alloc

Find and fix training bottlenecks. Zero code changes.

pip install alloc
alloc run python train.py

alloc v0.0.9 — Calibrate

 Run Summary
  Peak VRAM       31.2 GB / 40.0 GB (A100)
  VRAM used       78.0%
  Avg GPU util    72.3%
  Avg power       287 W
  Duration        24.1s (auto-stopped: metrics stable at 18.2s)
  Step time       148.5 ms (p50) / 152.1 ms (p90)
  Throughput      42.3 samples/sec

  Artifact: alloc_artifact.json.gz

That's it. No decorators, no config files, no code changes. Alloc wraps your command, profiles GPU usage, and tells you what's wrong.

What you get

alloc diagnose reads your training script and tells you exactly what to change:

alloc diagnose train.py

alloc diagnose — 3 findings in train.py

 CRITICAL  DL005 — DataLoader running in main thread
   train.py:47   num_workers=0 → num_workers=8
   num_workers=0 loads data in the main thread, blocking GPU computation entirely.
   Expected impact: ~30-50% faster training with parallel data loading

 WARNING   PREC002 — Using fp16, consider bf16
   train.py:56   dtype: float16 → dtype: bfloat16
   H100 supports bf16 natively — eliminates loss scaling overhead.
   Expected impact: ~5-10% speedup, eliminates GradScaler complexity

 INFO      THRU001 — cudnn.benchmark not enabled
   Add: torch.backends.cudnn.benchmark = True
   Expected impact: ~5-10% speedup for fixed-size inputs

Summary: 1 critical, 1 warning, 1 info
Run with --diff to generate patches | --json for CI output

alloc ghost estimates VRAM before you launch:

alloc ghost train.py --dtype bf16

 Ghost Scan — 7.0B params (bf16)

  Model weights       13.04 GB
  Gradients           13.04 GB
  Optimizer (Adam)    78.23 GB
  Activations (est.)   0.50 GB
  Buffer (10%)        10.48 GB

  Total VRAM         115.28 GB

alloc scan ranks GPU configs without a GPU:

alloc scan --model llama-3-70b --gpu H100-80GB --num-gpus 8

Works with everything

Alloc wraps your launch command. No framework-specific setup required.

alloc run python train.py
alloc run torchrun --nproc_per_node=4 train.py
alloc run accelerate launch train.py
alloc run srun python train.py           # Slurm
alloc run ray job submit -- python train.py

Multi-GPU detection is automatic (discovers all GPUs in the process tree).

Deeper signals (optional)

Add a one-line callback for step-level timing:

# HuggingFace
from alloc import HuggingFaceCallback
trainer = Trainer(..., callbacks=[HuggingFaceCallback()])

# Lightning
from alloc import LightningCallback
trainer = Trainer(..., callbacks=[LightningCallback()])

This unlocks step time p50/p90, throughput, and dataloader bottleneck detection.

All commands

Every command supports --json for CI/CD integration.

Dashboard

Log in to get team visibility, budget tracking, and optimization proposals:

alloc login --browser
alloc run python train.py    # auto-uploads when logged in

Dashboard at alloclabs.com

Design principles

1. Zero config — alloc run python train.py works out of the box
2. Never crash training — all Alloc failures are caught silently
3. No monkey-patching — external monitoring only, deeper signals opt-in
4. Local-first — works in air-gapped environments, no internet required

Links

• Website
• Documentation
• PyPI