Articles

Searchable, citable, durable content on LLMs and AI agents. 128 articles, page 2 / 13.

• LLM Pretraining Data Deduplication in Production: Using MinHash-LSH, Semantic Deduplication, and Contamination Gates to Reduce Memorization and Benchmark Leakage

  Template pages, reposted articles, and benchmark variants in LLM pretraining corpora amplify memorization, waste compute, and pollute evaluations. This article presents a complete production pipeline from exact deduplication and MinHash-LSH to semantic deduplication, benchmark isolation, and threshold rollback, covering a four-layer architecture, engineering pipeline, threshold calibration, and common pitfalls.

  07/14/2026 LLM / Pretraining Data / Data Deduplication / MinHash / Benchmark Contamination / Semantic Deduplication

• LLM Gradient Accumulation in Production: Token-Normalized Loss, DDP no_sync, and Scheduler Alignment to Avoid Equivalent Batch Distortion

  Gradient accumulation seems like simple batch summing, but variable-length samples, distributed sync, mixed precision, and learning rate schedulers silently alter the equivalent batch. This article dives into production-grade solutions like token normalization, DDP no_sync, AMP update boundaries, and tail window handling to achieve mathematically consistent gradient accumulation in LLM fine-tuning.

  07/14/2026 LLM Training / Gradient Accumulation / PyTorch DDP / Fine-Tuning / Mixed Precision

• LLM Activation Checkpointing in Production: Selective Recompute, Non-Reentrant Mode, and RNG Consistency to Reduce Training Memory

  During large model training, activations often exhaust GPU memory before parameters. This article explores selective recomputation, non-reentrant mode, RNG state consistency, and engineering gating strategies to reduce peak memory while preserving gradient correctness and controlling throughput loss.

  07/14/2026 LLM Training / Activation Checkpointing / PyTorch / GPU Memory / Distributed Training

• LLM FP8 Fine-Tuning in Production: Amax Scaling, Layer Whitelists, and BF16 Fallback for Stable Convergence

  FP8 reduces compute and memory for LLM fine-tuning, but poor scaling, outliers, and distributed sync issues can cause silent accuracy degradation. This guide provides a complete engineering solution with Amax monitoring, layer whitelists, BF16 fallback, dual-track evaluation, and release gates.

  07/14/2026 LLM Training / FP8 / Transformer Engine / Fine-tuning / MLOps / Mixed Precision Training / Distributed Training / Amax Scaling

• LLM Model Merging in Production: Reducing Parameter Interference with TIES/DARE, Calibration Sets, and Capability Regression Gates

  Model merging combines multiple fine-tuned models without retraining, but parameter conflicts, tokenizer mismatches, and capability overlap amplify production risks. This article covers TIES, DARE, calibration set selection, artifact fingerprints, and regression gates for production deployment.

  07/14/2026 LLM / Model Merging / TIES / DARE / MLOps

• LLM SFT Sequence Packing in Production: Boosting Throughput with Sample Boundaries, Loss Mask, and Length Buckets

  A practical guide to implementing safe Sequence Packing for SFT training, covering attention isolation, loss masking, position IDs, length bucketing, and validation gates to eliminate padding waste without cross-sample contamination.

  07/14/2026 LLM Training / SFT / Sequence Packing / FlashAttention / Data Collator

• LLM Unicode Output Governance in Production: Incremental UTF-8 Decoding, Grapheme Clusters, and NFC Boundary Buffering to Prevent Garbled Text and Truncation

  A systematic guide for streaming LLM applications covering three layers of boundaries—network bytes, API deltas, and user-visible characters—with solutions for incremental UTF-8 decoding, extended grapheme cluster buffering, NFC normalization, and replay testing to avoid garbled text, emoji splitting, and index mismatches.

  07/13/2026 LLM Engineering / Unicode / Streaming / Text Processing / Character Encoding / Grapheme Cluster

• Distributed LLM Fine-Tuning Checkpoint Resume in Production: Sharded Checkpoints, Atomic Commits, and World Size Changes

  A practical guide to production-grade checkpoint and resume for distributed LLM fine-tuning. Covers sharded checkpoints, atomic commits, optimizer/RNG state recovery, world size changes, and fault injection drills to prevent silent training drift.

  07/13/2026 LLM Training / Distributed Checkpoint / FSDP / Fault Tolerance

• LLM GPU Energy Efficiency in Production: Phase-Aware Clock Locking, Joules/Token, and Thermal Throttling Guardrails

  Setting a Power Cap alone often ignores the load differences between Prefill and Decode. This article explains how to use phase-aware clock locking, per-token energy metrics, and thermal throttling guardrails to manage inference energy efficiency without breaking latency SLOs.

  07/13/2026 LLM Serving / GPU Energy Efficiency / NVIDIA DCGM / Clock Lock / Performance Engineering

• LLM Distributed Inference Communication Fault Debugging: Locating Collective Communication Hangs with NCCL RAS, Flight Recorder, and Topology Baselines

  When NCCL collective communication hangs in large-scale distributed LLM inference, you're often left with only a vague timeout log. This article explains how to use RAS, Flight Recorder, topology baselines, and anomalous rank detection to build observable, rollback-capable communication gating for production deployments.

  07/13/2026 LLM Serving / NCCL / Distributed Inference / GPU / Reliability