Spaces:
Running
Running
Update app.py (#10)
Browse files
app.py
CHANGED
@@ -11,25 +11,41 @@ from huggingface_hub import ModelCard
|
|
11 |
|
12 |
from textwrap import dedent
|
13 |
|
14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
def process_model(model_id, q_method, hf_token):
|
17 |
-
|
18 |
MODEL_NAME = model_id.split('/')[-1]
|
19 |
fp16 = f"{MODEL_NAME}/{MODEL_NAME.lower()}.fp16.bin"
|
|
|
|
|
20 |
|
21 |
username = whoami(hf_token)["name"]
|
22 |
-
|
23 |
snapshot_download(repo_id=model_id, local_dir = f"{MODEL_NAME}", local_dir_use_symlinks=False)
|
24 |
print("Model downloaded successully!")
|
25 |
|
26 |
-
|
27 |
-
|
|
|
|
|
|
|
28 |
print("Model converted to fp16 successully!")
|
29 |
|
30 |
qtype = f"{MODEL_NAME}/{MODEL_NAME.lower()}.{q_method.upper()}.gguf"
|
31 |
quantise_ggml = f"./llama.cpp/quantize {fp16} {qtype} {q_method}"
|
32 |
-
subprocess.run(quantise_ggml, shell=True)
|
|
|
|
|
33 |
print("Quantised successfully!")
|
34 |
|
35 |
# Create empty repo
|
@@ -40,8 +56,7 @@ def process_model(model_id, q_method, hf_token):
|
|
40 |
exist_ok=True,
|
41 |
token=hf_token
|
42 |
)
|
43 |
-
print("
|
44 |
-
|
45 |
|
46 |
card = ModelCard.load(model_id)
|
47 |
card.data.tags = ["llama-cpp"] if card.data.tags is None else card.data.tags + ["llama-cpp"]
|
@@ -59,6 +74,10 @@ def process_model(model_id, q_method, hf_token):
|
|
59 |
```bash
|
60 |
llama-cli --hf-repo {repo_id} --model {qtype.split("/")[-1]} -p "The meaning to life and the universe is "
|
61 |
```
|
|
|
|
|
|
|
|
|
62 |
"""
|
63 |
)
|
64 |
card.save(os.path.join(MODEL_NAME, "README-new.md"))
|
@@ -93,17 +112,21 @@ iface = gr.Interface(
|
|
93 |
gr.Textbox(
|
94 |
lines=1,
|
95 |
label="Hub Model ID",
|
96 |
-
info="Model repo ID"
|
|
|
|
|
97 |
),
|
98 |
gr.Dropdown(
|
99 |
["Q2_K", "Q3_K_S", "Q3_K_M", "Q3_K_L", "Q4_0", "Q4_K_S", "Q4_K_M", "Q5_0", "Q5_K_S", "Q5_K_M", "Q6_K", "Q8_0"],
|
100 |
label="Quantization Method",
|
101 |
-
info="GGML quantisation type"
|
|
|
102 |
),
|
103 |
gr.Textbox(
|
104 |
lines=1,
|
105 |
label="HF Write Token",
|
106 |
-
info="https://hf.co/settings/token"
|
|
|
107 |
)
|
108 |
],
|
109 |
outputs=[
|
|
|
11 |
|
12 |
from textwrap import dedent
|
13 |
|
14 |
+
LLAMA_LIKE_ARCHS = ["MistralForCausalLM", "LlamaForCausalLM"]
|
15 |
+
|
16 |
+
def script_to_use(model_id, api):
|
17 |
+
info = api.model_info(model_id)
|
18 |
+
if info.config is None:
|
19 |
+
return None
|
20 |
+
arch = info.config.get("architectures", None)
|
21 |
+
if arch is None:
|
22 |
+
return None
|
23 |
+
arch = arch[0]
|
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 |
+
api = HfApi(token=hf_token)
|
31 |
|
32 |
username = whoami(hf_token)["name"]
|
33 |
+
|
34 |
snapshot_download(repo_id=model_id, local_dir = f"{MODEL_NAME}", local_dir_use_symlinks=False)
|
35 |
print("Model downloaded successully!")
|
36 |
|
37 |
+
conversion_script = script_to_use(model_id, api)
|
38 |
+
fp16_conversion = f"python llama.cpp/{conversion_script} {MODEL_NAME} --outtype f16 --outfile {fp16}"
|
39 |
+
result = subprocess.run(fp16_conversion, shell=True, capture_output=True)
|
40 |
+
if result.returncode != 0:
|
41 |
+
return (f"Error converting to fp16: {result.stderr}", "error.png")
|
42 |
print("Model converted to fp16 successully!")
|
43 |
|
44 |
qtype = f"{MODEL_NAME}/{MODEL_NAME.lower()}.{q_method.upper()}.gguf"
|
45 |
quantise_ggml = f"./llama.cpp/quantize {fp16} {qtype} {q_method}"
|
46 |
+
result = subprocess.run(quantise_ggml, shell=True, capture_output=True)
|
47 |
+
if result.returncode != 0:
|
48 |
+
return (f"Error quantizing: {result.stderr}", "error.png")
|
49 |
print("Quantised successfully!")
|
50 |
|
51 |
# Create empty repo
|
|
|
56 |
exist_ok=True,
|
57 |
token=hf_token
|
58 |
)
|
59 |
+
print("Repo created successfully!")
|
|
|
60 |
|
61 |
card = ModelCard.load(model_id)
|
62 |
card.data.tags = ["llama-cpp"] if card.data.tags is None else card.data.tags + ["llama-cpp"]
|
|
|
74 |
```bash
|
75 |
llama-cli --hf-repo {repo_id} --model {qtype.split("/")[-1]} -p "The meaning to life and the universe is "
|
76 |
```
|
77 |
+
|
78 |
+
```bash
|
79 |
+
llama-server --hf-repo {repo_id} --model {qtype.split("/")[-1]} -c 2048
|
80 |
+
```
|
81 |
"""
|
82 |
)
|
83 |
card.save(os.path.join(MODEL_NAME, "README-new.md"))
|
|
|
112 |
gr.Textbox(
|
113 |
lines=1,
|
114 |
label="Hub Model ID",
|
115 |
+
info="Model repo ID",
|
116 |
+
placeholder="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
|
117 |
+
value="TinyLlama/TinyLlama-1.1B-Chat-v1.0"
|
118 |
),
|
119 |
gr.Dropdown(
|
120 |
["Q2_K", "Q3_K_S", "Q3_K_M", "Q3_K_L", "Q4_0", "Q4_K_S", "Q4_K_M", "Q5_0", "Q5_K_S", "Q5_K_M", "Q6_K", "Q8_0"],
|
121 |
label="Quantization Method",
|
122 |
+
info="GGML quantisation type",
|
123 |
+
value="Q4_K_M",
|
124 |
),
|
125 |
gr.Textbox(
|
126 |
lines=1,
|
127 |
label="HF Write Token",
|
128 |
+
info="https://hf.co/settings/token",
|
129 |
+
type="password",
|
130 |
)
|
131 |
],
|
132 |
outputs=[
|