How tencent/angelslim works
AngelSlim is a broad large-model compression toolkit. A typical workflow configures a model and compression strategy, prepares calibration or training data, runs compression, and exports a deployable model; the README also documents sparse attention, token compression, diffusion optimization, and speculative decoding. The graph preserves these product branches. Several internal paths are represented at subsystem-directory level because their implementations were not excerpted; those relationships are therefore omitted or marked unverified rather than inferred from names alone.
Main entry points: engine.py Compression engine · core Quantization core · compressor_factory.py Compressor factory · ptq.py Post-training quantization
Components 19
Workflow and models
- Compression engine
angelslim/engine.py - Config parsing
angelslim/utils/config_parser.py - Model adapters
angelslim/models - Compressor factory
angelslim/compressor/compressor_factory.py
Quantization and transforms
- Post-training quantization
angelslim/compressor/quant/ptq.py - Quantization core
angelslim/compressor/quant/core - Quantization methods
angelslim/compressor/quant/modules - Model transforms
angelslim/compressor/transform - Quantized model export
angelslim/compressor/quant/core/save.py
Inference optimization
- Sparse attention
angelslim/compressor/sparsity - Token compression
angelslim/compressor/token_compressor - Diffusion optimization
angelslim/compressor/diffusion - Speculative decoding
angelslim/compressor/speculative
Training and distillation
- Megatron QAD
angelslim/compressor/mcore_qad - Quant-aware training
angelslim/compressor/qat/qat.py - Distillation
angelslim/compressor/distill
Data and calibration
- Data loaders
angelslim/data/dataloader.py - vLLM calibration
angelslim/compressor/quant/core/vllm_calibrate_utils
Other components
- Engineer
Connections 18
- Engineer → Compression engine: configures and runs
README.md - Compression engine → Config parsing: loads configuration
angelslim/engine.py - Compression engine → Model adapters: loads model
angelslim/engine.py - Compression engine → Data loaders: creates data loader
angelslim/engine.py - Compression engine → Compressor factory: creates compressor
angelslim/engine.py - Compressor factory → Post-training quantization: offers quantization
- Post-training quantization → Quantization core: uses quantization core
- Quantization core → Quantization methods: provides method modules
angelslim/compressor/quant/modules/daq/scale_search.py - Quantization core → Quantized model export: saves quantized models
angelslim/compressor/quant/core/__init__.py - Engineer → Megatron QAD: runs distributed QAD
README.md - Engineer → Quant-aware training: trains quant-aware models
README.md - Engineer → Distillation: distills models
README.md - Engineer → Model transforms: applies model transforms
README.md - Engineer → Sparse attention: uses sparse attention
README.md - Engineer → Token compression: compresses tokens
- Engineer → Diffusion optimization: optimizes diffusion
README.md - Engineer → Speculative decoding: trains speculative decoding
README.md - Engineer → vLLM calibration: calibrates activations