fixed the issue
Browse files- .DS_Store +0 -0
- app.py +9 -27
- requirements.txt +0 -1
.DS_Store
CHANGED
|
Binary files a/.DS_Store and b/.DS_Store differ
|
|
|
app.py
CHANGED
|
@@ -1,17 +1,10 @@
|
|
| 1 |
"""
|
| 2 |
-
CommitLens — Gradio UI
|
| 3 |
-
=======================
|
| 4 |
-
- Asks for a GitHub repo URL (and optional PAT)
|
| 5 |
-
- Runs commitlens.run_pipeline() to get per-file prompts
|
| 6 |
-
- Feeds each prompt to Mellum 2 for a per-file summary
|
| 7 |
-
- Combines all summaries and asks the model for a final .md report
|
| 8 |
-
- Displays everything in the browser
|
| 9 |
"""
|
| 10 |
|
| 11 |
from __future__ import annotations
|
| 12 |
|
| 13 |
import gradio as gr
|
| 14 |
-
import spaces
|
| 15 |
import torch
|
| 16 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
| 17 |
|
|
@@ -46,35 +39,32 @@ FINAL_SYSTEM_PROMPT = (
|
|
| 46 |
_model = None
|
| 47 |
_tokenizer = None
|
| 48 |
|
| 49 |
-
|
| 50 |
def _get_llm():
|
| 51 |
global _model, _tokenizer
|
| 52 |
if _model is None:
|
| 53 |
-
# 8-bit quantization to fit
|
| 54 |
quantization_config = BitsAndBytesConfig(
|
| 55 |
load_in_8bit=True,
|
| 56 |
)
|
| 57 |
|
| 58 |
_tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO_ID)
|
| 59 |
|
| 60 |
-
#
|
|
|
|
| 61 |
_model = AutoModelForCausalLM.from_pretrained(
|
| 62 |
MODEL_REPO_ID,
|
| 63 |
quantization_config=quantization_config,
|
| 64 |
device_map="auto",
|
| 65 |
-
torch_dtype=torch.
|
| 66 |
-
attn_implementation="flash_attention_2"
|
| 67 |
)
|
| 68 |
return _model, _tokenizer
|
| 69 |
|
| 70 |
-
|
| 71 |
def _extract_filename(prompt: str) -> str:
|
| 72 |
for line in prompt.splitlines():
|
| 73 |
if line.startswith("Filename :"):
|
| 74 |
return line.split(":", 1)[1].strip()
|
| 75 |
return "unknown"
|
| 76 |
|
| 77 |
-
|
| 78 |
def _generate_response(system_prompt: str, user_prompt: str, max_tokens: int) -> str:
|
| 79 |
model, tokenizer = _get_llm()
|
| 80 |
|
|
@@ -83,41 +73,34 @@ def _generate_response(system_prompt: str, user_prompt: str, max_tokens: int) ->
|
|
| 83 |
{"role": "user", "content": user_prompt},
|
| 84 |
]
|
| 85 |
|
| 86 |
-
# Format the prompt using the model's chat template
|
| 87 |
formatted_prompt = tokenizer.apply_chat_template(
|
| 88 |
messages, tokenize=False, add_generation_prompt=True
|
| 89 |
)
|
| 90 |
|
| 91 |
inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda")
|
| 92 |
|
| 93 |
-
# Generate with speed optimizations (KV Cache on, Greedy Decoding)
|
| 94 |
outputs = model.generate(
|
| 95 |
**inputs,
|
| 96 |
max_new_tokens=max_tokens,
|
| 97 |
-
use_cache=True,
|
| 98 |
-
do_sample=False,
|
| 99 |
-
pad_token_id=tokenizer.eos_token_id
|
| 100 |
)
|
| 101 |
|
| 102 |
-
# Decode and return just the generated response
|
| 103 |
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
|
| 104 |
return response.strip()
|
| 105 |
|
| 106 |
-
|
| 107 |
def _summarize(prompt: str) -> str:
|
| 108 |
return _generate_response(SUMMARY_SYSTEM_PROMPT, prompt, max_tokens=1024)
|
| 109 |
|
| 110 |
-
|
| 111 |
def _final_md(combined: str) -> str:
|
| 112 |
return _generate_response(FINAL_SYSTEM_PROMPT, combined, max_tokens=2048)
|
| 113 |
|
| 114 |
-
|
| 115 |
# ---------------------------------------------------------------------------
|
| 116 |
# Pipeline
|
| 117 |
# ---------------------------------------------------------------------------
|
| 118 |
|
| 119 |
-
#
|
| 120 |
-
@spaces.GPU(duration=300)
|
| 121 |
def process_repo(repo_url: str, token: str, progress: gr.Progress = gr.Progress()):
|
| 122 |
try:
|
| 123 |
progress(0, desc="Running CommitLens pipeline...")
|
|
@@ -146,7 +129,6 @@ def process_repo(repo_url: str, token: str, progress: gr.Progress = gr.Progress(
|
|
| 146 |
except Exception as e:
|
| 147 |
raise gr.Error(str(e))
|
| 148 |
|
| 149 |
-
|
| 150 |
# ---------------------------------------------------------------------------
|
| 151 |
# Gradio app
|
| 152 |
# ---------------------------------------------------------------------------
|
|
|
|
| 1 |
"""
|
| 2 |
+
CommitLens — Gradio UI (Dedicated T4 Medium Version)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
"""
|
| 4 |
|
| 5 |
from __future__ import annotations
|
| 6 |
|
| 7 |
import gradio as gr
|
|
|
|
| 8 |
import torch
|
| 9 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
| 10 |
|
|
|
|
| 39 |
_model = None
|
| 40 |
_tokenizer = None
|
| 41 |
|
|
|
|
| 42 |
def _get_llm():
|
| 43 |
global _model, _tokenizer
|
| 44 |
if _model is None:
|
| 45 |
+
# 8-bit quantization is mandatory to fit inside the T4's 16GB VRAM
|
| 46 |
quantization_config = BitsAndBytesConfig(
|
| 47 |
load_in_8bit=True,
|
| 48 |
)
|
| 49 |
|
| 50 |
_tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO_ID)
|
| 51 |
|
| 52 |
+
# CHANGED: torch_dtype set to float16 (T4 does not support bfloat16)
|
| 53 |
+
# CHANGED: Removed Flash Attention (T4 relies on native SDPA instead)
|
| 54 |
_model = AutoModelForCausalLM.from_pretrained(
|
| 55 |
MODEL_REPO_ID,
|
| 56 |
quantization_config=quantization_config,
|
| 57 |
device_map="auto",
|
| 58 |
+
torch_dtype=torch.float16,
|
|
|
|
| 59 |
)
|
| 60 |
return _model, _tokenizer
|
| 61 |
|
|
|
|
| 62 |
def _extract_filename(prompt: str) -> str:
|
| 63 |
for line in prompt.splitlines():
|
| 64 |
if line.startswith("Filename :"):
|
| 65 |
return line.split(":", 1)[1].strip()
|
| 66 |
return "unknown"
|
| 67 |
|
|
|
|
| 68 |
def _generate_response(system_prompt: str, user_prompt: str, max_tokens: int) -> str:
|
| 69 |
model, tokenizer = _get_llm()
|
| 70 |
|
|
|
|
| 73 |
{"role": "user", "content": user_prompt},
|
| 74 |
]
|
| 75 |
|
|
|
|
| 76 |
formatted_prompt = tokenizer.apply_chat_template(
|
| 77 |
messages, tokenize=False, add_generation_prompt=True
|
| 78 |
)
|
| 79 |
|
| 80 |
inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda")
|
| 81 |
|
|
|
|
| 82 |
outputs = model.generate(
|
| 83 |
**inputs,
|
| 84 |
max_new_tokens=max_tokens,
|
| 85 |
+
use_cache=True,
|
| 86 |
+
do_sample=False,
|
| 87 |
+
pad_token_id=tokenizer.eos_token_id
|
| 88 |
)
|
| 89 |
|
|
|
|
| 90 |
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
|
| 91 |
return response.strip()
|
| 92 |
|
|
|
|
| 93 |
def _summarize(prompt: str) -> str:
|
| 94 |
return _generate_response(SUMMARY_SYSTEM_PROMPT, prompt, max_tokens=1024)
|
| 95 |
|
|
|
|
| 96 |
def _final_md(combined: str) -> str:
|
| 97 |
return _generate_response(FINAL_SYSTEM_PROMPT, combined, max_tokens=2048)
|
| 98 |
|
|
|
|
| 99 |
# ---------------------------------------------------------------------------
|
| 100 |
# Pipeline
|
| 101 |
# ---------------------------------------------------------------------------
|
| 102 |
|
| 103 |
+
# CHANGED: Removed @spaces.GPU entirely. The T4 is dedicated and always active.
|
|
|
|
| 104 |
def process_repo(repo_url: str, token: str, progress: gr.Progress = gr.Progress()):
|
| 105 |
try:
|
| 106 |
progress(0, desc="Running CommitLens pipeline...")
|
|
|
|
| 129 |
except Exception as e:
|
| 130 |
raise gr.Error(str(e))
|
| 131 |
|
|
|
|
| 132 |
# ---------------------------------------------------------------------------
|
| 133 |
# Gradio app
|
| 134 |
# ---------------------------------------------------------------------------
|
requirements.txt
CHANGED
|
@@ -2,5 +2,4 @@ transformers
|
|
| 2 |
accelerate
|
| 3 |
bitsandbytes
|
| 4 |
torch
|
| 5 |
-
flash-attn
|
| 6 |
gradio
|
|
|
|
| 2 |
accelerate
|
| 3 |
bitsandbytes
|
| 4 |
torch
|
|
|
|
| 5 |
gradio
|