mx8 1.0.8rc1

MX8: bounded data runtime (Rust) exposed to Python.

mx8 (Python)

MX8 is a bounded-memory data runtime exposed to Python (built with PyO3 + maturin).

The v0 focus is “don’t OOM”: MX8 enforces backpressure with hard caps (so prefetch can’t runaway).

Further docs:

• Python API: ../../docs/python_api.md
• Vision labels/layout: ../../docs/vision_labels.md
• S3/runtime tuning: ../../docs/s3_runtime_tuning.md
• Memory contract: ../../docs/memory_contract.md
• Video GA checklist: ../../docs/video_ga_checklist.md
• Troubleshooting: ../../docs/troubleshooting.md

Install (from wheel)

Once you have a wheel (from CI or local build):

• python -m venv .venv && . .venv/bin/activate
• pip install mx8-*.whl

Install (from PyPI)

• python -m venv .venv && . .venv/bin/activate
• pip install mx8
• Optional vision/training deps: pip install pillow numpy torch

Quickstart (local, no S3)

import mx8

mx8.pack_dir(
    "/path/to/imagefolder",
    out="/path/to/mx8-dataset",
    shard_mb=512,
    label_mode="imagefolder",
    require_labels=True,
)

loader = mx8.image(
    "/path/to/mx8-dataset@refresh",
    batch=64,
    inflight=256 * 1024 * 1024,
    resize=(224, 224),  # (H,W); optional
)

print(loader.classes)  # ["cat", "dog", ...] if labels.tsv exists

for images, labels in loader:
    pass

Zero-manifest load (raw prefix)

import mx8

loader = mx8.load(
    "s3://bucket/raw-prefix/",
    recursive=True,  # default
    profile="balanced",
)

for batch in loader:
    pass

mx8.run(...) is the convenience wrapper that chooses local vs distributed mode from environment (WORLD_SIZE). mx8.resolve(...) is a short alias for mx8.resolve_manifest_hash(...).

Mix multiple loaders

mx8.mix(...) composes existing loaders into one deterministic stream. weights are sampling proportions (not model-loss weights).

import mx8

loader_a = mx8.load("s3://bucket/dataset_a/@refresh", profile="balanced", tune=True)
loader_b = mx8.load("s3://bucket/dataset_b/@refresh", profile="balanced", tune=True)

mixed = mx8.mix(
    [loader_a, loader_b],
    weights=[1, 1],   # fairness baseline (50:50)
    seed=0,
    epoch=0,
)

for batch in mixed:
    pass

print(mixed.stats())

Skewed example:

mixed = mx8.mix([loader_a, loader_b], weights=[7, 3], seed=0, epoch=0)

seed and epoch define deterministic schedule behavior:

• same seed + epoch => same source-pick sequence
• same seed, different epoch => controlled schedule variation

starvation is an optional watchdog threshold in scheduler ticks used for starvation counters in mixed.stats(). Set MX8_MIX_SNAPSHOT=1 (and optional MX8_MIX_SNAPSHOT_PERIOD_TICKS=64) to emit periodic mix_snapshot proof events.

Minimal API naming note:

• Top-level APIs use short kwargs (batch, ram_gb, coord, resume, ...).
• Advanced objects keep explicit names:
  • mx8.Constraints(max_inflight_bytes=..., max_ram_bytes=...)
  • mx8.RuntimeConfig(prefetch_batches=..., max_queue_batches=..., want=...)
  • mx8.DistributedDataLoader(..., autotune=..., resume_from=...)

Bounded memory (v0)

Set a hard cap and periodically print high-water marks:

import mx8

loader = mx8.image(
    "/path/to/mx8-dataset@refresh",
    batch=64,
    inflight=256 * 1024 * 1024,
    queue=8,
    prefetch=4,
)

for step, (images, labels) in enumerate(loader):
    if step % 100 == 0:
        print(loader.stats())  # includes ram_high_water_bytes

Avoid patterns that intentionally accumulate batches:

# ❌ Don't do this (will grow RSS regardless of any loader)
all_batches = list(loader)

Labels (optional)

label_mode="imagefolder" is designed to scale:

• Per-sample records reference a numeric label_id (u64), not a repeated string.
• The human-readable mapping is stored once at out/_mx8/labels.tsv.

If your input layout is mixed (files directly under the prefix and subfolders), label_mode="auto" may disable ImageFolder labeling. To enforce ImageFolder semantics, use:

mx8.pack_dir(..., label_mode="imagefolder", require_labels=True)