Wave 143: Router 6051L — reinit after corrupt objects
This commit is contained in:
Yanis Mahboub
99
blade-auto-colab.py
Executable file
99
blade-auto-colab.py
Executable file
@@ -0,0 +1,99 @@
|
||||
import asyncio, time
|
||||
from playwright.async_api import async_playwright
|
||||
|
||||
CELLS = [
|
||||
"!pip install -q unsloth datasets trl peft accelerate bitsandbytes",
|
||||
"!wget -q 'https://huggingface.co/datasets/yace222/weval-brain-dataset/resolve/main/train_chatml.jsonl' -O /content/dataset.jsonl\n!wc -l /content/dataset.jsonl",
|
||||
"from unsloth import FastLanguageModel\nimport torch\nmodel, tokenizer = FastLanguageModel.from_pretrained(model_name='unsloth/Qwen2.5-7B-Instruct-bnb-4bit', max_seq_length=2048, load_in_4bit=True)\nmodel = FastLanguageModel.get_peft_model(model, r=16, target_modules=['q_proj','k_proj','v_proj','o_proj','gate_proj','up_proj','down_proj'], lora_alpha=16, lora_dropout=0, bias='none', use_gradient_checkpointing='unsloth')\nprint('Model loaded!')",
|
||||
"from datasets import load_dataset\ndataset = load_dataset('json', data_files='/content/dataset.jsonl', split='train')\ndef format_chat(example):\n text = tokenizer.apply_chat_template(example['messages'], tokenize=False, add_generation_prompt=False)\n return {'text': text}\ndataset = dataset.map(format_chat)\nprint(f'Formatted: {len(dataset)} examples')",
|
||||
"from trl import SFTTrainer\nfrom transformers import TrainingArguments\ntrainer = SFTTrainer(model=model, tokenizer=tokenizer, train_dataset=dataset, dataset_text_field='text', max_seq_length=2048, args=TrainingArguments(per_device_train_batch_size=2, gradient_accumulation_steps=4, warmup_steps=5, num_train_epochs=3, learning_rate=2e-4, fp16=not torch.cuda.is_bf16_supported(), bf16=torch.cuda.is_bf16_supported(), logging_steps=1, output_dir='outputs', optim='adamw_8bit', seed=42))\nstats = trainer.train()\nprint(f'Done! Loss: {stats.training_loss:.4f}')",
|
||||
"model.save_pretrained_gguf('weval-brain-gguf', tokenizer, quantization_method='q4_k_m')\nprint('GGUF Q4_K_M saved!')\nimport os\nfor f in os.listdir('weval-brain-gguf'):\n print(f'{f}: {os.path.getsize(os.path.join(\"weval-brain-gguf\",f))//1024//1024}MB')",
|
||||
]
|
||||
|
||||
async def run():
|
||||
async with async_playwright() as p:
|
||||
ctx = await p.chromium.launch_persistent_context(
|
||||
"/tmp/chrome-weval", headless=True,
|
||||
args=["--no-sandbox","--disable-blink-features=AutomationControlled"],
|
||||
viewport={"width":1920,"height":1080},
|
||||
)
|
||||
page = ctx.pages[0] if ctx.pages else await ctx.new_page()
|
||||
await page.add_init_script("Object.defineProperty(navigator,'webdriver',{get:()=>undefined})")
|
||||
|
||||
# Open blank Colab
|
||||
print(f"[{time.strftime('%H:%M:%S')}] Opening Colab...")
|
||||
await page.goto("https://colab.research.google.com/#create=true", timeout=30000)
|
||||
await page.wait_for_timeout(10000)
|
||||
|
||||
# Try setting GPU via menu
|
||||
print(f"[{time.strftime('%H:%M:%S')}] Setting GPU runtime...")
|
||||
# Use Command Palette: Ctrl+Shift+P
|
||||
await page.keyboard.press("Control+Shift+p")
|
||||
await page.wait_for_timeout(2000)
|
||||
await page.keyboard.type("change runtime", delay=80)
|
||||
await page.wait_for_timeout(2000)
|
||||
await page.keyboard.press("Enter")
|
||||
await page.wait_for_timeout(3000)
|
||||
await page.screenshot(path="/tmp/colab-gpu-dialog.png")
|
||||
|
||||
# Try selecting T4 GPU
|
||||
try:
|
||||
# Look for GPU option in dialog
|
||||
t4_option = page.locator("text=T4")
|
||||
if await t4_option.count() > 0:
|
||||
await t4_option.first.click()
|
||||
await page.wait_for_timeout(1000)
|
||||
print("T4 GPU selected!")
|
||||
|
||||
# Click Save
|
||||
save_btn = page.locator("text=Save, button:has-text('Save')")
|
||||
if await save_btn.count() > 0:
|
||||
await save_btn.first.click()
|
||||
await page.wait_for_timeout(2000)
|
||||
print("GPU saved!")
|
||||
except Exception as e:
|
||||
print(f"GPU setup: {e}")
|
||||
|
||||
await page.keyboard.press("Escape")
|
||||
await page.wait_for_timeout(1000)
|
||||
|
||||
# Paste all cells
|
||||
print(f"[{time.strftime('%H:%M:%S')}] Pasting {len(CELLS)} cells...")
|
||||
|
||||
# Click on first cell
|
||||
cells = page.locator(".cell, .codecell, [class*='cell']")
|
||||
first_cell = page.locator("textarea, .CodeMirror, .monaco-editor, [contenteditable]").first
|
||||
|
||||
for i, code in enumerate(CELLS):
|
||||
if i > 0:
|
||||
# Add new cell: Ctrl+M then B
|
||||
await page.keyboard.press("Control+m")
|
||||
await page.wait_for_timeout(300)
|
||||
await page.keyboard.press("b")
|
||||
await page.wait_for_timeout(1000)
|
||||
|
||||
# Type code into cell
|
||||
await page.keyboard.type(code, delay=2)
|
||||
await page.wait_for_timeout(300)
|
||||
print(f" Cell {i+1}/{len(CELLS)} pasted ({len(code)} chars)")
|
||||
|
||||
await page.screenshot(path="/tmp/colab-all-cells.png")
|
||||
|
||||
# RUN ALL: Ctrl+F9
|
||||
print(f"[{time.strftime('%H:%M:%S')}] Running all cells (Ctrl+F9)...")
|
||||
await page.keyboard.press("Control+F9")
|
||||
await page.wait_for_timeout(10000)
|
||||
await page.screenshot(path="/tmp/colab-executing.png")
|
||||
|
||||
print(f"[{time.strftime('%H:%M:%S')}] TRAINING LAUNCHED!")
|
||||
print("Cells are executing. Training will take ~20 min on T4 GPU.")
|
||||
print("Check: /tmp/colab-executing.png")
|
||||
|
||||
# Copy screenshots
|
||||
import subprocess
|
||||
for f in ['colab-gpu-dialog.png','colab-all-cells.png','colab-executing.png']:
|
||||
subprocess.run(['cp', f'/tmp/{f}', f'/var/www/html/api/notebooks/{f}'], capture_output=True)
|
||||
|
||||
await ctx.close()
|
||||
|
||||
asyncio.run(run())
|
||||
Reference in New Issue
Block a user