forked from tinygrad/tinygrad
llama: correct optim_dtype for mxfp4 (#17482)
This commit is contained in:
+1
-1
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
|
||||
export SPLIT_W13=${SPLIT_W13:-0}
|
||||
export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
|
||||
|
||||
export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
|
||||
export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
|
||||
export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
|
||||
export GBS=$((BS * GRADIENT_ACC_STEPS))
|
||||
|
||||
|
||||
+1
-1
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
|
||||
export SPLIT_W13=${SPLIT_W13:-0}
|
||||
export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
|
||||
|
||||
export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
|
||||
export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
|
||||
export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
|
||||
export GBS=$((BS * GRADIENT_ACC_STEPS))
|
||||
|
||||
|
||||
Reference in New Issue
Block a user