GPUs for ML research · Crypto payment without KYC
IteraGPU
Laboratory / ML path: five questions, five starting points
IteraGPU / ML research

Which question is holding you up today?

From the first load to checkpoints, organize the points to verify before a compute campaign. Find guides to size, measure and keep results you can use.

Choose the question that's blocking you. Each path leads to a specific decision; the library gathers all the references.

My model hasn't run yet. Where do I start?

A memory envelope and the inputs for the first trial.

  1. Sizing GPU memory: computation, reservations and validation

    Compute weights in GiB, build an explicit reserve and validate the peak phase by phase. Worked examples, quantization pitfalls and per-card selection.

  2. GPU memory: explaining the gap between estimate and peak

    Distinguish weights, cache and activations, measure allocated and reserved by phase, then choose a margin with a notebook and a PyTorch protocol.

  3. KV cache: calculating memory by context, GQA and batch

    Calculate the KV cache using KV heads, context and concurrent sequences. Distinguish GQA, precision and static cache before a real trial.

The score is improving, but can I trust it?

An isolated evaluation set, reviewed errors and a qualified difference.

  1. Building an ML evaluation set without data leakage

    Separate groups, duplicates, and data uses to build a held-out ML evaluation set that is controllable and suited to your decision.

  2. Analyze ML errors to choose the next experiment

    Analyze classification and extraction errors: confusion matrix, references, taxonomy, regressions and verification of the next fix.

  3. Variability across seeds: is an ML gap convincing?

    Interpret ML repetitions with paired seeds, dispersion, and a minimum effect threshold. Worked example and limitations of a small sample.

I'm preparing a training run. What should I check before the run?

A full update, an explicit batch and usable traces.

  1. GPU training: validate the loop before the campaign

    Prepare the data, the batch and the resumption of a training run. One protocol per phase to control memory, gradients, validation and output files.

  2. Microbatch and accumulation: computing the effective batch

    Compute the effective batch, normalize the loss and verify gradient accumulation across one or more cards, along with its equivalence limits.

  3. Tracking your ML experiments: from run to decision

    Connect data, code, parameters, predictions and decision with an ML experiment manifest. An example of run and artifact tracking without a mandated tool.

Adapt or compress the model: how do I choose a variant?

A baseline, a controlled change and comparable quality.

  1. Fine-tuning: choosing an adaptation and verifying its contribution

    Prepare a fine-tuning with a reference, separate data, and a reload check. LoRA, quantized base, budget, and regression analysis.

  2. Comparing quantizations: memory, quality, and useful cost

    Compare reference, calibration, weight format, and KV cache with the same data. Measure memory and quality before deciding on the GPU budget.

  3. Planning ML ablations: control, effects and budget

    Build an ML ablation plan with a control, factors, interactions and a trial budget. A worked example to prioritize variants and reach a conclusion.

The budget is limited. Which trials should I run first?

An order of experiments and the plan's cost weighed against useful results.

  1. Budgeting an ML campaign: runs, decisions, and the full package

    Break down an ML campaign, prioritize variants, and tie the GPU package to useful results, with an example schedule and budget in USD.

  2. Planning ML ablations: control, effects and budget

    Build an ML ablation plan with a control, factors, interactions and a trial budget. A worked example to prioritize variants and reach a conclusion.

  3. ML benchmark: comparing cost at equivalent quality

    Fix the quality, measure the same corpus and compare the full cost of 3-, 7- or 30-day GPU plans. Protocol and raw table available for download.