CloudDock DeepSpeed 1.0.2

What’s new in 1.0.2 (vs 1.0.1)

• Stronger LLM training compatibility: better handling for modern training scripts and “real world” argument patterns.
• Dynamic runtime args: you can pass parameters at start time (no more “edit the script just to change lr/epochs/etc.”).
• Form override controls: you can disable front-end form parameters, or force override to guarantee deterministic runs.
• Built-in mini error tips: common failures get short hints so you can fix faster (without digging through 500 lines of logs).
• New indicators: in-site progress + VRAM usage + speed metrics to confirm the run is healthy at a glance.
• Deploy Helper (NEW): after training, test the model immediately inside the instance — no need to move files back to your own computer just to verify outputs.
• Ecosystem parity with Universal Usagi 4.1.3: VIP badge + end-time countdown + CloudDock Pass “Extend” status indicator.
• Bug fix: Appstore now avoids a class of interrupted installs that could get stuck mid-way.

Two ways to run training

1) CLI mode (terminal)

CLI is still the highest-control path: custom configs, debugging, distributed settings. 1.0.2 adds better compatibility for LLM-style scripts and makes it easier to keep arguments consistent.

# Example (single GPU)
deepspeed --num_gpus=1 train.py \
  --dataset /workspace/data \
  --output_dir /workspace/output \
  --lr 3e-5 --epochs 1 --batch_size 8


# Tip: keep dataset + outputs in /workspace (or your mounted storage) for clean transfers.

2) GUI mode (CloudDock DeepSpeed Console)

GUI is the fastest “blank instance → first run” path. In 1.0.2, you can dynamically pass args at launch, and you get real-time progress / VRAM / speed indicators plus small error hints when things fail.

Figure 2 — Console: progress + VRAM + speed indicators, log tail, and training status.

Figure 3 — Console: runtime args + form override controls (disable form args / force override).

Dynamic runtime args + override controls

1.0.2 introduces a clean policy for “where do arguments come from?” This solves two common problems: (1) people want quick GUI knobs, but also (2) advanced users want strict reproducibility and CLI-only control.

Argument sources

• Script args: arguments that are passed into your training script (train.py).
• Console form args: the GUI form fields (epochs, lr, batch size, etc.).
• Overrides: explicit rules that decide whether GUI fields apply or not.

Override modes (recommended)

• Normal (default): GUI form values apply (good for beginners and fast iteration).
• Disable form args: front-end form parameters are ignored (pure “script controls everything”).
• Force override: GUI form values always override script defaults (deterministic “console rules all” runs).

New indicators: progress + VRAM + speed

1.0.2 adds a quick-glance health panel so you can verify “is it actually training?” without staring at logs:

• Progress: step / total (or step-based progress when total is known)
• VRAM: current GPU memory usage (helps detect OOM risk early)
• Speed: tokens/sec or steps/sec style indicator (depending on the script / runner)

Figure 4 — Indicators panel: progress, VRAM usage, and speed metrics for fast sanity checks.

Deploy Helper (NEW): train → test instantly

The Deploy Helper is designed for the most annoying real-world loop: you finish training, then you have to move the model elsewhere just to test if it works. In 1.0.2, you can run a quick deployment test inside the instance and “take away” the final result with confidence.

What it does

• Spin up a lightweight test run for your trained output
• Feed a prompt and capture output immediately
• Show errors clearly (so you can fix right away and re-test)

Figure 5 — Deploy Helper: select model output + set test prompt.

Figure 6 — Deploy Helper: output preview + clear error reporting.

Mini error tips (common issues)

When something fails, 1.0.2 will surface small “what to check next” hints. Here are the most common patterns you’ll see during LLM runs:

• Out of memory (OOM): reduce batch size, sequence length, gradient accumulation, or enable a more memory-friendly precision.
• Loss becomes NaN / explodes: lower LR, verify dtype/mixed precision setup, check dataset quality, start with a tiny run first.
• Path not found: confirm your dataset/model paths match where you uploaded them (recommended: keep in /workspace/).
• Dependency/module errors: confirm you are using the container’s intended environment (the included venv / runtime), and that your script requirements match the image.

Ecosystem integration (Usagi parity)

DeepSpeed 1.0.2 aligns with CloudDock Universal Usagi 4.1.3 experience:

• VIP badge: consistent identity & perks display across containers
• End-time countdown: clearer remaining time awareness before shutdown
• CloudDock Pass “Extend” indicator: quick status light for extendability state

Bug fix: Appstore install interruptions

1.0.2 includes an Appstore reliability fix for cases where installs could be interrupted mid-way. The goal: fewer “stuck” installs and a cleaner “install → done” experience.

Recommended folder convention

/workspace/
  train.py
  data/
  output/
  configs/
  deploy_tests/

Upgrade notes (from 1.0.1)

• If you previously relied on “edit script defaults for every run” — switch to dynamic runtime args.
• If your results varied run-to-run because of mixed parameter sources — use Force override.
• If you want maximum purity (script decides everything) — use Disable form args.
• After training, validate with Deploy Helper before exporting or presenting results.
• Jupyter + TensorFlow: TensorFlow is now preinstalled in Jupyter — open a notebook and start training immediately.