From 77c3481f579d88df93d1fa8d54d0c6d2460df4ee Mon Sep 17 00:00:00 2001
From: "pierre.delaunay" <delaunap@rtx5.server.mila.quebec>
Date: Wed, 3 Jul 2024 15:15:58 -0400
Subject: [PATCH] Add dry command regression tests

---
 milabench/cli/dry.py                          | 236 +++++++++
 milabench/commands/__init__.py                |  41 +-
 milabench/common.py                           |   5 +-
 milabench/metadata.py                         |   2 +-
 milabench/system.py                           |  54 ++-
 tests/conftest.py                             |  13 +
 tests/test_command_reg.py                     |  44 ++
 .../test_command_reg_one_node.txt             | 452 +++++++++++++++++
 .../test_command_reg_two_nodes.txt            | 454 ++++++++++++++++++
 9 files changed, 1276 insertions(+), 25 deletions(-)
 create mode 100644 milabench/cli/dry.py
 create mode 100644 tests/test_command_reg.py
 create mode 100644 tests/test_command_reg/test_command_reg_one_node.txt
 create mode 100644 tests/test_command_reg/test_command_reg_two_nodes.txt

diff --git a/milabench/cli/dry.py b/milabench/cli/dry.py
new file mode 100644
index 000000000..dabc5a1de
--- /dev/null
+++ b/milabench/cli/dry.py
@@ -0,0 +1,236 @@
+import os
+import shlex
+from contextlib import contextmanager
+from dataclasses import dataclass
+
+import voir.instruments.gpu as voirgpu
+import yaml
+from coleo import Option, tooled
+
+from ..common import get_multipack
+from ..multi import make_execution_plan
+
+
+class MockDeviceSMI:
+    def __init__(self, ngpu, capacity) -> None:
+        self.devices = [i for i in range(ngpu)]
+        self.used = [1]
+        self.total = [capacity]
+        self.util = [40]
+        self.temp = [35]
+        self.power = [225]
+
+    def get_gpu_info(self, device):
+        return {
+            "device": device,
+            "product": "MockDevice",
+            "memory": {
+                "used": self.used[0] // (1024**2),
+                "total": self.total[0] // (1024**2),
+            },
+            "utilization": {
+                "compute": float(self.util[0]) / 100,
+                "memory": self.used[0] / self.total[0],
+            },
+            "temperature": self.temp[0],
+            "power": self.power[0],
+            "selection_variable": "CUDA_VISIBLE_DEVICES",
+        }
+
+    @property
+    def arch(self):
+        return "mock"
+
+    @property
+    def visible_devices(self):
+        return os.environ.get("CUDA_VISIBLE_DEVICES", None)
+
+    def get_gpus_info(self, selection=None):
+        gpus = dict()
+        for device in self.devices:
+            if (selection is None) or (selection and str(device) in selection):
+                gpus[device] = self.get_gpu_info(device)
+
+        return gpus
+
+    def close(self):
+        pass
+
+
+@contextmanager
+def assume_gpu(ngpu=1, capacity=80000, enabled=False):
+    if enabled:
+        old = voirgpu.DEVICESMI
+        voirgpu.DEVICESMI = MockDeviceSMI(ngpu, capacity)
+        yield
+        voirgpu.DEVICESMI = old
+    else:
+        yield
+
+
+class BashGenerator:
+    def __init__(self) -> None:
+        self.indent = 0
+        self.background_mode = False
+        self.print("#!/bin/sh")
+        self.print("")
+
+    def print(self, *args, **kwargs):
+        print("  " * self.indent, end="")
+        print(*args, **kwargs)
+
+    def section(self, title):
+        self.echo("---")
+        self.echo(title)
+        self.echo("=" * len(title))
+
+    def echo(self, msg):
+        self.print(f'echo "{msg}"')
+
+    def comment(self, cmt):
+        self.print(f"# {cmt}")
+
+    def env(self, env):
+        for k, v in env.items():
+            self.print(f"export {k}={shlex.quote(v)}")
+        self.print()
+
+    @contextmanager
+    def subshell(self):
+        self.print("time (")
+        self.indent += 1
+        yield
+        self.indent -= 1
+        self.print(")")
+
+    @contextmanager
+    def background(self):
+        self.background_mode = True
+        yield
+        self.print("wait")
+        self.background_mode = False
+
+    def command(self, *args, env=None, **kwargs):
+        prefix = []
+        if env is not None:
+            for k, v in env.items():
+                prefix.append(f"{k}={v}")
+
+        prefix = " ".join(prefix)
+        sufix = ""
+        if True:
+            sufix = "&"
+
+        frags = [prefix] + [str(a) for a in args] + [sufix]
+
+        self.print(" ".join(filter(lambda x: x != "", frags)))
+
+
+# fmt: off
+@dataclass
+class Arguments:
+    nnodes: int = 2
+    ngpu: int = 8
+    capacity: int = 80000
+    withenv: bool = True
+    usevoir: bool = False
+# fmt: on
+
+
+@tooled
+def arguments():
+    ngpu: Option & int = 8
+    capacity: Option & int = 80000
+    nnodes: Option & int = 2
+
+    # [negate]
+    withenv: Option & bool = True
+
+    # [negate]
+    usevoir: Option & bool = True
+    return Arguments(nnodes, ngpu, capacity, withenv, usevoir)
+
+
+@tooled
+def multipack_args(conf: Arguments):
+    from ..common import arguments as multiargs
+
+    args = multiargs()
+    args.system = "system_tmp.yaml"
+
+    system = {
+        "system": {
+            "arch": "cuda",
+            "nodes": [
+                {
+                    "name": str(i),
+                    "ip": f"192.168.0.{i + 10}" if i != 0 else "127.0.0.1",
+                    "user": "username",
+                    "main": i == 0,
+                    "port": 22,
+                }
+                for i in range(conf.nnodes)
+            ],
+        }
+    }
+
+    with open("system_tmp.yaml", "w") as file:
+        system = yaml.dump(system)
+        file.write(system)
+
+    return args
+
+
+@tooled
+def cli_dry(args=None):
+    """Generate dry commands to execute the bench standalone"""
+    from ..commands import disable_voir
+    from ..system import enable_offline
+    from ..sizer import resolve_argv, scale_argv
+
+    if args is None:
+        args = arguments()
+
+    with disable_voir(enabled=False), enable_offline(enabled=True):
+        with assume_gpu(args.ngpu, args.capacity, enabled=True):
+            repeat = 1
+            mp = get_multipack(multipack_args(args), run_name="dev")
+            gen = BashGenerator()
+
+            first_pack = True
+            for index in range(repeat):
+                for pack in mp.packs.values():
+                    if first_pack and args.withenv:
+                        first_pack = False
+                        gen.section("Virtual Env")
+
+                        venv = pack.core._nox_session.env["VIRTUAL_ENV"]
+                        gen.env({"VIRTUAL_ENV": venv})
+                        gen.print("source $VIRTUAL_ENV/bin/activate")
+
+                        gen.section("Milabench")
+                        gen.env(pack.make_env())
+
+                    exec_plan = make_execution_plan(pack, index, repeat)
+
+                    gen.section(pack.config["name"])
+                    with gen.subshell():
+                        with gen.background():
+                            for pack, argv, _ in exec_plan.commands():
+                                
+                                sized_args = scale_argv(pack, argv)
+                                final_args = resolve_argv(pack, sized_args)
+
+                                gen.command(*final_args, env=pack.config.get("env", {}))
+
+                    print()
+
+        try:
+            os.remove("system_tmp.yaml")
+        except:
+            pass
+
+
+
+if __name__ == "__main__":
+    cli_dry()
diff --git a/milabench/commands/__init__.py b/milabench/commands/__init__.py
index a1c6da78a..3fd602030 100644
--- a/milabench/commands/__init__.py
+++ b/milabench/commands/__init__.py
@@ -6,6 +6,8 @@
 from copy import deepcopy
 from hashlib import md5
 from typing import Dict, Generator, List, Tuple
+from contextlib import contextmanager
+import warnings
 
 from voir.instruments.gpu import get_gpu_info
 
@@ -217,9 +219,7 @@ def _argv(self, **kwargs) -> List:
             abs_main = script
 
         if not abs_main.exists():
-            raise FileNotFoundError(
-                f"Cannot run script or directory because it does not exist: {script}"
-            )
+            warnings.warn(f"Could not find `{script}` did you run `milabench install` ?")
 
         if abs_main.is_dir():
             script = ["-m", str(script)]
@@ -460,16 +460,39 @@ def _argv(self, **kwargs):
         devices = self.pack.config.get("devices", [])
         nproc = len(devices)
         if nproc > 1:
-            argv = [*super()._argv(**kwargs), f"--nproc_per_node={nproc}", "--"]
+            argv = [*super()._argv(**kwargs), f"--nproc_per_node={nproc}"]
+
             # Check if the sub-executor targets a module or not
             cmd = next(iter(self.exec.argv()), None)
-            # if the command exists and it is not a path assume it is a module
-            if cmd and not XPath(cmd).exists():
-                argv.append("-m")
+
+            if cmd:
+                # python or voir; tell it to not prepend python since we are doing it
+                if cmd in ("python", "voir"):
+                    argv.append("--no-python")
+
+                # if the command exists and it is not a path assume it is a module
+                # script is not a file, maybe it is a module
+                elif not XPath(cmd).exists():
+                    argv.append("-m")
+
+            # everything after torchrun args are script args
+            argv.append("--")
             return argv
         return []
 
 
+use_voir = True
+
+
+@contextmanager
+def disable_voir(enabled):
+    global use_voir
+    old = use_voir
+    use_voir = enabled
+    yield
+    use_voir = old
+
+
 class VoirCommand(WrapperCommand):
     """Execute an `Command` through voir
 
@@ -491,6 +514,10 @@ def __init__(self, executor: SingleCmdCommand, *voir_argv, **kwargs) -> None:
     def _argv(self, **kwargs) -> List:
         argv = super()._argv(**kwargs)
 
+        if not use_voir:
+            # voir replace python
+            return ["python"]
+        
         if voirconf := self.pack.config.get("voir", None):
             hsh = md5(str(voirconf).encode("utf8"))
             voirconf_file = (
diff --git a/milabench/common.py b/milabench/common.py
index eab593fbd..f08320f80 100644
--- a/milabench/common.py
+++ b/milabench/common.py
@@ -196,11 +196,14 @@ def is_selected(defn, args):
     )
 
 def _get_multipack(
-    args: CommonArguments = arguments(),
+    args: CommonArguments = None,
     run_name=None,
     overrides={},
     return_config=False,
 ):
+    if args is None:
+        args = arguments()
+
     if args.config is None:
         args.config = os.environ.get("MILABENCH_CONFIG", None)
 
diff --git a/milabench/metadata.py b/milabench/metadata.py
index 392f01f12..a3bacacf3 100644
--- a/milabench/metadata.py
+++ b/milabench/metadata.py
@@ -62,7 +62,7 @@ def machine_metadata(pack=None):
             "machine": uname.machine,
         },
         "accelerators": gpus,
-        "date": datetime.datetime.now(datetime.UTC).timestamp(),
+        "date": datetime.datetime.utcnow().timestamp(),
         "milabench": retrieve_git_versions(
             __tag__,
             __commit__,
diff --git a/milabench/system.py b/milabench/system.py
index 6edbba93d..846cebfce 100644
--- a/milabench/system.py
+++ b/milabench/system.py
@@ -3,7 +3,7 @@
 import socket
 from dataclasses import dataclass, field
 import sys
-
+from contextlib import contextmanager
 import psutil
 import yaml
 from voir.instruments.gpu import get_gpu_info
@@ -163,21 +163,28 @@ def get_remote_ip():
 
 
 def _resolve_ip(ip):
-    # Resolve the IP
-    try:
-        hostname, aliaslist, ipaddrlist = socket.gethostbyaddr(ip)
-        lazy_raise = None
-    except socket.gaierror as err:
-        # Get Addr Info (GAI) Error
-        #
-        # When we are connecting to a node through a ssh proxy jump
-        # the node IPs/Hostnames are not available until we reach
-        # the first node inside the cluster
-        #
-        hostname = ip
-        aliaslist = []
-        ipaddrlist = []
-        lazy_raise = err
+    hostname = ip
+    aliaslist = []
+    ipaddrlist = []
+    lazy_raise = None
+
+    if not offline:
+        # Resolve the IP
+        try:
+            hostname, aliaslist, ipaddrlist = socket.gethostbyaddr(ip)
+            lazy_raise = None
+        
+        except socket.herror as err:
+            lazy_raise = err
+
+        except socket.gaierror as err:
+            # Get Addr Info (GAI) Error
+            #
+            # When we are connecting to a node through a ssh proxy jump
+            # the node IPs/Hostnames are not available until we reach
+            # the first node inside the cluster
+            #
+            lazy_raise = err
 
     return hostname, aliaslist, ipaddrlist, lazy_raise
 
@@ -195,6 +202,21 @@ def _fix_weird(hostname):
     return hostname
 
 
+# If true that means we cannot resolve the ip addresses
+# so we ignore errors
+offline = False
+
+
+@contextmanager
+def enable_offline(enabled):
+    global offline
+    old = offline
+
+    offline = enabled
+    yield
+    offline = old
+
+
 def resolve_addresses(nodes):
     # Note: it is possible for self to be none
     # if we are running milabench on a node that is not part of the system
diff --git a/tests/conftest.py b/tests/conftest.py
index f2a1ddca2..3d021c570 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -25,6 +25,19 @@ def get_config(name):
     return get_config
 
 
+
+@pytest.fixture
+def official_config():
+    def get_config(name):
+        return here / ".." / "config" / f"{name}.yaml"
+    return get_config
+
+
+@pytest.fixture
+def standard_config(official_config):
+    return official_config("standard")
+
+
 @pytest.fixture
 def replayfolder():
     return here / "replays"
diff --git a/tests/test_command_reg.py b/tests/test_command_reg.py
new file mode 100644
index 000000000..f44b5a8db
--- /dev/null
+++ b/tests/test_command_reg.py
@@ -0,0 +1,44 @@
+import os
+from milabench.cli.dry import cli_dry, Arguments
+
+
+ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
+
+
+def compare(base, capsys, file_regression):
+    all = capsys.readouterr()
+    stdout = all.out
+    assert stdout != ""
+
+    stdout = stdout.replace(str(base), "$BASE")
+    stdout = stdout.replace(str(ROOT), "$SRC")
+    file_regression.check(stdout)
+
+
+
+def test_command_reg_one_node(standard_config, tmp_path, capsys, file_regression):
+    os.environ["MILABENCH_CONFIG"] = str(standard_config)
+    os.environ["MILABENCH_BASE"] = str(tmp_path)
+
+    args = Arguments()
+    args.ngpu = 8
+    args.capacity = 80000
+    args.nnodes = 1
+
+    cli_dry(args)
+    
+    compare(str(tmp_path), capsys, file_regression)
+
+
+def test_command_reg_two_nodes(standard_config, tmp_path, capsys, file_regression):
+    os.environ["MILABENCH_CONFIG"] = str(standard_config)
+    os.environ["MILABENCH_BASE"] = str(tmp_path)
+
+    args = Arguments()
+    args.ngpu = 8
+    args.capacity = 80000
+    args.nnodes = 2
+    
+    cli_dry(args)
+
+    compare(str(tmp_path), capsys, file_regression)
\ No newline at end of file
diff --git a/tests/test_command_reg/test_command_reg_one_node.txt b/tests/test_command_reg/test_command_reg_one_node.txt
new file mode 100644
index 000000000..02950c9ea
--- /dev/null
+++ b/tests/test_command_reg/test_command_reg_one_node.txt
@@ -0,0 +1,452 @@
+#!/bin/sh
+
+echo "---"
+echo "Virtual Env"
+echo "==========="
+export VIRTUAL_ENV=$BASE/venv/torch
+
+source $VIRTUAL_ENV/bin/activate
+echo "---"
+echo "Milabench"
+echo "========="
+export MILABENCH_DIR_BASE=$BASE
+export MILABENCH_DIR_VENV=$BASE/venv/torch
+export MILABENCH_DIR_DATA=$BASE/data
+export MILABENCH_DIR_RUNS=$BASE/runs
+export MILABENCH_DIR_EXTRA=$BASE/extra/llm
+export MILABENCH_DIR_CACHE=$BASE/cache
+export MILABENCH_CONFIG='{"system": {"arch": "cuda", "sshkey": null, "nodes": [{"ip": "127.0.0.1", "main": true, "name": "0", "port": 22, "user": "username", "hostname": "localhost", "aliaslist": [], "ipaddrlist": ["00:00:03:9b:fe:80:00:00:00:00:00:00:b8:ce:f6:03:00:0e:54:86", "fe80::3eec:efff:fe72:9f02%eno1np0", "00:00:00:00:00:00", "3c:ec:ef:72:9f:02", "10.20.10.13", "172.16.10.13", "::1", "fe80::bace:f603:e:5486%ibp65s0f0", "127.0.0.1"], "local": true}], "self": {"ip": "127.0.0.1", "main": true, "name": "0", "port": 22, "user": "username", "hostname": "localhost", "aliaslist": [], "ipaddrlist": ["00:00:03:9b:fe:80:00:00:00:00:00:00:b8:ce:f6:03:00:0e:54:86", "fe80::3eec:efff:fe72:9f02%eno1np0", "00:00:00:00:00:00", "3c:ec:ef:72:9f:02", "10.20.10.13", "172.16.10.13", "::1", "fe80::bace:f603:e:5486%ibp65s0f0", "127.0.0.1"], "local": true}}, "dirs": {"base": "$BASE", "venv": "$BASE/venv/torch", "data": "$BASE/data", "runs": "$BASE/runs", "extra": "$BASE/extra/llm", "cache": "$BASE/cache"}, "group": "llm", "install_group": "torch", "install_variant": "cuda", "run_name": "dev", "enabled": true, "capabilities": {"nodes": 1}, "max_duration": 800, "voir": {"options": {"stop": 30, "interval": "1s"}}, "validation": {"usage": {"gpu_load_threshold": 0.5, "gpu_mem_threshold": 0.5}}, "config_base": "$SRC/milabench/tests/../config", "config_file": "$SRC/milabench/tests/../config/standard.yaml", "definition": "$SRC/milabench/benchmarks/llama", "tags": ["inference", "llm", "nlp"], "plan": {"method": "per_gpu"}, "weight": 1.0, "name": "llama", "tag": ["llama"]}'
+export OMP_NUM_THREADS=8
+
+echo "---"
+echo "llama"
+echo "====="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "fp16"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  wait
+)
+
+echo "---"
+echo "bf16"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  wait
+)
+
+echo "---"
+echo "tf32"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  wait
+)
+
+echo "---"
+echo "fp32"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  wait
+)
+
+echo "---"
+echo "resnet50"
+echo "========"
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  wait
+)
+
+echo "---"
+echo "resnet50-noio"
+echo "============="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  wait
+)
+
+echo "---"
+echo "resnet152-ddp"
+echo "============="
+time (
+  $SRC/milabench/benchmarks/torchvision_ddp/activator $BASE/venv/torch $SRC/milabench/benchmarks/torchvision_ddp/main.py --epochs 10 --num-workers 8 --loader torch --data $BASE/data/FakeImageNet --model resnet152 --batch-size 256 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-fp32"
+echo "==================="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-fp16"
+echo "==================="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-tf32"
+echo "==================="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-tf32-fp16"
+echo "========================"
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "regnet_y_128gf"
+echo "=============="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  wait
+)
+
+echo "---"
+echo "bert-fp32"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "bert-fp16"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "bert-tf32"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "bert-tf32-fp16"
+echo "=============="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "t5"
+echo "=="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  wait
+)
+
+echo "---"
+echo "reformer"
+echo "========"
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  wait
+)
+
+echo "---"
+echo "whisper"
+echo "======="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  wait
+)
+
+echo "---"
+echo "resnet152"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D0 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D1 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D2 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D3 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D4 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D5 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D6 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D7 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "resnet152-multi"
+echo "==============="
+time (
+  $BASE/venv/torch/bin/torchrun --nproc_per_node=8 --no-python -- python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152-multi.0 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "davit_large"
+echo "==========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D0 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D1 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D2 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D3 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D4 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D5 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D6 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D7 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "davit_large-multi"
+echo "================="
+time (
+  $BASE/venv/torch/bin/torchrun --nproc_per_node=8 --no-python -- python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large-multi.0 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "focalnet"
+echo "========"
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D0 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D1 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D2 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D3 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D4 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D5 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D6 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D7 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "opt-1_3b"
+echo "========"
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=1 --multi_gpu --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=8 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-1.3b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "opt-1_3b-multinode"
+echo "=================="
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=1 --multi_gpu --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=8 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-1.3b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "opt-6_7b"
+echo "========"
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=1 --use_deepspeed --deepspeed_multinode_launcher=standard --zero_stage=2 --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=8 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-6.7b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "opt-6_7b-multinode"
+echo "=================="
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=1 --use_deepspeed --deepspeed_multinode_launcher=standard --zero_stage=2 --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=8 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-6.7b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "stargan"
+echo "======="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  wait
+)
+
+echo "---"
+echo "super-slomo"
+echo "==========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  wait
+)
+
+echo "---"
+echo "dlrm"
+echo "===="
+time (
+  python $SRC/milabench/benchmarks/dlrm/dlrm/dlrm_s_pytorch.py --num-batches 1000 --data-generation random --arch-mlp-bot 512-512-64 --arch-mlp-top 1024-1024-1024-1 --arch-sparse-feature-size 64 --arch-embedding-size 1000000-1000000-1000000-1000000-1000000-1000000-1000000-1000000 --num-indices-per-lookup 100 --arch-interaction-op dot --numpy-rand-seed 727 --print-freq 999999 --mini-batch-size 16384 --test-mini-batch-size 16384 --test-num-workers 0 --use-gpu --num-workers 8 &
+  wait
+)
+
+echo "---"
+echo "brax"
+echo "===="
+time (
+  python $SRC/milabench/benchmarks/brax/main.py --episode-length 20 --batch-size 1024 --num-minibatches 32 --num-envs 8192 &
+  wait
+)
+
diff --git a/tests/test_command_reg/test_command_reg_two_nodes.txt b/tests/test_command_reg/test_command_reg_two_nodes.txt
new file mode 100644
index 000000000..92942b264
--- /dev/null
+++ b/tests/test_command_reg/test_command_reg_two_nodes.txt
@@ -0,0 +1,454 @@
+#!/bin/sh
+
+echo "---"
+echo "Virtual Env"
+echo "==========="
+export VIRTUAL_ENV=$BASE/venv/torch
+
+source $VIRTUAL_ENV/bin/activate
+echo "---"
+echo "Milabench"
+echo "========="
+export MILABENCH_DIR_BASE=$BASE
+export MILABENCH_DIR_VENV=$BASE/venv/torch
+export MILABENCH_DIR_DATA=$BASE/data
+export MILABENCH_DIR_RUNS=$BASE/runs
+export MILABENCH_DIR_EXTRA=$BASE/extra/llm
+export MILABENCH_DIR_CACHE=$BASE/cache
+export MILABENCH_CONFIG='{"system": {"arch": "cuda", "sshkey": null, "nodes": [{"ip": "127.0.0.1", "main": true, "name": "0", "port": 22, "user": "username", "hostname": "localhost", "aliaslist": [], "ipaddrlist": ["00:00:03:9b:fe:80:00:00:00:00:00:00:b8:ce:f6:03:00:0e:54:86", "fe80::3eec:efff:fe72:9f02%eno1np0", "00:00:00:00:00:00", "3c:ec:ef:72:9f:02", "10.20.10.13", "172.16.10.13", "::1", "fe80::bace:f603:e:5486%ibp65s0f0", "127.0.0.1"], "local": true}, {"ip": "192.168.0.11", "main": false, "name": "1", "port": 22, "user": "username", "hostname": "192.168.0.11", "aliaslist": [], "ipaddrlist": [], "local": false}], "self": {"ip": "127.0.0.1", "main": true, "name": "0", "port": 22, "user": "username", "hostname": "localhost", "aliaslist": [], "ipaddrlist": ["00:00:03:9b:fe:80:00:00:00:00:00:00:b8:ce:f6:03:00:0e:54:86", "fe80::3eec:efff:fe72:9f02%eno1np0", "00:00:00:00:00:00", "3c:ec:ef:72:9f:02", "10.20.10.13", "172.16.10.13", "::1", "fe80::bace:f603:e:5486%ibp65s0f0", "127.0.0.1"], "local": true}}, "dirs": {"base": "$BASE", "venv": "$BASE/venv/torch", "data": "$BASE/data", "runs": "$BASE/runs", "extra": "$BASE/extra/llm", "cache": "$BASE/cache"}, "group": "llm", "install_group": "torch", "install_variant": "cuda", "run_name": "dev", "enabled": true, "capabilities": {"nodes": 1}, "max_duration": 800, "voir": {"options": {"stop": 30, "interval": "1s"}}, "validation": {"usage": {"gpu_load_threshold": 0.5, "gpu_mem_threshold": 0.5}}, "config_base": "$SRC/milabench/tests/../config", "config_file": "$SRC/milabench/tests/../config/standard.yaml", "definition": "$SRC/milabench/benchmarks/llama", "tags": ["inference", "llm", "nlp"], "plan": {"method": "per_gpu"}, "weight": 1.0, "name": "llama", "tag": ["llama"]}'
+export OMP_NUM_THREADS=8
+
+echo "---"
+echo "llama"
+echo "====="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/llama/main.py --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "fp16"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 30 --repeat 90 --m 8192 --n 8192 --dtype fp16 &
+  wait
+)
+
+echo "---"
+echo "bf16"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype bf16 &
+  wait
+)
+
+echo "---"
+echo "tf32"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 --tf32 &
+  wait
+)
+
+echo "---"
+echo "fp32"
+echo "===="
+time (
+  CUDA_VISIBLE_DEVICES=0 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=1 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=2 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=3 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=4 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=5 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=6 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  CUDA_VISIBLE_DEVICES=7 $SRC/milabench/benchmarks/flops/activator $BASE/venv/torch $SRC/milabench/benchmarks/flops/main.py --number 10 --repeat 90 --m 8192 --n 8192 --dtype fp32 &
+  wait
+)
+
+echo "---"
+echo "resnet50"
+echo "========"
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  wait
+)
+
+echo "---"
+echo "resnet50-noio"
+echo "============="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader synthetic_fixed --data $BASE/data/FakeImageNet --model resnet50 --batch-size 256 &
+  wait
+)
+
+echo "---"
+echo "resnet152-ddp"
+echo "============="
+time (
+  $SRC/milabench/benchmarks/torchvision_ddp/activator $BASE/venv/torch $SRC/milabench/benchmarks/torchvision_ddp/main.py --epochs 10 --num-workers 8 --loader torch --data $BASE/data/FakeImageNet --model resnet152 --batch-size 256 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-fp32"
+echo "==================="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-fp16"
+echo "==================="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-tf32"
+echo "==================="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "convnext_large-tf32-fp16"
+echo "========================"
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model convnext_large --batch-size 128 &
+  wait
+)
+
+echo "---"
+echo "regnet_y_128gf"
+echo "=============="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/torchvision/main.py --precision tf32-fp16 --lr 0.01 --no-stdout --epochs 50 --num-workers 8 --loader pytorch --data $BASE/data/FakeImageNet --model regnet_y_128gf --batch-size 64 &
+  wait
+)
+
+echo "---"
+echo "bert-fp32"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision fp32 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "bert-fp16"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision fp16 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "bert-tf32"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "bert-tf32-fp16"
+echo "=============="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model Bert --batch-size 32 &
+  wait
+)
+
+echo "---"
+echo "t5"
+echo "=="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model T5 --batch-size 16 &
+  wait
+)
+
+echo "---"
+echo "reformer"
+echo "========"
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model Reformer --batch-size 64 &
+  wait
+)
+
+echo "---"
+echo "whisper"
+echo "======="
+time (
+  CUDA_VISIBLE_DEVICES=0 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=1 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=2 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=3 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=4 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=5 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=6 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  CUDA_VISIBLE_DEVICES=7 python -m bench --precision tf32-fp16 --num-workers 8 --model Whisper --batch-size 64 &
+  wait
+)
+
+echo "---"
+echo "resnet152"
+echo "========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D0 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D1 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D2 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D3 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D4 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D5 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D6 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152.D7 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "resnet152-multi"
+echo "==============="
+time (
+  $BASE/venv/torch/bin/torchrun --nproc_per_node=8 --no-python -- python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model resnet152 --batch-size 256 --output $BASE/extra/timm/dev/resnet152-multi.0 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "davit_large"
+echo "==========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D0 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D1 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D2 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D3 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D4 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D5 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D6 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large.D7 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "davit_large-multi"
+echo "================="
+time (
+  $BASE/venv/torch/bin/torchrun --nproc_per_node=8 --no-python -- python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model davit_large --batch-size 128 --lr-base 0.01 --output $BASE/extra/timm/dev/davit_large-multi.0 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "focalnet"
+echo "========"
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D0 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D1 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D2 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D3 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D4 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D5 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D6 --checkpoint-hist 1 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/timm/pytorch-image-models/train.py --amp --amp-dtype bfloat16 --device cuda --val-split --data-dir $BASE/data --dataset FakeImageNet --workers 8 --model focalnet_base_lrf --output $BASE/extra/timm/dev/focalnet.D7 --checkpoint-hist 1 &
+  wait
+)
+
+echo "---"
+echo "opt-1_3b"
+echo "========"
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=1 --multi_gpu --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=8 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-1.3b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "opt-1_3b-multinode"
+echo "=================="
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=2 --multi_gpu --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=16 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-1.3b --cache $BASE/cache &
+  ssh -oCheckHostIP=no -oStrictHostKeyChecking=no -oPasswordAuthentication=no -oPasswordAuthentication=no -p 22 username@192.168.0.11 $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=1 --num_machines=2 --multi_gpu --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=16 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-1.3b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "opt-6_7b"
+echo "========"
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=1 --use_deepspeed --deepspeed_multinode_launcher=standard --zero_stage=2 --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=8 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-6.7b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "opt-6_7b-multinode"
+echo "=================="
+time (
+  $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=0 --num_machines=2 --use_deepspeed --deepspeed_multinode_launcher=standard --zero_stage=2 --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=16 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-6.7b --cache $BASE/cache &
+  ssh -oCheckHostIP=no -oStrictHostKeyChecking=no -oPasswordAuthentication=no -oPasswordAuthentication=no -p 22 username@192.168.0.11 $SRC/milabench/benchmarks/accelerate_opt/activator $BASE/venv/torch accelerate launch --mixed_precision=bf16 --dynamo_backend=no --machine_rank=1 --num_machines=2 --use_deepspeed --deepspeed_multinode_launcher=standard --zero_stage=2 --gradient_accumulation_steps=1 --num_cpu_threads_per_process=8 --main_process_ip=127.0.0.1 --main_process_port=22 --num_processes=16 $SRC/milabench/benchmarks/accelerate_opt/main.py --max_train_steps 100 --dataset_name wikitext --dataset_config_name wikitext-103-v1 --dataset_rev b08601e --validation_split_percentage 5 --per_gpu_batch_size 1 --cpus_per_gpu 8 --model_name facebook/opt-6.7b --cache $BASE/cache &
+  wait
+)
+
+echo "---"
+echo "stargan"
+echo "======="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/stargan/stargan/main.py --image_size 512 --c_dim 5 --batch_size 16 --dataset synth --celeba_image_dir $BASE/data --log_dir $BASE/extra/stargan/logs --model_save_dir $BASE/extra/stargan/models --sample_dir $BASE/extra/stargan/samples --result_dir $BASE/extra/stargan/results --num_workers 8 &
+  wait
+)
+
+echo "---"
+echo "super-slomo"
+echo "==========="
+time (
+  CUDA_VISIBLE_DEVICES=0 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=1 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=2 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=3 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=4 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=5 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=6 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  CUDA_VISIBLE_DEVICES=7 python $SRC/milabench/benchmarks/super-slomo/slomo/train.py --train_batch_size 64 --dataset_root $BASE/data/FakeImageNet --loader pytorch --num_workers 8 &
+  wait
+)
+
+echo "---"
+echo "dlrm"
+echo "===="
+time (
+  python $SRC/milabench/benchmarks/dlrm/dlrm/dlrm_s_pytorch.py --num-batches 1000 --data-generation random --arch-mlp-bot 512-512-64 --arch-mlp-top 1024-1024-1024-1 --arch-sparse-feature-size 64 --arch-embedding-size 1000000-1000000-1000000-1000000-1000000-1000000-1000000-1000000 --num-indices-per-lookup 100 --arch-interaction-op dot --numpy-rand-seed 727 --print-freq 999999 --mini-batch-size 16384 --test-mini-batch-size 16384 --test-num-workers 0 --use-gpu --num-workers 8 &
+  wait
+)
+
+echo "---"
+echo "brax"
+echo "===="
+time (
+  python $SRC/milabench/benchmarks/brax/main.py --episode-length 20 --batch-size 1024 --num-minibatches 32 --num-envs 8192 &
+  wait
+)
+