How ml-gsai/refusion works
ReFusion is a masked diffusion language model with training, parallel autoregressive-style inference, and benchmark evaluation. The inference script loads the Qwen3 model and performs block/slot generation; the evaluation entry point invokes lm-evaluation-harness, whose task definitions produce requests for registered model backends and whose evaluator aggregates and reports results. The graph also retains the harness’s task, filtering, and logging capabilities. Some source is unsampled, so unverified relationships are omitted rather than inferred.
Main entry points: evaluator.py Evaluation Orchestrator · modeling_qwen3_refusion.py ReFusion Model · generate.py Parallel Generation · __main__.py Harness CLI
Components 20
ReFusion Model
- Qwen3 Configuration
qwen3/configuration_qwen3.py - ReFusion Model
qwen3/modeling_qwen3_refusion.py - Diffusion Cache
qwen3/diffusion_cache_utils.py
Training and Inference
- Model Training
train.py - Parallel Generation
generate.py
Evaluation Harness
- Evaluation Entry
eval.py - Harness CLI
lm_eval/__main__.py - Evaluation Orchestrator
lm_eval/evaluator.py - Model Interface
lm_eval/api/model.py - Component Registry
lm_eval/api/registry.py - vLLM Backend
lm_eval/models/vllm_causallms.py - Request Cache
lm_eval/caching/cache.py
Tasks and Results
- Task Catalog
lm_eval/tasks/__init__.py - Task Runtime
lm_eval/api/task.py - Evaluation Requests
lm_eval/api/instance.py - Response Filters
lm_eval/filters/extraction.py - Metrics
lm_eval/api/metrics.py - Result Reporting
lm_eval/loggers/evaluation_tracker.py
Other components
- Engineer
- Training Data
Connections 23
- Engineer → Model Training: trains model
- Training Data → Model Training: provides examples
- Engineer → Parallel Generation: runs inference
- Parallel Generation → ReFusion Model: loads and runs
generate.py - Parallel Generation → Diffusion Cache: uses cache
generate.py - ReFusion Model → Qwen3 Configuration: uses configuration
qwen3/modeling_qwen3_refusion.py - ReFusion Model → Diffusion Cache: uses cache helpers
qwen3/modeling_qwen3_refusion.py - Engineer → Evaluation Entry: runs evaluation
- Evaluation Entry → Harness CLI: invokes
eval.py - Harness CLI → Evaluation Orchestrator: runs evaluation
lm_eval/__main__.py - Harness CLI → Task Catalog: loads tasks
lm_eval/__main__.py - Evaluation Orchestrator → Task Catalog: loads task definitions
lm_eval/evaluator.py - Task Catalog → Task Runtime: constructs tasks
lm_eval/tasks/__init__.py - Task Runtime → Evaluation Requests: builds requests
lm_eval/api/task.py - Evaluation Orchestrator → Evaluation Requests: processes requests
- Evaluation Orchestrator → Component Registry: selects model
lm_eval/evaluator.py - Component Registry → Model Interface: registers models
lm_eval/api/registry.py - Evaluation Orchestrator → Model Interface: runs model
lm_eval/evaluator.py - vLLM Backend → Model Interface: implements interface
lm_eval/models/vllm_causallms.py - Evaluation Orchestrator → Request Cache: uses request cache
lm_eval/evaluator.py - Evaluation Orchestrator → Response Filters: applies filters
- Evaluation Orchestrator → Metrics: aggregates metrics
lm_eval/evaluator.py - Evaluation Orchestrator → Result Reporting: logs results