Qwen35 397b GRPO open ruleset and constants

mlperf_logging/compliance_checker/training_6.1.0/open_common.yaml

@@ -2,5 +2,5 @@
 - KEY:
     NAME:  submission_benchmark
     REQ:   EXACTLY_ONE
-    CHECK: " v['value'] in ['flux1', 'llama31_8b', 'llama2_70b_lora', 'gpt_oss_20b', 'deepseekv3_671b'] "
+    CHECK: " v['value'] in ['flux1', 'llama31_8b', 'llama2_70b_lora', 'gpt_oss_20b', 'deepseekv3_671b', 'qwen35_397b_grpo'] "
     POST:  " enqueue_config('training_6.1.0/open_{}.yaml'.format(v['value'])) "

mlperf_logging/compliance_checker/training_6.1.0/open_qwen35_397b_grpo.yaml

@@ -0,0 +1,94 @@
+- KEY:
+    NAME:  global_batch_size
+    REQ:   EXACTLY_ONE
+    CHECK: " is_integer(v['value']) and v['value'] > 0 "
+
+- KEY:
+    NAME:  max_sequence_length
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 65536 "
+
+- KEY:
+    NAME:  eval_samples
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 256 "
+
+- KEY:
+    NAME:  init_checkpoint_step
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0 "
+
+- KEY:
+    NAME:  max_steps
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_name
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 'adamw' "
+
+- KEY:
+    NAME:  opt_base_learning_rate
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_adamw_beta_1
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_adamw_beta_2
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_adamw_epsilon
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_adamw_weight_decay
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_gradient_clip_norm
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_learning_rate_warmup_steps
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  opt_learning_rate_decay_schedule
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  generation_training_rollout_temperature
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  generation_training_rollout_top_p
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  generation_validation_rollout_temperature
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  generation_validation_rollout_top_p
+    REQ:   EXACTLY_ONE
+
+- KEY:
+    NAME:  num_prompts_per_step
+    REQ:   EXACTLY_ONE
+    CHECK: " is_integer(v['value']) and v['value'] > 0 "
+
+- KEY:
+    NAME:  num_generations_per_prompt
+    REQ:   EXACTLY_ONE
+    CHECK: " is_integer(v['value']) and v['value'] > 0 "
+
+- KEY:
+    NAME:  eval_accuracy
+    REQ:   AT_LEAST_ONE
+    CHECK:
+        - "'samples_count' in v['metadata']"
+        - "0.0 <= v['value'] <= 1.0"
+    ATLEAST_ONE_CHECK: "v['value'] > 0.69"

mlperf_logging/mllog/constants.py

@@ -58,6 +58,7 @@
 LLAMA31_8B = "llama31_8b"
 FLUX1 = "flux1"
 DEEPSEEKV3_671B = "deepseekv3_671b"
+QWEN35_397B_GRPO = "qwen35_397b_grpo"
 
 # Constant values - model info
 ADAGRAD = "adagrad"
@@ -178,6 +179,12 @@
 INIT_CHECKPOINT_STEP = "init_checkpoint_step"
 LORA_ALPHA = "lora_alpha"
 MOE_AUX_LOSS_COEFF = "moe_aux_loss_coeff"
+GENERATION_TRAINING_ROLLOUT_TEMPERATURE = "generation_training_rollout_temperature"
+GENERATION_TRAINING_ROLLOUT_TOP_P = "generation_training_rollout_top_p"
+GENERATION_VALIDATION_ROLLOUT_TEMPERATURE = "generation_validation_rollout_temperature"
+GENERATION_VALIDATION_ROLLOUT_TOP_P = "generation_validation_rollout_top_p"
+NUM_GENERATIONS_PER_PROMPT = "num_generations_per_prompt"
+NUM_PROMPTS_PER_STEP = "num_prompts_per_step"
 # Log keys - misc.
 BBOX = "bbox"
 SEGM = "segm"

mlperf_logging/result_summarizer/config.yaml

@@ -117,6 +117,7 @@ columns:
       llama31_8b: ["Benchmark results (minutes)", "Small LLM", "C4", "Llama31-8b"]
       gpt_oss_20b: ["Benchmark results (minutes)", "LLM", "C4", "GPT-OSS-20B"]
       deepseekv3_671b: ["Benchmark results (minutes)", "LLM", "C4", "DeepSeekV3-671B"]
+      qwen35_397b_grpo: ["Benchmark results (minutes)", "LLM post training", "R2E-Gym-Subset (easy split)", "Qwen3.5-397B-A17B"]
       default: [" ", " ", " "]
 
   hpc: