dkld": -0.0002834261395037174,
901
+ "dbits": 786432000
902
+ }
903
+ ]
904
+ },
905
+ {
906
+ "idx": 48,
907
+ "layers": [
908
+ "model.language_model.layers.24.self_attn.q_proj",
909
+ "model.language_model.layers.24.self_attn.k_proj",
910
+ "model.language_model.layers.24.self_attn.v_proj",
911
+ "model.language_model.layers.24.self_attn.o_proj"
912
+ ],
913
+ "candidates": [
914
+ {
915
+ "dkld": -0.00010582348331811264,
916
+ "dbits": 94371840
917
+ },
918
+ {
919
+ "dkld": 4.957020282744806e-05,
920
+ "dbits": 188743680
921
+ }
922
+ ]
923
+ },
924
+ {
925
+ "idx": 49,
926
+ "layers": [
927
+ "model.language_model.layers.24.mlp.gate_proj",
928
+ "model.language_model.layers.24.mlp.up_proj",
929
+ "model.language_model.layers.24.mlp.down_proj"
930
+ ],
931
+ "candidates": [
932
+ {
933
+ "dkld": -4.923101514578421e-05,
934
+ "dbits": 393216000
935
+ },
936
+ {
937
+ "dkld": -7.321541197599646e-05,
938
+ "dbits": 786432000
939
+ }
940
+ ]
941
+ },
942
+ {
943
+ "idx": 50,
944
+ "layers": [
945
+ "model.language_model.layers.25.self_attn.q_proj",
946
+ "model.language_model.layers.25.self_attn.k_proj",
947
+ "model.language_model.layers.25.self_attn.v_proj",
948
+ "model.language_model.layers.25.self_attn.o_proj"
949
+ ],
950
+ "candidates": [
951
+ {
952
+ "dkld": 1.7680786550039462e-05,
953
+ "dbits": 94371840
954
+ },
955
+ {
956
+ "dkld": -0.00017909421585501906,
957
+ "dbits": 188743680
958
+ }
959
+ ]
960
+ },
961
+ {
962
+ "idx": 51,
963
+ "layers": [
964
+ "model.language_model.layers.25.mlp.gate_proj",
965
+ "model.language_model.layers.25.mlp.up_proj",
966
+ "model.language_model.layers.25.mlp.down_proj"
967
+ ],
968
+ "candidates": [
969
+ {
970
+ "dkld": 0.0002028619870543258,
971
+ "dbits": 393216000
972
+ },
973
+ {
974
+ "dkld": 0.0002154889516532421,
975
+ "dbits": 786432000
976
+ }
977
+ ]
978
+ },
979
+ {
980
+ "idx": 52,
981
+ "layers": [
982
+ "model.language_model.layers.26.self_attn.q_proj",
983
+ "model.language_model.layers.26.self_attn.k_proj",
984
+ "model.language_model.layers.26.self_attn.v_proj",
985
+ "model.language_model.layers.26.self_attn.o_proj"
986
+ ],
987
+ "candidates": [
988
+ {
989
+ "dkld": -2.6849750429408514e-05,
990
+ "dbits": 94371840
991
+ },
992
+ {
993
+ "dkld": -0.00013140486553311903,
994
+ "dbits": 188743680
995
+ }
996
+ ]
997
+ },
998
+ {
999
+ "idx": 53,
1000
+ "layers": [
1001
+ "model.language_model.layers.26.mlp.gate_proj",
1002
+ "model.language_model.layers.26.mlp.up_proj",
1003
+ "model.language_model.layers.26.mlp.down_proj"
1004
+ ],
1005
+ "candidates": [
1006
+ {
1007
+ "dkld": -9.225751273334026e-05,
1008
+ "dbits": 393216000
1009
+ },
1010
+ {
1011
+ "dkld": -0.00036976165138186134,
1012
+ "dbits": 786432000
1013
+ }
1014
+ ]
1015
+ },
1016
+ {
1017
+ "idx": 54,
1018
+ "layers": [
1019
+ "model.language_model.layers.27.self_attn.q_proj",
1020
+ "model.language_model.layers.27.self_attn.k_proj",
1021
+ "model.language_model.layers.27.self_attn.v_proj",
1022
+ "model.language_model.layers.27.self_attn.o_proj"
1023
+ ],
1024
+ "candidates": [
1025
+ {
1026
+ "dkld": 0.00020883963443338316,
1027
+ "dbits": 94371840
1028
+ },
1029
+ {
1030
+ "dkld": 8.990662172436714e-05,
1031
+ "dbits": 188743680
1032
+ }
1033
+ ]
1034
+ },
1035
+ {
1036
+ "idx": 55,
1037
+ "layers": [
1038
+ "model.language_model.layers.27.mlp.gate_proj",
1039
+ "model.language_model.layers.27.mlp.up_proj",
1040
+ "model.language_model.layers.27.mlp.down_proj"
1041
+ ],
1042
+ "candidates": [
1043
+ {
1044
+ "dkld": 2.523777075111311e-05,
1045
+ "dbits": 393216000
1046
+ },
1047
+ {
1048
+ "dkld": -0.00022528180852532387,
1049
+ "dbits": 786432000
1050
+ }
1051
+ ]
1052
+ },
1053
+ {
1054
+ "idx": 56,
1055
+ "layers": [
1056
+ "model.language_model.layers.28.self_attn.q_proj",
1057
+ "model.language_model.layers.28.self_attn.k_proj",
1058
+ "model.language_model.layers.28.self_attn.v_proj",
1059
+ "model.language_model.layers.28.self_attn.o_proj"
1060
+ ],
1061
+ "candidates": [
1062
+ {
1063
+ "dkld": -4.28589060902651e-05,
1064
+ "dbits": 94371840
1065
+ },
1066
+ {
1067
+ "dkld": 0.00010976642370222334,
1068
+ "dbits": 188743680
1069
+ }
1070
+ ]
1071
+ },
1072
+ {
1073
+ "idx": 57,
1074
+ "layers": [
1075
+ "model.language_model.layers.28.mlp.gate_proj",
1076
+ "model.language_model.layers.28.mlp.up_proj",
1077
+ "model.language_model.layers.28.mlp.down_proj"
1078
+ ],
1079
+ "candidates": [
1080
+ {
1081
+ "dkld": -0.00026168590411546067,
1082
+ "dbits": 393216000
1083
+ },
1084
+ {
1085
+ "dkld": -0.0003445466049015744,
1086
+ "dbits": 786432000
1087
+ }
1088
+ ]
1089
+ },
1090
+ {
1091
+ "idx": 58,
1092
+ "layers": [
1093
+ "model.language_model.layers.29.self_attn.q_proj",
1094
+ "model.language_model.layers.29.self_attn.k_proj",
1095
+ "model.language_model.layers.29.self_attn.v_proj",
1096
+ "model.language_model.layers.29.self_attn.o_proj"
1097
+ ],
1098
+ "candidates": [
1099
+ {
1100
+ "dkld": 6.120707839726847e-05,
1101
+ "dbits": 94371840
1102
+ },
1103
+ {
1104
+ "dkld": 0.00010455059818922918,
1105
+ "dbits": 188743680
1106
+ }
1107
+ ]
1108
+ },
1109
+ {
1110
+ "idx": 59,
1111
+ "layers": [
1112
+ "model.language_model.layers.29.mlp.gate_proj",
1113
+ "model.language_model.layers.29.mlp.up_proj",
1114
+ "model.language_model.layers.29.mlp.down_proj"
1115
+ ],
1116
+ "candidates": [
1117
+ {
1118
+ "dkld": -0.000123786833137296,
1119
+ "dbits": 393216000
1120
+ },
1121
+ {
1122
+ "dkld": -0.0003314546775072813,
1123
+ "dbits": 786432000
1124
+ }
1125
+ ]
1126
+ },
1127
+ {
1128
+ "idx": 60,
1129
+ "layers": [
1130
+ "model.language_model.layers.30.self_attn.q_proj",
1131
+ "model.language_model.layers.30.self_attn.k_proj",
1132
+ "model.language_model.layers.30.self_attn.v_proj",
1133
+ "model.language_model.layers.30.self_attn.o_proj"
1134
+ ],
1135
+ "candidates": [
1136
+ {
1137
+ "dkld": 0.00029394258745013557,
1138
+ "dbits": 94371840
1139
+ },
1140
+ {
1141
+ "dkld": 0.00022767381742594406,
1142
+ "dbits": 188743680
1143
+ }
1144
+ ]
1145
+ },
1146
+ {
1147
+ "idx": 61,
1148
+ "layers": [
1149
+ "model.language_model.layers.30.mlp.gate_proj",
1150
+ "model.language_model.layers.30.mlp.up_proj",
1151
+ "model.language_model.layers.30.mlp.down_proj"
1152
+ ],
1153
+ "candidates": [
1154
+ {
1155
+ "dkld": -0.00012490730732681032,
1156
+ "dbits": 393216000
1157
+ },
1158
+ {
1159
+ "dkld": -0.0001571818720549345,
1160
+ "dbits": 786432000
1161
+ }
1162
+ ]
1163
+ },
1164
+ {
1165
+ "idx": 62,
1166
+ "layers": [
1167
+ "model.language_model.layers.31.self_attn.q_proj",
1168
+ "model.language_model.layers.31.self_attn.k_proj",
1169
+ "model.language_model.layers.31.self_attn.v_proj",
1170
+ "model.language_model.layers.31.self_attn.o_proj"
1171
+ ],
1172
+ "candidates": [
1173
+ {
1174
+ "dkld": -0.0004900929052382708,
1175
+ "dbits": 94371840
1176
+ },
1177
+ {
1178
+ "dkld": -0.0005018283613026364,
1179
+ "dbits": 188743680
1180
+ }
1181
+ ]
1182
+ },
1183
+ {
1184
+ "idx": 63,
1185
+ "layers": [
1186
+ "model.language_model.layers.31.mlp.gate_proj",
1187
+ "model.language_model.layers.31.mlp.up_proj",
1188
+ "model.language_model.layers.31.mlp.down_proj"
1189
+ ],
1190
+ "candidates": [
1191
+ {
1192
+ "dkld": 0.00022143623791634481,
1193
+ "dbits": 393216000
1194
+ },
1195
+ {
1196
+ "dkld": 0.00010823202319443226,
1197
+ "dbits": 786432000
1198
+ }
1199
+ ]
1200
+ },
1201
+ {
1202
+ "idx": 64,
1203
+ "layers": [
1204
+ "model.language_model.layers.32.self_attn.q_proj",
1205
+ "model.language_model.layers.32.self_attn.k_proj",
1206
+ "model.language_model.layers.32.self_attn.v_proj",
1207
+ "model.language_model.layers.32.self_attn.o_proj"
1208
+ ],
1209
+ "candidates": [
1210
+ {
1211
+ "dkld": -0.0003077854868024643,
1212
+ "dbits": 94371840
1213
+ },
1214
+ {
1215
+ "dkld": -0.00022793915122748531,
1216
+ "dbits": 188743680
1217
+ }
1218
+ ]
1219
+ },
1220
+ {
1221
+ "idx": 65,
1222
+ "layers": [
1223
+ "model.language_model.layers.32.mlp.gate_proj",
1224
+ "model.language_model.layers.32.mlp.up_proj",
1225
+ "model.language_model.layers.32.mlp.down_proj"
1226
+ ],
1227
+ "candidates": [
1228
+ {
1229
+ "dkld": -1.8477952107798234e-05,
1230
+ "dbits": 393216000
1231
+ },
1232
+ {
1233
+ "dkld": -0.00013991552405059893,
1234
+ "dbits": 786432000
1235
+ }
1236
+ ]
1237
+ },
1238
+ {
1239
+ "idx": 66,
1240
+ "layers": [
1241
+ "model.language_model.layers.33.self_attn.q_proj",
1242
+ "model.language_model.layers.33.self_attn.k_proj",
1243
+ "model.language_model.layers.33.self_attn.v_proj",
1244
+ "model.language_model.layers.33.self_attn.o_proj"
1245
+ ],
1246
+ "candidates": [
1247
+ {
1248
+ "dkld": 0.0001139423344284074,
1249
+ "dbits": 94371840
1250
+ },
1251
+ {
1252
+ "dkld": 0.00010652639903127592,
1253
+ "dbits": 188743680
1254
+ }
1255
+ ]
1256
+ },
1257
+ {
1258
+ "idx": 67,
1259
+ "layers": [
1260
+ "model.language_model.layers.33.mlp.gate_proj",
1261
+ "model.language_model.layers.33.mlp.up_proj",
1262
+ "model.language_model.layers.33.mlp.down_proj"
1263
+ ],
1264
+ "candidates": [
1265
+ {
1266
+ "dkld": -0.00015837834216655966,
1267
+ "dbits": 393216000
1268
+ },
1269
+ {
1270
+ "dkld": -0.00017344346269965172,
1271
+ "dbits": 786432000
1272
+ }
1273
+ ]
1274
+ },
1275
+ {
1276
+ "idx": 68,
1277
+ "layers": [
1278
+ "model.language_model.layers.34.self_attn.q_proj",
1279
+ "model.language_model.layers.34.self_attn.k_proj",
1280
+ "model.language_model.layers.34.self_attn.v_proj",
1281
+ "model.language_model.layers.34.self_attn.o_proj"
1282
+ ],
1283
+ "candidates": [
1284
+ {
1285
+ "dkld": -2.0092679187677653e-05,
1286
+ "dbits": 94371840
1287
+ },
1288
+ {
1289
+ "dkld": -4.768059588969309e-05,
1290
+ "dbits": 188743680
1291
+ }
1292
+ ]
1293
+ },
1294
+ {
1295
+ "idx": 69,
1296
+ "layers": [
1297
+ "model.language_model.layers.34.mlp.gate_proj",
1298
+ "model.language_model.layers.34.mlp.up_proj",
1299
+ "model.language_model.layers.34.mlp.down_proj"
1300
+ ],
1301
+ "candidates": [
1302
+ {
1303
+ "dkld": -0.00011453106999398388,
1304
+ "dbits": 393216000
1305
+ },
1306
+ {
1307
+ "dkld": -0.00014291619881989792,
1308
+ "dbits": 786432000
1309
+ }
1310
+ ]
1311
+ },
1312
+ {
1313
+ "idx": 70,
1314
+ "layers": [
1315
+ "model.language_model.layers.35.self_attn.q_proj",
1316
+ "model.language_model.layers.35.self_attn.k_proj",
1317
+ "model.language_model.layers.35.self_attn.v_proj",
1318
+ "model.language_model.layers.35.self_attn.o_proj"
1319
+ ],
1320
+ "candidates": [
1321
+ {
1322
+ "dkld": 0.00015730569139121853,
1323
+ "dbits": 94371840
1324
+ },
1325
+ {
1326
+ "dkld": 0.0002371510956436229,
1327
+ "dbits": 188743680
1328
+ }
1329
+ ]
1330
+ },
1331
+ {
1332
+ "idx": 71,
1333
+ "layers": [
1334
+ "model.language_model.layers.35.mlp.gate_proj",
1335
+ "model.language_model.layers.35.mlp.up_proj",
1336
+ "model.language_model.layers.35.mlp.down_proj"
1337
+ ],
1338
+ "candidates": [
1339
+ {
1340
+ "dkld": -0.00023716585710645277,
1341
+ "dbits": 393216000
1342
+ },
1343
+ {
1344
+ "dkld": -0.0004655693657696358,
1345
+ "dbits": 786432000
1346
+ }
1347
+ ]
1348
+ },
1349
+ {
1350
+ "idx": 72,
1351
+ "layers": [
1352
+ "model.language_model.layers.36.self_attn.q_proj",
1353
+ "model.language_model.layers.36.self_attn.k_proj",
1354
+ "model.language_model.layers.36.self_attn.v_proj",
1355
+ "model.language_model.layers.36.self_attn.o_proj"
1356
+ ],
1357
+ "candidates": [
1358
+ {
1359
+ "dkld": -0.00032587922178209383,
1360
+ "dbits": 94371840
1361
+ },
1362
+ {
1363
+ "dkld": -0.0005042652599513697,
1364
+ "dbits": 188743680
1365
+ }
1366
+ ]
1367
+ },
1368
+ {
1369
+ "idx": 73,
1370
+ "layers": [
1371
+ "model.language_model.layers.36.mlp.gate_proj",
1372
+ "model.language_model.layers.36.mlp.up_proj",
1373
+ "model.language_model.layers.36.mlp.down_proj"
1374
+ ],
1375
+ "candidates": [
1376
+ {
1377
+ "dkld": -0.0006116618402302321,
1378
+ "dbits": 393216000
1379
+ },
1380
+ {
1381
+ "dkld": -0.0006361586041748746,
1382
+ "dbits": 786432000
1383
+ }
1384
+ ]
1385
+ },
1386
+ {
1387
+ "idx": 74,
1388
+ "layers": [
1389
+ "model.language_model.layers.37.self_attn.q_proj",
1390
+ "model.language_model.layers.37.self_attn.k_proj",
1391
+ "model.language_model.layers.37.self_attn.v_proj",
1392
+ "model.language_model.layers.37.self_attn.o_proj"
1393
+ ],
1394
+ "candidates": [
1395
+ {
1396
+ "dkld": 5.171378143131733e-05,
1397
+ "dbits": 94371840
1398
+ },
1399
+ {
1400
+ "dkld": -1.6777683049445935e-05,
1401
+ "dbits": 188743680
1402
+ }
1403
+ ]
1404
+ },
1405
+ {
1406
+ "idx": 75,
1407
+ "layers": [
1408
+ "model.language_model.layers.37.mlp.gate_proj",
1409
+ "model.language_model.layers.37.mlp.up_proj",
1410
+ "model.language_model.layers.37.mlp.down_proj"
1411
+ ],
1412
+ "candidates": [
1413
+ {
1414
+ "dkld": -0.0006116549018770623,
1415
+ "dbits": 393216000
1416
+ },
1417
+ {
1418
+ "dkld": -0.0006575850769877656,
1419
+ "dbits": 786432000
1420
+ }
1421
+ ]
1422
+ },
1423
+ {
1424
+ "idx": 76,
1425
+ "layers": [
1426
+ "model.language_model.layers.38.self_attn.q_proj",
1427
+ "model.language_model.layers.38.self_attn.k_proj",
1428
+ "model.language_model.layers.38.self_attn.v_proj",
1429
+ "model.language_model.layers.38.self_attn.o_proj"
1430
+ ],
1431
+ "candidates": [
1432
+ {
1433
+ "dkld": -0.00017759450711310465,
1434
+ "dbits": 94371840
1435
+ },
1436
+ {
1437
+ "dkld": -0.0003292498644441588,
1438
+ "dbits": 188743680
1439
+ }
1440
+ ]
1441
+ },
1442
+ {
1443
+ "idx": 77,
1444
+ "layers": [
1445
+ "model.language_model.layers.38.mlp.gate_proj",
1446
+ "model.language_model.layers.38.mlp.up_proj",
1447
+ "model.language_model.layers.38.mlp.down_proj"
1448
+ ],
1449
+ "candidates": [
1450
+ {
1451
+ "dkld": 9.779524989425581e-05,
1452
+ "dbits": 393216000
1453
+ },
1454
+ {
1455
+ "dkld": 0.00022586658596990272,
1456
+ "dbits": 786432000
1457
+ }
1458
+ ]
1459
+ },
1460
+ {
1461
+ "idx": 78,
1462
+ "layers": [
1463
+ "model.language_model.layers.39.self_attn.q_proj",
1464
+ "model.language_model.layers.39.self_attn.k_proj",
1465
+ "model.language_model.layers.39.self_attn.v_proj",
1466
+ "model.language_model.layers.39.self_attn.o_proj"
1467
+ ],
1468
+ "candidates": [
1469
+ {
1470
+ "dkld": -0.00019017886370420456,
1471
+ "dbits": 94371840
1472
+ },
1473
+ {
1474
+ "dkld": -0.00014174073003234144,
1475
+ "dbits": 188743680
1476
+ }
1477
+ ]
1478
+ },
1479
+ {
1480
+ "idx": 79,
1481
+ "layers": [
1482
+ "model.language_model.layers.39.mlp.gate_proj",
1483
+ "model.language_model.layers.39.mlp.up_proj",
1484
+ "model.language_model.layers.39.mlp.down_proj"
1485
+ ],
1486
+ "candidates": [
1487
+ {
1488
+ "dkld": 0.00011877082288264118,
1489
+ "dbits": 393216000
1490
+ },
1491
+ {
1492
+ "dkld": -0.00013138423673808575,
1493
+ "dbits": 786432000
1494
+ }
1495
+ ]
1496
+ },
1497
+ {
1498
+ "idx": 80,
1499
+ "layers": [
1500
+ "model.language_model.layers.40.self_attn.q_proj",
1501
+ "model.language_model.layers.40.self_attn.k_proj",
1502
+ "model.language_model.layers.40.self_attn.v_proj",
1503
+ "model.language_model.layers.40.self_attn.o_proj"
1504
+ ],
1505
+ "candidates": [
1506
+ {
1507
+ "dkld": 0.00014259032905100666,
1508
+ "dbits": 94371840
1509
+ },
1510
+ {
1511
+ "dkld": 0.0003622303716838138,
1512
+ "dbits": 188743680
1513
+ }
1514
+ ]
1515
+ },
1516
+ {
1517
+ "idx": 81,
1518
+ "layers": [
1519
+ "model.language_model.layers.40.mlp.gate_proj",
1520
+ "model.language_model.layers.40.mlp.up_proj",
1521
+ "model.language_model.layers.40.mlp.down_proj"
1522
+ ],
1523
+ "candidates": [
1524
+ {
1525
+ "dkld": -0.0004363835323601961,
1526
+ "dbits": 393216000
1527
+ },
1528
+ {
1529
+ "dkld": -0.0004892456810921542,
1530
+ "dbits": 786432000
1531
+ }
1532
+ ]
1533
+ },
1534
+ {
1535
+ "idx": 82,
1536
+ "layers": [
1537
+ "model.language_model.layers.41.self_attn.q_proj",
1538
+ "model.language_model.layers.41.self_attn.k_proj",
1539
+ "model.language_model.layers.41.self_attn.v_proj",
1540
+ "model.language_model.layers.41.self_attn.o_proj"
1541
+ ],
1542
+ "candidates": [
1543
+ {
1544
+ "dkld": -0.00037515051662923016,
1545
+ "dbits": 94371840
1546
+ },
1547
+ {
1548
+ "dkld": -0.00036390214227141615,
1549
+ "dbits": 188743680
1550
+ }
1551
+ ]
1552
+ },
1553
+ {
1554
+ "idx": 83,
1555
+ "layers": [
1556
+ "model.language_model.layers.41.mlp.gate_proj",
1557
+ "model.language_model.layers.41.mlp.up_proj",
1558
+ "model.language_model.layers.41.mlp.down_proj"
1559
+ ],
1560
+ "candidates": [
1561
+ {
1562
+ "dkld": -2.6784790679829085e-05,
1563
+ "dbits": 393216000
1564
+ },
1565
+ {
1566
+ "dkld": -0.0002820216584950852,
1567
+ "dbits": 786432000
1568
+ }
1569
+ ]
1570
+ },
1571
+ {
1572
+ "idx": 84,
1573
+ "layers": [
1574
+ "model.language_model.layers.42.self_attn.q_proj",
1575
+ "model.language_model.layers.42.self_attn.k_proj",
1576
+ "model.language_model.layers.42.self_attn.v_proj",
1577
+ "model.language_model.layers.42.self_attn.o_proj"
1578
+ ],
1579
+ "candidates": [
1580
+ {
1581
+ "dkld": -0.0002433180809021107,
1582
+ "dbits": 94371840
1583
+ },
1584
+ {
1585
+ "dkld": -0.00022505610249937136,
1586
+ "dbits": 188743680
1587
+ }
1588
+ ]
1589
+ },
1590
+ {
1591
+ "idx": 85,
1592
+ "layers": [
1593
+ "model.language_model.layers.42.mlp.gate_proj",
1594
+ "model.language_model.layers.42.mlp.up_proj",
1595
+ "model.language_model.layers.42.mlp.down_proj"
1596
+ ],
1597
+ "candidates": [
1598
+ {
1599
+ "dkld": -0.00030364627018572565,
1600
+ "dbits": 393216000
1601
+ },
1602
+ {
1603
+ "dkld": -0.0002478116191923785,
1604
+ "dbits": 786432000
1605
+ }
1606
+ ]
1607
+ },
1608
+ {
1609
+ "idx": 86,
1610
+ "layers": [
1611
+ "model.language_model.layers.43.self_attn.q_proj",
1612
+ "model.language_model.layers.43.self_attn.k_proj",
1613
+ "model.language_model.layers.43.self_attn.v_proj",
1614
+ "model.language_model.layers.43.self_attn.o_proj"
1615
+ ],
1616
+ "candidates": [
1617
+ {
1618
+ "dkld": 0.00013466081582008682,
1619
+ "dbits": 94371840
1620
+ },
1621
+ {
1622
+ "dkld": 7.684668526047878e-06,
1623
+ "dbits": 188743680
1624
+ }
1625
+ ]
1626
+ },
1627
+ {
1628
+ "idx": 87,
1629
+ "layers": [
1630
+ "model.language_model.layers.43.mlp.gate_proj",
1631
+ "model.language_model.layers.43.mlp.up_proj",
1632
+ "model.language_model.layers.43.mlp.down_proj"
1633
+ ],
1634
+ "candidates": [
1635
+ {
1636
+ "dkld": 2.2285990416981427e-05,
1637
+ "dbits": 393216000
1638
+ },
1639
+ {
1640
+ "dkld": 9.187171235658376e-06,
1641
+ "dbits": 786432000
1642
+ }
1643
+ ]
1644
+ },
1645
+ {
1646
+ "idx": 88,
1647
+ "layers": [
1648
+ "model.language_model.layers.44.self_attn.q_proj",
1649
+ "model.language_model.layers.44.self_attn.k_proj",
1650
+ "model.language_model.layers.44.self_attn.v_proj",
1651
+ "model.language_model.layers.44.self_attn.o_proj"
1652
+ ],
1653
+ "candidates": [
1654
+ {
1655
+ "dkld": 0.0001239406876265836,
1656
+ "dbits": 94371840
1657
+ },
1658
+ {
1659
+ "dkld": 0.00019352706149219912,
1660
+ "dbits": 188743680
1661
+ }
1662
+ ]
1663
+ },
1664
+ {
1665
+ "idx": 89,
1666
+ "layers": [
1667
+ "model.language_model.layers.44.mlp.gate_proj",
1668
+ "model.language_model.layers.44.mlp.up_proj",
1669
+ "model.language_model.layers.44.mlp.down_proj"
1670
+ ],
1671
+ "candidates": [
1672
+ {
1673
+ "dkld": -0.00016605677083136872,
1674
+ "dbits": 393216000
1675
+ },
1676
+ {
1677
+ "dkld": -0.00035673477686942734,
1678
+ "dbits": 786432000
1679
+ }
1680
+ ]
1681
+ },
1682
+ {
1683
+ "idx": 90,
1684
+ "layers": [
1685
+ "model.language_model.layers.45.self_attn.q_proj",
1686
+ "model.language_model.layers.45.self_attn.k_proj",
1687
+ "model.language_model.layers.45.self_attn.v_proj",
1688
+ "model.language_model.layers.45.self_attn.o_proj"
1689
+ ],
1690
+ "candidates": [
1691
+ {
1692
+ "dkld": 0.00018179877661167465,
1693
+ "dbits": 94371840
1694
+ },
1695
+ {
1696
+ "dkld": -9.876396507041418e-06,
1697
+ "dbits": 188743680
1698
+ }
1699
+ ]
1700
+ },
1701
+ {
1702
+ "idx": 91,
1703
+ "layers": [
1704
+ "model.language_model.layers.45.mlp.gate_proj",
1705
+ "model.language_model.layers.45.mlp.up_proj",
1706
+ "model.language_model.layers.45.mlp.down_proj"
1707
+ ],
1708
+ "candidates": [
1709
+ {
1710
+ "dkld": -0.0004525623749941643,
1711
+ "dbits": 393216000
1712
+ },
1713
+ {
1714
+ "dkld": -0.00030217254534364857,
1715
+ "dbits": 786432000
1716
+ }
1717
+ ]
1718
+ },
1719
+ {
1720
+ "idx": 92,
1721
+ "layers": [
1722
+ "model.language_model.layers.46.self_attn.q_proj",
1723
+ "model.language_model.layers.46.self_attn.k_proj",
1724
+ "model.language_model.layers.46.self_attn.v_proj",
1725
+ "model.language_model.layers.46.self_attn.o_proj"
1726
+ ],
1727
+ "candidates": [
1728
+ {
1729
+ "dkld": -0.00010313107632101137,
1730
+ "dbits": 94371840
1731
+ },
1732
+ {
1733
+ "dkld": -1.639225520194687e-05,
1734
+ "dbits": 188743680
1735
+ }
1736
+ ]
1737
+ },
1738
+ {
1739
+ "idx": 93,
1740
+ "layers": [
1741
+ "model.language_model.layers.46.mlp.gate_proj",
1742
+ "model.language_model.layers.46.mlp.up_proj",
1743
+ "model.language_model.layers.46.mlp.down_proj"
1744
+ ],
1745
+ "candidates": [
1746
+ {
1747
+ "dkld": -0.00023827492259442806,
1748
+ "dbits": 393216000
1749
+ },
1750
+ {
1751
+ "dkld": -0.0001679945271462202,
1752
+ "dbits": 786432000
1753
+ }
1754
+ ]
1755
+ },
1756
+ {
1757
+ "idx": 94,
1758
+ "layers": [
1759
+ "model.language_model.layers.47.self_attn.q_proj",
1760
+ "model.language_model.layers.47.self_attn.k_proj",
1761
+ "model.language_model.layers.47.self_attn.v_proj",
1762
+ "model.language_model.layers.47.self_attn.o_proj"
1763
+ ],
1764
+ "candidates": [
1765
+ {
1766
+ "dkld": 0.00047768964432179373,
1767
+ "dbits": 94371840
1768
+ },
1769
+ {
1770
+ "dkld": 0.0004286430310457834,
1771
+ "dbits": 188743680
1772
+ }
1773
+ ]
1774
+ },
1775
+ {
1776
+ "idx": 95,
1777
+ "layers": [
1778
+ "model.language_model.layers.47.mlp.gate_proj",
1779
+ "model.language_model.layers.47.mlp.up_proj",
1780
+ "model.language_model.layers.47.mlp.down_proj"
1781
+ ],
1782
+ "candidates": [
1783
+ {
1784
+ "dkld": -0.0003748951479792817,
1785
+ "dbits": 393216000
1786
+ },
1787
+ {
1788
+ "dkld": -0.00041852202266456917,
1789
+ "dbits": 786432000
1790
+ }
1791
+ ]
1792
+ },
1793
+ {
1794
+ "idx": 96,
1795
+ "layers": [
1796
+ "model.language_model.layers.48.self_attn.q_proj",
1797
+ "model.language_model.layers.48.self_attn.k_proj",
1798
+ "model.language_model.layers.48.self_attn.v_proj",
1799
+ "model.language_model.layers.48.self_attn.o_proj"
1800
+ ],
1801
+ "candidates": [
1802
+ {
1803
+ "dkld": -0.00015185866504907608,
1804
+ "dbits": 94371840
1805
+ },
1806
+ {
1807
+ "dkld": 2.9604835435736998e-05,
1808
+ "dbits": 188743680
1809
+ }
1810
+ ]
1811
+ },
1812
+ {
1813
+ "idx": 97,
1814
+ "layers": [
1815
+ "model.language_model.layers.48.mlp.gate_proj",
1816
+ "model.language_model.layers.48.mlp.up_proj",
1817
+ "model.language_model.layers.48.mlp.down_proj"
1818
+ ],
1819
+ "candidates": [
1820
+ {
1821
+ "dkld": -0.000497997272759676,
1822
+ "dbits": 393216000
1823
+ },
1824
+ {
1825
+ "dkld": -0.0005106110125780161,
1826
+ "dbits": 786432000
1827
+ }
1828
+ ]
1829
+ },
1830
+ {
1831
+ "idx": 98,
1832
+ "layers": [
1833
+ "model.language_model.layers.49.self_attn.q_proj",
1834
+ "model.language_model.layers.49.self_attn.k_proj",
1835
+ "model.language_model.layers.49.self_attn.v_proj",
1836
+ "model.language_model.layers.49.self_attn.o_proj"
1837
+ ],
1838
+ "candidates": [
1839
+ {
1840
+ "dkld": -0.00013194279745221138,
1841
+ "dbits": 94371840
1842
+ },
1843
+ {
1844
+ "dkld": -0.00026086703874173,
1845
+ "dbits": 188743680
1846
+ }
1847
+ ]
1848
+ },
1849
+ {
1850
+ "idx": 99,
1851
+ "layers": [
1852
+ "model.language_model.layers.49.mlp.gate_proj",
1853
+ "model.language_model.layers.49.mlp.up_proj",
1854
+ "model.language_model.layers.49.mlp.down_proj"
1855
+ ],
1856
+ "candidates": [
1857
+ {
1858
+ "dkld": 0.00028887651860712293,
1859
+ "dbits": 393216000
1860
+ },
1861
+ {
1862
+ "dkld": 8.507091552018009e-05,
1863
+ "dbits": 786432000
1864
+ }
1865
+ ]
1866
+ },
1867
+ {
1868
+ "idx": 100,
1869
+ "layers": [
1870
+ "model.language_model.layers.50.self_attn.q_proj",
1871
+ "model.language_model.layers.50.self_attn.k_proj",
1872
+ "model.language_model.layers.50.self_attn.v_proj",
1873
+ "model.language_model.layers.50.self_attn.o_proj"
1874
+ ],
1875
+ "candidates": [
1876
+ {
1877
+ "dkld": 0.00041282796300945446,
1878
+ "dbits": 94371840
1879
+ },
1880
+ {
1881
+ "dkld": 0.00034878966398535494,
1882
+ "dbits": 188743680
1883
+ }
1884
+ ]
1885
+ },
1886
+ {
1887
+ "idx": 101,
1888
+ "layers": [
1889
+ "model.language_model.layers.50.mlp.gate_proj",
1890
+ "model.language_model.layers.50.mlp.up_proj",
1891
+ "model.language_model.layers.50.mlp.down_proj"
1892
+ ],
1893
+ "candidates": [
1894
+ {
1895
+ "dkld": -0.0003386356867849938,
1896
+ "dbits": 393216000
1897
+ },
1898
+ {
1899
+ "dkld": -0.0005072795320302248,
1900
+ "dbits": 786432000
1901
+ }
1902
+ ]
1903
+ },
1904
+ {
1905
+ "idx": 102,
1906
+ "layers": [
1907
+ "model.language_model.layers.51.self_attn.q_proj",
1908
+ "model.language_model.layers.51.self_attn.k_proj",
1909
+ "model.language_model.layers.51.self_attn.v_proj",
1910
+ "model.language_model.layers.51.self_attn.o_proj"
1911
+ ],
1912
+ "candidates": [
1913
+ {
1914
+ "dkld": 8.664890192447428e-05,
1915
+ "dbits": 94371840
1916
+ },
1917
+ {
1918
+ "dkld": 0.000201887823641278,
1919
+ "dbits": 188743680
1920
+ }
1921
+ ]
1922
+ },
1923
+ {
1924
+ "idx": 103,
1925
+ "layers": [
1926
+ "model.language_model.layers.51.mlp.gate_proj",
1927
+ "model.language_model.layers.51.mlp.up_proj",
1928
+ "model.language_model.layers.51.mlp.down_proj"
1929
+ ],
1930
+ "candidates": [
1931
+ {
1932
+ "dkld": -0.0009231174364686123,
1933
+ "dbits": 393216000
1934
+ },
1935
+ {
1936
+ "dkld": -0.0008698628749698456,
1937
+ "dbits": 786432000
1938
+ }
1939
+ ]
1940
+ },
1941
+ {
1942
+ "idx": 104,
1943
+ "layers": [
1944
+ "model.language_model.layers.52.self_attn.q_proj",
1945
+ "model.language_model.layers.52.self_attn.k_proj",
1946
+ "model.language_model.layers.52.self_attn.v_proj",
1947
+ "model.language_model.layers.52.self_attn.o_proj"
1948
+ ],
1949
+ "candidates": [
1950
+ {
1951
+ "dkld": 0.0006682481616735458,
1952
+ "dbits": 94371840
1953
+ },
1954
+ {
1955
+ "dkld": 0.00046845399774608776,
1956
+ "dbits": 188743680
1957
+ }
1958
+ ]
1959
+ },
1960
+ {
1961
+ "idx": 105,
1962
+ "layers": [
1963
+ "model.language_model.layers.52.mlp.gate_proj",
1964
+ "model.language_model.layers.52.mlp.up_proj",
1965
+ "model.language_model.layers.52.mlp.down_proj"
1966
+ ],
1967
+ "candidates": [
1968
+ {
1969
+ "dkld": -0.0005143266171216965,
1970
+ "dbits": 393216000
1971
+ },
1972
+ {
1973
+ "dkld": -0.00028531067073345184,
1974
+ "dbits": 786432000
1975
+ }
1976
+ ]
1977
+ },
1978
+ {
1979
+ "idx": 106,
1980
+ "layers": [
1981
+ "model.language_model.layers.53.self_attn.q_proj",
1982
+ "model.language_model.layers.53.self_attn.k_proj",
1983
+ "model.language_model.layers.53.self_attn.v_proj",
1984
+ "model.language_model.layers.53.self_attn.o_proj"
1985
+ ],
1986
+ "candidates": [
1987
+ {
1988
+ "dkld": -0.0011058779433369692,
1989
+ "dbits": 94371840
1990
+ },
1991
+ {
1992
+ "dkld": -0.0008262845687568299,
1993
+ "dbits": 188743680
1994
+ }
1995
+ ]
1996
+ },
1997
+ {
1998
+ "idx": 107,
1999
+ "layers": [
2000
+ "model.language_model.layers.53.mlp.gate_proj",
2001
+ "model.language_model.layers.53.mlp.up_proj",
2002
+ "model.language_model.layers.53.mlp.down_proj"
2003
+ ],
2004
+ "candidates": [
2005
+ {
2006
+ "dkld": -0.0005857495125383139,
2007
+ "dbits": 393216000
2008
+ },
2009
+ {
2010
+ "dkld": -0.0007907640654593873,
2011
+ "dbits": 786432000
2012
+ }
2013
+ ]
2014
+ },
2015
+ {
2016
+ "idx": 108,
2017
+ "layers": [
2018
+ "model.language_model.layers.54.self_attn.q_proj",
2019
+ "model.language_model.layers.54.self_attn.k_proj",
2020
+ "model.language_model.layers.54.self_attn.v_proj",
2021
+ "model.language_model.layers.54.self_attn.o_proj"
2022
+ ],
2023
+ "candidates": [
2024
+ {
2025
+ "dkld": -0.0003247374668717551,
2026
+ "dbits": 94371840
2027
+ },
2028
+ {
2029
+ "dkld": -0.00047106938436627943,
2030
+ "dbits": 188743680
2031
+ }
2032
+ ]
2033
+ },
2034
+ {
2035
+ "idx": 109,
2036
+ "layers": [
2037
+ "model.language_model.layers.54.mlp.gate_proj",
2038
+ "model.language_model.layers.54.mlp.up_proj",
2039
+ "model.language_model.layers.54.mlp.down_proj"
2040
+ ],
2041
+ "candidates": [
2042
+ {
2043
+ "dkld": -0.0006814375985413901,
2044
+ "dbits": 393216000
2045
+ },
2046
+ {
2047
+ "dkld": -0.0006852156948298216,
2048
+ "dbits": 786432000
2049
+ }
2050
+ ]
2051
+ },
2052
+ {
2053
+ "idx": 110,
2054
+ "layers": [
2055
+ "model.language_model.layers.55.self_attn.q_proj",
2056
+ "model.language_model.layers.55.self_attn.k_proj",
2057
+ "model.language_model.layers.55.self_attn.v_proj",
2058
+ "model.language_model.layers.55.self_attn.o_proj"
2059
+ ],
2060
+ "candidates": [
2061
+ {
2062
+ "dkld": -0.0008712090086192004,
2063
+ "dbits": 94371840
2064
+ },
2065
+ {
2066
+ "dkld": -0.000794619973748939,
2067
+ "dbits": 188743680
2068
+ }
2069
+ ]
2070
+ },
2071
+ {
2072
+ "idx": 111,
2073
+ "layers": [
2074
+ "model.language_model.layers.55.mlp.gate_proj",
2075
+ "model.language_model.layers.55.mlp.up_proj",
2076
+ "model.language_model.layers.55.mlp.down_proj"
2077
+ ],
2078
+ "candidates": [
2079
+ {
2080
+ "dkld": -0.0004104899242520499,
2081
+ "dbits": 393216000
2082
+ },
2083
+ {
2084
+ "dkld": -0.0003673459868878126,
2085
+ "dbits": 786432000
2086
+ }
2087
+ ]
2088
+ },
2089
+ {
2090
+ "idx": 112,
2091
+ "layers": [
2092
+ "model.language_model.layers.56.self_attn.q_proj",
2093
+ "model.language_model.layers.56.self_attn.k_proj",
2094
+ "model.language_model.layers.56.self_attn.v_proj",
2095
+ "model.language_model.layers.56.self_attn.o_proj"
2096
+ ],
2097
+ "candidates": [
2098
+ {
2099
+ "dkld": -8.137007243932959e-05,
2100
+ "dbits": 94371840
2101
+ },
2102
+ {
2103
+ "dkld": -0.0005230564158409945,
2104
+ "dbits": 188743680
2105
+ }
2106
+ ]
2107
+ },
2108
+ {
2109
+ "idx": 113,
2110
+ "layers": [
2111
+ "model.language_model.layers.56.mlp.gate_proj",
2112
+ "model.language_model.layers.56.mlp.up_proj",
2113
+ "model.language_model.layers.56.mlp.down_proj"
2114
+ ],
2115
+ "candidates": [
2116
+ {
2117
+ "dkld": -0.0007926002144813538,
2118
+ "dbits": 393216000
2119
+ },
2120
+ {
2121
+ "dkld": -0.0004942396190017462,
2122
+ "dbits": 786432000
2123
+ }
2124
+ ]
2125
+ },
2126
+ {
2127
+ "idx": 114,
2128
+ "layers": [
2129
+ "model.language_model.layers.57.self_attn.q_proj",
2130
+ "model.language_model.layers.57.self_attn.k_proj",
2131
+ "model.language_model.layers.57.self_attn.v_proj",
2132
+ "model.language_model.layers.57.self_attn.o_proj"
2133
+ ],
2134
+ "candidates": [
2135
+ {
2136
+ "dkld": -4.884987138212282e-05,
2137
+ "dbits": 94371840
2138
+ },
2139
+ {
2140
+ "dkld": -0.00027763238176703453,
2141
+ "dbits": 188743680
2142
+ }
2143
+ ]
2144
+ },
2145
+ {
2146
+ "idx": 115,
2147
+ "layers": [
2148
+ "model.language_model.layers.57.mlp.gate_proj",
2149
+ "model.language_model.layers.57.mlp.up_proj",
2150
+ "model.language_model.layers.57.mlp.down_proj"
2151
+ ],
2152
+ "candidates": [
2153
+ {
2154
+ "dkld": -0.0007438508328050375,
2155
+ "dbits": 393216000
2156
+ },
2157
+ {
2158
+ "dkld": -0.0007115922868251856,
2159
+ "dbits": 786432000
2160
+ }
2161
+ ]
2162
+ },
2163
+ {
2164
+ "idx": 116,
2165
+ "layers": [
2166
+ "model.language_model.layers.58.self_attn.q_proj",
2167
+ "model.language_model.layers.58.self_attn.k_proj",
2168
+ "model.language_model.layers.58.self_attn.v_proj",
2169
+ "model.language_model.layers.58.self_attn.o_proj"
2170
+ ],
2171
+ "candidates": [
2172
+ {
2173
+ "dkld": -0.00016308482736349106,
2174
+ "dbits": 94371840
2175
+ },
2176
+ {
2177
+ "dkld": -0.0003048637881875038,
2178
+ "dbits": 188743680
2179
+ }
2180
+ ]
2181
+ },
2182
+ {
2183
+ "idx": 117,
2184
+ "layers": [
2185
+ "model.language_model.layers.58.mlp.gate_proj",
2186
+ "model.language_model.layers.58.mlp.up_proj",
2187
+ "model.language_model.layers.58.mlp.down_proj"
2188
+ ],
2189
+ "candidates": [
2190
+ {
2191
+ "dkld": -0.0008469442371279168,
2192
+ "dbits": 393216000
2193
+ },
2194
+ {
2195
+ "dkld": -0.000874680187553184,
2196
+ "dbits": 786432000
2197
+ }
2198
+ ]
2199
+ },
2200
+ {
2201
+ "idx": 118,
2202
+ "layers": [
2203
+ "model.language_model.layers.59.self_attn.q_proj",
2204
+ "model.language_model.layers.59.self_attn.k_proj",
2205
+ "model.language_model.layers.59.self_attn.v_proj",
2206
+ "model.language_model.layers.59.self_attn.o_proj"
2207
+ ],
2208
+ "candidates": [
2209
+ {
2210
+ "dkld": 2.2697122767562083e-05,
2211
+ "dbits": 94371840
2212
+ },
2213
+ {
2214
+ "dkld": -0.0004360306076705567,
2215
+ "dbits": 188743680
2216
+ }
2217
+ ]
2218
+ },
2219
+ {
2220
+ "idx": 119,
2221
+ "layers": [
2222
+ "model.language_model.layers.59.mlp.gate_proj",
2223
+ "model.language_model.layers.59.mlp.up_proj",
2224
+ "model.language_model.layers.59.mlp.down_proj"
2225
+ ],
2226
+ "candidates": [
2227
+ {
2228
+ "dkld": -0.0005720941349864062,
2229
+ "dbits": 393216000
2230
+ },
2231
+ {
2232
+ "dkld": -0.0006580702494830004,
2233
+ "dbits": 786432000
2234
+ }
2235
+ ]
2236
+ },
2237
+ {
2238
+ "idx": 120,
2239
+ "layers": [
2240
+ "model.language_model.layers.60.self_attn.q_proj",
2241
+ "model.language_model.layers.60.self_attn.k_proj",
2242
+ "model.language_model.layers.60.self_attn.v_proj",
2243
+ "model.language_model.layers.60.self_attn.o_proj"
2244
+ ],
2245
+ "candidates": [
2246
+ {
2247
+ "dkld": -0.001185041572898643,
2248
+ "dbits": 94371840
2249
+ },
2250
+ {
2251
+ "dkld": -0.0012297863606363646,
2252
+ "dbits": 188743680
2253
+ }
2254
+ ]
2255
+ },
2256
+ {
2257
+ "idx": 121,
2258
+ "layers": [
2259
+ "model.language_model.layers.60.mlp.gate_proj",
2260
+ "model.language_model.layers.60.mlp.up_proj",
2261
+ "model.language_model.layers.60.mlp.down_proj"
2262
+ ],
2263
+ "candidates": [
2264
+ {
2265
+ "dkld": -0.00043576946482062895,
2266
+ "dbits": 393216000
2267
+ },
2268
+ {
2269
+ "dkld": -0.00037752632051707424,
2270
+ "dbits": 786432000
2271
+ }
2272
+ ]
2273
+ },
2274
+ {
2275
+ "idx": 122,
2276
+ "layers": [
2277
+ "model.language_model.layers.61.self_attn.q_proj",
2278
+ "model.language_model.layers.61.self_attn.k_proj",
2279
+ "model.language_model.layers.61.self_attn.v_proj",
2280
+ "model.language_model.layers.61.self_attn.o_proj"
2281
+ ],
2282
+ "candidates": [
2283
+ {
2284
+ "dkld": 0.00023652003146706901,
2285
+ "dbits": 94371840
2286
+ },
2287
+ {
2288
+ "dkld": -0.00035312902182341177,
2289
+ "dbits": 188743680
2290
+ }
2291
+ ]
2292
+ },
2293
+ {
2294
+ "idx": 123,
2295
+ "layers": [
2296
+ "model.language_model.layers.61.mlp.gate_proj",
2297
+ "model.language_model.layers.61.mlp.up_proj",
2298
+ "model.language_model.layers.61.mlp.down_proj"
2299
+ ],
2300
+ "candidates": [
2301
+ {
2302
+ "dkld": -0.00031162113882601816,
2303
+ "dbits": 393216000
2304
+ },
2305
+ {
2306
+ "dkld": -0.0005344925448298454,
2307
+ "dbits": 786432000
2308
+ }
2309
+ ]
2310
+ },
2311
+ {
2312
+ "idx": 124,
2313
+ "layers": [
2314
+ "model.language_model.layers.62.self_attn.q_proj",
2315
+ "model.language_model.layers.62.self_attn.k_proj",
2316
+ "model.language_model.layers.62.self_attn.v_proj",
2317
+ "model.language_model.layers.62.self_attn.o_proj"
2318
+ ],
2319
+ "candidates": [
2320
+ {
2321
+ "dkld": -0.0011726192198693863,
2322
+ "dbits": 94371840
2323
+ },
2324
+ {
2325
+ "dkld": -0.0011081209406256787,
2326
+ "dbits": 188743680
2327
+ }
2328
+ ]
2329
+ },
2330
+ {
2331
+ "idx": 125,
2332
+ "layers": [
2333
+ "model.language_model.layers.62.mlp.gate_proj",
2334
+ "model.language_model.layers.62.mlp.up_proj",
2335
+ "model.language_model.layers.62.mlp.down_proj"
2336
+ ],
2337
+ "candidates": [
2338
+ {
2339
+ "dkld": -0.0033292342443019263,
2340
+ "dbits": 393216000
2341
+ },
2342
+ {
2343
+ "dkld": -0.003356328653171675,
2344
+ "dbits": 786432000
2345
+ }
2346
+ ]
2347
+ },
2348
+ {
2349
+ "idx": 126,
2350
+ "layers": [
2351
+ "model.language_model.layers.63.self_attn.q_proj",
2352
+ "model.language_model.layers.63.self_attn.k_proj",
2353
+ "model.language_model.layers.63.self_attn.v_proj",
2354
+ "model.language_model.layers.63.self_attn.o_proj"
2355
+ ],
2356
+ "candidates": [
2357
+ {
2358
+ "dkld": 0.00011754971928892854,
2359
+ "dbits": 94371840
2360
+ },
2361
+ {
2362
+ "dkld": 0.00028725313022731624,
2363
+ "dbits": 188743680
2364
+ }
2365
+ ]
2366
+ },
2367
+ {
2368
+ "idx": 127,
2369
+ "layers": [
2370
+ "model.language_model.layers.63.mlp.gate_proj",
2371
+ "model.language_model.layers.63.mlp.up_proj",
2372
+ "model.language_model.layers.63.mlp.down_proj"
2373
+ ],
2374
+ "candidates": [
2375
+ {
2376
+ "dkld": -0.0020378781482577324,
2377
+ "dbits": 393216000
2378
+ },
2379
+ {
2380
+ "dkld": -0.002213719766587019,
2381
+ "dbits": 786432000
2382
+ }
2383
+ ]
2384
+ }
2385
+ ],
2386
+ "base_kld": 0.19816926405765117,
2387
+ "arch_string": "Qwen3VLForConditionalGeneration"
2388
+ }