How darkdatter/gb10-repo works
This repository is a measured deployment and benchmarking recipe for Qwen3.8-27B on DGX Spark, not the model server or SparkStation implementation itself. The documented workflow verifies GPU access, prepares the image and weights, serves the model standalone or through optional SparkStation, then benchmarks throughput and HumanEval quality. Tuning guidance and recorded results are also first-class product outputs. Relationships below are based on README statements; no source excerpts were available to verify implementation-level calls.
Main entry points: sparkstation-models.yaml.md SparkStation gateway · 01-build-and-fetch.sh Image and weights · scripts Standalone serving · run-humaneval.sh HumanEval runner
Components 20
Setup
- GPU verification
scripts/00-verify-gpu.sh - Image and weights
scripts/01-build-and-fetch.sh
Serving
- Standalone serving
scripts - SparkStation gateway
patches/sparkstation-models.yaml.md
Benchmarks
- Performance benchmark
bench/perf.py - Long-context benchmark
bench/longctx.py - HumanEval runner
scripts/run-humaneval.sh - HumanEval generation
bench/humaneval/generate.py - HumanEval execution
bench/humaneval/execute.py - HumanEval report
bench/humaneval/report.py - Benchmark helpers
bench/common.py
Tuning and results
- Measured results
results/RESULTS.md - Reproduction notes
results/REPRODUCTION.md - Build manifest
results/BUILD-MANIFEST.md - Model tuning notes
patches/sparkstation-models.yaml.md - Health patch
patches/sparkstation-gateway-health.patch - Memory override patch
patches/sparkstation-mem-fraction-override.patch - Model patch notes
patches/sparkstation-models.yaml.md
Other components
- Operator
- Model repositories
Connections 17
- Operator → GPU verification: verifies GPU
README.md - Operator → Image and weights: prepares artifacts
README.md - Image and weights → Model repositories: fetches weights
README.md - Operator → Standalone serving: starts server
README.md - Operator → SparkStation gateway: starts gateway
README.md - Standalone serving → Performance benchmark: serves benchmark
README.md - Standalone serving → HumanEval runner: serves evaluations
README.md - HumanEval runner → HumanEval execution: executes problems
README.md - Operator → Long-context benchmark: runs benchmark
- Performance benchmark → Measured results: informs measurements
README.md - HumanEval runner → Measured results: records quality
README.md - Image and weights → Build manifest: documents inputs
README.md - SparkStation gateway → Model tuning notes: uses configuration
README.md - SparkStation gateway → Health patch: documents mitigation
README.md - SparkStation gateway → Memory override patch: documents override
README.md - Model tuning notes → Model patch notes: describes settings
- Measured results → Reproduction notes: compares runs
README.md