Requested changes

app.py

@@ -1,99 +1,227 @@
+import os
 from pathlib import Path
 from tempfile import TemporaryDirectory
 
 import gradio as gr
-from huggingface_hub import CommitOperationAdd, HfApi
-from huggingface_hub.utils import RepositoryNotFoundError
+from huggingface_hub import HfApi, ModelCard
+from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError
 
 from convert import convert
 
+# Repo with files totalling more than 24GB are not converted. Avoid to have a memory issue.
+try:
+    MAX_REPO_SIZE = int(os.environ.get("MAX_REPO_SIZE"))
+except:
+    MAX_REPO_SIZE = 24 * 1000 * 1000 * 1000
+
+
+class Generator:
+    # Taken from https://stackoverflow.com/a/34073559
+    # Allows to log process in Gradio
+    def __init__(self, gen):
+        self.gen = gen
+
+    def __iter__(self):
+        self.value = yield from self.gen
+
 
 def run(
     token: str, model_id: str, precision: str, quantization: bool, destination: str
-) -> str:
+):
+    _all_logs = []
+
+    def _log(msg: str):
+        print(msg)  # for container logs
+        _all_logs.append(msg)
+        return "\n\n".join(_all_logs)  # for Gradio output
+
     if token == "" or model_id == "":
-        return """
+        yield _log(
+            """
         ### Invalid input 🐞
         
         Please fill a token and model_id.
         """
+        )
+        return
     if destination == "":
+        _log("Destination not provided. Will default to the initial repo.")
         destination = model_id
 
     api = HfApi(token=token)
     try:
         # TODO: make a PR to bloomz.cpp to be able to pass a token
-        api.model_info(
-            repo_id=model_id, token=False
-        )  # only public repos are accessible
+        model_info = api.model_info(repo_id=model_id, files_metadata=True, token=False)
+        _log(f"Model {model_id} exists.")
     except RepositoryNotFoundError:
-        return f"""
+        yield _log(
+            f"""
         ### Error 😢😢😢
         
         Repository {model_id} not found. Only public models are convertible at the moment.
         """
+        )
+        return
+
+    total_size = sum(
+        file.size
+        for file in model_info.siblings
+        if file.rfilename.endswith(".pt") or file.rfilename.endswith(".bin")
+    )
+    if total_size > MAX_REPO_SIZE:
+        yield _log(
+            f"""
+        ### Unprocessable 😢😢😢
+        
+        Model {model_id} is too big and cannot be processed in this Space. This Space needs to be able to load the model in memory before converting it. To avoid a memory issue, we do not process models bigger than {MAX_REPO_SIZE}b.
+
+        You have 2 options:
+        - [Duplicate this Space](https://huggingface.co/spaces/Wauplin/bloomz.cpp-converter?duplicate=true) and assign a bigger machine. You will need to set 'MAX_REPO_SIZE' as a secret to overwrite the default value. Once you are done, remove the upgraded hardware and/or delete the Space.
+        - Manually convert the weights by following [this guide](https://github.com/NouamaneTazi/bloomz.cpp#usage).
+        """
+        )
+        return
 
     try:
         with TemporaryDirectory() as cache_folder:
-            model_path = convert(
-                cache_folder=Path(cache_folder),
-                model_id=model_id,
-                precision=precision,
-                quantization=quantization,
+            convert_progress = Generator(
+                convert(
+                    cache_folder=Path(cache_folder),
+                    model_id=model_id,
+                    precision=precision,
+                    quantization=quantization,
+                )
+            )
+            for msg in convert_progress:
+                yield _log(msg)
+            model_path = convert_progress.value
+            yield _log(f"Model converted: {model_path}")
+
+            destination_url = api.create_repo(repo_id=destination, exist_ok=True)
+            destination = destination_url.repo_id
+            yield _log(f"Destination model: {destination_url}")
+            pr = api.create_pull_request(
+                repo_id=destination_url.repo_id,
+                title=f"[WIP] Add {model_path.name} from bloomz.cpp converter.",
+                description="This PR has been created using the [bloomz.cpp converter Space](https://huggingface.co/spaces/Wauplin/bloomz.cpp-converter). It adds weights compatible with the [bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp#usage) project.",
+            )
+            pr_url = f"https://huggingface.co/{destination}/discussions/{pr.num}"
+            yield _log(f"Created PR: {pr_url} (empty)")
+
+            yield _log(f"Uploading model to PR")
+            api.upload_file(
+                repo_id=destination,
+                path_or_fileobj=model_path,
+                path_in_repo=model_path.name,
+                revision=pr.git_reference,
+            )
+            yield _log(f"Model uploaded to PR")
+
+            yield _log(f"Modifying model card in PR (add `bloom` and `ggml` tags)")
+            try:
+                card = ModelCard.load(repo_id_or_path=destination)
+            except EntryNotFoundError:  # new repo => no model card yet
+                card = ModelCard(
+                    "This model contains a model based on the Bloom architecture with weights compatible with [bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp). This model card has been automatically generated [by the bloomz.cpp converter Space](https://huggingface.co/spaces/Wauplin/bloomz.cpp-converter) and must be completed."
+                )
+            if card.data.tags is None:
+                card.data.tags = []
+            tags = card.data.tags
+            if "ggml" not in tags:
+                tags.append("ggml")
+            if "bloom" not in tags:
+                tags.append("bloom")
+            card.push_to_hub(
+                repo_id=destination, token=token, revision=pr.git_reference
             )
-            print("[model_path]", model_path)
+            yield _log(f"Model card modified in PR.")
 
-            commit_info = api.create_commit(
+            api.change_discussion_status(
                 repo_id=destination,
-                operations=[
-                    CommitOperationAdd(
-                        path_or_fileobj=model_path, path_in_repo=model_path.name
-                    )
-                ],
-                create_pr=True,
-                commit_message=f"Add {model_path.name} from bloomz.cpp converter.",
+                discussion_num=pr.num,
+                new_status="open",
+                comment="PR is now complete and ready to be reviewed.",
             )
+            yield _log(f"[PR]({pr_url}) is complete and ready to be reviewed.")
 
-            return f"""
+            yield _log(
+                f"""
             ### Success 🔥
-            Yay! This model was successfully converted and a PR was open using your token, here:
-            
-            # [{commit_info.pr_url}]({commit_info.pr_url})
+            Yay! This model was successfully converted! Make sure to let the repo owner know about it and review your PR. You might need to complete the PR manually, especially to add information in the model card.
             """
+            )
+            return
     except Exception as e:
-        return f"""
+        yield _log(
+            f"""
         ### Error 😢😢😢
         
         {e}
         """
+        )
+        return
 
 
+TITLE = """
+<h1 style="font-weight: 900; font-size: 32px; margin-bottom: 10px; margin-top: 10px; text-align: center;">
+    Make any BLOOM-like model compatible with bloomz.cpp
+</h1>
+"""
+
 DESCRIPTION = """
-The steps are the following:
-- Paste your HF token. You can create one in your [settings page](https://huggingface.co/settings/tokens).
-- Input a model id from the Hub. This model must be public.
-- Choose which precision you want to use (default to FP16).
-- (optional) Opt-in for 4-bit quantization.
-- (optional) By default a PR to the initial repo will be created. You can choose a different destination repo if you want. The destination repo must exist.
-- Click Submit:
+This Space allows you to automatically export any Bloom-like model hosted on the 🤗 Hub to be compatible with [bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp). Converted weights are either exported to a repo you own (or that we create for you) or to the original repo by opening a PR on the target model. Once exported, the model can run with bloomz.cpp. Check out [this guide](https://github.com/NouamaneTazi/bloomz.cpp#usage) to see how!
+
+Don't know which Bloom model are available on the 🤗 Hub? Find a complete list at https://huggingface.co/models?other=bloom.
+
+To use this Space, please follow these steps:
+
+1. Paste your HF token. You can create one in your [settings page](https://huggingface.co/settings/tokens). The token requires a write-access token to create a PR and upload the weights.
+1. Input a model id from the Hub. This model must be public.
+1. Choose which precision you want to use (default to FP16).
+1. (optional) Opt-in for 4-bit quantization.
+1. (optional) By default a PR to the initial repo will be created. You can choose a different destination repo if you want. The destination repo will be created if it doesn't exist.
+1. Click "Convert!"
 
 That's it! You'll get feedback if it works or not, and if it worked, you'll get the URL of the opened PR 🔥
+If you encounter any issues please let us know [by opening a Discussion](https://huggingface.co/spaces/Wauplin/bloomz.cpp-converter/discussions/new).
 """
 
-demo = gr.Interface(
-    title="Convert any BLOOM-like model to be compatible with bloomz.cpp",
-    description=DESCRIPTION,
-    allow_flagging="never",
-    article="Check out the [bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp) repo on GitHub",
-    inputs=[
-        gr.Text(max_lines=1, label="your hf_token"),
-        gr.Text(max_lines=1, label="model_id (e.g.: bigscience/bloomz-7b1)"),
-        gr.Radio(choices=["FP16", "FP32"], label="Precision", value="FP16"),
-        gr.Checkbox(value=False, label="4-bits quantization"),
-        gr.Text(max_lines=1, label="destination (e.g.: my-username/bloomz-7b1.cpp)"),
-    ],
-    outputs=[gr.Markdown(label="output")],
-    fn=run,
-).queue()
-
-demo.launch()
+
+with gr.Blocks() as demo:
+    gr.HTML(TITLE)
+
+    with gr.Row():
+        with gr.Column(scale=50):
+            gr.Markdown(DESCRIPTION)
+
+        with gr.Column(scale=50):
+            input_token = gr.Text(max_lines=1, label="Hugging Face token")
+            input_model = gr.Text(
+                max_lines=1, label="Model id (e.g.: bigscience/bloomz-7b1)"
+            )
+            input_precision = gr.Radio(
+                choices=["FP16", "FP32"], label="Precision", value="FP16"
+            )
+            input_quantization = gr.Checkbox(value=False, label="4-bits quantization")
+            input_destination = gr.Text(
+                max_lines=1,
+                label="Destination (e.g.: my-username/bloomz-7b1.cpp) - optional",
+            )
+            btn = gr.Button("Convert!")
+
+    output = gr.Markdown(label="Output")
+
+    btn.click(
+        fn=run,
+        inputs=[
+            input_token,
+            input_model,
+            input_precision,
+            input_quantization,
+            input_destination,
+        ],
+        outputs=output,
+    )
+
+
+demo.queue().launch()

convert.py

@@ -1,12 +1,13 @@
 from pathlib import Path
 from subprocess import run
+from typing import Generator
 
 BLOOMZ_FOLDER = Path(__file__).parent / "bloomz.cpp"
 
 
 def convert(
     cache_folder: Path, model_id: str, precision: str, quantization: bool
-) -> Path:
+) -> Generator[str, Path, None]:
     # Conversion
     cmd = [
         "python",
@@ -16,6 +17,7 @@ def convert(
     ]
     if precision == "FP32":
         cmd.append("--use-fp32")
+    yield f"Running command: `{' '.join(cmd)}`"
     run(cmd, check=True)
 
     # Model file should exist
@@ -23,6 +25,7 @@ def convert(
     _, model_name = model_id.split("/")
     model_path = cache_folder / f"ggml-model-{model_name}-{f_suffix}.bin"
     assert model_path.is_file()
+    yield f"Model successfully converted to ggml: {model_path}"
 
     # Quantization
     if quantization:
@@ -35,9 +38,11 @@ def convert(
             str(q_model_path),
             "2",
         ]
+        yield f"Running command: `{' '.join(cmd)}`"
         run(cmd, check=True)
         assert q_model_path.is_file()
         model_path = q_model_path
+        yield f"Model successfully quantized: {model_path}"
 
     # Return
     return model_path