v0.1.2-pre5 queued up/downloads, better quantization and conversion h…

…andling

Homepage.py

@@ -1,5 +1,5 @@
 # main.py
-# Version 0.1.2-pre5: Aesthetics and minor fixes
+# Version 0.1.2-pre5: Bugfixes, better quantization and conversion handling, and queued up/downloads.
 import streamlit as st
 st.set_page_config(layout="wide")
 from st_pages import Page, Section, show_pages, add_indentation

TODO.md

@@ -1,12 +1,10 @@
 TODO allow gpu offload for certain tasks
-TODO be able to queue later jobs for files that don't yet exist (i.e. convert then quantize) and handle errors in ordering
-TODO add a page that allows you to make one great big queued op for download/convert/quantize/(imatrix)/upload/etc
+TODO be able to queue later jobs for files that don't yet exist (i.e. convert then quantize) and handle errors in ordering/make a page for it
 TODO LLaMA-Factory integration
 TODO Lilac integration?
 TODO support further cli options for relevant commands/pages
 
 # ON DECK
-TODO create scripts for download/upload and use those as commands for download/upload (shocker)
 TODO imatrix creation
 TODO imatrix quantization support
 TODO readme creation from template!

pages/Convert_Safetensors.py

@@ -1,4 +1,4 @@
-# FILESTATUS: pretty much done, needs testing. Last updated v0.1.2-pre3
+# FILESTATUS: needs expanded functionality. Last updated v0.1.2-pre5
 # IMPORTS ---------------------------------------------------------------------------------
 import streamlit as st
 st.set_page_config(layout="wide")

pages/Hugging_Face_Downloader.py

@@ -1,4 +1,4 @@
-# FILESTATUS: completed, needs testing. Last updated v0.1.2-pre3
+# FILESTATUS: completed, needs testing. Last updated v0.1.2-pre5
 # IMPORTS ---------------------------------------------------------------------------------
 import requests, streamlit as st
 st.set_page_config(layout="wide")
@@ -8,29 +8,10 @@
 from util.paths import *
 
 # FUNCTIONS ---------------------------------------------------------------------------------
-# TODO be able to change output directory
-# TODO write new scheduler and parallelize
-# TODO maybe you need to make another scheduler for just download jobs while the other is used for conversion/quantization/finetuning jobs
-
-# write the download task to the queue
-def queue_command(file_url, download_path, filename):
-    command = [
-        "aria2c", file_url,
-        "--max-connection-per-server=16", "--split=8", "--min-split-size=25M", "--allow-overwrite=true",
-        "-d", str(download_path), "-o", filename,
-        "--continue=true"
-    ]
-    get_scheduler().add_job(command)
 
 # queues a download task for each file in the file_links_dict
-def trigger_command(file_links_dict, model_name):
-    folder_name = model_name.split("/")[-1]
-    download_path = models_dir() / folder_name
-    download_path.mkdir(parents=True, exist_ok=True)
-
-    for file_name, file_url in file_links_dict.items():
-        filename = Path(file_name).name
-        queue_command(file_url, download_path, filename)
+def trigger_command(model_name):
+    get_scheduler().add_job(["python3", "util/download.py", model_name])
 
     return "Download tasks have been queued."
 
@@ -77,7 +58,7 @@ def get_files_from_repo(url, repo_name):
 
 if st.button("Download Files"):
     if 'file_links_dict' in st.session_state and st.session_state['file_links_dict']:
-        queue_message = trigger_command(st.session_state['file_links_dict'], model_name)
+        queue_message = trigger_command(model_name)
         st.text(queue_message)
     else:
         st.error("No files to download. Please get the file list first.")

pages/Quantize_GGUF.py

@@ -1,4 +1,4 @@
-# FILESTATUS: migrated implementation but need to finish testing pre1 before this can be used and tested. Last updated v0.1.2-pre4
+# FILESTATUS: needs expanded functionality. Last updated v0.1.2-pre5
 # IMPORTS ---------------------------------------------------------------------------------
 import os, streamlit as st
 st.set_page_config(layout="wide")

pages/Upload_Converted_To_HF.py

@@ -1,66 +1,35 @@
-# FILESTATUS: needs to be migrated to the New Way of Doing Things. Last updated v0.1.2-pre3
+# FILESTATUS: needs potentially expanded functionality. Last updated v0.1.2-pre5
 # IMPORTS ---------------------------------------------------------------------------------
 import os, streamlit as st
 st.set_page_config(layout="wide")
 from st_pages import add_indentation
 from huggingface_hub import HfApi
 from requests.exceptions import HTTPError
 from util.paths import *
+from util.scheduler import *
 from util.key import decrypt_token
 
 # FUNCTIONS ---------------------------------------------------------------------------------
 
-# Search for the llama.cpp directory
-if not llama_cpp_dir():
-    st.error("llama.cpp directory not found. Please check the file structure.")
+# Gathers files and uploads to HuggingFace
+def trigger_command(token, repo_name, files_to_upload, high_precision_files, medium_precision_files, selected_model):
+    final_upload = []
 
-## Uses username from HF Token
-def get_username_from_token(token):
-    api = HfApi()
-    user_info = api.whoami(token=token)
-    return user_info['name']
+    # Rename particular filepaths
+    for file_name in files_to_upload:
+        if file_name in high_precision_files.get(selected_model, []):
+            folder_path = models_dir() / selected_model / "High-Precision-Quantization"
+        elif file_name in medium_precision_files.get(selected_model, []):
+            folder_path = models_dir() / selected_model / "Medium-Precision-Quantization"
+        else:
+            continue
 
+        file_path = folder_path / file_name
+        final_upload.append(str(file_path))
 
-# Gathers files and uploads to HuggingFace
-def upload_files_to_repo(token, repo_name, files_to_upload, readme_content, high_precision_files, medium_precision_files, selected_model):
-    try:
-        api = HfApi()
-        username = get_username_from_token(token)
-        repo_id = f"{username}/{repo_name}"
-
-        # Check if the repository exists, if not create it
-        try:
-            api.repo_info(repo_id=repo_id, token=token)
-        except HTTPError as e:
-            if e.response.status_code == 404:
-                api.create_repo(repo_id=repo_id, token=token, repo_type="model")
-            else:
-                raise
-
-        # Upload README.md if content is provided
-        if readme_content:
-            readme_path = models_dir() / 'README.md'
-            with open(str(readme_path), 'w') as readme_file:
-                readme_file.write(readme_content)
-            api.upload_file(path_or_fileobj=str(readme_path), path_in_repo='README.md', repo_id=repo_id, token=token)
-            os.remove(str(readme_path))
-
-        # Upload selected files
-        for file_name in files_to_upload:
-            if file_name in high_precision_files.get(selected_model, []):
-                folder_path = models_dir() / selected_model / "High-Precision-Quantization"
-            elif file_name in medium_precision_files.get(selected_model, []):
-                folder_path = models_dir() / selected_model / "Medium-Precision-Quantization"
-            else:
-                continue
-
-            file_path = folder_path / file_name
-            if file_path.is_file():
-                api.upload_file(path_or_fileobj=str(file_path), path_in_repo=file_name, repo_id=repo_id, token=token)
-
-        return f"Files uploaded successfully. View at: https://huggingface.co/{repo_id}"
-    except Exception as e:
-        return f"An error occurred: {str(e)}"
+    command = ["python3", "util/upload.py", token, repo_name, *final_upload]
+    get_scheduler().add_job(command)
+    return "Upload tasks have been queued."
 
 # Cache the function to improve performance
 @st.cache_data
@@ -103,8 +72,7 @@ def get_combined_files(model):
 
 
 # Repository details and README content
-repo_name = st.text_input("Repository Name", value=f"{selected_model}-GGUF")
-readme_content = st.text_area("README.md Content", "Enter content for README.md")
+repo_name = st.text_input("Repository Name", value=f"{selected_model}-gguf")
 
 # Token input
 use_unencrypted_token = st.checkbox("Unencrypted Token")
@@ -117,15 +85,7 @@ def get_combined_files(model):
 
 # Upload button
 if st.button("Upload Selected Files") and hf_token:
-    upload_message = upload_files_to_repo(
-        token=hf_token, 
-        repo_name=repo_name, 
-        files_to_upload=selected_files, 
-        readme_content=readme_content,
-        high_precision_files=high_precision_files,
-        medium_precision_files=medium_precision_files,
-        selected_model=selected_model
-    )
+    upload_message = trigger_command(hf_token, repo_name, selected_files, high_precision_files, medium_precision_files, selected_model)
     st.info(upload_message)
 
 if 'HUGGINGFACE_TOKEN' in os.environ:

requirements.txt

@@ -4,4 +4,9 @@ cryptography
 st-pages
 streamlit-sortables
 torch~=2.1.1
-einops~=0.7.0
+einops~=0.7.0
+numpy~=1.24.4
+sentencepiece~=0.1.98
+transformers>=4.35.2,<5.0.0
+gguf>=0.1.0
+protobuf>=4.21.0,<5.0.0

util/download.py

@@ -0,0 +1,11 @@
+# FILESTATUS Complete and tested. Last updated 0.1.2-pre5
+import sys
+from huggingface_hub import snapshot_download
+from pathlib import Path
+
+# Simple download script. full_model_name comes in the form of "model_creator/model_name"
+full_model_name = sys.argv[1]
+model_creator, model_name = full_model_name.split("/")
+download_path = Path("llama.cpp/models") / model_name
+download_path.mkdir(parents=True, exist_ok=True)
+snapshot_download(repo_id=f"{model_creator}/{model_name}", local_dir=str(download_path))

util/scheduler.py

@@ -1,4 +1,4 @@
-# FILESTATUS: fully implemented but needs significant testing for efficiency, thread safety, bugfixing, overall functionality. Last updated v0.1.2-pre4
+# FILESTATUS: fully implemented but needs significant testing for efficiency, thread safety, bugfixing, overall functionality. Last updated v0.1.2-pre5
 # IMPORTS ---------------------------------------------------------------------------------
 from pathlib import Path
 from datetime import datetime
@@ -7,7 +7,6 @@
 
 # FUNCTIONS ---------------------------------------------------------------------------------
 
-# TODO perhaps make download tasks independent of other tasks? but then you run into thread safety issues I think
 # the scheduler singleton class (do i need st.cache_resource here?)
 @st.cache_resource
 class Scheduler:

util/upload.py

@@ -0,0 +1,23 @@
+# FILESTATUS Complete and tested. Last updated 0.1.2-pre5
+import sys
+from huggingface_hub import HfApi
+from requests.exceptions import HTTPError
+
+# Simple upload script. full_model_name comes in the form of "model_creator/model_name"
+token = sys.argv[1]
+api = HfApi()
+username = api.whoami(token=token)['name']
+repo_id = f"{username}/{sys.argv[2]}"
+
+# Check if the repository exists, if not create it
+try:
+    api.repo_info(repo_id=repo_id, token=token)
+except HTTPError as e:
+    if e.response.status_code == 404:
+        api.create_repo(repo_id=repo_id, token=token, repo_type="model")
+    else:
+        raise
+
+# Upload selected files
+for file_name in sys.argv[3:]:
+    api.upload_file(path_or_fileobj=file_name, path_in_repo=file_name, repo_id=repo_id, token=token)