Using Training Optimization

Using Training Optimization

Overview

This meta-skill routes you to the right training optimization specialist based on symptoms. Training issues often have multiple potential causes—this skill helps diagnose symptoms and route to the appropriate specialist. Load this skill when you encounter training problems but aren't sure which specific technique to apply.

Core Principle: Diagnose before routing. Training issues often have multiple causes. Ask clarifying questions to understand symptoms before routing to specific skills. Wrong diagnosis wastes time—systematic routing saves it.

Knowledge cutoff (2026-05): The optimizer / schedule / precision landscape covered in this pack is calibrated to early-2026. Modern entries (Lion, Sophia, Muon, AdEMAMix, Schedule-Free, Prodigy, AdamW8bit / paged optimizers, FP8, WSD schedules, muP/mu-Transfer, ZeRO/FSDP strategy choice) are all routable. AdamW + cosine schedule + BF16 mixed precision remains the boring-and-correct default for most workloads; modern alternatives are pointed-tool replacements with documented trade-offs, not blanket upgrades. The router does not hardcode model IDs or vendor-specific recipes — capability tiers and symptom signatures drive routing.

Cross-pack frame (frozen vocabulary):

• training-optimization (this pack) covers training dynamics: optimizers, schedules, precision, batch size, gradients, loss objectives, regularization, in-flight experiment tracking.
• yzmir-llm-specialist (llm-finetuning-strategies.md) owns preference-tuning method choice: DPO, GRPO, SimPO, ORPO, KTO, IPO, RLHF/RLAIF. The boundary: "should I use DPO or SFT for my task?" is a method question → llm-specialist. "Does Lion beat AdamW for my DPO run?" is a training-dynamics question → here (optimizer choice) plus llm-specialist (preference method).
• yzmir-pytorch-engineering owns the implementation surface: FSDP API, torch.compile, torch.amp, DataLoader internals, distributed bring-up. We make the strategy call (e.g. "you need ZeRO-3"); they handle the API.
• yzmir-ml-production owns the production registry / lineage. Training-side experiment tracking (in-flight logging, run comparison) is here; long-lived model registry, lineage, and deployment artifacts are there.

When to Use

Load this skill when:

• Model not learning (loss stuck, not decreasing, poor accuracy)
• Training instability (loss spikes, NaN values, divergence)
• Overfitting (large train/val gap, poor generalization)
• Training too slow (throughput issues, time constraints)
• Hyperparameter selection (optimizer, learning rate, batch size, regularization)
• Experiment management (tracking runs, comparing configurations)
• Convergence issues (slow learning, plateaus, local minima)
• Setting up new training pipeline
• Modern-optimizer questions (Lion, Sophia, Muon, AdEMAMix, Schedule-Free, Prodigy, 8-bit / paged optimizers)
• Modern-schedule questions (WSD / warmup-stable-decay, infinite-LR)
• Precision questions (FP8, BF16, mixed precision, loss scaling)
• Distributed-strategy choice (ZeRO stages, FSDP sharding strategy — the what, not the API)
• muP / mu-Transfer for hyperparameter transfer across scale

Don't use for: PyTorch implementation bugs and APIs (use pytorch-engineering), model architecture selection (use neural-architectures), production deployment / model registry (use ml-production), RL-specific training (use deep-rl), LLM preference-tuning method choice (use llm-specialist).

How to Access Reference Sheets

IMPORTANT: All reference sheets are located in the SAME DIRECTORY as this SKILL.md file.

When this skill is loaded from: skills/using-training-optimization/SKILL.md

Reference sheets like optimization-algorithms.md are at: skills/using-training-optimization/optimization-algorithms.md

NOT at: skills/optimization-algorithms.md ← WRONG PATH

When you see a link like [optimization-algorithms.md](optimization-algorithms.md), read the file from the same directory as this SKILL.md.

Routing by Primary Symptom

Symptom: "Model Not Learning" / "Loss Not Decreasing"

Keywords: stuck, flat loss, not improving, not learning, accuracy not increasing

Diagnostic questions (ask BEFORE routing):

1. "Is loss completely flat from the start, decreasing very slowly, or was it learning then stopped?"
2. "Any NaN or Inf values in your loss?"
3. "What optimizer and learning rate are you using?"
4. "What does your loss curve look like over time?"

Route based on answers:

Multi-skill routing: Often need optimization-algorithms (choose optimizer) + learning-rate-scheduling (choose LR/schedule) together for "not learning" issues.

Symptom: "Training Unstable" / "Loss Spikes" / "NaN Values"

Keywords: NaN, Inf, exploding, diverging, unstable, spikes, FP8 underflow, loss scale

Diagnostic questions:

1. "When do NaN values appear - immediately at start, or after N epochs?"
2. "What's your learning rate and schedule?"
3. "Are you using mixed precision training? FP16, BF16, or FP8?"
4. "What loss function are you using?"

Route to (in priority order):

1. gradient-management (PRIMARY)

   • When: Gradient explosion, NaN gradients
   • Why: Gradient issues cause instability. Must stabilize gradients first.
   • Techniques: Gradient clipping, gradient scaling, NaN debugging

2. learning-rate-scheduling (SECONDARY)

   • When: LR too high causes instability; or WSD decay phase mistuned
   • Why: High LR can cause divergence even with clipped gradients
   • Check: If NaN appears later in training, LR schedule might be increasing too much

3. batch-size-and-memory-tradeoffs (precision section)

   • When: Using FP8 / FP16 / BF16 and seeing under/overflow
   • Why: Mixed-precision loss-scale tuning and FP8 scaling factors live here
   • Cross-ref: pytorch-engineering for the torch.amp / FP8 API itself

4. loss-functions-and-objectives (if numerical issues)

   • When: Loss computation has numerical instability (log(0), division by zero)
   • Why: Numerical instability in loss propagates to gradients
   • Check: Custom loss functions especially prone to this

Cross-pack note: The strategy (which precision, when to clip, what loss-scale window) is here. The API (torch.amp.autocast, GradScaler, FP8 hooks) is in pytorch-engineering.

Symptom: "Model Overfits" / "Train/Val Gap Large"

Keywords: overfitting, train/val gap, poor generalization, memorizing training data

Diagnostic questions:

1. "How large is your dataset (number of training examples)?"
2. "What's the train accuracy vs validation accuracy?"
3. "What regularization are you currently using?"
4. "What's your model size (parameters) relative to dataset size?"

Route to (multi-skill approach):

1. overfitting-prevention (PRIMARY)

   • Techniques: Dropout, weight decay, early stopping, L1/L2 regularization
   • When: Always the first stop for overfitting
   • Why: Comprehensive regularization strategy

2. data-augmentation-strategies (HIGHLY RECOMMENDED)

   • When: Dataset is small (< 10K examples) or moderate (10K-100K)
   • Why: Increases effective dataset size, teaches invariances
   • Priority: Higher priority for smaller datasets

3. hyperparameter-tuning

   • When: Need to find optimal regularization strength
   • Why: Balance between underfitting and overfitting
   • Use: After implementing regularization techniques

Decision factors:

• Small dataset (< 1K): data-augmentation is CRITICAL + overfitting-prevention
• Medium dataset (1K-10K): overfitting-prevention + data-augmentation
• Large dataset (> 10K) but still overfitting: overfitting-prevention + check model capacity
• Model too large: Consider neural-architectures for model size discussion

Symptom: "Training Too Slow" / "Low Throughput"

Keywords: slow training, low GPU utilization, takes too long, time per epoch

CRITICAL: Diagnose bottleneck before routing

Diagnostic questions:

1. "Is it slow per-step (low throughput) or just need many steps?"
2. "What's your GPU utilization percentage?"
3. "Are you using data augmentation? How heavy?"
4. "What's your current batch size? Precision (FP32/BF16/FP8)?"

Route based on bottleneck:

Cross-pack boundaries:

• Data loading issues → pytorch-engineering (DataLoader, prefetching, num_workers)
• Distributed training API (FSDP wrap, DDP setup) → pytorch-engineering (distributed-training-strategies)
• Distributed strategy choice (ZeRO-1 vs 2 vs 3, FSDP full-shard vs hybrid) → optimization-algorithms (strategy section here) + batch-size-and-memory-tradeoffs
• Batch size optimization for speed/memory → training-optimization (batch-size-and-memory-tradeoffs)
• Profiling to identify bottleneck → pytorch-engineering (performance-profiling)

Key principle: Profile FIRST before optimizing. Low GPU utilization = wrong optimization target.

Symptom: "Which X Should I Use?" (Direct Questions)

Direct hyperparameter questions route to specific skills:

For new project setup, route to MULTIPLE in sequence:

1. optimization-algorithms - Choose optimizer (default: AdamW; consider modern alternatives only with cause)
2. learning-rate-scheduling - Choose initial LR and schedule (default: cosine + warmup; WSD when continual training)
3. batch-size-and-memory-tradeoffs - Choose batch size and precision (default: BF16 mixed precision)
4. experiment-tracking - Set up tracking
5. training-loop-architecture - Design training loop

Symptom: "Modern Optimizer / Schedule / Precision Question"

Keywords: Lion, Sophia, Muon, Shampoo, AdEMAMix, Schedule-Free, Prodigy, 8-bit optimizer, bitsandbytes, paged optimizer, WSD schedule, warmup-stable-decay, FP8, BF16, mixed precision, muP, mu-Transfer, ZeRO, FSDP strategy

Routing table:

Cross-pack split (preference tuning):

• "Should I use DPO or SFT?" → llm-specialist (method choice)
• "Lion vs AdamW for my DPO run?" → here for optimizer; llm-specialist for the DPO method itself
• "DPO is unstable, what do I do?" → BOTH: llm-specialist (β, reference model, pair quality) AND here (gradient-management, learning-rate-scheduling, optimization-algorithms)

Symptom: "Need to Track Experiments" / "Compare Configurations"

Keywords: experiment tracking, MLflow, wandb, tensorboard, compare runs, log metrics

Route to:

1. experiment-tracking (PRIMARY)

   • Tools: MLflow, Weights & Biases, TensorBoard, Neptune
   • When: Setting up tracking, comparing runs, organizing experiments during training
   • Why: Systematic in-flight experiment management

2. hyperparameter-tuning (if systematic search)

   • When: Running many configurations systematically
   • Why: Automated hyperparameter search integrates with tracking

3. training-loop-architecture (for integration)

   • When: Need to integrate tracking into training loop
   • Why: Proper callback and logging design

Cross-pack boundary: Long-lived model registry, model lineage, and production artifact tracking → yzmir-ml-production / experiment-tracking-and-versioning.md. This pack covers the training-time logging surface; ml-production owns what happens after the run is done.

Cross-Cutting Multi-Skill Scenarios

Scenario: New Training Setup (First Time)

Route to (in order):

1. optimization-algorithms - Select optimizer (boring default: AdamW; modern alternatives only with documented cause)
2. learning-rate-scheduling - Choose LR and warmup strategy (default: cosine; WSD if you'll resume/continue)
3. batch-size-and-memory-tradeoffs - Determine batch size and precision (default: BF16)
4. loss-functions-and-objectives - Verify loss function appropriate for task
5. experiment-tracking - Set up experiment logging
6. training-loop-architecture - Design training loop with checkpointing

Why this order: Foundation (optimizer/LR/batch/precision) → Objective (loss) → Infrastructure (tracking/loop)

Scenario: Convergence Issues

Route to (diagnose first, then in order):

1. gradient-management - Check gradients not vanishing/exploding (use gradient checking)
2. learning-rate-scheduling - Adjust LR schedule (might be too high/low/wrong schedule)
3. optimization-algorithms - Consider different optimizer if current one unsuitable

Why this order: Stability (gradients) → Adaptation (LR) → Algorithm (optimizer)

Common mistakes:

• Jumping to change optimizer without checking gradients
• Blaming optimizer when LR is the issue
• Not using gradient monitoring to diagnose
• Switching to Lion/Sophia/Muon because it's modern, without verifying AdamW was tuned

Scenario: Overfitting Issues

Route to (multi-pronged approach):

1. overfitting-prevention - Implement regularization (dropout, weight decay, early stopping)
2. data-augmentation-strategies - Increase effective dataset size
3. hyperparameter-tuning - Find optimal regularization strength

Why all three: Overfitting needs comprehensive strategy, not single technique.

Prioritization:

• Small dataset (< 1K): data-augmentation is MOST critical
• Medium dataset (1K-10K): overfitting-prevention + augmentation balanced
• Large dataset but still overfitting: overfitting-prevention + check model size

Scenario: Training Speed + Memory Constraints

Route to:

1. batch-size-and-memory-tradeoffs (PRIMARY) - Gradient accumulation, 8-bit/paged optimizers, BF16/FP8 precision, ZeRO/FSDP memory math
2. optimization-algorithms - Strategy choice for distributed (ZeRO stage, FSDP sharding)
3. pytorch-engineering (cross-pack) - FSDP/DDP API, memory profiling, torch.compile
4. data-augmentation-strategies - Reduce augmentation overhead if bottleneck

Why: Speed and memory are coupled. We make the strategy call (precision, optimizer memory, sharding choice); pytorch-engineering provides the API.

Scenario: Multi-Task Learning or Custom Loss

Route to:

1. loss-functions-and-objectives (PRIMARY) - Multi-task loss design, uncertainty weighting
2. gradient-management - Check gradients per task, gradient balancing
3. hyperparameter-tuning - Tune task weights, loss coefficients

Why: Custom losses need careful design + gradient analysis + tuning.

Scenario: Preference Tuning (DPO / GRPO / SimPO / etc.)

This is intentionally a cross-pack scenario.

1. yzmir-llm-specialist / llm-finetuning-strategies.md (PRIMARY) - Method choice (DPO vs GRPO vs SimPO vs ORPO vs KTO vs IPO vs RLHF), β / reference-model handling, pair construction, reward modeling.
2. optimization-algorithms (here) - Optimizer choice for the run (AdamW default; Lion/Sophia/Muon only if symptom-driven).
3. learning-rate-scheduling (here) - LR / warmup / schedule for preference-tuning runs (typically smaller LR than SFT).
4. gradient-management (here) - Preference losses can spike; clipping and monitoring are pack-here concerns.
5. batch-size-and-memory-tradeoffs (here) - Reference-model memory cost shapes batch/precision choices.

Routing rule: If the user asks "which preference method?" the answer comes from llm-specialist. If they ask "my DPO run is unstable, what optimizer / LR / clipping?" the answer is both packs jointly — never assume optimizer choice alone fixes a preference-tuning instability.

Ambiguous Queries - Clarification Protocol

When symptom unclear, ASK ONE diagnostic question before routing:

Never guess when ambiguous. Ask once, route accurately.

Common Routing Mistakes

Key principle: Diagnosis before solutions, clarification before routing, multi-skill for complex issues.

When NOT to Use Training-Optimization Pack

Skip training-optimization when:

Training-optimization pack is for: Framework-agnostic training algorithms, hyperparameters, optimization techniques, and training strategies that apply across architectures.

Boundaries:

• PyTorch implementation/infrastructure/API → pytorch-engineering
• Architecture selection → neural-architectures
• Production/inference/registry → ml-production
• RL-specific algorithms → deep-rl
• LLM preference-tuning method choice → llm-specialist

Red Flags - Stop and Reconsider

If you catch yourself about to:

• ❌ Suggest specific optimizer without routing → Route to optimization-algorithms
• ❌ Suggest learning rate value without diagnosis → Route to learning-rate-scheduling
• ❌ Say "add dropout" without comprehensive strategy → Route to overfitting-prevention
• ❌ Suggest "reduce batch size" without profiling → Check if memory or speed issue, route appropriately
• ❌ Give trial-and-error fixes for NaN → Route to gradient-management for systematic debugging
• ❌ Recommend a modern optimizer (Lion/Sophia/Muon) because it's "newer" → Route to optimization-algorithms and consult the comparison table; AdamW is the default
• ❌ Treat a DPO/GRPO/SimPO instability as a pure optimizer problem → Route to llm-specialist (method) jointly with this pack (dynamics)
• ❌ Recommend FP8 because it's fastest, ignoring loss-scale risk → Route to batch-size-and-memory-tradeoffs (precision) + gradient-management
• ❌ Provide generic training advice → Identify specific symptom and route to specialist
• ❌ Accept user's self-diagnosis without verification → Ask diagnostic questions first

All of these mean: You're about to give incomplete advice. Route to specialist instead.

Common Rationalizations (Don't Do These)

If you catch yourself thinking ANY of these, STOP and route to specialist or ask diagnostic question.

Pressure Resistance - Critical Discipline

Time/Emergency Pressure

Emergency protocol: Fast clarification (30 sec) → Correct routing (60 sec) → Specialist handles efficiently

Authority/Hierarchy Pressure

Authority protocol: Acknowledge → Verify symptoms → Route based on evidence, not opinion

User Self-Diagnosis Pressure

Verification protocol: User attribution is hypothesis, not diagnosis. Verify with symptoms.

Red Flags Checklist - Self-Check Before Routing

Before giving ANY training advice or routing, ask yourself:

1. ❓ Did I identify specific symptoms?

   • If no → Ask clarifying question
   • If yes → Proceed

2. ❓ Is this symptom in my routing table?

   • If yes → Route to specialist skill
   • If no → Ask diagnostic question

3. ❓ Am I about to give direct advice?

   • If yes → STOP. Why am I not routing?
   • Check rationalization table—am I making excuses?

4. ❓ Could this symptom have multiple causes?

   • If yes → Ask diagnostic question to narrow down
   • If no → Route confidently

5. ❓ Is this training-optimization or another pack?

   • PyTorch errors / API → pytorch-engineering
   • Architecture choice → neural-architectures
   • Deployment / model registry → ml-production
   • Preference-tuning method choice → llm-specialist

6. ❓ Am I feeling pressure to skip routing?

   • Time pressure → Route anyway (faster overall)
   • Authority pressure → Verify symptoms anyway
   • User self-diagnosis → Confirm with questions anyway
   • Expert user → Route anyway (specialists help experts too)

If you failed ANY check, do NOT give direct advice. Route to specialist or ask clarifying question.

Training Optimization Specialist Skills

After routing, load the appropriate specialist skill for detailed guidance. The pack contains exactly 10 specialist sheets alongside this router:

1. optimization-algorithms.md — Optimizer selection: SGD / Adam / AdamW core, plus modern landscape (Lion, Sophia, AdEMAMix, Muon, Shampoo, Schedule-Free, Prodigy, AdamW8bit / paged optimizers), and ZeRO/FSDP strategy choice (cross-ref pytorch-engineering for API).
2. learning-rate-scheduling.md — LR schedulers (cosine, step, exponential, WSD / warmup-stable-decay), warmup strategies, infinite-LR variants, cyclical learning rates.
3. loss-functions-and-objectives.md — Custom losses, multi-task learning, weighted objectives, numerical stability. Preference losses (DPO/GRPO/etc.) are pointers to yzmir-llm-specialist.
4. gradient-management.md — Gradient clipping, accumulation, scaling, vanishing/exploding gradient diagnosis, NaN debugging, AMP/FP8 loss-scale interaction.
5. batch-size-and-memory-tradeoffs.md — Batch size effects on convergence, gradient accumulation, precision strategy (FP32 / BF16 / FP16 / FP8), memory math for ZeRO/FSDP. (No separate precision sheet exists; precision lives here.)
6. data-augmentation-strategies.md — Augmentation techniques (geometric, color, mixing), policy design, AutoAugment / RandAugment / TrivialAugment, MixUp / CutMix.
7. overfitting-prevention.md — Regularization (L1/L2, dropout, weight decay), early stopping, label smoothing, generalization techniques.
8. training-loop-architecture.md — Training loop design, monitoring, logging, checkpointing integration, callbacks, gradient accumulation orchestration.
9. hyperparameter-tuning.md — Search strategies (grid, random, Bayesian, ASHA, BOHB), AutoML, Optuna, Ray Tune, muP / mu-Transfer for scale-up.
10. experiment-tracking.md — In-flight tracking with MLflow, Weights & Biases, TensorBoard, Neptune, run comparison. (Production registry / lineage → yzmir-ml-production / experiment-tracking-and-versioning.md.)

Quick Reference: Symptom → Skills

Integration Notes

Cross-pack boundaries (bidirectional):

• yzmir-pytorch-engineering — FSDP API, torch.compile, torch.amp, DataLoader, distributed bring-up, memory profiling. We pick the strategy; they implement it.
• yzmir-llm-specialist (llm-finetuning-strategies.md) — Preference-tuning method choice (DPO / GRPO / SimPO / ORPO / KTO / IPO), RLHF/RLAIF, fine-tuning recipes. We supply training dynamics for those runs (optimizer / LR / clipping / precision).
• yzmir-ml-production (experiment-tracking-and-versioning.md) — Long-lived model registry, lineage, deployment artifacts. We own in-flight logging during training; they own the artifact lifecycle after.
• yzmir-neural-architectures — Architecture selection precedes training. We don't choose architectures.
• yzmir-deep-rl — RL-specific training (PPO/SAC/etc.) lives there. General optimizer/LR/precision still cross-refs here.

Summary

This meta-skill: Diagnose symptoms → Route to specialists → Resist pressure to give direct advice. Training issues often have multiple causes. Clarify symptoms, route to specialists. Wrong routing wastes more time than asking one clarifying question.

Modern landscape (2026-05) is integrated: Lion, Sophia, Muon, AdEMAMix, Schedule-Free, Prodigy, AdamW8bit, paged optimizers, WSD schedules, FP8, muP, ZeRO/FSDP strategy. AdamW + cosine + BF16 remains the boring-and-correct default — modern alternatives are symptom-driven choices, not blanket upgrades.

Cross-pack discipline: Preference-tuning method choice (DPO/GRPO/SimPO/etc.) lives in yzmir-llm-specialist; PyTorch APIs live in yzmir-pytorch-engineering; production registry lives in yzmir-ml-production. Route deliberately across pack boundaries when the question spans them.

Route based on symptoms, not guesses.