Upload 4 files
Browse files- README.md +42 -8
- app.py +383 -0
- indexer.py +149 -0
- requirements.txt +2 -0
README.md
CHANGED
|
@@ -1,15 +1,49 @@
|
|
| 1 |
---
|
| 2 |
-
title: Transformers
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: gradio
|
| 7 |
sdk_version: 6.30.0
|
| 8 |
-
python_version: '3.13'
|
| 9 |
app_file: app.py
|
| 10 |
pinned: false
|
| 11 |
-
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
---
|
| 14 |
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Transformers Playground
|
| 3 |
+
emoji: 🛝
|
| 4 |
+
colorFrom: yellow
|
| 5 |
+
colorTo: gray
|
| 6 |
sdk: gradio
|
| 7 |
sdk_version: 6.30.0
|
|
|
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
+
hf_oauth: true
|
| 11 |
+
hf_oauth_expiration_minutes: 480
|
| 12 |
+
hf_oauth_scopes:
|
| 13 |
+
- jobs
|
| 14 |
+
- read-repos
|
| 15 |
+
short_description: Copy official transformers snippets or run them on HF Jobs
|
| 16 |
---
|
| 17 |
|
| 18 |
+
# Transformers Playground
|
| 19 |
+
|
| 20 |
+
Pick any transformers model on the Hub, get the usage snippets from its
|
| 21 |
+
`model_doc` page (adapted to your checkpoint), copy them, or run them on
|
| 22 |
+
Hugging Face Jobs with your own account.
|
| 23 |
+
|
| 24 |
+
## How it works
|
| 25 |
+
|
| 26 |
+
- `indexer.py` sparse-clones `docs/source/en/model_doc` from transformers,
|
| 27 |
+
extracts the `<hfoptions id="usage">` tabs and usage code blocks, drops
|
| 28 |
+
training and quantization examples (v1 is inference only), and writes
|
| 29 |
+
`snippets.json`.
|
| 30 |
+
- `app.py` matches a model's `config.json` `model_type` to its doc page,
|
| 31 |
+
swaps the doc's checkpoint for the selected model, adds a generic
|
| 32 |
+
`pipeline()` snippet from the model's task, and recommends the cheapest
|
| 33 |
+
Jobs flavor that fits the weights (live prices from `list_jobs_hardware`).
|
| 34 |
+
- Runs go through `run_job` with the user's OAuth token, so each user pays
|
| 35 |
+
for their own compute. Jobs need a PRO, Team or Enterprise account.
|
| 36 |
+
|
| 37 |
+
## Setup
|
| 38 |
+
|
| 39 |
+
1. Create the Space from these files. On first boot it builds the index
|
| 40 |
+
itself (about 30 seconds).
|
| 41 |
+
2. Optional but recommended: publish the index as a dataset and refresh it
|
| 42 |
+
nightly, then set a Space variable `INDEX_REPO=<you>/transformers-snippets`.
|
| 43 |
+
|
| 44 |
+
```bash
|
| 45 |
+
hf jobs scheduled run @daily python:3.12 --secrets HF_TOKEN -- bash -c \
|
| 46 |
+
"pip install -q huggingface_hub && \
|
| 47 |
+
curl -sL https://huggingface.co/spaces/<you>/transformers-playground/raw/main/indexer.py -o indexer.py && \
|
| 48 |
+
python indexer.py --push <you>/transformers-snippets"
|
| 49 |
+
```
|
app.py
ADDED
|
@@ -0,0 +1,383 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Transformers Playground: find a model, grab its doc snippets, run them on HF Jobs."""
|
| 2 |
+
import json
|
| 3 |
+
import os
|
| 4 |
+
import re
|
| 5 |
+
import time
|
| 6 |
+
|
| 7 |
+
import gradio as gr
|
| 8 |
+
from huggingface_hub import (
|
| 9 |
+
HfApi, cancel_job, fetch_job_logs, hf_hub_download, inspect_job, run_job,
|
| 10 |
+
)
|
| 11 |
+
from huggingface_hub.utils import GatedRepoError, HfHubHTTPError, RepositoryNotFoundError
|
| 12 |
+
|
| 13 |
+
from indexer import build_index
|
| 14 |
+
|
| 15 |
+
INDEX_REPO = os.getenv("INDEX_REPO", "") # e.g. "your-org/transformers-snippets"
|
| 16 |
+
UV_IMAGE = "ghcr.io/astral-sh/uv:python3.12-bookworm"
|
| 17 |
+
api = HfApi()
|
| 18 |
+
|
| 19 |
+
# ---------------------------------------------------------------- index ----
|
| 20 |
+
|
| 21 |
+
def load_index():
|
| 22 |
+
if INDEX_REPO:
|
| 23 |
+
try:
|
| 24 |
+
path = hf_hub_download(INDEX_REPO, "snippets.json", repo_type="dataset")
|
| 25 |
+
with open(path) as f:
|
| 26 |
+
return json.load(f)
|
| 27 |
+
except Exception as e: # fall through to a local build
|
| 28 |
+
print(f"Could not load {INDEX_REPO}: {e}")
|
| 29 |
+
if os.path.exists("snippets.json"):
|
| 30 |
+
with open("snippets.json") as f:
|
| 31 |
+
return json.load(f)
|
| 32 |
+
print("Building snippet index from transformers@main ...")
|
| 33 |
+
index = build_index("main")
|
| 34 |
+
with open("snippets.json", "w") as f:
|
| 35 |
+
json.dump(index, f)
|
| 36 |
+
return index
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
INDEX = load_index()
|
| 40 |
+
_norm = lambda s: re.sub(r"[^a-z0-9]", "", (s or "").lower())
|
| 41 |
+
DOC_BY_NORM = {_norm(k): k for k in INDEX["docs"]}
|
| 42 |
+
|
| 43 |
+
# ------------------------------------------------------------- hardware ----
|
| 44 |
+
|
| 45 |
+
def _gb(s):
|
| 46 |
+
m = re.search(r"([\d.]+)", s or "")
|
| 47 |
+
return float(m.group(1)) if m else 0.0
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def load_hardware():
|
| 51 |
+
try:
|
| 52 |
+
hw = api.list_jobs_hardware()
|
| 53 |
+
return [{
|
| 54 |
+
"name": h.name,
|
| 55 |
+
"ram": _gb(h.ram),
|
| 56 |
+
"vram": _gb(h.accelerator.vram) if h.accelerator else 0.0,
|
| 57 |
+
"gpus": int(h.accelerator.quantity) if h.accelerator else 0,
|
| 58 |
+
"gpu": h.accelerator.model if h.accelerator else None,
|
| 59 |
+
"usd_hr": h.unit_cost_usd * 60,
|
| 60 |
+
} for h in hw]
|
| 61 |
+
except Exception as e:
|
| 62 |
+
print(f"Hardware list unavailable ({e}); using a minimal fallback.")
|
| 63 |
+
return [
|
| 64 |
+
{"name": "cpu-basic", "ram": 16, "vram": 0, "gpus": 0, "gpu": None, "usd_hr": 0.01},
|
| 65 |
+
{"name": "t4-small", "ram": 15, "vram": 16, "gpus": 1, "gpu": "T4", "usd_hr": 0.40},
|
| 66 |
+
{"name": "l4x1", "ram": 30, "vram": 24, "gpus": 1, "gpu": "L4", "usd_hr": 0.80},
|
| 67 |
+
{"name": "a100-large", "ram": 142, "vram": 80, "gpus": 1, "gpu": "A100", "usd_hr": 2.50},
|
| 68 |
+
]
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
HARDWARE = sorted(load_hardware(), key=lambda h: h["usd_hr"])
|
| 72 |
+
HW_BY_NAME = {h["name"]: h for h in HARDWARE}
|
| 73 |
+
BYTES = {"F64": 8, "F32": 4, "BF16": 2, "F16": 2, "F8_E4M3": 1, "F8_E5M2": 1, "I8": 1, "U8": 1, "I32": 4}
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def estimate_gb(meta):
|
| 77 |
+
if not meta.get("params"):
|
| 78 |
+
return None
|
| 79 |
+
return meta["params"] * meta["bytes"] / 1e9 * 1.2 + 1.0 # weights + KV/activation headroom
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
def recommend(meta, needs_gpu, multi_gpu_ok):
|
| 83 |
+
need = estimate_gb(meta)
|
| 84 |
+
if need is None:
|
| 85 |
+
return "t4-small", "Model size unknown (no safetensors metadata), starting with a T4."
|
| 86 |
+
if not needs_gpu and meta["params"] <= 1.5e9:
|
| 87 |
+
for h in HARDWARE:
|
| 88 |
+
if h["gpus"] == 0 and h["ram"] >= need * 1.5:
|
| 89 |
+
return h["name"], f"~{need:.1f} GB needed; small enough for CPU."
|
| 90 |
+
for h in HARDWARE:
|
| 91 |
+
if h["gpus"] and h["vram"] >= need and h["ram"] >= need and (h["gpus"] == 1 or multi_gpu_ok):
|
| 92 |
+
return h["name"], f"~{need:.1f} GB needed; cheapest GPU that fits."
|
| 93 |
+
return HARDWARE[-1]["name"], f"~{need:.1f} GB needed; larger than any single flavor, expect OOM."
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
def hw_choices(meta):
|
| 97 |
+
need = estimate_gb(meta) or 0
|
| 98 |
+
out = []
|
| 99 |
+
for h in HARDWARE:
|
| 100 |
+
spec = f"{h['gpus']}x {h['gpu']}, {h['vram']:.0f} GB" if h["gpus"] else f"CPU, {h['ram']:.0f} GB RAM"
|
| 101 |
+
small = " (too small)" if (h["gpus"] and h["vram"] < need) else ""
|
| 102 |
+
out.append((f"{h['name']} [{spec}] ${h['usd_hr']:.2f}/h{small}", h["name"]))
|
| 103 |
+
return out
|
| 104 |
+
|
| 105 |
+
# ------------------------------------------------------------- snippets ----
|
| 106 |
+
|
| 107 |
+
SAMPLE_IMG = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 108 |
+
SAMPLE_AUDIO = "https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac"
|
| 109 |
+
GENERIC_INPUTS = {
|
| 110 |
+
"text-generation": '"Plants create energy through a process known as", max_new_tokens=50',
|
| 111 |
+
"text-classification": '"I love how easy this library makes things!"',
|
| 112 |
+
"token-classification": '"My name is Sarah and I live in London."',
|
| 113 |
+
"fill-mask": 'f"Paris is the {pipe.tokenizer.mask_token} of France."',
|
| 114 |
+
"question-answering": 'question="Where do I live?", context="My name is Sarah and I live in London."',
|
| 115 |
+
"zero-shot-classification": '"I just bought a new laptop", candidate_labels=["tech", "cooking", "sports"]',
|
| 116 |
+
"summarization": '"The tower is 324 metres tall, about the same height as an 81-storey building, and the tallest structure in Paris."',
|
| 117 |
+
"translation": '"How old are you?"',
|
| 118 |
+
"feature-extraction": '"Transformers is great."',
|
| 119 |
+
"sentence-similarity": '"Transformers is great."',
|
| 120 |
+
"image-classification": f'"{SAMPLE_IMG}"',
|
| 121 |
+
"object-detection": f'"{SAMPLE_IMG}"',
|
| 122 |
+
"image-segmentation": f'"{SAMPLE_IMG}"',
|
| 123 |
+
"depth-estimation": f'"{SAMPLE_IMG}"',
|
| 124 |
+
"image-to-text": f'"{SAMPLE_IMG}"',
|
| 125 |
+
"zero-shot-image-classification": f'"{SAMPLE_IMG}", candidate_labels=["cats", "dogs", "birds"]',
|
| 126 |
+
"image-feature-extraction": f'"{SAMPLE_IMG}"',
|
| 127 |
+
"automatic-speech-recognition": f'"{SAMPLE_AUDIO}"',
|
| 128 |
+
"audio-classification": f'"{SAMPLE_AUDIO}"',
|
| 129 |
+
"image-text-to-text": (
|
| 130 |
+
f'text=[{{"role": "user", "content": [{{"type": "image", "url": "{SAMPLE_IMG}"}}, '
|
| 131 |
+
f'{{"type": "text", "text": "Describe this image."}}]}}], max_new_tokens=50'
|
| 132 |
+
),
|
| 133 |
+
}
|
| 134 |
+
TASK_ALIAS = {"sentence-similarity": "feature-extraction"}
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
def generic_snippet(task):
|
| 138 |
+
if task not in GENERIC_INPUTS:
|
| 139 |
+
return None
|
| 140 |
+
code = (
|
| 141 |
+
"from transformers import pipeline\n\n"
|
| 142 |
+
f'pipe = pipeline(task="{TASK_ALIAS.get(task, task)}", model="MODEL_ID", device_map="auto")\n'
|
| 143 |
+
f"print(pipe({GENERIC_INPUTS[task]}))\n"
|
| 144 |
+
)
|
| 145 |
+
return {"label": "Generic pipeline (auto-generated)", "section": "", "language": "python",
|
| 146 |
+
"code": code, "checkpoint": "MODEL_ID", "needs_gpu": False}
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
def adapt(snippet, model_id):
|
| 150 |
+
code = snippet["code"]
|
| 151 |
+
if snippet.get("checkpoint"):
|
| 152 |
+
code = code.replace(snippet["checkpoint"], model_id)
|
| 153 |
+
return code
|
| 154 |
+
|
| 155 |
+
|
| 156 |
+
def snippet_task(code):
|
| 157 |
+
m = re.search(r'pipeline\(\s*(?:task\s*=\s*)?["\']([\w-]+)["\']', code)
|
| 158 |
+
return m.group(1) if m else None
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
def infer_deps(code, task):
|
| 162 |
+
deps = ["transformers", "torch", "accelerate"]
|
| 163 |
+
checks = {
|
| 164 |
+
"pillow": r"\bPIL\b|Image\.open|image",
|
| 165 |
+
"torchvision": r"torchvision|image|Image|vision|video",
|
| 166 |
+
"requests": r"\brequests\b",
|
| 167 |
+
"datasets": r"load_dataset|from datasets|import datasets",
|
| 168 |
+
"soundfile": r"audio|soundfile|\.flac|\.wav",
|
| 169 |
+
"librosa": r"librosa|audio|\.flac|\.wav",
|
| 170 |
+
"sentencepiece": r"sentencepiece|T5|Marian",
|
| 171 |
+
"num2words": r"SmolVLM|Idefics3",
|
| 172 |
+
}
|
| 173 |
+
blob = code + " " + (task or "")
|
| 174 |
+
deps += [d for d, pat in checks.items() if re.search(pat, blob)]
|
| 175 |
+
return ", ".join(deps)
|
| 176 |
+
|
| 177 |
+
# ------------------------------------------------------------ callbacks ----
|
| 178 |
+
|
| 179 |
+
def search_models(query, task):
|
| 180 |
+
models = api.list_models(search=query or None, filter="transformers",
|
| 181 |
+
pipeline_tag=task or None, sort="downloads", limit=30)
|
| 182 |
+
ids = [m.id for m in models]
|
| 183 |
+
return gr.update(choices=ids, value=ids[0] if ids else None)
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
def _token(oauth_token):
|
| 187 |
+
return oauth_token.token if oauth_token else None
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
def select_model(model_id, oauth_token: gr.OAuthToken | None):
|
| 191 |
+
empty = ("", gr.update(choices=[], value=None), [], {}, "", "", gr.update(), "", "")
|
| 192 |
+
if not model_id:
|
| 193 |
+
return empty
|
| 194 |
+
try:
|
| 195 |
+
info = api.model_info(model_id, token=_token(oauth_token))
|
| 196 |
+
except RepositoryNotFoundError:
|
| 197 |
+
return ("Model not found, or it is private and you are not signed in.",) + empty[1:]
|
| 198 |
+
|
| 199 |
+
cfg = info.config or {}
|
| 200 |
+
model_type = cfg.get("model_type")
|
| 201 |
+
arch = DOC_BY_NORM.get(_norm(model_type))
|
| 202 |
+
params, dtype_bytes = None, 2
|
| 203 |
+
if info.safetensors:
|
| 204 |
+
params = info.safetensors.total
|
| 205 |
+
main_dtype = max(info.safetensors.parameters, key=info.safetensors.parameters.get)
|
| 206 |
+
dtype_bytes = BYTES.get(main_dtype, 2)
|
| 207 |
+
meta = {"model_id": model_id, "params": params, "bytes": dtype_bytes,
|
| 208 |
+
"task": info.pipeline_tag, "gated": bool(info.gated)}
|
| 209 |
+
|
| 210 |
+
snippets = list(INDEX["docs"][arch]["snippets"]) if arch else []
|
| 211 |
+
generic = generic_snippet(info.pipeline_tag)
|
| 212 |
+
if generic:
|
| 213 |
+
snippets.append(generic)
|
| 214 |
+
snippets = [dict(s, adapted=adapt(s, model_id)) for s in snippets]
|
| 215 |
+
|
| 216 |
+
size = f"{params / 1e9:.2f}B params" if params else "size unknown"
|
| 217 |
+
lines = [f"### [{model_id}](https://huggingface.co/{model_id})",
|
| 218 |
+
f"`{model_type or 'unknown arch'}` | {info.pipeline_tag or 'no task tag'} | {size}"]
|
| 219 |
+
if arch:
|
| 220 |
+
d = INDEX["docs"][arch]
|
| 221 |
+
lines.append(f"Snippets from [{d['title']} docs]({d['doc_url']}) "
|
| 222 |
+
f"at transformers [`{INDEX['sha'][:8]}`]({d['source_url']})")
|
| 223 |
+
else:
|
| 224 |
+
lines.append("No model_doc page matched this architecture.")
|
| 225 |
+
if info.gated:
|
| 226 |
+
lines.append("This model is gated. Accept its license on the model page before running.")
|
| 227 |
+
if not snippets:
|
| 228 |
+
lines.append("No runnable snippets found for this model.")
|
| 229 |
+
|
| 230 |
+
choices = [(f"{s['label']} ({s['language']})", i) for i, s in enumerate(snippets)]
|
| 231 |
+
first = 0 if snippets else None
|
| 232 |
+
return ("\n\n".join(lines), gr.update(choices=choices, value=first), snippets, meta,
|
| 233 |
+
*snippet_outputs(first, snippets, meta))
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
def snippet_outputs(idx, snippets, meta):
|
| 237 |
+
if idx is None or not snippets:
|
| 238 |
+
return "", "", gr.update(choices=hw_choices(meta), value="cpu-basic"), "", ""
|
| 239 |
+
s = snippets[idx]
|
| 240 |
+
code = s["adapted"]
|
| 241 |
+
flavor, why = recommend(meta, s["needs_gpu"], "device_map" in code)
|
| 242 |
+
notes = [f"Recommended hardware: **{flavor}**. {why}"]
|
| 243 |
+
st = snippet_task(code)
|
| 244 |
+
if st and meta.get("task") and TASK_ALIAS.get(meta["task"], meta["task"]) != st:
|
| 245 |
+
notes.append(f"Heads up: this snippet uses the `{st}` task but the model is tagged "
|
| 246 |
+
f"`{meta['task']}`. It may need edits.")
|
| 247 |
+
if s["language"] != "python":
|
| 248 |
+
notes.append("Shell snippets are copy-only; run Python snippets on Jobs.")
|
| 249 |
+
return (code, s["code"], gr.update(choices=hw_choices(meta), value=flavor),
|
| 250 |
+
"\n\n".join(notes), infer_deps(code, meta.get("task")))
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
def cost_note(flavor, timeout):
|
| 254 |
+
h = HW_BY_NAME.get(flavor)
|
| 255 |
+
if not h:
|
| 256 |
+
return ""
|
| 257 |
+
return f"{flavor}: ${h['usd_hr']:.2f}/h. Worst case at a {int(timeout)} min timeout: ${h['usd_hr'] * timeout / 60:.2f}"
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
def build_script(code, deps, tf_source):
|
| 261 |
+
deps = [d.strip() for d in deps.split(",") if d.strip()]
|
| 262 |
+
if tf_source == "main":
|
| 263 |
+
deps = [d for d in deps if d != "transformers"]
|
| 264 |
+
deps.append("transformers @ https://github.com/huggingface/transformers/archive/refs/heads/main.zip")
|
| 265 |
+
header = "# /// script\n# requires-python = \">=3.10\"\n# dependencies = [\n"
|
| 266 |
+
header += "".join(f'# "{d}",\n' for d in deps) + "# ]\n# ///\n\n"
|
| 267 |
+
return header + code
|
| 268 |
+
|
| 269 |
+
|
| 270 |
+
def run_snippet(code, idx, snippets, meta, flavor, deps, tf_source, timeout,
|
| 271 |
+
oauth_token: gr.OAuthToken | None):
|
| 272 |
+
token = _token(oauth_token)
|
| 273 |
+
if not token:
|
| 274 |
+
yield "Sign in with Hugging Face first. Jobs run on your account.", "", None
|
| 275 |
+
return
|
| 276 |
+
if idx is not None and snippets and snippets[idx]["language"] != "python":
|
| 277 |
+
yield "This is a shell snippet. Copy it to run locally.", "", None
|
| 278 |
+
return
|
| 279 |
+
try:
|
| 280 |
+
api.auth_check(meta["model_id"], token=token)
|
| 281 |
+
except GatedRepoError:
|
| 282 |
+
yield f"You don't have access to {meta['model_id']} yet. Accept its license on the model page.", "", None
|
| 283 |
+
return
|
| 284 |
+
except Exception:
|
| 285 |
+
pass # let the job surface any other access problem
|
| 286 |
+
|
| 287 |
+
script = build_script(code, deps, tf_source)
|
| 288 |
+
try:
|
| 289 |
+
job = run_job(
|
| 290 |
+
image=UV_IMAGE,
|
| 291 |
+
command=["bash", "-c", 'printf "%s" "$PLAYGROUND_SCRIPT" > /tmp/playground.py && uv run /tmp/playground.py'],
|
| 292 |
+
env={"PLAYGROUND_SCRIPT": script, "PYTHONUNBUFFERED": "1"},
|
| 293 |
+
secrets={"HF_TOKEN": token},
|
| 294 |
+
flavor=flavor,
|
| 295 |
+
timeout=f"{int(timeout)}m",
|
| 296 |
+
labels={"app": "transformers-playground"},
|
| 297 |
+
token=token,
|
| 298 |
+
)
|
| 299 |
+
except HfHubHTTPError as e:
|
| 300 |
+
hint = " Jobs need a PRO, Team or Enterprise account with credits." if e.response is not None and e.response.status_code in (402, 403) else ""
|
| 301 |
+
yield f"Could not start the job: {e}.{hint}", "", None
|
| 302 |
+
return
|
| 303 |
+
|
| 304 |
+
status = f"Job started on **{flavor}**. [Open job page]({job.url})"
|
| 305 |
+
logs, last = "", 0.0
|
| 306 |
+
yield status + "\n\nWaiting for logs...", logs, job.id
|
| 307 |
+
try:
|
| 308 |
+
for line in fetch_job_logs(job_id=job.id, follow=True, token=token):
|
| 309 |
+
logs += line.rstrip("\n") + "\n"
|
| 310 |
+
if time.time() - last > 0.5:
|
| 311 |
+
last = time.time()
|
| 312 |
+
yield status + "\n\nRunning...", logs, job.id
|
| 313 |
+
except Exception as e:
|
| 314 |
+
logs += f"\n[log stream interrupted: {e}]\n"
|
| 315 |
+
stage = inspect_job(job_id=job.id, token=token).status.stage
|
| 316 |
+
yield f"{status}\n\nFinished with status **{stage}**.", logs, job.id
|
| 317 |
+
|
| 318 |
+
|
| 319 |
+
def stop_job(job_id, oauth_token: gr.OAuthToken | None):
|
| 320 |
+
if not job_id or not oauth_token:
|
| 321 |
+
return "No running job to cancel."
|
| 322 |
+
cancel_job(job_id=job_id, token=oauth_token.token)
|
| 323 |
+
return f"Cancel requested for job `{job_id}`."
|
| 324 |
+
|
| 325 |
+
# ------------------------------------------------------------------- UI ----
|
| 326 |
+
|
| 327 |
+
TASKS = ["", "text-generation", "image-text-to-text", "text-classification", "token-classification",
|
| 328 |
+
"fill-mask", "question-answering", "summarization", "translation", "feature-extraction",
|
| 329 |
+
"automatic-speech-recognition", "audio-classification", "image-classification",
|
| 330 |
+
"object-detection", "image-segmentation", "depth-estimation", "zero-shot-classification"]
|
| 331 |
+
|
| 332 |
+
with gr.Blocks(title="Transformers Playground") as demo:
|
| 333 |
+
snippets_state, meta_state, job_state = gr.State([]), gr.State({}), gr.State(None)
|
| 334 |
+
|
| 335 |
+
with gr.Row(equal_height=True):
|
| 336 |
+
gr.Markdown("# Transformers Playground\nPick a model, copy the official usage snippet, "
|
| 337 |
+
"or run it on Hugging Face Jobs with your own account.")
|
| 338 |
+
gr.LoginButton(scale=0, min_width=240)
|
| 339 |
+
|
| 340 |
+
with gr.Row():
|
| 341 |
+
query = gr.Textbox(label="Search models", placeholder="llama, whisper, Qwen/Qwen2.5-0.5B-Instruct ...", scale=3)
|
| 342 |
+
task = gr.Dropdown(TASKS, value="", label="Task", scale=2)
|
| 343 |
+
search_btn = gr.Button("Search", scale=1)
|
| 344 |
+
model = gr.Dropdown(label="Model", choices=[], allow_custom_value=True,
|
| 345 |
+
info="Pick from results, or paste any model ID.")
|
| 346 |
+
model_md = gr.Markdown()
|
| 347 |
+
|
| 348 |
+
with gr.Row():
|
| 349 |
+
with gr.Column(scale=3):
|
| 350 |
+
snippet_pick = gr.Radio(label="Snippet", choices=[])
|
| 351 |
+
code = gr.Code(language="python", interactive=True, lines=18,
|
| 352 |
+
label="Snippet, adapted to your model (editable)", buttons=["copy", "download"])
|
| 353 |
+
with gr.Accordion("Original snippet from the docs", open=False):
|
| 354 |
+
original = gr.Code(language="python", interactive=False, buttons=["copy"])
|
| 355 |
+
with gr.Column(scale=2):
|
| 356 |
+
notes = gr.Markdown()
|
| 357 |
+
flavor = gr.Dropdown(label="Hardware", choices=hw_choices({}), value="cpu-basic")
|
| 358 |
+
timeout = gr.Slider(5, 120, value=30, step=5, label="Timeout (minutes)")
|
| 359 |
+
cost = gr.Markdown(cost_note("cpu-basic", 30))
|
| 360 |
+
deps = gr.Textbox(label="Dependencies", info="Comma-separated pip specs added to the job.")
|
| 361 |
+
tf_source = gr.Radio(["latest release", "main"], value="latest release",
|
| 362 |
+
label="transformers version",
|
| 363 |
+
info="Snippets come from the main docs; pick main if a newer API fails.")
|
| 364 |
+
with gr.Row():
|
| 365 |
+
run_btn = gr.Button("Run on HF Jobs", variant="primary")
|
| 366 |
+
stop_btn = gr.Button("Cancel job", variant="stop")
|
| 367 |
+
status = gr.Markdown()
|
| 368 |
+
logs = gr.Code(language="shell", label="Job logs", lines=14, interactive=False)
|
| 369 |
+
|
| 370 |
+
search_btn.click(search_models, [query, task], model)
|
| 371 |
+
query.submit(search_models, [query, task], model)
|
| 372 |
+
model.change(select_model, [model],
|
| 373 |
+
[model_md, snippet_pick, snippets_state, meta_state, code, original, flavor, notes, deps])
|
| 374 |
+
snippet_pick.input(snippet_outputs, [snippet_pick, snippets_state, meta_state],
|
| 375 |
+
[code, original, flavor, notes, deps])
|
| 376 |
+
flavor.change(cost_note, [flavor, timeout], cost)
|
| 377 |
+
timeout.change(cost_note, [flavor, timeout], cost)
|
| 378 |
+
run_btn.click(run_snippet, [code, snippet_pick, snippets_state, meta_state, flavor, deps, tf_source, timeout],
|
| 379 |
+
[status, logs, job_state])
|
| 380 |
+
stop_btn.click(stop_job, [job_state], status)
|
| 381 |
+
|
| 382 |
+
if __name__ == "__main__":
|
| 383 |
+
demo.launch(theme=gr.themes.Soft(primary_hue="amber", secondary_hue="slate"))
|
indexer.py
ADDED
|
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Build the snippet index from transformers' docs/source/en/model_doc.
|
| 2 |
+
|
| 3 |
+
Run locally: python indexer.py --out snippets.json
|
| 4 |
+
Push to the Hub: python indexer.py --push your-org/transformers-snippets
|
| 5 |
+
(Schedule it as a daily HF Job to keep the index fresh.)
|
| 6 |
+
"""
|
| 7 |
+
import argparse
|
| 8 |
+
import json
|
| 9 |
+
import os
|
| 10 |
+
import re
|
| 11 |
+
import shutil
|
| 12 |
+
import subprocess
|
| 13 |
+
import time
|
| 14 |
+
|
| 15 |
+
REPO = "huggingface/transformers"
|
| 16 |
+
DOC_DIR = "docs/source/en/model_doc"
|
| 17 |
+
|
| 18 |
+
FENCE = re.compile(r"```(py|python|bash|sh)\s*\n(.*?)```", re.S)
|
| 19 |
+
HFOPTIONS = re.compile(r'<hfoptions id="([^"]+)">(.*?)</hfoptions>', re.S)
|
| 20 |
+
HFOPTION = re.compile(r'<hfoption id="([^"]+)">(.*?)</hfoption>', re.S)
|
| 21 |
+
HEADING = re.compile(r"^(#{2,3})\s+(.+)$", re.M)
|
| 22 |
+
CHECKPOINT = re.compile(
|
| 23 |
+
r'(?:from_pretrained\(\s*|model\s*=\s*|--model\s+)["\']([\w.\-]+/[\w.\-]+)["\']?'
|
| 24 |
+
)
|
| 25 |
+
|
| 26 |
+
# v1 is inference-only: drop training and quantization snippets.
|
| 27 |
+
EXCLUDE = re.compile(
|
| 28 |
+
r"\bTrainer\b|\.train\(|TrainingArguments|SFTTrainer|trl\b|loss\.backward|optimizer\."
|
| 29 |
+
r"|BitsAndBytesConfig|TorchAoConfig|quantization_config|GPTQConfig|AwqConfig|HqqConfig"
|
| 30 |
+
)
|
| 31 |
+
INFERENCE_HINT = re.compile(r"pipeline\(|\.generate\(|from_pretrained|transformers (chat|run|serve)")
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def fetch_docs(ref="main", workdir="/tmp/transformers-docs"):
|
| 35 |
+
"""Shallow, sparse clone of just model_doc. No GitHub API calls, so no rate limits."""
|
| 36 |
+
shutil.rmtree(workdir, ignore_errors=True)
|
| 37 |
+
run = lambda *a, **kw: subprocess.run(a, check=True, capture_output=True, text=True, **kw)
|
| 38 |
+
run("git", "clone", "--quiet", "--depth", "1", "--filter=blob:none", "--sparse",
|
| 39 |
+
"--branch", ref, f"https://github.com/{REPO}", workdir)
|
| 40 |
+
run("git", "sparse-checkout", "set", DOC_DIR, cwd=workdir)
|
| 41 |
+
sha = run("git", "rev-parse", "HEAD", cwd=workdir).stdout.strip()
|
| 42 |
+
doc_root = os.path.join(workdir, DOC_DIR)
|
| 43 |
+
files = {}
|
| 44 |
+
for name in sorted(os.listdir(doc_root)):
|
| 45 |
+
if name.endswith(".md"):
|
| 46 |
+
with open(os.path.join(doc_root, name), encoding="utf-8") as f:
|
| 47 |
+
files[f"{DOC_DIR}/{name}"] = f.read()
|
| 48 |
+
return sha, files
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def _section_for(text, pos):
|
| 52 |
+
"""Nearest heading above a position, used as context for loose code blocks."""
|
| 53 |
+
title = ""
|
| 54 |
+
for m in HEADING.finditer(text):
|
| 55 |
+
if m.start() > pos:
|
| 56 |
+
break
|
| 57 |
+
title = m.group(2).strip()
|
| 58 |
+
return title
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def _clean(code):
|
| 62 |
+
code = code.strip("\n")
|
| 63 |
+
return re.sub(r"^\s*<!--.*?-->\s*$", "", code, flags=re.M).strip("\n")
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
def _snippet(code, lang, label, section):
|
| 67 |
+
code = _clean(code)
|
| 68 |
+
if not code or EXCLUDE.search(code):
|
| 69 |
+
return None
|
| 70 |
+
if not INFERENCE_HINT.search(code):
|
| 71 |
+
return None
|
| 72 |
+
lang = "python" if lang in ("py", "python") else "bash"
|
| 73 |
+
ckpt = CHECKPOINT.search(code)
|
| 74 |
+
return {
|
| 75 |
+
"label": label,
|
| 76 |
+
"section": section,
|
| 77 |
+
"language": lang,
|
| 78 |
+
"code": code,
|
| 79 |
+
"checkpoint": ckpt.group(1) if ckpt else None,
|
| 80 |
+
"needs_gpu": bool(re.search(r'cuda|device_map|attn_implementation="flash', code)),
|
| 81 |
+
}
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def parse_doc(text):
|
| 85 |
+
snippets, covered = [], []
|
| 86 |
+
# 1. Structured usage tabs: <hfoptions id="usage"><hfoption id="Pipeline">...
|
| 87 |
+
for group in HFOPTIONS.finditer(text):
|
| 88 |
+
covered.append(group.span())
|
| 89 |
+
section = _section_for(text, group.start())
|
| 90 |
+
for opt in HFOPTION.finditer(group.group(2)):
|
| 91 |
+
for fence in FENCE.finditer(opt.group(2)):
|
| 92 |
+
s = _snippet(fence.group(2), fence.group(1), opt.group(1), section)
|
| 93 |
+
if s:
|
| 94 |
+
snippets.append(s)
|
| 95 |
+
# 2. Loose code blocks under usage-like headings (older doc pages).
|
| 96 |
+
for fence in FENCE.finditer(text):
|
| 97 |
+
if any(a <= fence.start() < b for a, b in covered):
|
| 98 |
+
continue
|
| 99 |
+
section = _section_for(text, fence.start())
|
| 100 |
+
if not re.search(r"usage|example|inference|generat|quick|tips|notes", section, re.I):
|
| 101 |
+
continue
|
| 102 |
+
s = _snippet(fence.group(2), fence.group(1), section or "Example", section)
|
| 103 |
+
if s:
|
| 104 |
+
snippets.append(s)
|
| 105 |
+
# De-duplicate identical code.
|
| 106 |
+
seen, out = set(), []
|
| 107 |
+
for s in snippets:
|
| 108 |
+
if s["code"] not in seen:
|
| 109 |
+
seen.add(s["code"])
|
| 110 |
+
out.append(s)
|
| 111 |
+
return out
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def build_index(ref="main"):
|
| 115 |
+
sha, files = fetch_docs(ref)
|
| 116 |
+
docs = {}
|
| 117 |
+
for path, text in files.items():
|
| 118 |
+
arch = os.path.basename(path)[:-3]
|
| 119 |
+
title = re.search(r"^#\s+(.+)$", text, re.M)
|
| 120 |
+
docs[arch] = {
|
| 121 |
+
"title": title.group(1).strip() if title else arch,
|
| 122 |
+
"doc_url": f"https://huggingface.co/docs/transformers/model_doc/{arch}",
|
| 123 |
+
"source_url": f"https://github.com/{REPO}/blob/{sha}/{path}",
|
| 124 |
+
"snippets": parse_doc(text),
|
| 125 |
+
}
|
| 126 |
+
return {"ref": ref, "sha": sha, "built_at": int(time.time()), "docs": docs}
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
if __name__ == "__main__":
|
| 130 |
+
ap = argparse.ArgumentParser()
|
| 131 |
+
ap.add_argument("--ref", default="main")
|
| 132 |
+
ap.add_argument("--out", default="snippets.json")
|
| 133 |
+
ap.add_argument("--push", help="dataset repo id to upload snippets.json to")
|
| 134 |
+
args = ap.parse_args()
|
| 135 |
+
|
| 136 |
+
index = build_index(args.ref)
|
| 137 |
+
with open(args.out, "w") as f:
|
| 138 |
+
json.dump(index, f)
|
| 139 |
+
n = sum(len(d["snippets"]) for d in index["docs"].values())
|
| 140 |
+
print(f"{len(index['docs'])} docs, {n} snippets @ {index['sha'][:8]}")
|
| 141 |
+
|
| 142 |
+
if args.push:
|
| 143 |
+
from huggingface_hub import HfApi
|
| 144 |
+
api = HfApi()
|
| 145 |
+
api.create_repo(args.push, repo_type="dataset", exist_ok=True)
|
| 146 |
+
api.upload_file(path_or_fileobj=args.out, path_in_repo="snippets.json",
|
| 147 |
+
repo_id=args.push, repo_type="dataset",
|
| 148 |
+
commit_message=f"Index transformers@{index['sha'][:8]}")
|
| 149 |
+
print(f"Pushed to https://huggingface.co/datasets/{args.push}")
|
requirements.txt
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
huggingface_hub>=2.2,<3
|
| 2 |
+
requests
|