mljourney

Off-the-shelf embedding models underperform on domain-specific retrieval. A practical guide to synthetic query generation, hard negative mining, contrastive fine-tuning, and Matryoshka training — with realistic expectations on NDCG@10 improvement.

AWQ, GPTQ, and bitsandbytes each quantize LLM weights differently. A practical comparison covering how each works, quality at 4-bit and 8-bit, inference speed, and when to use which — including the right choice for QLoRA training vs production serving.

FlashAttention-2 delivers 2–4x faster attention and eliminates O(N²) memory with a single argument change. A practical guide to enabling it in HuggingFace, PyTorch SDPA, and fine-tuning pipelines — including attention mask compatibility and where it helps most.

RLHF with PPO is powerful but complex. DPO achieves comparable alignment with a fraction of the engineering overhead. A practical comparison of PPO, DPO, and newer variants like SimPO and ORPO — covering when to use each and how to build a preference dataset that actually works.

Serving multiple fine-tuned model variants with separate deployments wastes GPU memory proportional to the number of adapters. A guide to multi-adapter serving with vLLM, S-LoRA architecture, adapter routing strategies, and GPU memory planning.

Feast, Tecton, and Hopsworks each take a different approach to feature store architecture. A practical comparison covering offline and online stores, streaming feature support, managed vs self-hosted trade-offs, and how to choose based on your actual requirements.

GPU sitting at 40-60% utilization while the model code looks fine? The dataloader is likely the bottleneck. A systematic guide to diagnosing and fixing slow data loading in PyTorch training pipelines.

Gradient accumulation and gradient checkpointing are frequently confused but solve different problems. A precise guide to how each works, when to use them, how to combine them, and how to reason about GPU memory during training.

A practical guide to how attention actually works — scaled dot-product, multi-head, MQA, GQA, Flash Attention, and RoPE — with the implications for memory, throughput, and context length that matter for production deployments.

Handwritten test cases give false confidence. Building an eval dataset from production logs — with stratified sampling, proper labeling, and slice-based reporting — produces evaluations that actually catch regressions before they reach users.