[Kernel Feature] Accelerator-Aware Resource Limits - cgroups v2 Wrapper for AI

[mikejmorgan-ai]

Accelerator-Aware Resource Limits - cgroups v2 Wrapper for AI Workloads

Part of: Kernel-Level AI Enhancements (Tier 1 - User-Space)

Description

Wraps Linux cgroups v2 with an AI-friendly CLI for managing GPU, CPU, and memory resources with workload presets.

Effort: 1-2 weeks | Bounty: $100

The Solution

cortex limits create inference-job --preset inference --gpus 2
cortex limits apply inference-job --pid 12345
eval $(cortex limits env inference-job)  # for new processes
cortex limits status inference-job

Workload Presets

Preset	CPU	Memory	GPU	OOM Score	Use Case
inference	4 cores	32G	100%	-500	Low-latency serving
training	16 cores	128G	100%	-800	Long training jobs
batch	8 cores	64G	80%	0	Background processing
interactive	2 cores	16G	50%	-200	Development

Features

• cgroups v2 unified hierarchy support
• Workload presets with sensible defaults
• NVIDIA GPU isolation via environment variables
• OOM score adjustment for priority
• CPU quota, weight, and affinity
• Memory hard and soft limits

Acceptance Criteria

• Profiles can be created with presets
• Limits are correctly applied to cgroups
• GPU environment variables are generated
• Works in user mode with delegation
• Unit tests pass

Files

Complete implementation available:

• accelerator_limits.py (~700 lines)
• test_accelerator_limits.py (~300 lines)
• README_ACCELERATOR_LIMITS.md

Priority

Medium - Important for multi-tenant environments

[yaya1738]

Also interested in this bounty. Relevant experience:

• Linux cgroups v2 and systemd resource management
• Python CLI tools with argparse/click
• Docker and container resource isolation

Approach: Wrap cgroups v2 controllers with AI workload presets, expose via clean CLI.

Can work on this alongside #223 (/dev/llm) since they're complementary.

[yaya1738]

Update: Prototype Complete

Working implementation with all features:

Test Results: 32/32 passing

TestPresets (5 tests) ✓
TestResourceProfile (7 tests) ✓
TestProfileStore (6 tests) ✓
TestGPUManager (5 tests) ✓
TestCgroupsController (3 tests) ✓
TestCLI (3 tests) ✓
TestEndToEnd (3 tests) ✓

Implementation:

• accelerator_limits.py (~530 lines)
• All 4 presets (inference, training, batch, interactive)
• cgroups v2 with user delegation support
• GPU env vars (CUDA_VISIBLE_DEVICES, TF, PyTorch)
• Profile persistence to ~/.config/cortex/limits/

Demo:

# Create profile
cortex limits create llm-serve --preset inference --gpus 2

# Apply to process
cortex limits apply llm-serve --pid 12345

# Get environment
eval $(cortex limits env llm-serve)
# exports: CUDA_VISIBLE_DEVICES=0,1

# Check status
cortex limits status llm-serve

Ready to submit PR. Happy to adjust based on feedback.

[mikejmorgan-ai]

Update: Prototype Complete

Working implementation with all features:

Test Results: 32/32 passing

TestPresets (5 tests) ✓
TestResourceProfile (7 tests) ✓
TestProfileStore (6 tests) ✓
TestGPUManager (5 tests) ✓
TestCgroupsController (3 tests) ✓
TestCLI (3 tests) ✓
TestEndToEnd (3 tests) ✓

Implementation:

• accelerator_limits.py (~530 lines)
• All 4 presets (inference, training, batch, interactive)
• cgroups v2 with user delegation support
• GPU env vars (CUDA_VISIBLE_DEVICES, TF, PyTorch)
• Profile persistence to ~/.config/cortex/limits/

Demo:

# Create profile
cortex limits create llm-serve --preset inference --gpus 2

# Apply to process
cortex limits apply llm-serve --pid 12345

# Get environment
eval $(cortex limits env llm-serve)
# exports: CUDA_VISIBLE_DEVICES=0,1

# Check status
cortex limits status llm-serve

Ready to submit PR. Happy to adjust based on feedback.

Welcome aboard @yaya1738 you're assigned to #222.

Before diving in:

1. ⭐ Star the repo — helps with visibility
2. 💬 Join Discord — discord.gg/uCqHvxjU83 (coordination happens there)
3. 👥 Invite 2 developers you know are excellent — we're building the team and your network matters

To get started:

git clone https://github.com/cortexlinux/cortex
cd cortex
pip install -e .
# Your work goes in cortex/kernel_features/

[yaya1738]

Thanks @mikejmorgan-ai! Excited to contribute.

✅ Starred the repo
✅ Will join Discord shortly

I'll fork and submit the PR for #222 today. Already have #223 ready too - happy to take both ($225).

Quick question: should I submit the code to cortex/kernel_features/accelerator_limits/ or a different path?