Archives
- 11 Aug Megatron-LM Fine-Tuning Deep Dive: Qwen Recipes, Datasets, and the Training Loop
- 29 Jun Miles Deep Dive: Main Loop, Rollout, and Reward
- 26 Jun TVM FFI Deep Dive: Shipping Native Kernels Without Tying Them to One PyTorch Build
- 08 Jun Serving MoE Models: A Deep Dive into Parallelism Strategies (TP, EP, DP)
- 08 Jun NVSHMEM Deep Dive: One-Sided GPU Communication, and How It Differs from NCCL
- 05 Jun Mixture-of-Experts Across the Stack: From a 20-Line Reference to Wide Expert Parallelism
- 04 Jun Where gpu_memory_utilization Actually Goes: vLLM's Memory Budget, Line by Line
- 19 May GRPO from First Principles, and How verl Implements It
- 14 May Fused Linear Cross-Entropy: The Math, the Memory Wall, and Two Ways Around It
- 12 May Tracing the HuggingFace Trainer Dataloader Path — From `trainer.train()` to a `next(epoch_iterator)`
- 11 May PyTorch DDP Deep Dive — Initialization, Buffer Broadcast, the Reducer, and the torch.compile Story
- 10 May How vLLM's `torch.compile` Backend Handles Dynamic Batch Sizes — A Code Walkthrough
- 08 May Qwen3 Architecture Layer-by-Layer — Reading the HuggingFace Implementation
- 30 Apr Torch.compile Deep Dive II — Inductor Codegen and Buffer Lifetimes, End to End
- 28 Apr SGLang Deep Dive: How Engine.generate() Boots and Runs
- 27 Apr vLLM v1 Deep Dive: How LLM(...) and the Server Boot and Generate
- 24 Apr CUTLASS Deep Dive: From CuTe Layouts to Blackwell tcgen05
- 10 Apr NCCL Deep Dive: From P2P/CUMEM to torchrun Init