Merge pull request #315 from flexaihq/main

various updates for CM tools for Windows and better debugging

CHANGES.md

@@ -1,3 +1,8 @@
+### 20240927
+ * added "test dummy" script to test Docker containers
+ * added more standard Nvidia Docker configuration for PyTorch
+ * added better support to select Docker configurations via UID
+
 ### 20240916
  * fixed "cm add script"
 

COPYRIGHT.txt

@@ -1,5 +1,5 @@
 Copyright (c) 2021-2024 MLCommons
 
-The cTuning foundation and OctoML donated this project to MLCommons to benefit everyone.
+Grigori Fursin, the cTuning foundation and OctoML donated this project to MLCommons to benefit everyone.
 
 Copyright (c) 2014-2021 cTuning foundation

README.md

@@ -139,6 +139,11 @@ cm run script \
 
 [Apache 2.0](LICENSE.md)
 
+## CM concepts
+
+* https://doi.org/10.5281/zenodo.8105339
+* https://arxiv.org/abs/2406.16791
+
 ## Authors
 
 [Grigori Fursin](https://cKnowledge.org/gfursin) and [Arjun Suresh](https://www.linkedin.com/in/arjunsuresh)

automation/script/module.py

@@ -4028,14 +4028,58 @@ def docker(self, i):
 
  (out) (str): if 'con', output to console
 
- parsed_artifact (list): prepared in CM CLI or CM access function
- [ (artifact alias, artifact UID) ] or
- [ (artifact alias, artifact UID), (artifact repo alias, artifact repo UID) ]
-
  (repos) (str): list of repositories to search for automations
 
  (output_dir) (str): output directory (./ by default)
 
+ (docker) (dict): convert keys into docker_{key} strings for CM >= 2.3.8.1
+
+
+ (docker_skip_build) (bool): do not generate Dockerfiles and do not recreate Docker image (must exist)
+ (docker_noregenerate) (bool): do not generate Dockerfiles
+ (docker_norecreate) (bool): do not recreate Docker image
+
+ (docker_cfg) (str): if True, show all available basic docker configurations, otherwise pre-select one
+ (docker_cfg_uid) (str): if True, select docker configuration with this UID
+
+ (docker_path) (str): where to create or find Dockerfile
+ (docker_gh_token) (str): GitHub token for private repositories
+ (docker_save_script) (str): if !='' name of script to save docker command
+ (docker_interactive) (bool): if True, run in interactive mode
+ (docker_it) (bool): the same as `docker_interactive`
+ (docker_detached) (bool): detach Docker
+ (docker_dt) (bool) the same as `docker_detached`
+
+ (docker_base_image) (str): force base image
+ (docker_os) (str): force docker OS (default: ubuntu)
+ (docker_os_version) (str): force docker OS version (default: 22.04)
+ (docker_image_tag_extra) (str): add extra tag (default:-latest)
+
+ (docker_cm_repo) (str): force CM automation repository when building Docker (default: cm4mlops)
+ (docker_cm_repos)
+ (docker_cm_repo_flags)
+
+ (dockerfile_env)
+
+ (docker_skip_cm_sys_upgrade) (bool): if True, do not install CM sys deps
+
+ (docker_extra_sys_deps)
+
+ (fake_run_deps)
+ (docker_run_final_cmds)
+
+ (all_gpus)
+ (num_gpus)
+
+ (docker_device)
+
+ (docker_port_maps)
+
+ (docker_shm_size)
+
+ (docker_extra_run_args)
+ 
+
  Returns:
  (CM return dict):
 

automation/script/module_misc.py

@@ -1335,15 +1335,9 @@ def dockerfile(i):
  Args:
  (CM input dict):
 
- (out) (str): if 'con', output to console
-
- parsed_artifact (list): prepared in CM CLI or CM access function
- [ (artifact alias, artifact UID) ] or
- [ (artifact alias, artifact UID), (artifact repo alias, artifact repo UID) ]
-
- (repos) (str): list of repositories to search for automations
-
- (output_dir) (str): output directory (./ by default)
+ (out) (str): if 'con', output to console
+ (repos) (str): list of repositories to search for automations
+ (output_dir) (str): output directory (./ by default)
 
  Returns:
  (CM return dict):
@@ -1632,15 +1626,6 @@ def docker(i):
 
  (out) (str): if 'con', output to console
 
- (docker_skip_build) (bool): do not generate Dockerfiles and do not recreate Docker image (must exist)
- (docker_noregenerate) (bool): do not generate Dockerfiles
- (docker_norecreate) (bool): do not recreate Docker image
-
- (docker_path) (str): where to create or find Dockerfile
- (docker_gh_token) (str): GitHub token for private repositories
- (docker_save_script) (str): if !='' name of script to save docker command
- (docker_interactive) (bool): if True, run in interactive mode
- (docker_cfg) (str): if True, show all available basic docker configurations, otherwise pre-select one
 
  Returns:
  (CM return dict):
@@ -1653,6 +1638,20 @@ def docker(i):
  import copy
  import re
 
+ from cmind import __version__ as current_cm_version
+
+ self_module = i['self_module']
+
+ if type(i.get('docker', None)) == dict:
+ # Grigori started cleaning and refactoring this code on 20240929
+ # 
+ # 1. use --docker dictionary instead of --docker_{keys}
+
+ if utils.compare_versions(current_cm_version, '2.3.8.1') >= 0:
+ docker_params = utils.convert_dictionary(i['docker'], 'docker')
+ i.update(docker_params)
+ del(i['docker'])
+
  quiet = i.get('quiet', False)
 
  detached = i.get('docker_detached', '')
@@ -1670,13 +1669,12 @@ def docker(i):
 
  # Check simplified CMD: cm docker script "python app image-classification onnx"
  # If artifact has spaces, treat them as tags!
- self_module = i['self_module']
  self_module.cmind.access({'action':'detect_tags_in_artifact', 'automation':'utils', 'input':i})
 
  # CAREFUL -> artifacts and parsed_artifacts are not supported in input (and should not be?)
  if 'artifacts' in i: del(i['artifacts'])
  if 'parsed_artifacts' in i: del(i['parsed_artifacts'])
- 
+
  # Prepare "clean" input to replicate command
  r = self_module.cmind.access({'action':'prune_input', 'automation':'utils', 'input':i, 'extra_keys_starts_with':['docker_']})
  i_run_cmd_arc = r['new_input']
@@ -1693,13 +1691,19 @@ def docker(i):
 
  # Check available configurations
  docker_cfg = i.get('docker_cfg', '')
- if docker_cfg != '':
+ docker_cfg_uid = i.get('docker_cfg_uid', '')
+
+ if docker_cfg != '' or docker_cfg_uid != '':
  # Check if docker_cfg is turned on but not selected
  if type(docker_cfg) == bool or str(docker_cfg).lower() in ['true','yes']:
  docker_cfg= ''
-
- r = self_module.cmind.access({'action':'select_cfg', 'automation':'utils,dc2743f8450541e3', 
- 'tags':'basic,docker,configurations', 'title':'docker', 'alias':docker_cfg})
+
+ r = self_module.cmind.access({'action':'select_cfg', 
+ 'automation':'utils,dc2743f8450541e3', 
+ 'tags':'basic,docker,configurations', 
+ 'title':'docker', 
+ 'alias':docker_cfg,
+ 'uid':docker_cfg_uid})
  if r['return'] > 0: 
  if r['return'] == 16:
  return {'return':1, 'error':'Docker configuration {} was not found'.format(docker_cfg)}
@@ -1708,10 +1712,9 @@ def docker(i):
  selection = r['selection']
 
  docker_input_update = selection['meta']['input']
- 
+
  i.update(docker_input_update)
 
-
  ########################################################################################
  # Run dockerfile
  if not noregenerate_docker_file:
@@ -1722,7 +1725,7 @@ def docker(i):
  cur_dir = os.getcwd()
 
  console = i.get('out') == 'con'
- 
+
  # Search for script(s)
  r = aux_search({'self_module': self_module, 'input': i})
  if r['return']>0: return r

automation/utils/module_cfg.py

@@ -230,16 +230,18 @@ def select_cfg(i):
  self_module = i['self_module']
  tags = i['tags']
  alias = i.get('alias', '')
+ uid = i.get('uid', '')
  title = i.get('title', '')
 
  # Check if alias is not provided 
  r = self_module.cmind.access({'action':'find', 'automation':'cfg', 'tags':'basic,docker,configurations'})
  if r['return'] > 0: return r
- 
+
  lst = r['list']
 
  selector = []
 
+ # Do coarse-grain search for CM artifacts
  for l in lst:
  p = l.path
 
@@ -257,45 +259,53 @@ def select_cfg(i):
  if not f.startswith('_cm') and (f.endswith('.json') or f.endswith('.yaml')):
  selector.append({'path':os.path.join(p, f), 'alias':f[:-5]})
 
- if len(selector) == 0:
- return {'return':16, 'error':'configuration was not found'}
 
- select = 0
- if len(selector) > 1:
- xtitle = ' ' + title if title!='' else ''
- print ('')
- print ('Available{} configurations:'.format(xtitle))
-
- print ('')
+ # Load meta for name and UID
+ selector_with_meta = []
+ for s in range(0, len(selector)):
+ ss = selector[s]
 
- for s in range(0, len(selector)):
- ss = selector[s]
+ path = ss['path']
 
-  path = ss['path']
+ full_path_without_ext = path[:-5]
 
- full_path_without_ext = path[:-5]
+ r = cmind.utils.load_yaml_and_json(full_path_without_ext)
+ if r['return']>0:
+ print ('Warning: problem loading configuration file {}'.format(path))
 
- r = cmind.utils.load_yaml_and_json(full_path_without_ext)
- if r['return']>0:
- print ('Warning: problem loading configuration file {}'.format(path))
+ meta = r['meta']
 
-  meta = r['meta']
+ if uid == '' or meta.get('uid', '') == uid:
  ss['meta'] = meta
+ selector_with_meta.append(ss)
+
+ # Quit if no configurations found
+ if len(selector_with_meta) == 0:
+ return {'return':16, 'error':'configuration was not found'}
 
- selector = sorted(selector, key = lambda x: x['meta'].get('name',''))
+ select = 0
+ if len(selector_with_meta) > 1:
+ xtitle = ' ' + title if title!='' else ''
+ print ('')
+ print ('Available{} configurations:'.format(xtitle))
+
+ print ('')
+
+ selector_with_meta = sorted(selector_with_meta, key = lambda x: x['meta'].get('name',''))
  s = 0
- for ss in selector:
+ for ss in selector_with_meta:
  alias = ss['alias']
- name = ss['meta'].get('name','')
+ uid = ss['meta'].get('uid', '')
+ name = ss['meta'].get('name', '')
 
  x = name
  if x!='': x+=' '
- x += '('+alias+')'
- 
- print ('{}) {}'.format(s, x))
+ x += '(' + uid + ')'
+
+ print (f'{s}) {x}'.format(s, x))
 
  s+=1
- 
+
  print ('')
  select = input ('Enter configuration number of press Enter for 0: ')
 
@@ -306,6 +316,6 @@ def select_cfg(i):
  if select<0 or select>=len(selector):
  return {'return':1, 'error':'selection is out of range'}
 
- ss = selector[select]
+ ss = selector_with_meta[select]
 
  return {'return':0, 'selection':ss}

cfg/benchmark-run-mlperf-inference-v4.1/_cm.yaml

@@ -0,0 +1,39 @@
+alias: benchmark-run-mlperf-inference-v4.1
+uid: b7e89771987d4168
+
+automation_alias: cfg
+automation_uid: 88dce9c160324c5d
+
+tags:
+- benchmark
+- run
+- mlperf
+- inference
+- v4.1
+
+name: "MLPerf inference - v4.1"
+
+supported_compute:
+- ee8c568e0ac44f2b
+- fe379ecd1e054a00
+- d8f06040f7294319
+
+bench_uid: 39877bb63fb54725
+
+view_dimensions:
+- - input.device
+ - "MLPerf device"
+- - input.implementation
+ - "MLPerf implementation"
+- - input.backend
+ - "MLPerf backend"
+- - input.model
+ - "MLPerf model"
+- - input.scenario
+ - "MLPerf scenario"
+- - input.host_os
+ - "Host OS"
+- - output.state.cm-mlperf-inference-results-last.performance
+ - "Got performance"
+- - output.state.cm-mlperf-inference-results-last.accuracy
+ - "Got accuracy"

cfg/docker-basic-configurations/basic-ubuntu-24.04.yaml

@@ -0,0 +1,9 @@
+uid: 12e86eb386314866
+
+name: "Basic Ubuntu 24.04"
+
+input:
+ docker_base_image: 'ubuntu:24.04'
+ docker_os: ubuntu
+ docker_os_version: '24.04'
+

cfg/docker-basic-configurations/nvidia-ubuntu-20.04-cuda-11.8-cudnn-8.6.0-pytorch-1.13.0.yaml

@@ -1,6 +1,8 @@
 uid: 854e65fb31584d63
 
-name: "Nvidia Ubuntu 20.04 CUDA 11.8 cuDNN 8.6.0 PyTorch 1.13.0"
+name: "Nvidia Ubuntu 20.04 CUDA 11.8 cuDNN 8.6.0 PyTorch 1.13.0 (pytorch:22.10)"
+
+ref_url: https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-22-10.html
 
 input:
  docker_base_image: 'nvcr.io/nvidia/pytorch:22.10-py3'

cfg/docker-basic-configurations/nvidia-ubuntu-22.04-cuda-12.1-cudnn-8.9.1-pytorch-2.0.0.yaml

@@ -1,6 +1,8 @@
 uid: e0e7167139a74e36
 
-name: "Nvidia Ubuntu 22.04 CUDA 12.1 cuDNN 8.9.1 PyTorch 2.0.0"
+name: "Nvidia Ubuntu 22.04 CUDA 12.1 cuDNN 8.9.1 PyTorch 2.0.0 (pytorch:23.05)"
+
+ref_url: https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-23-05.html
 
 input:
  docker_base_image: 'nvcr.io/nvidia/pytorch:23.05-py3'

cfg/docker-basic-configurations/nvidia-ubuntu-22.04-cuda-12.4-cudnn-9.0.0-pytorch-2.3.0.yaml

@@ -1,6 +1,8 @@
 uid: 49fc51f2999b4545
 
-name: "Nvidia Ubuntu 22.04 CUDA 12.4 cuDNN 9.0.0 PyTorch 2.3.0"
+name: "Nvidia Ubuntu 22.04 CUDA 12.4 cuDNN 9.0.0 PyTorch 2.3.0 (pytorch:24.03)"
+
+ref_url: https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-24-03.html
 
 input:
  docker_base_image: 'nvcr.io/nvidia/pytorch:24.03-py3'