mlperf/resnet: update beam params to increase time and quality (#4396)

* mlperf/resnet: update beam params to increase time and quality

* revert upcast 8 in search space and add rocm setup function

* refactor to independent setup.sh script
This commit is contained in:
Francis Lam
2024-05-02 17:14:46 -07:00
committed by GitHub
parent ca6c8ae739
commit 3cf8291f2f
7 changed files with 15 additions and 6 deletions

View File

@@ -4,7 +4,7 @@ export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=48 LR=7
export SPLIT_REDUCEOP=1 LAZYCACHE=0 RESET_STEP=0
export TRAIN_BEAM=3 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=128 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=50
export TRAIN_BEAM=3 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
export BENCHMARK=10 DEBUG=2

View File

@@ -4,7 +4,7 @@ export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=48 LR=7
export SPLIT_REDUCEOP=1 LAZYCACHE=0 RESET_STEP=0
export TRAIN_BEAM=3 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=128 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=50
export TRAIN_BEAM=3 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
export EVAL_START_EPOCH=3 EVAL_FREQ=4

View File

@@ -7,7 +7,7 @@ export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=48 LR=7
export SPLIT_REDUCEOP=1 LAZYCACHE=0 RESET_STEP=0
export TRAIN_BEAM=3 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=128 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=50
export TRAIN_BEAM=3 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
# pip install -e ".[mlperf]"
export LOGMLPERF=1

View File

@@ -1,10 +1,12 @@
#!/bin/bash
export PYTHONPATH="."
export MODEL="resnet"
export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=48 LR=7
export SPLIT_REDUCEOP=1 LAZYCACHE=0 RESET_STEP=0
export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=128 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=25
export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
export BENCHMARK=10 DEBUG=2

View File

@@ -1,10 +1,12 @@
#!/bin/bash
export PYTHONPATH="."
export MODEL="resnet"
export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=48 LR=7
export SPLIT_REDUCEOP=1 LAZYCACHE=0 RESET_STEP=0
export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=128 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=25
export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
export EVAL_START_EPOCH=3 EVAL_FREQ=4

View File

@@ -7,7 +7,7 @@ export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=48 LR=7
export SPLIT_REDUCEOP=1 LAZYCACHE=0 RESET_STEP=0
export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=128 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=25
export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
# pip install -e ".[mlperf]"
export LOGMLPERF=1

View File

@@ -0,0 +1,5 @@
#!/bin/bash
rocm-smi --setprofile compute
rocm-smi --setmclk 3
rocm-smi --setperflevel high