ShivanshiNigam/loglens-vllm-ready

LogLens AI — Custom Sliding Window Log Analyzer

A custom-architected causal LM with Sliding Window Attention, trained on real production log data.
Memory usage: 61 MB fixed regardless of log file size.
10× cheaper than GPT-4 for log analysis.

⚠️ PROOF OF CONCEPT (PoC) DISCLAIMER
This model proves the architecture and memory math of Sliding Window Attention. This custom 16M parameter model was trained on a small synthetic dataset of log files to prove the plumbing (RAG pipeline, vLLM deployment). It is NOT a general-purpose AI like ChatGPT. To get true natural language generation in production, this architectural concept must be injected into a large pre-trained model (like Llama-3 8B) rather than using our tiny custom-built weights.

---

What's in This Repo

File	What it does
model.safetensors	Trained weights (16M params)
config.json	Model architecture config
tokenizer.json + vocab	GPT-2 tokenizer
log_attention.py	Custom sliding window attention module
mini_deepseek_model.py	Full causal LM architecture (RMSNorm, SwiGLU, RoPE)
train_mini_deepseek.py	Training pipeline — run this to retrain on your own logs
save_and_upload.py	Verify saved model + upload to HuggingFace
ingest_logs.py	Embed log files into ChromaDB vector store
rag_query.py	Query logs via ChromaDB + this model (RAG pipeline)
requirements.txt	All Python dependencies

---

Quick Start

1. Run the live demo (no setup needed)

👉 https://huggingface.co/spaces/ShivanshiNigam/loglens-demo

2. Use the RAG pipeline locally

# Clone the GitHub repo
git clone https://github.com/shivanshinigam/DeepseekLogLens
cd DeepseekLogLens

# Install dependencies
pip install -r requirements.txt

# Generate sample logs
python3 generate_logs.py
python3 generate_3_log_files.py

# Step 1 — Embed your log files into the vector DB
python3 ingest_logs.py

# Step 2 — Ask a question (local keyword analysis, no GPU needed)
python3 rag_query.py --query "Why did the payments fail?"

# Step 3 — With a live vLLM endpoint (RunPod)
python3 rag_query.py --query "Why did the payments fail?" \
  --endpoint https://YOUR_POD_ID-8000.proxy.runpod.net/v1 \
  --model ShivanshiNigam/loglens-mini-deepseek

3. Retrain on your own log data

# Put your .log files in logs/
python3 train_mini_deepseek.py --epochs 5 --device cuda

# Save and upload
python3 save_and_upload.py --repo-id YOUR_USERNAME/loglens-mini-deepseek

---

Architecture

Input log tokens
      ↓
Embedding (vocab=50257, dim=256)
      ↓
4× DeepSeekDecoderLayer:
   ├── RMSNorm
   ├── LogWindowAttention (window=128, RoPE)
   ├── RMSNorm
   └── SwiGLU FFN (dim=256 → 1024 → 256)
      ↓
RMSNorm
      ↓
LM Head (tied to embedding weights)
      ↓
Logits (vocab=50257)

Memory: 61 MB · Parameters: 16M · Training loss: 10.8 → 2.9 (2 epochs, CPU)

---

Full RAG System Flow

Your Log Files (any size)
        ↓
ingest_logs.py  →  ChromaDB vector store
                         ↓  (semantic search)
User Query  →  rag_query.py  →  vLLM endpoint  →  Answer

---

Deploy on RunPod

See the full guide: docs/runpod_deployment.md

One-line vLLM command:

--model ShivanshiNigam/loglens-mini-deepseek --port 8000 --trust-remote-code

Cheapest GPU: RTX 4000 Ada @ $0.28/hr (20 GB VRAM — 20× more than needed)