usematcha 0.2.4

GPU energy observability for AI training workloads

matcha

GPU energy observability for AI training

Measure GPU energy per training run and per step.
No code changes. Structured output for any observability stack.

---

Install

pip install usematcha

Requires Linux + an NVIDIA GPU with drivers installed. Python 3.9+.

---

Quick start

Prefix your training command with matcha run:

matcha run torchrun --standalone --nproc_per_node=8 train_gpt.py

Training runs at full speed. One line at the end:

matcha_energy gpus:8x NVIDIA H100 80GB HBM3 total:778168J (216.16Wh) duration:203.1s avg_power:3832W peak_power:4120W samples:2031

No code changes. No config files. Works with any training script.

---

Commands

matcha run

Total energy for a whole command. Zero overhead — matcha doesn't touch stdout.

matcha run python train.py
matcha run torchrun --standalone --nproc_per_node=8 train_gpt.py

matcha wrap

Per-step energy. matcha parses the training stdout for step markers (step 10, iter 10, [10/1000], etc.) and appends energy per step.

matcha wrap torchrun --standalone --nproc_per_node=8 train_gpt.py

step:1/20000 train_loss:6.9357 train_time:612ms step_avg:612.00ms energy:2354.0J/step avg_power:3847W peak_power:4120W
step:2/20000 train_loss:16.7414 train_time:831ms step_avg:721.50ms energy:3012.6J/step avg_power:3625W peak_power:3998W
...
matcha_energy gpus:8x NVIDIA H100 80GB HBM3 total:778168J (216.16Wh) duration:203.1s avg_power:3832W peak_power:4120W samples:2031

matcha monitor

A drop-in replacement for running watch nvidia-smi in a second terminal.

matcha monitor
matcha monitor --gpus 0,1,2,3 --interval 500

---

Observability

matcha plugs into any stack you already have. Three outputs, pick one or several:

JSONL for files & ingestion

matcha wrap --output run.jsonl \
    --label team=capacity --label config=lr_3e-4 \
    torchrun --standalone --nproc_per_node=8 train_gpt.py

run.jsonl contains session_start, one step record per step, and session_end. Per-GPU breakdowns included.

{"type":"step","step":1,"energy_j":2354.0,"duration_s":0.612,"avg_power_w":3847,"peak_power_w":4120,
 "train_metrics":{"train_loss":6.9357,"train_time_ms":612,"step_avg_ms":612.0},
 "gpus":[{"idx":0,"energy_j":323.5,"avg_power_w":528.6,"peak_power_w":540.0}, ... ]}

Stream into ClickHouse, DuckDB, Loki, or anything that reads JSON lines:

cat run.jsonl | clickhouse-client --query "INSERT INTO energy_steps FORMAT JSONEachRow"

Prometheus (pull)

matcha wrap --prometheus :9400 torchrun train.py

Exposes /metrics for any Prometheus scraper. Both GPU-live gauges and step-level metrics.

matcha_step_energy_joules{run_id="...",gpu="0"}      2354.0
matcha_step_duration_seconds{run_id="..."}           0.612
matcha_step_peak_power_watts{run_id="..."}           4120
matcha_step_gpu_energy_deviation_ratio{gpu="3"}     -0.18   # straggler signal
matcha_metric_train_loss{run_id="..."}               6.9357
matcha_metric_step_avg_ms{run_id="..."}              612.0

One-line alerts:

- alert: GpuStraggler
  expr: matcha_step_gpu_energy_deviation_ratio < -0.15
  for: 5m

OpenTelemetry / OTLP (push)

pip install 'usematcha[otlp]'

matcha wrap \
    --otlp https://otlp-gateway.grafana.net/otlp \
    --otlp-header "Authorization=Basic <token>" \
    torchrun train.py

Pushes the same metric set to Grafana Cloud, Honeycomb, Datadog, or any OTel collector. Metric names match the Prometheus endpoint so dashboards port across deployments.

Training metrics (automatic)

In wrap mode, matcha parses numeric fields from your stdout — train_loss:6.94, lr=1.23e-4, HuggingFace {'loss': 2.3, ...} — and surfaces them alongside energy. Works out of the box for nanoGPT, modded-nanogpt, parameter-golf, DeepSpeed, HF Trainer.

No config. No code changes. In Grafana:

matcha_metric_train_loss        +
matcha_step_energy_joules       → efficiency curve: J per unit loss reduction

---

Flags

Flag	Description
--output PATH	Write JSONL records to a file.
--json	Emit JSONL to stdout/stderr (use --output when running wrap).
--prometheus :PORT	Expose a Prometheus /metrics endpoint.
--otlp URL	Push metrics to an OTLP/HTTP collector.
--otlp-header K=V	Auth header for OTLP (repeatable).
--label KEY=VALUE	Attach a label to the run (repeatable).
--run-id ID	Stable run identifier. Honors MATCHA_RUN_ID.
--gpus	all, a single index (0), or a list (0,1,2,3).
--interval	Peak-power poll interval in ms. Default: 100.

---

Multi-GPU

matcha auto-detects every visible GPU and reports summed totals plus per-GPU breakdowns. The per-GPU arrays make straggler detection a one-query affair:

• Straggler — one rank consistently drawing ~30% less power usually means a stuck collective, a thermal-throttled card, or a PCIe link degraded to Gen3.
• DP / PP / TP fingerprinting — the per-GPU power pattern tells you what parallelism strategy is actually running.

matcha run torchrun --standalone --nproc_per_node=8 train_gpt.py
matcha run --gpus 0,1,2,3 torchrun ...
matcha run --gpus 0 torchrun ...

---

How it works

matcha reads energy directly from NVML's hardware accumulator (nvmlDeviceGetTotalEnergyConsumption, Volta+). Per-step and session energy are exact counter deltas — millijoule-precise, no integration error. A background poller (default 100 ms) plus boundary reads at each step transition track peak power. Pre-Volta GPUs fall back to trapezoidal integration of polled samples.

The training process runs natively — matcha never touches your model or your training loop.

• run does not intercept stdout. Close to zero-overhead as reasonable.
• wrap pipes stdout to detect step boundaries, then appends energy data inline or emits structured records.
• monitor samples directly without launching a child process.

---

Why

Frontier training runs burn hundreds of MWh. The gap between teams that optimize for energy-per-step and those that don't is measured in millions of dollars per training run. matcha makes that number visible without changing your training code.

8xH100 training run — 1 hour:
  Energy cost:   $0.26 (2.16 kWh @ $0.12/kWh)
  Compute cost:  $23.20 (RunPod @ $23.20/hr)

  → Optimizing energy per step == faster training == less rental time

---

Built by

Keeya Labs · Docs · GitHub

License

Apache 2.0 — see LICENSE.