fix: use pipeline parallelism instead of TP for reasoning-predictor (bnb prequant blocks TP)
This commit is contained in:
@@ -28,7 +28,8 @@ spec:
|
||||
- --quantization=bitsandbytes
|
||||
- --dtype=float16
|
||||
- --kv-cache-dtype=auto
|
||||
- --tensor-parallel-size=2
|
||||
- --tensor-parallel-size=1
|
||||
- --pipeline-parallel-size=2
|
||||
- --max-model-len=16384
|
||||
- --gpu-memory-utilization=0.90
|
||||
- --max-num-seqs=4
|
||||
|
||||
Reference in New Issue
Block a user