Flux update, v6.0 added

mlperf_logging/benchmark_meta.py

@@ -155,7 +155,14 @@
         'llama2_70b_lora',
         'rgat',
         'llama31_405b'
-    ]  
+    ],
+    '6.0': [
+        'llama31_8b',
+        'dlrm_dcnv2',   
+        'flux1',
+        'llama2_70b_lora',
+        'llama31_405b'
+    ]    
     },
 
     'hpc': {

mlperf_logging/compliance_checker/README.md

@@ -10,9 +10,9 @@ To check a log file for compliance:
 
     python -m mlperf_logging.compliance_checker [--config YAML] [--usage training/hpc] [--ruleset MLPERF_EDITION] FILENAME
 
-By default, 5.1.0 training edition rules are used and the default config is set to `5.1.0/common.yaml`.
+By default, 6.0.0 training edition rules are used and the default config is set to `6.0.0/common.yaml`.
 This config will check all common keys and enqueue benchmark specific config to be checked as well.
-Old training editions, still supported are 5.0.0, 4.1.0, 4.0.0, 3.1.0, 3.0.0, 2.1.0, 2.0.0, 1.1.0, 1.0.0, 0.7.0 and 0.6.0
+Old training editions, still supported are 6.0.0, 5.1.0, 5.0.0, 4.1.0, 4.0.0, 3.1.0, 3.0.0, 2.1.0, 2.0.0, 1.1.0, 1.0.0, 0.7.0 and 0.6.0
 
 To check hpc compliance rules (only 1.0.0 ruleset is supported), set --usage hpc --ruleset 1.0.0.
 
@@ -22,23 +22,19 @@ As log examples use [NVIDIA's training logs](https://github.com/mlperf/training_
 
 ### Existing config files for training submissions
 
-    5.1.0/common.yaml          - currently the default config file, checks common fields complience and equeues benchmark-specific config file
-    5.1.0/closed_common.yaml   - the common rules file for closed submissions. These rules apply to all benchmarks
-    5.1.0/open_common.yaml     - the common rules file for open submissions. These rules apply to all benchmarks
-    5.1.0/closed_retinanet.yaml      - Per-benchmark rules, closed submissions.    
-    5.1.0/closed_llama31_8b.yaml
-    5.1.0/closed_llama31_405b.yaml
-    5.1.0/closed_dlrm_dcnv2.yaml
-    5.1.0/closed_rgat.yaml
-    5.1.0/closed_llama2_70b_lora.yaml
-    5.1.0/closed_flux1.yaml
-    5.1.0/open_retinanet.yaml        - Per-benchmark rules, open submissions.    
-    5.1.0/open_llama31_8b.yaml
-    5.1.0/open_llama31_405b.yaml
-    5.1.0/open_dlrm_dcnv2.yaml
-    5.1.0/open_rgat.yaml
-    5.1.0/open_llama2_70b_lora.yaml
-    5.1.0/open_flux1.yaml
+    6.0.0/common.yaml          - currently the default config file, checks common fields complience and equeues benchmark-specific config file
+    6.0.0/closed_common.yaml   - the common rules file for closed submissions. These rules apply to all benchmarks
+    6.0.0/open_common.yaml     - the common rules file for open submissions. These rules apply to all benchmarks
+    6.0.0/closed_llama31_8b.yaml
+    6.0.0/closed_llama31_405b.yaml
+    6.0.0/closed_dlrm_dcnv2.yaml
+    6.0.0/closed_llama2_70b_lora.yaml
+    6.0.0/closed_flux1.yaml
+    6.0.0/open_llama31_8b.yaml
+    6.0.0/open_llama31_405b.yaml
+    6.0.0/open_dlrm_dcnv2.yaml
+    6.0.0/open_llama2_70b_lora.yaml
+    6.0.0/open_flux1.yaml
 
 ### Existing config files for HPC submissions
 

mlperf_logging/compliance_checker/mlp_compliance.py

@@ -315,7 +315,7 @@ def get_parser():
     parser.add_argument('--usage', type=str, default='training',
                     choices=usage_choices(),
                     help='what WG do the benchmarks come from')
-    parser.add_argument('--ruleset', type=str, default='5.1.0',
+    parser.add_argument('--ruleset', type=str, default='6.0.0',
                     choices=rule_choices(),
                     help='what version of rules to check the log against')
     parser.add_argument('--config',  type=str,

mlperf_logging/compliance_checker/mlp_parser/__init__.py

@@ -10,6 +10,7 @@
 from .ruleset_410 import parse_file as parse_file_410
 from .ruleset_500 import parse_file as parse_file_500
 from .ruleset_510 import parse_file as parse_file_510
+from .ruleset_600 import parse_file as parse_file_600
 
 def parse_file(filename, ruleset='0.6.0'):
     if ruleset == '0.6.0':
@@ -36,5 +37,7 @@ def parse_file(filename, ruleset='0.6.0'):
         return parse_file_500(filename)
     elif ruleset == '5.1.0':
         return parse_file_510(filename)
+    elif ruleset == '6.0.0':
+        return parse_file_600(filename)
     else:
         raise Exception(f'Ruleset "{ruleset}" is not supported')

mlperf_logging/compliance_checker/mlp_parser/ruleset_600.py

@@ -0,0 +1,105 @@
+'''
+Parses a text MLPerf log into a structured format.
+'''
+
+from __future__ import print_function
+
+import collections
+import json
+import re
+import sys
+from dataclasses import dataclass
+
+from io import open
+
+@dataclass
+class LogLine:
+    """Class for keeping track of an item in inventory."""
+    full_string: str
+    timestamp: float
+    key: str
+    value: str
+    lineno: int
+
+TOKEN = ':::MLLOG '
+
+
+def parse_line(line):
+    if not line.startswith(TOKEN):
+        return None
+
+    return json.loads(line[len(TOKEN):])
+
+
+def string_to_logline(lineno, string):
+    ''' Returns a LogLine or raises a ValueError '''
+    m = parse_line(string)
+
+    if m is None:
+        raise ValueError('does not match regex')
+
+    args = []
+    args.append(string) # full string
+
+    ts = float(m['time_ms']) # may raise error, e.g. "1.2.3"
+    # TODO check for weird values
+    args.append(ts)
+
+    args.append(m['key']) # key
+
+    j = { 'value': m['value'], 'metadata': m['metadata'] }
+    args.append(j)
+
+    args.append(lineno)
+    return LogLine(*args)
+
+
+def parse_file(filename):
+    ''' Reads a file by name and returns list of loglines and list of errors'''
+    with open(filename, encoding='latin-1') as f:
+        return parse_generator(f)
+
+
+def strip_and_dedup(gen):
+    lines = []
+    for l in gen:
+        if TOKEN not in l:
+            continue
+        lines.append(re.sub(".*"+TOKEN, TOKEN, l))
+    return lines
+
+
+
+def parse_generator(gen):
+    ''' Reads a generator of lines and returns (loglines, errors)
+    The list of errors are any parsing issues as a tuple (str_line, error_msg)
+    '''
+    loglines = []
+    failed = []
+    for lineno, line in enumerate(strip_and_dedup(gen)):
+        line = line.strip()
+        try:
+            ll = string_to_logline(lineno, line)
+            loglines.append(ll)
+        except ValueError as e:
+            failed.append((line, str(e)))
+    return loglines, failed
+
+
+if __name__ == '__main__':
+    if len(sys.argv) != 2:
+        print('usage: mlp_parser.py FILENAME')
+        print('       tests parsing on the file.')
+        sys.exit(1)
+
+    filename = sys.argv[1]
+    lines, errors = parse_file(filename)
+
+    print('Parsed {} log lines with {} errors.'.format(len(lines), len(errors)))
+
+    if len(errors) > 0:
+        print('Lines which failed to parse:')
+        for line, error in errors:
+            print('  Following line failed: {}'.format(error))
+            print(line)
+

mlperf_logging/compliance_checker/training_6.0.0/closed_common.yaml

@@ -0,0 +1,11 @@
+
+- KEY:
+    NAME:  submission_benchmark
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] in ['flux1', 'dlrm_dcnv2', 'llama31_8b', 'llama2_70b_lora', 'llama31_405b'] "
+    POST:  " enqueue_config('training_6.0.0/closed_{}.yaml'.format(v['value'])) "
+
+- KEY:
+    NAME: gradient_accumulation_steps
+    REQ: EXACTLY_ONE
+    CHECK: " v['value'] > 0 "

mlperf_logging/compliance_checker/training_6.0.0/closed_dlrm_dcnv2.yaml

@@ -0,0 +1,59 @@
+- KEY:
+    NAME:  global_batch_size
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_name
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 'adagrad' "
+
+- KEY:
+    NAME:  opt_base_learning_rate
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_adagrad_learning_rate_decay
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  opt_weight_decay
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  opt_adagrad_initial_accumulator_value
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  opt_adagrad_epsilon
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 1e-8 "
+
+- KEY:
+    NAME:  opt_learning_rate_warmup_steps
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  opt_learning_rate_decay_start_step
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  opt_learning_rate_decay_steps
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  eval_accuracy
+    REQ:   AT_LEAST_ONE
+    CHECK:
+        - "'epoch_num' in v['metadata']"
+    ATLEAST_ONE_CHECK: "v['value'] >= 0.80275 and v['value'] <= 1.0"
+
+- KEY:
+    NAME:  eval_samples
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 89137319 "

mlperf_logging/compliance_checker/training_6.0.0/closed_flux1.yaml

@@ -0,0 +1,56 @@
+- KEY:
+    NAME:  global_batch_size
+    REQ:   AT_LEAST_ONE
+    CHECK: " v['value'] >= 0 "
+
+- KEY:
+    NAME:  evaluation_frequency
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 262144"
+
+- KEY:
+    NAME:  opt_name
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 'adamw' "
+
+- KEY:
+    NAME:  opt_adamw_beta_1
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.9 "
+
+- KEY:
+    NAME:  opt_adamw_beta_2
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.95 "
+
+- KEY:
+    NAME:  opt_adamw_epsilon
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 1e-08 "
+
+- KEY:
+    NAME:  opt_adamw_weight_decay
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.1 "
+
+- KEY:
+    NAME:  opt_base_learning_rate
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] >= 0.0 "
+
+- KEY:
+    NAME:  opt_learning_rate_warmup_steps
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] >= 0 "
+
+- KEY:
+    NAME:  opt_gradient_clip_norm
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 1.0 "
+
+- KEY:
+    NAME: eval_accuracy
+    REQ: AT_LEAST_ONE
+    CHECK:
+        - "'samples_count' in v['metadata']"
+    ATLEAST_ONE_CHECK: "v['value'] <= 0.586 and v['value'] > 0.0"

mlperf_logging/compliance_checker/training_6.0.0/closed_llama2_70b_lora.yaml

@@ -0,0 +1,42 @@
+- KEY:
+    NAME:  global_batch_size
+    REQ:   EXACTLY_ONE
+    POST: >
+        s['global_batch_size'] = v['value']
+
+- KEY:
+    NAME:  opt_base_learning_rate
+    REQ:   EXACTLY_ONE
+
+
+- KEY:
+    NAME:  opt_learning_rate_training_steps
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_gradient_clip_norm
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_adamw_weight_decay
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  gradient_accumulation_steps
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  lora_alpha
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  lora_rank
+    REQ:   EXACTLY_ONE   
+    CHECK: " v['value'] == 16"
+
+- KEY:
+    NAME:  eval_accuracy
+    REQ:   AT_LEAST_ONE
+    CHECK:
+        - "'samples_count' in v['metadata']"
+    ATLEAST_ONE_CHECK: "(v['value'] <= 0.925) and v['value'] > 0.0"