ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
unsloth/tests/metrics
History
Rachel Li 078b76281a Add vLLM-style runtime metrics for inference and training 
- Comprehensive metrics collection system (inference + training)
- Prometheus-compatible export with optional HTTP server
- Programmatic access to metrics
- Automatic instrumentation of inference and training loops
- All tests passing
2026-03-12 23:10:55 +00:00
..
__init__.py Add vLLM-style runtime metrics for inference and training 2026-03-12 23:10:55 +00:00
README.md Add vLLM-style runtime metrics for inference and training 2026-03-12 23:10:55 +00:00
test_metrics_standalone.py Add vLLM-style runtime metrics for inference and training 2026-03-12 23:10:55 +00:00
TEST_RESULTS.txt Add vLLM-style runtime metrics for inference and training 2026-03-12 23:10:55 +00:00

README.md

Metrics Tests

Running Tests

python3 tests/metrics/test_metrics_standalone.py

Test Coverage

The test suite verifies:

  1. InferenceStats - Request tracking, token counts, latencies
  2. TrainingStats - Batch tracking, loss, throughput
  3. StatsCollector - Singleton pattern, enable/disable
  4. Prometheus Export - Client availability check

Test Results

See TEST_RESULTS.txt for the latest test run output.

Expected Output

============================================================
Unsloth Metrics Collection - Standalone Test
============================================================

1. Testing InferenceStats...
   ✅ Total requests: 1
   ✅ Prompt tokens: 10
   ✅ Generation tokens: 5
   ✅ Avg latency: 0.087s

2. Testing TrainingStats...
   ✅ Total steps: 3
   ✅ Total samples: 12
   ✅ Avg loss: 0.4500

3. Testing StatsCollector...
   ✅ StatsCollector working

4. Testing Prometheus export...
   ✅ Prometheus client available
   ℹ️  Full Prometheus export test requires GPU environment

============================================================
✅ All metrics tests passed!
============================================================