Docs/Kaggle_Deployment_Guide.md

🏆 Kaggle & Cloud GPU Deployment Guide

This guide outlines how to upload, configure, and execute the Efficient Attention Research Framework on Kaggle GPUs (P100 / T4 / L4 / A100), Google Colab, or any remote Linux GPU cluster.

---

📁 Repository Zip & Upload Options

Option A: Direct Git Clone (Recommended if public repo)

In a Kaggle Notebook cell:

!git clone https://github.com/your-username/SVD_linear_Attention.git
%cd SVD_linear_Attention/Codes/framework
!python run_kaggle.py list

Option B: Kaggle Dataset Upload

1. Zip the Codes/framework folder:

   cd Codes/framework
   zip -r framework_code.zip .
   

2. Upload framework_code.zip as a Kaggle Dataset named framework-code.
3. In your Kaggle Notebook:

   !cp -r /kaggle/input/framework-code /kaggle/working/framework
   %cd /kaggle/working/framework
   !python run_kaggle.py train --attn_type full --task sst2 --epochs 3
   

---

⚡ Execution Commands on Kaggle

All commands run via python run_kaggle.py (which auto-configures Kaggle paths and installs missing packages):

1. List All Components

!python run_kaggle.py list

2. Fine-Tune Attention Model

!python run_kaggle.py train \
    --attn_type full \
    --task sst2 \
    --epochs 3 \
    --batch_size 32 \
    --lr 2e-5 \
    --exp_name distilbert_sst2_baseline

3. Apply SVD Model Compression

!python run_kaggle.py compress \
    --compressor aasvd \
    --task sst2 \
    --compression_rank 64 \
    --exp_name distilbert_aasvd_compressed

4. Evaluate Checkpoint

!python run_kaggle.py eval \
    --attn_type full \
    --task sst2 \
    --model_path /kaggle/working/results/distilbert_aasvd_compressed/model \
    --exp_name eval_aasvd_compressed

5. Multi-Seed Grid Sweep

!python run_kaggle.py sweep \
    --attn_types full nha lsh \
    --seeds 42 123 456 \
    --task sst2

---

📊 Kaggle Output Artifacts

Results are saved to /kaggle/working/results/<exp_name>/:

• config.json: Hardware, hyperparameters, and model config
• metrics/: Final evaluation metrics (eval_final.json, compression.json)
• history/: Full training loss history (log_history.json, log_history.csv)
• plots/: loss_curve.png & metric_curve.png
• model/: HuggingFace saved model weights & tokenizer
• SUMMARY.md: Formatted Markdown report