Spaces:
Restarting
on
A10G
Restarting
on
A10G
Some cleaning in huggingface_hub integration (#13)
Browse files- Some cleaning in huggingface_hub integration (85fe931f3911f59e26cc6c4ca2f28c5d1affa26a)
Co-authored-by: Lucain Pouget <Wauplin@users.noreply.huggingface.co>
app.py
CHANGED
@@ -24,25 +24,23 @@ def script_to_use(model_id, api):
|
|
24 |
return "convert.py" if arch in LLAMA_LIKE_ARCHS else "convert-hf-to-gguf.py"
|
25 |
|
26 |
def process_model(model_id, q_method, hf_token):
|
27 |
-
|
28 |
-
fp16 = f"{
|
29 |
|
30 |
try:
|
31 |
api = HfApi(token=hf_token)
|
32 |
|
33 |
-
|
34 |
-
|
35 |
-
snapshot_download(repo_id=model_id, local_dir = f"{MODEL_NAME}", local_dir_use_symlinks=False)
|
36 |
print("Model downloaded successully!")
|
37 |
|
38 |
conversion_script = script_to_use(model_id, api)
|
39 |
-
fp16_conversion = f"python llama.cpp/{conversion_script} {
|
40 |
result = subprocess.run(fp16_conversion, shell=True, capture_output=True)
|
41 |
if result.returncode != 0:
|
42 |
raise Exception(f"Error converting to fp16: {result.stderr}")
|
43 |
print("Model converted to fp16 successully!")
|
44 |
|
45 |
-
qtype = f"{
|
46 |
quantise_ggml = f"./llama.cpp/quantize {fp16} {qtype} {q_method}"
|
47 |
result = subprocess.run(quantise_ggml, shell=True, capture_output=True)
|
48 |
if result.returncode != 0:
|
@@ -50,20 +48,15 @@ def process_model(model_id, q_method, hf_token):
|
|
50 |
print("Quantised successfully!")
|
51 |
|
52 |
# Create empty repo
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
repo_type="model",
|
57 |
-
exist_ok=True,
|
58 |
-
token=hf_token
|
59 |
-
)
|
60 |
-
print("Repo created successfully!")
|
61 |
|
62 |
card = ModelCard.load(model_id)
|
63 |
card.data.tags = ["llama-cpp"] if card.data.tags is None else card.data.tags + ["llama-cpp"]
|
64 |
card.text = dedent(
|
65 |
f"""
|
66 |
-
# {
|
67 |
This model was converted to GGUF format from [`{model_id}`](https://huggingface.co/{model_id}) using llama.cpp.
|
68 |
Refer to the [original model card](https://huggingface.co/{model_id}) for more details on the model.
|
69 |
## Use with llama.cpp
|
@@ -73,39 +66,37 @@ def process_model(model_id, q_method, hf_token):
|
|
73 |
```
|
74 |
|
75 |
```bash
|
76 |
-
llama-cli --hf-repo {
|
77 |
```
|
78 |
|
79 |
```bash
|
80 |
-
llama-server --hf-repo {
|
81 |
```
|
82 |
"""
|
83 |
)
|
84 |
-
card.save(os.path.join(
|
85 |
|
86 |
api.upload_file(
|
87 |
path_or_fileobj=qtype,
|
88 |
path_in_repo=qtype.split("/")[-1],
|
89 |
-
repo_id=
|
90 |
-
repo_type="model",
|
91 |
)
|
92 |
|
93 |
api.upload_file(
|
94 |
-
path_or_fileobj=f"{
|
95 |
path_in_repo="README.md",
|
96 |
-
repo_id=
|
97 |
-
repo_type="model",
|
98 |
)
|
99 |
print("Uploaded successfully!")
|
100 |
|
101 |
return (
|
102 |
-
f'Find your repo <a href=\'{
|
103 |
"llama.png",
|
104 |
)
|
105 |
except Exception as e:
|
106 |
return (f"Error: {e}", "error.png")
|
107 |
finally:
|
108 |
-
shutil.rmtree(
|
109 |
print("Folder cleaned up successfully!")
|
110 |
|
111 |
|
|
|
24 |
return "convert.py" if arch in LLAMA_LIKE_ARCHS else "convert-hf-to-gguf.py"
|
25 |
|
26 |
def process_model(model_id, q_method, hf_token):
|
27 |
+
model_name = model_id.split('/')[-1]
|
28 |
+
fp16 = f"{model_name}/{model_name.lower()}.fp16.bin"
|
29 |
|
30 |
try:
|
31 |
api = HfApi(token=hf_token)
|
32 |
|
33 |
+
snapshot_download(repo_id=model_id, local_dir=model_name, local_dir_use_symlinks=False)
|
|
|
|
|
34 |
print("Model downloaded successully!")
|
35 |
|
36 |
conversion_script = script_to_use(model_id, api)
|
37 |
+
fp16_conversion = f"python llama.cpp/{conversion_script} {model_name} --outtype f16 --outfile {fp16}"
|
38 |
result = subprocess.run(fp16_conversion, shell=True, capture_output=True)
|
39 |
if result.returncode != 0:
|
40 |
raise Exception(f"Error converting to fp16: {result.stderr}")
|
41 |
print("Model converted to fp16 successully!")
|
42 |
|
43 |
+
qtype = f"{model_name}/{model_name.lower()}.{q_method.upper()}.gguf"
|
44 |
quantise_ggml = f"./llama.cpp/quantize {fp16} {qtype} {q_method}"
|
45 |
result = subprocess.run(quantise_ggml, shell=True, capture_output=True)
|
46 |
if result.returncode != 0:
|
|
|
48 |
print("Quantised successfully!")
|
49 |
|
50 |
# Create empty repo
|
51 |
+
new_repo_url = api.create_repo(repo_id=f"{model_name}-{q_method}-GGUF", exist_ok=True)
|
52 |
+
new_repo_id = new_repo_url.repo_id
|
53 |
+
print("Repo created successfully!", new_repo_url)
|
|
|
|
|
|
|
|
|
|
|
54 |
|
55 |
card = ModelCard.load(model_id)
|
56 |
card.data.tags = ["llama-cpp"] if card.data.tags is None else card.data.tags + ["llama-cpp"]
|
57 |
card.text = dedent(
|
58 |
f"""
|
59 |
+
# {new_repo_id}
|
60 |
This model was converted to GGUF format from [`{model_id}`](https://huggingface.co/{model_id}) using llama.cpp.
|
61 |
Refer to the [original model card](https://huggingface.co/{model_id}) for more details on the model.
|
62 |
## Use with llama.cpp
|
|
|
66 |
```
|
67 |
|
68 |
```bash
|
69 |
+
llama-cli --hf-repo {new_repo_id} --model {qtype.split("/")[-1]} -p "The meaning to life and the universe is "
|
70 |
```
|
71 |
|
72 |
```bash
|
73 |
+
llama-server --hf-repo {new_repo_id} --model {qtype.split("/")[-1]} -c 2048
|
74 |
```
|
75 |
"""
|
76 |
)
|
77 |
+
card.save(os.path.join(model_name, "README-new.md"))
|
78 |
|
79 |
api.upload_file(
|
80 |
path_or_fileobj=qtype,
|
81 |
path_in_repo=qtype.split("/")[-1],
|
82 |
+
repo_id=new_repo_id,
|
|
|
83 |
)
|
84 |
|
85 |
api.upload_file(
|
86 |
+
path_or_fileobj=f"{model_name}/README-new.md",
|
87 |
path_in_repo="README.md",
|
88 |
+
repo_id=new_repo_id,
|
|
|
89 |
)
|
90 |
print("Uploaded successfully!")
|
91 |
|
92 |
return (
|
93 |
+
f'Find your repo <a href=\'{new_repo_url}\' target="_blank" style="text-decoration:underline">here</a>',
|
94 |
"llama.png",
|
95 |
)
|
96 |
except Exception as e:
|
97 |
return (f"Error: {e}", "error.png")
|
98 |
finally:
|
99 |
+
shutil.rmtree(model_name, ignore_errors=True)
|
100 |
print("Folder cleaned up successfully!")
|
101 |
|
102 |
|