-
Notifications
You must be signed in to change notification settings - Fork 0
gpu acceleration
Syed Ibrahim Omer edited this page Apr 12, 2026
·
5 revisions
Overview of GPU-accelerated components throughout the pipeline.
| Component | Library | GPU Benefit |
|---|---|---|
| Random Forest training | cuML | 10–100× vs sklearn on large datasets |
| LSTM / Transformer training | PyTorch CUDA | Standard deep learning acceleration |
| Article embeddings | TensorRT INT8 (Jina v5) | ~4× vs FP32 inference |
| Cluster labeling | TensorRT INT8 (Flan-T5) | ~4× vs FP32 inference |
| Event labeling | TensorRT INT8 (LED) | ~4× vs FP32 inference |
| HDBSCAN clustering | cuML | GPU-accelerated density clustering |
| Ensemble inference | PyTorch CUDA | Parallel 3-model forward pass |
All three NLP engines use INT8 quantization:
- 4× memory reduction vs FP32 (32 bits → 8 bits per weight)
- ~4× throughput increase via reduced memory bandwidth
- Calibrated using representative datasets (not post-training quantization)
- See TensorRT Engines for engine-specific details
-
cuml.ensemble.RandomForestClassifierfor random forest (replaces sklearn) -
cuml.cluster.HDBSCANfor news article clustering (replaces hdbscan) - Both require NVIDIA GPU with CUDA toolkit
| Resource | Minimum | Recommended |
|---|---|---|
| GPU VRAM | 8 GB | 16+ GB |
| CUDA version | 11.8 | 12.x |
| TensorRT | 8.x | Latest |
| System RAM | 16 GB | 32+ GB |
scripts/profile_trt_engines.py provides benchmarking across batch sizes, measuring:
- Throughput (tokens/sec, embeddings/sec)
- Latency (ms per call)
- Peak GPU memory (MB)
- GPU utilization (%)
- TensorRT Engines — Engine implementation details
- Random Forest — cuML random forest
- Jina v5 Embeddings — TRT embedding engine
- Flan-T5 Summarization — TRT labeling engine
- Event Clustering — GPU HDBSCAN
- Reproducibility — Environment setup
- Project Overview — Goals, research questions, methodology, and team
- Glossary — Key terms used throughout this wiki
Raw inputs that feed the prediction system.
| Page | Description |
|---|---|
| Visa Data | US Department of State visa issuance statistics (108 monthly PDFs) |
| Encounter Data | CBP Southwest border encounter statistics (FY2019–2026) |
| Google News | 170K+ news articles across 15 countries × 8 topics |
| Google Trends | Monthly search-interest time series (15 countries × 8 keywords) |
| Exchange Rates | IMF Real Effective Exchange Rate for 6 countries |
The end-to-end flow from raw data to production forecasts.
| Page | Description |
|---|---|
| Data Collection | Ingestion layer: async scraping, bounded concurrency, retry logic |
| Data Processing | PDF parsing, JSON→Parquet, encounter merging |
| NLP Enrichment | Embedding → Clustering → Labeling → Sentiment |
| Panel Construction | Feature engineering: 18 lag features, 6 lead targets |
| Training Pipeline | Out-of-time train/test split, 4 architectures |
| Inference Pipeline | Horizon-aware ensemble, production prediction flow |
Machine learning architectures and their roles in the ensemble.
| Page | Description |
|---|---|
| Random Forest | cuML GPU Random Forest — best at short horizons (Lead 1–2) |
| LSTM | MigrationLSTM — country-aware with SurgeJointLoss |
| Transformer | MigrationTransformer — best at long horizons (Lead 5–6) |
| Horizon-Aware Ensemble | Dynamic weighting: RF→short, Transformer→long |
| SurgeJointLoss | Dual-objective loss: Huber + BCE for crisis detection |
| Jina v5 Embeddings | TensorRT INT8 news article embeddings (768-dim) |
| Flan-T5 Summarization | TensorRT INT8 cluster labeling engine |
Statistical techniques driving the lead-lag and surge analysis.
| Page | Description |
|---|---|
| Lead-Lag Analysis | Pearson correlation at 0–6 month offsets |
| Surge Detection | Quantile-based and σ-threshold spike identification |
| Sentiment Analysis | Rule-based lexicon scoring for migration-relevant news |
| Event Clustering | HDBSCAN GPU clustering + LED label generation |
| Cross-Correlation Analysis | CCF analysis, VAR benchmarking, ADF stationarity tests |
| Multiple Comparison Correction | Benjamini-Hochberg FDR for 58 significant signals |
What the system discovered about migration predictability.
| Page | Description |
|---|---|
| Event-Visa Findings | News events as leading indicators (r=0.617 at 3-month lag) |
| Exchange Rate Findings | Exchange rate signals (DR r=0.498 at 2-month lag) |
| Model Performance | Ensemble results: F1=0.96 at Lead 1, F1=0.86 at Lead 6 |
Reference documentation for every src/ subpackage and key files.
| Page | Description |
|---|---|
| Main Entry Point |
src/main.py CLI: bootstrap, collect-live, sync-data |
| Collection Module |
src/collection/* — visa, encounter, news, trends, HF sync |
| Processing Module |
src/processing/* — parse, merge, build_panel, summarize |
| Analysis Module |
src/analysis/* — events, exchange_rate, trends_analysis, plots |
| Models Module |
src/models/* — surge_model, train_and_evaluate, inference |
| News Scraper | Deep dive: batch decoding, checkpoint recovery, throttling |
| PDF Parser | Deep dive: PyMuPDF table extraction, VISA_MAP normalization |
| TensorRT Engines | Deep dive: Jina-v5, Flan-T5, LED TensorRT engines |
| Build Panel Detail | Deep dive: lag/lead construction, forward-fill strategies |
| HF Sync | Deep dive: bidirectional Hugging Face Hub sync |
Compute, reproducibility, and operational details.
| Page | Description |
|---|---|
| GPU Acceleration | TensorRT INT8, cuML, CUDA streams, NVML profiling |
| Reproducibility | HF bootstrap, run.sh pipeline, dependency checking |