How pixelml/club-170hx works
Club CMP 170HX is a community-tested collection of hardware diagnostics, AI workload recipes, and reproducible benchmark experiments—not a single application. A typical operator runs a model recipe against a live serving endpoint, checks runtime and card telemetry, records usage-accounted measurements, and validates results. The repository also preserves independent long-context and multimodal benchmarks, Jev-compatible classification experiments, and model-specific vLLM patches. The tree and excerpts establish only selected connections; unsampled experiment families are represented by their directories, without inferred call chains.
Main entry points: build_qwen_recipe_notebook.py Recipe runner · jev_probe.py Jev probes · mimo_v2.py MiMo model · live_benchmark.py Usage benchmark
Components 22
Live recipe
- Recipe runner
scripts/build_qwen_recipe_notebook.py - Usage benchmark
recipes/qwen3.8-27b-dflash2/live_benchmark.py - Runtime safety
recipes/qwen3.8-27b-dflash2/live_support.py - Model preparation
recipes/qwen3.8-27b-dflash2/prepare_pinned_models.py - Result validation
recipes/qwen3.8-27b-dflash2/result_pipeline.py
Benchmark experiments
- Jev classifier
results/2026-09-20-qwen3.8-27b-w4a16-jev-1card-vllm/harness/jev_server.py - Jev probes
results/2026-09-20-qwen3.8-27b-w4a16-jev-1card-vllm/harness/jev_probe.py - Jev measurements
results/2026-09-20-qwen3.8-27b-w4a16-jev-1card-vllm/harness - Long-context driver
results/2026-09-02-deepseek-v4-flash-vision-exp-4card-longctx-262k/run_longctx.py - Long-context tests
results/2026-09-02-deepseek-v4-flash-vision-exp-4card-longctx-262k/longctx_bench.py - Other benchmarks
results - Executed experiments
notebooks
Model extensions
- MiMo model
results/2026-09-22-mimo-v2.6-flash-4card-pp4-vllm/patches/mimo_v2.py - MiMo predictor
results/2026-09-22-mimo-v2.6-flash-4card-pp4-vllm/patches/mimo_v2_mtp.py - MiMo vision
results/2026-09-22-mimo-v2.6-flash-4card-pp4-vllm/patches/mimo_v2_omni.py - Other model patches
results/2026-09-22-mimo-v2.6-flash-4card-pp4-vllm/patches
Hardware diagnostics
- Card diagnostics
scripts/qc/memtest-select.py - Compute health gate
results/2026-09-02-health-gate/receipts/matmul_check.py - Power experiments
results/2026-09-02-deepseek-v4-flash-vision-exp-4card-repro-power/bench_repro_power.py - Cost analysis
results/2026-09-02-cross-platform-cost-watt/make_chart.py
Other components
- Operator
- Serving endpoint
Connections 20
- Operator → Recipe runner: runs
scripts/build_qwen_recipe_notebook.py - Recipe runner → Usage benchmark: uses source
scripts/build_qwen_recipe_notebook.py - Recipe runner → Runtime safety: uses source
scripts/build_qwen_recipe_notebook.py - Recipe runner → Result validation: uses source
scripts/build_qwen_recipe_notebook.py - Usage benchmark → Serving endpoint: sends requests
recipes/qwen3.8-27b-dflash2/live_benchmark.py - Operator → Jev probes: runs probes
results/2026-09-20-qwen3.8-27b-w4a16-jev-1card-vllm/harness/jev_probe.py - Jev probes → Jev classifier: calls endpoint
results/2026-09-20-qwen3.8-27b-w4a16-jev-1card-vllm/harness/jev_probe.py - Jev probes → Serving endpoint: queries upstream
- Operator → Long-context driver: runs tests
results/2026-09-02-deepseek-v4-flash-vision-exp-4card-longctx-262k/run_longctx.py - Long-context driver → Long-context tests: calls tests
results/2026-09-02-deepseek-v4-flash-vision-exp-4card-longctx-262k/run_longctx.py - Operator → Other benchmarks: runs experiments
- Operator → Executed experiments: reproduces results
- Operator → Card diagnostics: checks card
- Operator → Compute health gate: checks compute
- Operator → Power experiments: measures power
- Operator → Cost analysis: compares cost
- MiMo predictor → MiMo model: imports model
results/2026-09-22-mimo-v2.6-flash-4card-pp4-vllm/patches/mimo_v2_mtp.py - MiMo vision → MiMo model: imports model
results/2026-09-22-mimo-v2.6-flash-4card-pp4-vllm/patches/mimo_v2_omni.py - Operator → MiMo model: uses patch
- Operator → Other model patches: uses patches