[`refactor`]: Tab & URL syncing; parameter counts as model size; filtering; search (#89)

- Compute model size based on number of parameters (bd6a61b46998cdcc89eeb567850b734a48b87737)
- Refactor gradio Tabs initialization (56def8ad644fed5c5ef653500101899fe01c6876)
- Introduce tabs <-> URL relation for easier sharing (2d428eb2835e975816ff72bcda38cfe96323bcd3)
- Reintroduce missing markdown (e2b41c870d6e112347e1ba44c8620fcaf88bd719)
- Add search bar/filtering; always show Model Size (ab565badcfc1c2b02ddce0c2a0618875b5f8982f)
- Ignore voyage-lite model size (82d58936b4312ab7fb0ef838b7bac1df70a46d45)
- Remove debugging statement (d81785e017b21579643aab840f584d59d0c95179)
- Introduce new intervals: 100M & 250M and 250M & 500M (2eb890dc40c9a4e71475f42b024d703a9f45db7b)
- API -> Proprietary (cfacdeee424f776fb60c10adbe3f4a9bb04d9f2a)
- Add Sentence Transformers model type option (6c6aac59f6d2f5f1744c9618e30e7be17a82d0b9)
- Fix embedding dimensions if Dense module exists (e82960d3779fe6bec1618330b1cbf63c4d383e9d)
- Use separate proprietary models list (561360760efd385ce82ea67ebca7674d6526a515)
- Fix proprietary models disappearing after model size toggling (485f27b4dcf9a19d4c963c858641da9da2da4a33)
- Increment gradio SDK version (fefaea64210d697cb4a0cf4a24a0c156784fdb1d)
- Base "Open" models on the proprietary model list (2db25dc3a419e68be78dca9ef0e997e8e38588ae)
- Update e5-mistral-7b model size from 7110 to 7111 (418d26a052f7f729cdabf09aeba288b02581ba4b)
- Merge branch 'main' into model_size_parameters (7d3a9f6218e33c393a5d886b415a64c740f593ff)
- Add Memory Usage column to all tables (970b6a5470a7aef9a30ed14a27ae957f82c2b131)
- List Cohere-embed-english-v3.0 as proprietary (5bd316f1be45e51845d64c6908cfa7c1b18ba88a)
- Merge branch 'main' into model_size_parameters (0ebd4b87bbe1abf36714d2d16bbb08a528912334)
- Move globals around slightly (a8ba8f1f0e653322bb5447a2abfa437e3e74baba)
- Add parameter count for Google Gecko (4de60b86d8f1da6f6c10a7b9bb631f4165e18724)

.gitignore

@@ -0,0 +1 @@
+*.pyc

README.md

@@ -4,7 +4,7 @@ emoji: 🥇
 colorFrom: blue
 colorTo: indigo
 sdk: gradio
-sdk_version: 4.0.2
+sdk_version: 4.20.0
 app_file: app.py
 pinned: false
 tags:

utils/model_size.py

@@ -0,0 +1,40 @@
+import json
+import re
+from huggingface_hub.hf_api import ModelInfo, get_safetensors_metadata, model_info as get_model_info, get_hf_file_metadata, hf_hub_url
+from huggingface_hub import hf_hub_download
+
+# Map model IDs to the number of bytes used for one parameter. So, 4 bytes for fp32, 2 bytes for fp16, etc.
+# By default, we assume that the model is stored in fp32.
+KNOWN_BYTES_PER_PARAM = {}
+
+
+def get_model_parameters_memory(model_info: ModelInfo):
+    '''Get the size of the model in million of parameters.'''
+    try:
+        safetensors = get_safetensors_metadata(model_info.id)
+        num_parameters = sum(safetensors.parameter_count.values())
+        return round(num_parameters / 1e6), round(num_parameters * 4 / 1024**3, 2)
+    except Exception as e:
+        pass
+
+    filenames = [sib.rfilename for sib in model_info.siblings]
+    if "pytorch_model.bin" in filenames:
+        url = hf_hub_url(model_info.id, filename="pytorch_model.bin")
+        meta = get_hf_file_metadata(url)
+        bytes_per_param = KNOWN_BYTES_PER_PARAM.get(model_info.id, 4)
+        return round(meta.size / bytes_per_param / 1e6), round(meta.size / 1024**3, 2)
+    
+    if "pytorch_model.bin.index.json" in filenames:
+        index_path = hf_hub_download(model_info.id, filename="pytorch_model.bin.index.json")
+        """
+        {
+        "metadata": {
+            "total_size": 28272820224
+        },....
+        """
+        size = json.load(open(index_path))
+        bytes_per_param = KNOWN_BYTES_PER_PARAM.get(model_info.id, 4)
+        if ("metadata" in size) and ("total_size" in size["metadata"]):
+            return round(size["metadata"]["total_size"] / bytes_per_param / 1e6), round(size["metadata"]["total_size"] / 1024**3, 2)
+
+    return None, None