Docs

Author: benoit-cty

docs/advanced/adastra.md

@@ -1,6 +1,6 @@
-# CodeCarbon on CINES Adastra HPC with AMD ROCM
+# ROCm and PyTorch on SLURM SuperComputer
 
-This project was provided with computing and storage resources by GENCI at CINES thanks to the grant AD010615147R1 on the supercomputer Adastra's MI250x/MI300 partition.
+This project was provided with computing and storage resources by GENCI at CINES thanks to the grant AD010615147R1 on the [supercomputer Adastra](https://dci.dci-gitlab.cines.fr/webextranet/architecture/index.html)'s MI250x/MI300 partition.
 
 Thanks to this grant we were able to develop and test the AMD ROCM support in CodeCarbon, and provide this quick start guide to help other users of Adastra HPC to easily monitor the carbon emissions of their machine learning workloads running on AMD GPUs.
 

docs/introduction/power_estimation.md

@@ -7,6 +7,7 @@ While energy is the metric primarily responsible for CO₂ emissions estimations
 ## 1. Energy as the Source of Truth
 
 The most accurate tracking methods rely on built-in hardware energy counters rather than instantaneous power draw. For example:
+
 - **NVIDIA GPUs** using `nvmlDeviceGetTotalEnergyConsumption` return accumulated energy in millijoules.
 - **AMD GPUs** using `amdsmi_get_energy_count` yield a counter that is multiplied by its resolution and converted into millijoules.
 - **CPUs** using the RAPL interface read from files like `energy_uj` to get accumulated microjoules.
@@ -40,6 +41,7 @@ The tracker has designated logic blocks for different components (e.g., CPU, RAM
 Inside the main `EmissionsTracker`, the energy values are securely accumulated over the session's lifespan.
 
 For recording the power, a running sum is maintained:
+
 - As CodeCarbon sequentially takes measurements, it tracks the output of `power.W`.
 - It dynamically increments running variables like `_gpu_power_sum`, `_cpu_power_sum`, `_ram_power_sum`.
 - It increments a global counter `_power_measurement_count`.
@@ -53,6 +55,7 @@ This smoothing process prevents singular short measurement anomalies from skewin
 ## Summary Pipeline
 
 In short:
+
 1. **Hardware Counters (Accumulated Energy)**
 2. Subtract `last_energy` = **Energy Delta**
 3. Divide Energy Delta by `last_duration` = **Interval Average Power**

mkdocs.yml

@@ -113,6 +113,7 @@ nav:
     - Motivation: introduction/motivation.md
     - Methodology: introduction/methodology.md
     - RAPL Metrics: introduction/rapl.md
+    - Power Estimation: introduction/power_estimation.md
     - Model Comparisons: introduction/model_examples.md
     - Frequently Asked Questions: introduction/faq.md
   - Getting Started:
@@ -121,11 +122,11 @@ nav:
     - CodeCarbon API: getting-started/api.md
     - Parameters: getting-started/parameters.md
     - Examples: getting-started/examples.md
-    - Comet Integration: getting-started/comet.md
     - Test on Scaleway: getting-started/test_on_scaleway.md
   - Advanced Usage:
-    - Linux service: advanced/advanced_installation.md
+    - Install CodeCarbon as a Linux service: advanced/linux_service.md
     - Deploy with Ansible: advanced/ansible.md
+    - Comet Integration: getting-started/comet.md
     - ROCm and PyTorch on SLURM SuperComputer: advanced/adastra.md
   - Logging:
     - Output: logging/output.md