{"id":37442756,"url":"https://github.com/lexsi-labs/dlbacktrace","last_synced_at":"2026-04-13T08:01:17.429Z","repository":{"id":264824177,"uuid":"890291943","full_name":"Lexsi-Labs/DLBacktrace","owner":"Lexsi-Labs","description":"DL Backtrace is a new explainablity technique for deep learning models that works for any modality and model type. ","archived":false,"fork":false,"pushed_at":"2026-01-13T11:17:46.000Z","size":30727,"stargazers_count":20,"open_issues_count":0,"forks_count":5,"subscribers_count":1,"default_branch":"dlb_v2","last_synced_at":"2026-01-16T20:12:08.863Z","etag":null,"topics":["deep-learning","explainability","xai"],"latest_commit_sha":null,"homepage":"https://dlbacktrace.lexsi.ai/","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"other","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/Lexsi-Labs.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE.md","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":"docs/support/changelog.md","governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2024-11-18T10:21:54.000Z","updated_at":"2026-01-13T11:17:40.000Z","dependencies_parsed_at":null,"dependency_job_id":"d305db75-331b-4854-b586-fb1138ba3a58","html_url":"https://github.com/Lexsi-Labs/DLBacktrace","commit_stats":null,"previous_names":["aryaxai/dlbacktrace","lexsi-labs/dlbacktrace"],"tags_count":8,"template":false,"template_full_name":null,"purl":"pkg:github/Lexsi-Labs/DLBacktrace","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lexsi-Labs%2FDLBacktrace","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lexsi-Labs%2FDLBacktrace/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lexsi-Labs%2FDLBacktrace/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lexsi-Labs%2FDLBacktrace/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/Lexsi-Labs","download_url":"https://codeload.github.com/Lexsi-Labs/DLBacktrace/tar.gz/refs/heads/dlb_v2","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lexsi-Labs%2FDLBacktrace/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":28565001,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-01-19T08:53:44.001Z","status":"ssl_error","status_checked_at":"2026-01-19T08:52:40.245Z","response_time":67,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.6:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["deep-learning","explainability","xai"],"created_at":"2026-01-16T06:46:15.733Z","updated_at":"2026-01-19T09:00:42.822Z","avatar_url":"https://github.com/Lexsi-Labs.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003cp align=\"center\"\u003e\n  \u003cimg src=\"https://raw.githubusercontent.com/Lexsi-Labs/DLBacktrace/dlb_v2/assets/images/dlb_logo.png\"\n       alt=\"DLB Logo\" width=\"300px\"/\u003e\n\u003c/p\u003e\n\n\n\u003ch2 align=\"center\"\u003eDLBacktrace: Model-Agnostic Explainability for Any Deep Learning Model\u003c/h2\u003e\n\n\u003cp align=\"center\"\u003e\n  \u003ca href=\"https://python.org\"\u003e\u003cimg src=\"https://img.shields.io/badge/Python-3.8%2B-blue.svg\"\u003e\u003c/a\u003e\n  \u003ca href=\"https://pytorch.org\"\u003e\u003cimg src=\"https://img.shields.io/badge/PyTorch-2.6%2B-red.svg\"\u003e\u003c/a\u003e\n  \u003ca href=\"https://dlbacktrace.lexsi.ai/\"\u003e\u003cimg src=\"https://img.shields.io/badge/docs-available-green.svg\"\u003e\u003c/a\u003e\n  \u003ca href=\"LICENSE.md\"\u003e\u003cimg src=\"https://img.shields.io/badge/license-Lexsi%20Labs%20Source%20Available-orange\"\u003e\u003c/a\u003e\n\u003c/p\u003e\n\n\u003cp align=\"center\"\u003e\n  \u003ci\u003eMaking AI Transparent and Explainable — Developed by \u003ca href=\"https://lexsi.ai/\"\u003eLexsi Labs\u003c/a\u003e\u003c/i\u003e 🚀\n\u003c/p\u003e\n\n---\n\n## Overview\n\nDLBacktrace is a model-agnostic explainability framework developed by *Lexsi Labs*. It provides comprehensive layerwise importance values (relevance scores) and model tracing capabilities across a wide range of model architectures — including Transformers, Large Language Models (LLMs), Mixture-of-Experts (MoEs), and more — as well as diverse task types such as Tabular, Vision, and Text. The framework is designed for robust and efficient execution on both CPU and GPU environments.\n\n## Key Features\n\n### Core Capabilities\n- **🔍 Deep Model Interpretability:** Gain comprehensive insights into your AI models using advanced relevance propagation algorithms\n- **🎯 Multi-Task Support:** Binary/Multi-class classification, object detection, segmentation, and text generation\n- **🏗️ Architecture Agnostic:** Support for Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Transformer, and custom architectures including Mixture-of-Experts (MoEs)\n- **⚡ High Performance:** Optimized execution engine with CUDA acceleration and deterministic tracing\n- **🔧 Robust Operations:** Full support for negative indexing and complex tensor operations\n- **📊 Comprehensive Tracing:** Layer-wise activation and relevance analysis with detailed execution tracking\n- **🛡️ Production Ready:** Deterministic execution environment with comprehensive error handling\n\n### Advanced Features\n- **🚀 High-Level Pipeline Interface:** Simplified API for text/image classification and generation with automatic model loading and configuration\n- **🎲 DLB Auto Sampler:** Advanced text generation with multiple decoding strategies — including greedy decoding, beam search (deterministic), and stochastic sampling methods such as temperature, top-k, and top-p — along with token-level relevance tracking\n- **🧠 Mixture-of-Experts (MoEs) Support:** Built-in support for MoEs architectures (JetMoE, OLMoE, Qwen3-MoE, GPT-Oss) with expert-level relevance analysis\n- **🌡️ Temperature Scaling:** Control generation diversity and model confidence with flexible temperature parameters\n- **🔄 Enhanced Execution Engine:** Critical fixes for RoBERTa, LLaMA, and other transformer models \n\n---\n\n## ⚡ Performance Improvements in DLB v2\n\nDLB v2 introduces **major architectural upgrades** to the explainability engine — resulting in *orders of magnitude faster performance* compared to v1.\n\n\u003e 📘 **Note:** All benchmarks below were conducted on the **LLaMA-3.2–3B model** using the MMLU dataset on an NVIDIA RTX 4090.\n\n---\n\n### 📊 Benchmark Summary\n\n| Metric | DLB v1 | DLB v2 | Improvement |\n|:-------|:-------|:-------|:-------------|\n| ⏱️ **Explainability Time** | 250–30,000 s (grows exponentially) | 🕒 12–18 s (nearly constant) | 🔥 **20× → 1400× faster** depending on sequence length |\n| 🚀 **Throughput** | ~0.01–0.03 tokens/s | 🧩 3–75 tokens/s | ⚡ **100× → 2500× higher** |\n| 📈 **Scalability** | Time increases ~8× every doubling | Grows slowly with input size | ✅ **Stable \u0026 linear-like scaling** |\n\n---\n\n### 📈 Performance Graphs\n\n| Metric | Comparison Plot |\n|:--------|:----------------|\n| 🕒 **Total Time vs Sequence Length** | \u003cimg src=\"https://raw.githubusercontent.com/Lexsi-Labs/DLBacktrace/dlb_v2/assets/images/total_dlb_time_comparison.png\" width=\"500\"/\u003e |\n| 🚀 **Token Throughput** | \u003cimg src=\"https://raw.githubusercontent.com/Lexsi-Labs/DLBacktrace/dlb_v2/assets/images/throughput_comparison.png\" width=\"500\"/\u003e |\n| ⚙️ **Speedup (v2 / v1)** | \u003cimg src=\"https://raw.githubusercontent.com/Lexsi-Labs/DLBacktrace/dlb_v2/assets/images/speedup_comparison.png\" width=\"500\"/\u003e |\n\n---\n\n\u003e 💡 **DLB v2** provides **consistent low latency** and achieves **20×–1400× speedup** over DLB v1 as sequence length increases. \n\n## Installation\n\n### From Source (Recommended)\n\n```bash\npip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu126\npip install dl-backtrace\n```\n\n### Requirements\n\n- Python 3.8+\n- PyTorch 2.6+ (with CUDA 12.6 support recommended)\n- Additional dependencies: transformers, matplotlib, seaborn, graphviz, joblib, zstandard\n\nSee `requirements.txt` for the complete list of dependencies.\n\n### Hugging Face Setup\n\nFor accessing models from Hugging Face Hub (required for BERT, RoBERTa, LLaMA, etc.):\n\n```bash\n# Install Hugging Face CLI\npip install huggingface_hub\n\n# Login to Hugging Face (required for gated models)\nhuggingface-cli login\n```\n\nYou'll need a Hugging Face account and access token. Get your token from [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens).\n\n## Quick Start\n\n### PyTorch Models (Recommended)\n\n```python\nimport torch\nimport torch.nn as nn\nfrom dl_backtrace.pytorch_backtrace import DLBacktrace\n\n# Define your model\nclass MyModel(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.linear = nn.Linear(10, 1)\n    \n    def forward(self, x):\n        return self.linear(x)\n\n# Initialize model and DLBacktrace\nmodel = MyModel()\nx = torch.randn(1, 10)  # Example input\n\n# Create DLBacktrace instance\ndlb = DLBacktrace(\n    model=model,\n    input_for_graph=(x,),\n    device='cuda',    # 'cpu',\n    verbose=False\n)\n\n# Get layer-wise outputs\nnode_io = dlb.predict(x)\n\n# Calculate relevance propagation\nrelevance = dlb.evaluation(\n    mode=\"default\",\n    multiplier=100.0,\n    task=\"binary-classification\"\n)\n```\n\n## Advanced Features\n\n### 🚀 High-Level Pipeline Interface\nSimplified API for common ML tasks with automatic model loading and configuration:\n\n```python\nfrom dl_backtrace.pytorch_backtrace import DLBacktrace\nfrom transformers import AutoTokenizer, AutoModelForSequenceClassification\nimport torch\n\n# Load model and tokenizer\nmodel = AutoModelForSequenceClassification.from_pretrained(\"textattack/bert-base-uncased-SST-2\")\ntokenizer = AutoTokenizer.from_pretrained(\"textattack/bert-base-uncased-SST-2\")\n\n# Prepare input\ntext = \"This movie is fantastic!\"\ntokens = tokenizer(text, return_tensors=\"pt\")\ninput_ids = tokens[\"input_ids\"]\nattention_mask = tokens[\"attention_mask\"]\n\n# Initialize DLBacktrace\ndlb = DLBacktrace(\n    model,\n    (input_ids, attention_mask),\n    device='cuda',  # or 'cpu'\n    verbose=False\n)\n\n# Run text classification with run_task() - ONE CALL!\nresults = dlb.run_task(\n    task=\"text-classification\",  # or \"auto\" for automatic detection\n    inputs={'input_ids': input_ids, 'attention_mask': attention_mask},\n    debug=False\n)\n\n# Access predictions and relevance\npredictions = results['predictions']\nrelevance = results['relevance']\nprint(f\"Predicted class: {predictions.argmax(axis=-1)}\")\nprint(f\"Token relevance shape: {relevance['input_ids'].shape}\")\n```\n\n### 🎲 DLB Auto Sampler\nAdvanced text generation with multiple sampling strategies and token-level relevance tracking:\n\n```python\nfrom dl_backtrace.pytorch_backtrace.dlbacktrace.core.dlb_auto_sampler import DLBAutoSampler\n\nsampler = DLBAutoSampler(model, tokenizer)\n\n# Greedy sampling\noutput = sampler.generate(\"Prompt\", strategy=\"greedy\", max_length=50)\n\n# Top-k and Top-p sampling\noutput = sampler.generate(\"Prompt\", strategy=\"top_k\", top_k=50, temperature=0.8)\noutput = sampler.generate(\"Prompt\", strategy=\"top_p\", top_p=0.9, temperature=0.8)\n\n# Beam search\noutput = sampler.generate(\"Prompt\", strategy=\"beam_search\", num_beams=5)\n\n# Access token-level relevance\nprint(output['relevance_scores'])\n```\n\n### 🧠 Mixture-of-Experts (MoEs) Support\nBuilt-in support for MoE architectures with expert-level relevance analysis:\n\n```python\nfrom dl_backtrace.moe_pytorch_backtrace import Backtrace\nfrom transformers import AutoTokenizer, AutoModelForCausalLM\n\n# Load MoE model (GPT-OSS, JetMoE, OLMoE, Qwen3-MoE)\nmodel = AutoModelForCausalLM.from_pretrained(\"openai/gpt-oss-20b\")\ntokenizer = AutoTokenizer.from_pretrained(\"openai/gpt-oss-20b\")\n\n# Initialize MoE Backtrace\nbt = Backtrace(\n    model=model,\n    model_type='gpt_oss',  # or 'jetmoe', 'olmoe', 'qwen'\n    device='cpu'  # or 'cuda'\n)\n\n# Prepare input\nprompt = \"What is the capital of France?\"\ntokens = tokenizer(prompt, return_tensors=\"pt\")\ninput_ids = tokens[\"input_ids\"]\nattention_mask = tokens[\"attention_mask\"]\n\n# Run generation with run_task() - ONE CALL!\nresults = bt.run_task(\n    task=\"generation\",\n    inputs={'input_ids': input_ids, 'attention_mask': attention_mask},\n    tokenizer=tokenizer,\n    max_new_tokens=10,\n    return_relevance=True,\n    return_scores=True,\n    debug=False\n)\n\n# Access generated text and relevance\ngenerated_text = tokenizer.decode(results['generated_ids'][0], skip_special_tokens=True)\nprint(f\"Generated: {generated_text}\")\n\n# Get expert-level relevance\nexpert_relevance = bt.all_layer_expert_relevance\nprint(f\"Expert routing across {len(expert_relevance)} layers\")\n```\n\n### 🌡️ Temperature Scaling\nControl generation diversity and model confidence:\n\n```python\nfrom dl_backtrace.pytorch_backtrace.dlbacktrace import DLBacktrace\n\ndlb = DLBacktrace(model, input_for_graph=(input_tensor,))\n\n# Apply temperature scaling for generation\noutput = dlb.generate_with_temperature(\n    input_ids,\n    temperature=0.7,  # Lower = more focused, Higher = more diverse\n    max_length=100\n)\n```\n\n### ⚡ Execution Engines\nDLBacktrace provides optimized execution engines:\n\n#### ExecutionEngineNoCache \n- **Memory-efficient**: Runs entirely in RAM for faster execution\n- **Enhanced Operations**: Supports 100+ PyTorch operations with robust error handling\n- **Recent Improvements**: Critical fixes for transformer models (RoBERTa, LLaMA, BERT)\n\n### 🛡️ Deterministic Execution Environment\nDLBacktrace automatically sets up a deterministic environment for consistent results:\n- ✅ CUDA memory management and synchronization\n- ✅ Deterministic algorithms and cuDNN settings\n- ✅ Random seed control and environment variables\n- ✅ Warning suppression for cleaner output\n\n### 🔧 Robust Tensor Operations\nFull support for PyTorch's negative indexing and complex operations:\n- ✅ `transpose(-1, -2)`, `permute([-1, -2, 0])`\n- ✅ `unsqueeze(-1)`, `squeeze(-1)`\n- ✅ `slice(dim=-1, ...)`, `cat(tensors, dim=-1)`\n- ✅ `index_select(dim=-1, ...)`\n\n### Evaluation Parameters\n\n| Parameter    | Description | Values |\n|--------------|-------------|--------|\n| `mode`       | Evaluation algorithm mode | `default`, `contrastive` |\n| `multiplier` | Starting relevance at output layer | Float (default: 100.0) |\n| `scaler`     | Relevance scaling factor | Float (default: 1.0) |\n| `thresholding` | Pixel selection threshold for segmentation | Float (default: 0.5) |\n| `task`       | Model task type | `binary-classification`, `multi-class classification` |\n\n## Example Notebooks\n\n| Name        | Task        | Colab Link                          | HTML Link                                       |\n|-------------|-------------|-------------------------------| -------------------------------|\n| Custom Tabular Model | Binary Classification | [Colab Link](https://colab.research.google.com/drive/13N2sfAxA_7GJsZ7VAKS5WIOg6GusQghI?usp=sharing)| [HTML Version](https://drive.google.com/file/d/1_5DZWIe0Q5eB4CtVhkbO4zTCRbe6yW5D/view?usp=sharing) |\n| VGG Model | Multi-Class Classification | [Colab Link](https://colab.research.google.com/drive/1LxSazvy1Y2i0Ho1Qw7K905aNmvh66GzR?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1AO7cBJ9o6-Xtaqnall8HWG9ZccwtY-QZ/view?usp=sharing) |\n| ResNet Model | Multi-Class Classification | [Colab Link](https://colab.research.google.com/drive/1O8Is0X-IrKxXzgJeR1Xxy21OOR-UGfm7?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1tU573KFG2HPwM0bhFnGSueSc1uYm1ZCb/view?usp=sharing) |\n| ViT Model | Multi-Class Classification | [Colab Link](https://colab.research.google.com/drive/1B1xN5w51-tRnHyycbHV6lxi0tKxONnoS?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1E59Jlsv-v65V4qRL1q0xb0f5gmIsgPeg/view?usp=sharing) |\n| DenseNet Model | Multi-Class Classification | [Colab Link](https://colab.research.google.com/drive/1TspNRXd-qf81iZepUpU_TrUokZynynbu?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1B-6YwaA97ZVk693bQVukRZjVmZeaJt5Z/view?usp=sharing) |\n| EfficientNet Model | Multi-Class Classification | [Colab Link](https://colab.research.google.com/drive/1Xv9ghxp0OXcfYOseoq37KiJbz-UkLvWq?usp=sharing) | [HTML Version](https://drive.google.com/file/d/19sl0_U9VZbFooX21osmKj35lwHNI6_uz/view?usp=sharing) |\n| MobileNet Model | Multi-Class Classification | [Colab Link](https://colab.research.google.com/drive/12D6NgYTf4Gud3BXLRfnUK1m5-RLZY6dv?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1Bi01z-_9Gejy7_eA7MsA6-fk9JTRfOQs/view?usp=sharing) |\n| BERT-Base Model | Sentiment Classification | [Colab Link](https://colab.research.google.com/drive/1eBAQ8TToJUs6EN4-md08cjfZUmAArX0s?usp=drive_link) | [HTML Version](https://drive.google.com/file/d/11YmWQ4KtztAk8TEIwFU3kajtM-VLc2O_/view?usp=sharing) |\n| LLaMA-3.2-1B Model | Text Generation | [Colab Link](https://colab.research.google.com/drive/1eZUK-PwI6iZeKnfcQHxIXO7xo8wYDEek?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1QJMWMcxD5ifS8ygoySXgFo9OxhXZTgVY/view?usp=sharing) |\n| Qwen-3-0.6B Model | Text Generation | [Colab Link](https://colab.research.google.com/drive/1nJsxk7HPhiZ1DUIXn0eOnu0yjye-WOQT?usp=drive_link) | [HTML Version](https://drive.google.com/file/d/1yBqe34IExmfD-YYZF_U3wuOK8JiEBDM2/view?usp=sharing) |\n| JetMoE | Text Generation | [Colab Link](https://colab.research.google.com/drive/1MKxR60Mf1F_TNMuE31T3GRWHFPTm5IhP?usp=drive_link) | [HTML Version](https://drive.google.com/file/d/1z39LZVa88YCFKgQPTXKgZIVoav7t1PBH/view?usp=sharing) |\n| OLMoE | Text Generation | [Colab Link](https://colab.research.google.com/drive/1e13QAFw4A8jKhZVb1jh-C_16A2u1vZIH?usp=drive_link) | [HTML Version](https://drive.google.com/file/d/1cHByb9EcR-7k2BAL-HmbprNS5oiL2k8q/view?usp=sharing) |\n| Qwen-3-MoE | Text Generation | [Colab Link](https://colab.research.google.com/drive/1ja26JGNt22rOMT0HCyjOMASXxfnMpkm8?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1Upk1l-SnDGT_Mw3Bm6mNmdAVWWYpyW-S/view?usp=sharing) |\n| GPT-Oss | Text Generation | [Colab Link](https://colab.research.google.com/drive/1zZXTOV4NOrauNDXjjYB8rausbBGoLYtX?usp=sharing) | [HTML Version](https://drive.google.com/file/d/1zs_XyjqmKiT-W3YEqFFTqEKL7z0N_bXq/view?usp=sharing) |\n\n\nFor more detailed examples and use cases, check out our documentation.\n\n## Supported Layers\n\n### PyTorch\n\n**Core Operations:**\n- [x] **Linear (Fully Connected) Layer**\n- [x] **Convolutional Layer** (Conv2D)\n- [x] **Reshape \u0026 Flatten Layers**\n- [x] **Pooling Layers** (AdaptiveAvgPool2d, MaxPool2d, AvgPool2d, AdaptiveMaxPool2d)\n- [x] **1D Pooling Layers** (AvgPool1d, MaxPool1d, AdaptiveAvgPool1d, AdaptiveMaxPool1d)\n- [x] **Concatenate \u0026 Add Layers**\n- [x] **LSTM Layer**\n- [x] **Dropout Layer**\n- [x] **Embedding Layer**\n\n**Advanced Operations:**\n- [x] **Tensor Manipulation** (transpose, permute, unsqueeze, squeeze, slice, cat, index_select)\n- [x] **Negative Indexing Support** (all operations support PyTorch's negative indexing)\n- [x] **Layer Normalization**\n- [x] **Batch Normalization**\n- [x] **View \u0026 Reshape Operations**\n\n## Testing \u0026 Validation\n\n### Supported Models\nDLBacktrace has been extensively tested with:\n- **Vision Models**: ResNet, VGG, DenseNet, EfficientNet, MobileNet, ViT\n- **NLP Models**: BERT, ALBERT, RoBERTa, DistilBERT, ELECTRA, XLNet, LLaMA-3.2, Qwen\n- **MoE Models**: JetMoE, OLMoE (Open Language Model with Experts), Qwen3-MoE, GPT-OSS\n- **Tasks**: Classification, Object Detection, Segmentation, Text Generation, Expert-Level Analysis\n\n## Getting Started\n\nIf you're new to DLBacktrace:\n\n1. **📖 Read the Documentation**: [https://dlbacktrace.lexsi.ai/](https://dlbacktrace.lexsi.ai/)\n2. **🚀 Try the Quick Start**: See examples above for PyTorch models\n3. **💻 Explore Notebooks**: Check out our comprehensive example notebooks for various use cases\n4. **🧪 Run Tests**: Validate your installation with the benchmark scripts\n\nFor advanced features like the Pipeline Interface, Auto Sampler, MoE models, and Temperature Scaling, refer to the full documentation.\n\n## Contributing\n\nWe welcome contributions from the community! Please follow our contribution guidelines and submit pull requests for any improvements.\n\n## License\n\nThis project is licensed under a custom License - see the [LICENSE](LICENSE.md) file for details.\n\n## Recent Updates \u0026 New Features\n\n### Latest Release (2025)\n\n**New Features:**\n- 🚀 **High-Level Pipeline Interface**: Simplified API for text/image classification and generation\n- 🎲 **DLB Auto Sampler**: Advanced text generation with multiple sampling strategies\n- 🧠 **MoE Model Support**: Built-in support for Mixture of Experts architectures (JetMoE, OLMoE, Qwen3-MoE, GPT-OSS)\n- 🌡️ **Temperature Scaling**: Flexible control over generation diversity and model confidence\n\n**Critical Fixes \u0026 Improvements:**\n- 🔧 Enhanced execution engine with robust handling of complex tensor operations\n- ⚡ Deterministic environment setup for consistent, reproducible results\n- 🛡️ Comprehensive error handling for production use\n- 🚨 Critical fixes for transformer models (RoBERTa, LLaMA, BERT)\n- 🧠 Smart attention detection for bidirectional vs causal attention\n- 💾 Memory optimization and improved OOM error handling\n\n## Contact\n\nFor any inquiries, support, or collaboration opportunities:\n\n- **Email**: [support@lexsi.ai](mailto:support@lexsi.ai)\n- **Website**: [https://lexsi.ai/](https://lexsi.ai/)\n- **GitHub Issues**: [https://github.com/Lexsi-Labs/DLBacktrace/issues](https://github.com/Lexsi-Labs/DLBacktrace/issues)\n- **Documentation**: [https://dlbacktrace.lexsi.ai/](https://dlbacktrace.lexsi.ai/)\n\n## Citations\n```\n@misc{sankarapu2024dlbacktracemodelagnosticexplainability,\n      title={DLBacktrace: A Model Agnostic Explainability for any Deep Learning Models}, \n      author={Vinay Kumar Sankarapu and Chintan Chitroda and Yashwardhan Rathore and Neeraj Kumar Singh and Pratinav Seth},\n      year={2024},\n      eprint={2411.12643},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2411.12643}, \n}\n```\n---\n\n\u003cp align=\"center\"\u003e \u003cb\u003eDLBacktrace\u003c/b\u003e — Bridging Performance and Explainability 🔍\u003cbr\u003e \u003ca href=\"https://lexsi.ai/\"\u003eLexsi Labs\u003c/a\u003e | \u003ca href=\"mailto:support@lexsi.ai\"\u003esupport@lexsi.ai\u003c/a\u003e \u003c/p\u003e \n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flexsi-labs%2Fdlbacktrace","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Flexsi-labs%2Fdlbacktrace","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flexsi-labs%2Fdlbacktrace/lists"}