Skip to content

CLAUDE.md

Latest commit

 

History

History
28 lines (23 loc) · 1.2 KB

CLAUDE.md

File metadata and controls

28 lines (23 loc) · 1.2 KB
Raw

Etha

Run everything with pixi run. dev and vllm both define a submit task, so the env flag is mandatory: pixi run -e dev <task>.

Verifying changes to src/etha/comm/ or src/etha/tensor_bus/

Unit tests only cover disjoint-mesh CPU paths — run all three layers:

  1. Unit tests (CPU/gloo, ~9 min): 
    pixi run -e dev pytest tests/test_communication_cpu.py \
  tests/test_communication_symmetric_mesh.py \
  tests/test_communication_replicate_shard.py tests/test_partial_chunk_reduce.py
  2. Transfer benchmark (GPU/NCCL): pixi run -e dev submit bench/run_benchmark.sh. Quick single-node check (P2P/BROADCAST/SHADOW paths): 
    ETHA_BENCH_SMOKE=1 pixi run -e dev torchrun --nnodes=1 --nproc-per-node=8 \
  --master-addr=localhost --master-port=29555 ./bench/transfer_benchmark.py
  3. vLLM weight-sync example (8 GPUs, the only end-to-end production-path check): pixi run -e vllm vllm_weight_sync. Set HF_HOME=/data/hf (cached 30B model). Long-running — monitor under an agent, not inline.

Gotchas:

  • vLLM env build failing with Permission denied in .pixi/build/work/: a stale read-only git checkout — chmod -R u+w .pixi/build/work/vllm-* then retry.