Manish Biswal · Portfolio

01Deep learningdocs/dl/

The 32-Bit Trap: how a 4 GB tensor becomes a wild pointer
aug 2026megakernelscute dslfeatured22 min

Fixing the 32-bit addressing trap of CuTe DSL encountered while writing qwen2.5 14b megakernel
Vectorization and coalescing: the two ways to move memory fast
jul 2026gpu kernelscutlass cute4 min

Two independent knobs decide kernel bandwidth: how wide each thread's load is, and whether adjacent lanes touch adjacent addresses. CuTe gives you exactly one call for each.
Optimizer state: the memory bank of deep learning
jun 2026optimizersadamw6 min

What AdamW remembers per parameter, why bias correction exists, and why mixed precision forces an FP32 master copy of the weights.
Visualizing LLM tensors: from tokens to batch matrices
may 2026transformerstensors6 min

From tokenizer IDs to padded batches and attention masks: every shape a prompt passes through, drawn so every number fits on screen.
PyTorch basics: tensors, autograd, and the training loop
may 2026pytorchcheat sheet8 min

Tensor mechanics, autograd with a hand-checked gradient example, the five-line training loop, and the pitfalls behind ninety percent of runtime errors.

Deploying agents on scale: what nobody tells you about scaling coding sandboxes
may 2026ai agentsinfrastructure7 min

Queues instead of open connections, ephemeral sandbox pods instead of bare shells, JSONL logs instead of memory pressure. What breaks when ten agents share one VM.
Building a tiny pub/sub system in TypeScript
may 2026typescriptpub/sub5 min

One Map, one Set, async iterators: a GraphQL-ready event bus in a single dependency-free class.

How to start contributing to open source: a guide for newcomers
may 2026mentorshipgit8 min

Tiny doc fixes this week, Hacktoberfest in October, GSoC or LFX within a year: contribution types, program calendars, skills, and a repeatable routine.