diff --git a/docs/src/snippets/configs/meta-models/muse-spark.jsx b/docs/src/snippets/configs/meta-models/muse-spark.jsx index ff9c49f59c..a2d7e1ffa0 100644 --- a/docs/src/snippets/configs/meta-models/muse-spark.jsx +++ b/docs/src/snippets/configs/meta-models/muse-spark.jsx @@ -390,7 +390,9 @@ export const config = { { match: { hw: "gb300", quant: "bf16" }, nnodes: 2, - env: [], + // Expandable segments admit the two-node NVL72 TP8 group to custom + // all-reduce v2; without them decode all-reduce falls back to NCCL. + env: ["PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True"], flags: [ "--model-path {{MODEL_NAME}}", "--tp-size 8",