Local Control Surface

STRIX HALO X

CPU / CUDA / MPS / ROCm

Universal LLM Fine-Tuning Dashboard

Architecture, System Design, Conceptual Framework, and Core Development by Franz Ayestaran / Enhanced Pair Programming with GitHub Copilot (GPT-5.3-Codex), Claude Code, and OpenAI models

Backend ...

Python ...

Workspace ...

Training Run

Loading runtime rules...

Model

Dataset

Type Strategy

Batch Size Gradient Accumulation

Epochs Max Length

Learning Rate Flags

Benchmark Run

Models

Types Batches

Max Length World Size

Safety Margin Mode

Loading benchmark profile...

On Apple Silicon, this benchmark is most useful as an estimate/plan surface. The multi-node cluster workflow remains Linux + AMD oriented.

Benchmark Charts

Show Rolling Average Show Latency Delta Normalize to Zero Baseline

Latency Trend (Instantaneous vs Smoothed)

EMA Smoothed Instantaneous Latency Rolling Avg (N=20)

Draft vs Target Latency Delta

Latency Delta (target - draft)

Acceptance Timeline

Accepted Rejected

Inference

Load a saved adapter directory or trained model directory from the workspace, then run prompts directly from the browser.

Saved Artifact

Base Model Override

Prompt

Max New Tokens Temperature

Top P Resolved Base Model

Completion

No inference run yet.

GGUF Export

Export a saved adapter/model directory to GGUF using export_gguf.py. This runs as a cancellable background job and streams logs in the Jobs panel.

Example Preset Apply

Saved Artifact Source

Source Kind Base Model Override

Manual Source Path (optional override)

Export Name Quantization

Output Dir llama.cpp Dir

Trust Remote Code Keep Merged Model

GGUF Outputs

Jobs

Job history is persisted to `workspace/dashboard-jobs.json` so completed runs survive dashboard restarts.

Job Output

No job selected.