Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 27 additions & 0 deletions docs/blog/tensor-parallelism-optimization.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
# Optimizing Tensor Parallelism in JAX and XLA

> **Draft** — this page is a placeholder. Content is in progress.

## Background

{/* Why TP matters: model sizes past single-GPU memory, the compute/communication
tradeoff versus FSDP and pipeline parallelism. */}

## Where tensor parallelism costs performance

{/* All-gather / reduce-scatter on the critical path, exposed communication,
small GEMM tiles at high TP degree, NVLink versus inter-node bandwidth. */}

## Optimizations

{/* One subsection per optimization, with the mechanism and how to enable it. */}

## Results

{/* Model, hardware, TP degrees swept, throughput and step-time deltas. */}

## Reproducing

{/* Container tag, framework config flags, launch command. */}

## Takeaways
6 changes: 6 additions & 0 deletions docs/index.yml
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,12 @@ navigation:
- page: nsys-jax
path: nsys-jax.md

# ==================== Tech Blogs ====================
- section: Tech Blogs
contents:
- page: Optimizing Tensor Parallelism in JAX and XLA
path: blog/tensor-parallelism-optimization.md

# ==================== Resiliency ====================
- section: Resiliency
contents:
Expand Down
Loading