coremateai 0.1.1

Torch and CUDA helper toolkit for AI training and inference workflows.

coremate

CoreMate is a lightweight Python toolkit focused on Torch and CUDA optimization for AI workflows.

PyPI release page: https://pypi.org/project/coremateai/0.1.1/

Installation

pip install coremateai==0.1.1

Optional extras:

pip install "coremateai[ai]"

Install PyTorch separately (matching your CUDA runtime):

pip install torch

What It Does

• Detects AI stack details (torch, CUDA, GPU, driver).
• Recommends training presets by task (cv, nlp, llm, etc.).
• Builds full training plans (build_training_plan) with LR, batch and accumulation heuristics.
• Applies reproducibility helpers (set_global_seed).
• Tunes Torch runtime (cudnn, TF32, matmul precision, threads).
• Applies CUDA allocator/env defaults.
• Provides one-shot optimization (optimize_torch_cuda).

Python Usage

from coremate import (
    detect_ai_stack,
    recommend_training_preset,
    build_training_plan,
    set_global_seed,
    tune_torch_runtime,
    apply_cuda_env_defaults,
    optimize_torch_cuda,
)

stack = detect_ai_stack()
print(stack["cuda_available"], stack["torch_version"])

preset = recommend_training_preset(task="llm", aggressive=True, hardware=stack)
print(preset.batch_size, preset.precision)

plan = build_training_plan(
    task="llm",
    model_scale="base",
    target_global_batch_size=64,
    aggressive=True,
    hardware=stack,
)
print(plan.global_batch_size, plan.learning_rate)

set_global_seed(42, deterministic=True)
apply_cuda_env_defaults(max_split_size_mb=256, expandable_segments=True)
tune_torch_runtime(seed=42, deterministic=True, benchmark=True, allow_tf32=True)

result = optimize_torch_cuda(
    task="llm",
    model_scale="base",
    target_global_batch_size=64,
    aggressive=True,
    seed=42,
    deterministic=True,
)
print(result["plan"]["global_batch_size"], result["tune_error"])

CLI Usage

coremate ai report
coremate ai recommend --task llm --aggressive
coremate ai plan --task llm --model-scale base --target-global-batch 64 --aggressive
coremate ai seed --seed 42 --deterministic
coremate ai env --max-split-size-mb 256 --expandable-segments
coremate ai tune --seed 42 --deterministic --benchmark --allow-tf32
coremate ai optimize --task llm --model-scale base --target-global-batch 64 --aggressive --seed 42 --deterministic

Testing

python -m pytest