Spaces:
Running
on
A100
Running
on
A100
feat: Choosable CLI, Custom Output Shard Size, LORA extraction
Browse files- app.py +142 -29
- requirements.txt +1 -0
app.py
CHANGED
@@ -11,6 +11,7 @@ import gradio as gr
|
|
11 |
import huggingface_hub
|
12 |
import torch
|
13 |
import yaml
|
|
|
14 |
from gradio_logsview.logsview import Log, LogsView, LogsViewRunner
|
15 |
from mergekit.config import MergeConfiguration
|
16 |
|
@@ -43,7 +44,7 @@ has_gpu = torch.cuda.is_available()
|
|
43 |
# )
|
44 |
|
45 |
cli = "mergekit-yaml config.yaml merge --copy-tokenizer" + (
|
46 |
-
" --cuda --low-cpu-memory --allow-crimes" if has_gpu else " --allow-crimes --
|
47 |
)
|
48 |
|
49 |
MARKDOWN_DESCRIPTION = """
|
@@ -111,17 +112,19 @@ examples = [[str(f)] for f in pathlib.Path("examples").glob("*.yaml")]
|
|
111 |
COMMUNITY_HF_TOKEN = os.getenv("COMMUNITY_HF_TOKEN")
|
112 |
|
113 |
|
114 |
-
def merge(yaml_config: str, hf_token: str, repo_name: str) -> Iterable[List[Log]]:
|
115 |
runner = LogsViewRunner()
|
116 |
|
117 |
if not yaml_config:
|
118 |
yield runner.log("Empty yaml, pick an example below", level="ERROR")
|
119 |
return
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
|
|
|
|
125 |
|
126 |
is_community_model = False
|
127 |
if not hf_token:
|
@@ -170,7 +173,7 @@ def merge(yaml_config: str, hf_token: str, repo_name: str) -> Iterable[List[Log]
|
|
170 |
# Set tmp HF_HOME to avoid filling up disk Space
|
171 |
tmp_env = os.environ.copy() # taken from https://stackoverflow.com/a/4453495
|
172 |
tmp_env["HF_HOME"] = f"{tmpdirname}/.cache"
|
173 |
-
full_cli =
|
174 |
yield from runner.run_command(full_cli.split(), cwd=merged_path, env=tmp_env)
|
175 |
|
176 |
if runner.exit_code != 0:
|
@@ -187,27 +190,139 @@ def merge(yaml_config: str, hf_token: str, repo_name: str) -> Iterable[List[Log]
|
|
187 |
yield runner.log(f"Model successfully uploaded to HF: {repo_url.repo_id}")
|
188 |
|
189 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
190 |
with gr.Blocks() as demo:
|
191 |
gr.Markdown(MARKDOWN_DESCRIPTION)
|
192 |
|
193 |
-
with gr.
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
211 |
gr.Examples(
|
212 |
examples,
|
213 |
fn=lambda s: (s,),
|
@@ -218,11 +333,9 @@ with gr.Blocks() as demo:
|
|
218 |
)
|
219 |
gr.Markdown(MARKDOWN_ARTICLE)
|
220 |
|
221 |
-
button.click(fn=merge, inputs=[config, token, repo_name], outputs=[logs])
|
222 |
-
|
223 |
|
224 |
# Run garbage collection every hour to keep the community org clean.
|
225 |
-
# Empty models might
|
226 |
def _garbage_collect_every_hour():
|
227 |
while True:
|
228 |
try:
|
|
|
11 |
import huggingface_hub
|
12 |
import torch
|
13 |
import yaml
|
14 |
+
import bitsandbytes
|
15 |
from gradio_logsview.logsview import Log, LogsView, LogsViewRunner
|
16 |
from mergekit.config import MergeConfiguration
|
17 |
|
|
|
44 |
# )
|
45 |
|
46 |
cli = "mergekit-yaml config.yaml merge --copy-tokenizer" + (
|
47 |
+
" --cuda --low-cpu-memory --allow-crimes" if has_gpu else " --allow-crimes --lazy-unpickle"
|
48 |
)
|
49 |
|
50 |
MARKDOWN_DESCRIPTION = """
|
|
|
112 |
COMMUNITY_HF_TOKEN = os.getenv("COMMUNITY_HF_TOKEN")
|
113 |
|
114 |
|
115 |
+
def merge(program: str, yaml_config: str, out_shard_size: str, hf_token: str, repo_name: str) -> Iterable[List[Log]]:
|
116 |
runner = LogsViewRunner()
|
117 |
|
118 |
if not yaml_config:
|
119 |
yield runner.log("Empty yaml, pick an example below", level="ERROR")
|
120 |
return
|
121 |
+
# TODO: validate moe config and mega config?
|
122 |
+
if program not in ("mergekit-moe", "mergekit-mega"):
|
123 |
+
try:
|
124 |
+
merge_config = MergeConfiguration.model_validate(yaml.safe_load(yaml_config))
|
125 |
+
except Exception as e:
|
126 |
+
yield runner.log(f"Invalid yaml {e}", level="ERROR")
|
127 |
+
return
|
128 |
|
129 |
is_community_model = False
|
130 |
if not hf_token:
|
|
|
173 |
# Set tmp HF_HOME to avoid filling up disk Space
|
174 |
tmp_env = os.environ.copy() # taken from https://stackoverflow.com/a/4453495
|
175 |
tmp_env["HF_HOME"] = f"{tmpdirname}/.cache"
|
176 |
+
full_cli = f"{program} {cli} --lora-merge-cache {tmpdirname}/.lora_cache --out-shard-size {out_shard_size}"
|
177 |
yield from runner.run_command(full_cli.split(), cwd=merged_path, env=tmp_env)
|
178 |
|
179 |
if runner.exit_code != 0:
|
|
|
190 |
yield runner.log(f"Model successfully uploaded to HF: {repo_url.repo_id}")
|
191 |
|
192 |
|
193 |
+
def extract(finetuned_model: str, base_model: str, rank: int, hf_token: str, repo_name: str) -> Iterable[List[Log]]:
|
194 |
+
runner = LogsViewRunner()
|
195 |
+
if not finetuned_model or not base_model:
|
196 |
+
yield runner.log("All field should be filled")
|
197 |
+
|
198 |
+
is_community_model = False
|
199 |
+
if not hf_token:
|
200 |
+
if "/" in repo_name and not repo_name.startswith("mergekit-community/"):
|
201 |
+
yield runner.log(
|
202 |
+
f"Cannot upload merge model to namespace {repo_name.split('/')[0]}: you must provide a valid token.",
|
203 |
+
level="ERROR",
|
204 |
+
)
|
205 |
+
return
|
206 |
+
yield runner.log(
|
207 |
+
"No HF token provided. Your lora will be uploaded to the https://huggingface.co/mergekit-community organization."
|
208 |
+
)
|
209 |
+
is_community_model = True
|
210 |
+
if not COMMUNITY_HF_TOKEN:
|
211 |
+
raise gr.Error("Cannot upload to community org: community token not set by Space owner.")
|
212 |
+
hf_token = COMMUNITY_HF_TOKEN
|
213 |
+
|
214 |
+
api = huggingface_hub.HfApi(token=hf_token)
|
215 |
+
|
216 |
+
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
|
217 |
+
tmpdir = pathlib.Path(tmpdirname)
|
218 |
+
merged_path = tmpdir / "merged"
|
219 |
+
merged_path.mkdir(parents=True, exist_ok=True)
|
220 |
+
|
221 |
+
if not repo_name:
|
222 |
+
yield runner.log("No repo name provided. Generating a random one.")
|
223 |
+
repo_name = "lora"
|
224 |
+
# Make repo_name "unique" (no need to be extra careful on uniqueness)
|
225 |
+
repo_name += "-" + "".join(random.choices(string.ascii_lowercase, k=7))
|
226 |
+
repo_name = repo_name.replace("/", "-").strip("-")
|
227 |
+
|
228 |
+
if is_community_model and not repo_name.startswith("mergekit-community/"):
|
229 |
+
repo_name = f"mergekit-community/{repo_name}"
|
230 |
+
|
231 |
+
try:
|
232 |
+
yield runner.log(f"Creating repo {repo_name}")
|
233 |
+
repo_url = api.create_repo(repo_name, exist_ok=True)
|
234 |
+
yield runner.log(f"Repo created: {repo_url}")
|
235 |
+
except Exception as e:
|
236 |
+
yield runner.log(f"Error creating repo {e}", level="ERROR")
|
237 |
+
return
|
238 |
+
|
239 |
+
# Set tmp HF_HOME to avoid filling up disk Space
|
240 |
+
tmp_env = os.environ.copy() # taken from https://stackoverflow.com/a/4453495
|
241 |
+
tmp_env["HF_HOME"] = f"{tmpdirname}/.cache"
|
242 |
+
full_cli = f"mergekit-extract-lora {finetuned_model} {base_model} lora --rank={rank}"
|
243 |
+
yield from runner.run_command(full_cli.split(), cwd=merged_path, env=tmp_env)
|
244 |
+
|
245 |
+
if runner.exit_code != 0:
|
246 |
+
yield runner.log("Lora extraction failed. Deleting repo as no lora is uploaded.", level="ERROR")
|
247 |
+
api.delete_repo(repo_url.repo_id)
|
248 |
+
return
|
249 |
+
|
250 |
+
yield runner.log("Lora extracted successfully. Uploading to HF.")
|
251 |
+
yield from runner.run_python(
|
252 |
+
api.upload_folder,
|
253 |
+
repo_id=repo_url.repo_id,
|
254 |
+
folder_path=merged_path / "lora",
|
255 |
+
)
|
256 |
+
yield runner.log(f"Lora successfully uploaded to HF: {repo_url.repo_id}")
|
257 |
+
|
258 |
+
|
259 |
with gr.Blocks() as demo:
|
260 |
gr.Markdown(MARKDOWN_DESCRIPTION)
|
261 |
|
262 |
+
with gr.Tabs():
|
263 |
+
with gr.TabItem("Merge Model"):
|
264 |
+
with gr.Row():
|
265 |
+
filename = gr.Textbox(visible=False, label="filename")
|
266 |
+
config = gr.Code(language="yaml", lines=10, label="config.yaml")
|
267 |
+
with gr.Column():
|
268 |
+
program = gr.Dropdown(
|
269 |
+
["mergekit-yaml", "mergekit-mega", "mergekit-moe"],
|
270 |
+
label="Mergekit Command",
|
271 |
+
info="Choose CLI",
|
272 |
+
)
|
273 |
+
out_shard_size = gr.Dropdown(
|
274 |
+
["500M", "1B", "2B", "3B", "4B", "5B"],
|
275 |
+
label="Output Shard Size",
|
276 |
+
value="500M",
|
277 |
+
)
|
278 |
+
token = gr.Textbox(
|
279 |
+
lines=1,
|
280 |
+
label="HF Write Token",
|
281 |
+
info="https://hf.co/settings/token",
|
282 |
+
type="password",
|
283 |
+
placeholder="Optional. Will upload merged model to MergeKit Community if empty.",
|
284 |
+
)
|
285 |
+
repo_name = gr.Textbox(
|
286 |
+
lines=1,
|
287 |
+
label="Repo name",
|
288 |
+
placeholder="Optional. Will create a random name if empty.",
|
289 |
+
)
|
290 |
+
button = gr.Button("Merge", variant="primary")
|
291 |
+
logs = LogsView(label="Terminal output")
|
292 |
+
button.click(fn=merge, inputs=[program, config, out_shard_size, token, repo_name], outputs=[logs])
|
293 |
+
|
294 |
+
with gr.TabItem("LORA Extraction"):
|
295 |
+
with gr.Row():
|
296 |
+
with gr.Column():
|
297 |
+
finetuned_model = gr.Textbox(
|
298 |
+
lines=1,
|
299 |
+
label="Finetuned Model",
|
300 |
+
)
|
301 |
+
base_model = gr.Textbox(
|
302 |
+
lines=1,
|
303 |
+
label="Base Model",
|
304 |
+
)
|
305 |
+
rank = gr.Dropdown(
|
306 |
+
[32, 64, 128],
|
307 |
+
label="Rank level",
|
308 |
+
value=32,
|
309 |
+
)
|
310 |
+
with gr.Column():
|
311 |
+
token = gr.Textbox(
|
312 |
+
lines=1,
|
313 |
+
label="HF Write Token",
|
314 |
+
info="https://hf.co/settings/token",
|
315 |
+
type="password",
|
316 |
+
placeholder="Optional. Will upload merged model to MergeKit Community if empty.",
|
317 |
+
)
|
318 |
+
repo_name = gr.Textbox(
|
319 |
+
lines=1,
|
320 |
+
label="Repo name",
|
321 |
+
placeholder="Optional. Will create a random name if empty.",
|
322 |
+
)
|
323 |
+
button = gr.Button("Extract LORA", variant="primary")
|
324 |
+
logs = LogsView(label="Terminal output")
|
325 |
+
button.click(fn=extract, inputs=[finetuned_model, base_model, rank, token, repo_name], outputs=[logs])
|
326 |
gr.Examples(
|
327 |
examples,
|
328 |
fn=lambda s: (s,),
|
|
|
333 |
)
|
334 |
gr.Markdown(MARKDOWN_ARTICLE)
|
335 |
|
|
|
|
|
336 |
|
337 |
# Run garbage collection every hour to keep the community org clean.
|
338 |
+
# Empty models might exist if the merge fails abruptly (e.g. if user leaves the Space).
|
339 |
def _garbage_collect_every_hour():
|
340 |
while True:
|
341 |
try:
|
requirements.txt
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
torch
|
|
|
2 |
git+https://github.com/arcee-ai/mergekit.git
|
3 |
# see https://huggingface.co/spaces/Wauplin/gradio_logsview
|
4 |
gradio_logsview@https://huggingface.co/spaces/Wauplin/gradio_logsview/resolve/main/gradio_logsview-0.0.5-py3-none-any.whl
|
|
|
1 |
torch
|
2 |
+
bitsandbytes
|
3 |
git+https://github.com/arcee-ai/mergekit.git
|
4 |
# see https://huggingface.co/spaces/Wauplin/gradio_logsview
|
5 |
gradio_logsview@https://huggingface.co/spaces/Wauplin/gradio_logsview/resolve/main/gradio_logsview-0.0.5-py3-none-any.whl
|