Skip to content

Add TPU model performance optimization use case (tpu_performance_autoresearch_wiki)#15

Open
vlasenkoalexey wants to merge 2 commits into
WecoAI:masterfrom
vlasenkoalexey:add-tpu-perf-autoresearch
Open

Add TPU model performance optimization use case (tpu_performance_autoresearch_wiki)#15
vlasenkoalexey wants to merge 2 commits into
WecoAI:masterfrom
vlasenkoalexey:add-tpu-perf-autoresearch

Conversation

@vlasenkoalexey

Copy link
Copy Markdown

Adds a use-case row to the Use Cases table for vlasenkoalexey/tpu_performance_autoresearch_wiki.

It applies the autoresearch keep/discard loop to TPU model performance (MFU / tokens-per-sec) on v6e-8 — profiling each run through an XProf MCP server, making one model-code change per experiment, and keeping or reverting against measured MFU. Includes Llama3-8B and Qwen3-8B case studies across JAX and torchax lanes; several agent+harness stacks surpass the MaxText reference.

Per the contributing note on verifiable traces, the full exploration trace is public:

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant