llama: use ALL2ALL (#14353)

This commit is contained in:
nimlgen
2026-01-26 22:01:53 +03:00
committed by GitHub
parent 3f25eb3026
commit e152f1b0f5

View File

@@ -7,6 +7,7 @@ export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
export DEBUG=${DEBUG:-0}
export FLASH_ATTENTION=${FLASH_ATTENTION:-1}
export ALL2ALL=${ALL2ALL:-1}
export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
export DP=8 BS=8 EVAL_BS=8 GRADIENT_ACC_STEPS=1