{ "reference": "alltp8-t4-r4c1-v3", "completed": 14, "completed_runs": 14, "all_passed": true, "sampling": { "temperature": 0.7, "top_p": 0.95, "top_k": -1, "max_response_len": 128, "steps": 2, "kl_coef": 0.01 }, "scope": "Eight H100, Qwen3-8B; two real updates per topology, all responses truncated, PG=0. Finite matrix; ROCm and long training are not certified.", "rows": [ { "run": "alltp8-t1-r1c8-v6", "training_tp": 1, "training_cp": 8, "rollout_tp": 1, "rollout_cp": 8, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t1-r1c8-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "errors": [], "expected_workers_per_cp_rank": 1, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 122.91588592529297, 125.93977975845337 ], "mean_step_seconds": 124.42783284187317, "metrics": { "0": { "train/loss": 3.906373240170069e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042620241641998, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762205718318e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985952287912369, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/sleep_time": 28.886922597885132, "perf/wake_up_time": 3.757960557937622, "perf/update_weights_time": 33.03117918968201, "perf/data_preprocess_time": 0.0030951499938964844, "perf/train_wait_time": 110.1331102848053, "perf/ref_log_probs_time": 3.0428411960601807, "perf/log_probs_time": 1.2708029747009277, "perf/actor_train_time": 7.848167419433594, "perf/train_time": 12.782775640487671, "perf/log_probs_tflops": 3.19670948995382, "perf/ref_log_probs_tflops": 1.3350640626096133, "perf/actor_train_tflops": 1.5528674575782115, "perf/actor_train_tok_per_s": 272.1654477847717, "perf/step_time": 122.91588592529297, "perf/wait_time_ratio": 0.896003876600158 }, "1": { "perf/sleep_time": 9.615916728973389, "perf/wake_up_time": 4.389890432357788, "perf/update_weights_time": 36.614415884017944, "perf/data_preprocess_time": 0.003058910369873047, "perf/train_wait_time": 119.18095397949219, "perf/ref_log_probs_time": 1.1862952709197998, "perf/log_probs_time": 0.641533374786377, "perf/actor_train_time": 4.314104080200195, "perf/train_time": 6.758825778961182, "perf/log_probs_tflops": 5.760432242513589, "perf/ref_log_probs_tflops": 3.115168396399885, "perf/actor_train_tflops": 2.5698333661411183, "perf/actor_train_tok_per_s": 450.6149976589784, "perf/step_time": 125.93977975845337, "perf/wait_time_ratio": 0.9463328759830747 } } }, { "run": "alltp8-t1-r4c1-v3", "training_tp": 1, "training_cp": 8, "rollout_tp": 4, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t1-r4c1-v3", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 48.87170624732971, 40.814088106155396 ], "mean_step_seconds": 44.842897176742554, "metrics": { "0": { "train/loss": 3.906373240170069e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042620241641998, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762205718318e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985952287912369, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.750502824783325, "perf/data_preprocess_time": 0.011228084564208984, "perf/train_wait_time": 36.77781105041504, "perf/ref_log_probs_time": 3.103057384490967, "perf/log_probs_time": 1.2790584564208984, "perf/actor_train_time": 7.089299917221069, "perf/train_time": 12.093895196914673, "perf/log_probs_tflops": 3.1760768311211525, "perf/ref_log_probs_tflops": 1.3091565593958243, "perf/actor_train_tflops": 1.7190927072586368, "perf/actor_train_tok_per_s": 301.29914447705994, "perf/step_time": 48.87170624732971, "perf/wait_time_ratio": 0.7525378971687638 }, "1": { "perf/update_weights_time": 27.36841106414795, "perf/data_preprocess_time": 0.011683940887451172, "perf/train_wait_time": 35.136491537094116, "perf/ref_log_probs_time": 0.6575005054473877, "perf/log_probs_time": 0.6429979801177979, "perf/actor_train_time": 3.7618889808654785, "perf/train_time": 5.677596569061279, "perf/log_probs_tflops": 5.747311268522149, "perf/ref_log_probs_tflops": 5.620542503238744, "perf/actor_train_tflops": 2.9470642718843285, "perf/actor_train_tok_per_s": 516.761661465287, "perf/step_time": 40.814088106155396, "perf/wait_time_ratio": 0.8608912551397906 } } }, { "run": "alltp8-t2-r4c1-v4", "training_tp": 2, "training_cp": 4, "rollout_tp": 4, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t2-r4c1-v4", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 45.659247159957886, 40.97211980819702 ], "mean_step_seconds": 43.315683484077454, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042620241641998, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762205718318e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985952287912369, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.44197106361389, "perf/data_preprocess_time": 0.011423826217651367, "perf/train_wait_time": 36.18894839286804, "perf/ref_log_probs_time": 2.480095624923706, "perf/log_probs_time": 0.9902727603912354, "perf/actor_train_time": 5.677483081817627, "perf/train_time": 9.470298767089844, "perf/log_probs_tflops": 4.102291905396891, "perf/ref_log_probs_tflops": 1.6379964902413668, "perf/actor_train_tflops": 2.146578617960816, "perf/actor_train_tok_per_s": 376.223049759607, "perf/step_time": 45.659247159957886, "perf/wait_time_ratio": 0.7925874963747742 }, "1": { "perf/update_weights_time": 27.418487787246704, "perf/data_preprocess_time": 0.011552572250366211, "perf/train_wait_time": 35.22965478897095, "perf/ref_log_probs_time": 0.6383514404296875, "perf/log_probs_time": 0.6353287696838379, "perf/actor_train_time": 4.153909206390381, "perf/train_time": 5.742465019226074, "perf/log_probs_tflops": 5.81668848178718, "perf/ref_log_probs_tflops": 5.789145763155914, "perf/actor_train_tflops": 2.668938597225107, "perf/actor_train_tok_per_s": 467.9928961878481, "perf/step_time": 40.97211980819702, "perf/wait_time_ratio": 0.8598445712326259 } } }, { "run": "alltp8-t4-r1c1-v6", "training_tp": 4, "training_cp": 2, "rollout_tp": 1, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r1c1-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 50.588900566101074, 46.66252160072327 ], "mean_step_seconds": 48.62571108341217, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.69325542449951, "perf/data_preprocess_time": 0.01121068000793457, "perf/train_wait_time": 42.30711889266968, "perf/ref_log_probs_time": 2.09855055809021, "perf/log_probs_time": 0.8908624649047852, "perf/actor_train_time": 5.11951756477356, "perf/train_time": 8.281781673431396, "perf/log_probs_tflops": 4.560061837965286, "perf/ref_log_probs_tflops": 1.93580655630474, "perf/actor_train_tflops": 2.380529734114321, "perf/actor_train_tok_per_s": 417.2268134594196, "perf/step_time": 50.588900566101074, "perf/wait_time_ratio": 0.8362925151415347 }, "1": { "perf/update_weights_time": 27.530277490615845, "perf/data_preprocess_time": 0.009261131286621094, "perf/train_wait_time": 41.500365018844604, "perf/ref_log_probs_time": 0.6014037132263184, "perf/log_probs_time": 0.5608925819396973, "perf/actor_train_time": 3.836705207824707, "perf/train_time": 5.162156581878662, "perf/log_probs_tflops": 6.588622591491703, "perf/ref_log_probs_tflops": 6.144806650665487, "perf/actor_train_tflops": 2.8895961534114627, "perf/actor_train_tok_per_s": 506.6847450347085, "perf/step_time": 46.66252160072327, "perf/wait_time_ratio": 0.8893725327137347 } } }, { "run": "alltp8-t4-r1c2-v6", "training_tp": 4, "training_cp": 2, "rollout_tp": 1, "rollout_cp": 2, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r1c2-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 4, 4 ], "errors": [], "expected_workers_per_cp_rank": 4, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 57.245049238204956, 52.52477693557739 ], "mean_step_seconds": 54.884913086891174, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.560138940811157, "perf/data_preprocess_time": 0.011485099792480469, "perf/train_wait_time": 48.90938973426819, "perf/ref_log_probs_time": 2.153376579284668, "perf/log_probs_time": 0.8877286911010742, "perf/actor_train_time": 5.120829343795776, "perf/train_time": 8.335659503936768, "perf/log_probs_tflops": 4.576159326392063, "perf/ref_log_probs_tflops": 1.886519974336067, "perf/actor_train_tflops": 2.379919924890595, "perf/actor_train_tok_per_s": 417.11993440826245, "perf/step_time": 57.245049238204956, "perf/wait_time_ratio": 0.8543863685180726 }, "1": { "perf/update_weights_time": 27.389793872833252, "perf/data_preprocess_time": 0.011810302734375, "perf/train_wait_time": 47.62924385070801, "perf/ref_log_probs_time": 0.5513417720794678, "perf/log_probs_time": 0.5643515586853027, "perf/actor_train_time": 3.617114782333374, "perf/train_time": 4.895533084869385, "perf/log_probs_tflops": 6.54824015260444, "perf/ref_log_probs_tflops": 6.702756300923535, "perf/actor_train_tflops": 3.0650198507530257, "perf/actor_train_tok_per_s": 537.4449297254371, "perf/step_time": 52.52477693557739, "perf/wait_time_ratio": 0.9067957377358528 } } }, { "run": "alltp8-t4-r1c4-v6", "training_tp": 4, "training_cp": 2, "rollout_tp": 1, "rollout_cp": 4, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r1c4-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 2, 2, 2, 2 ], "errors": [], "expected_workers_per_cp_rank": 2, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 70.85922479629517, 65.27129006385803 ], "mean_step_seconds": 68.0652574300766, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 28.99709701538086, "perf/data_preprocess_time": 0.01773691177368164, "perf/train_wait_time": 62.66554141044617, "perf/ref_log_probs_time": 2.0900919437408447, "perf/log_probs_time": 0.8763141632080078, "perf/actor_train_time": 5.054697275161743, "perf/train_time": 8.193683385848999, "perf/log_probs_tflops": 4.635766600206967, "perf/ref_log_probs_tflops": 1.943640776786662, "perf/actor_train_tflops": 2.411057106654132, "perf/actor_train_tok_per_s": 422.57723533634385, "perf/step_time": 70.85922479629517, "perf/wait_time_ratio": 0.884366736872947 }, "1": { "perf/update_weights_time": 27.736430168151855, "perf/data_preprocess_time": 0.010935068130493164, "perf/train_wait_time": 60.335859060287476, "perf/ref_log_probs_time": 0.5455343723297119, "perf/log_probs_time": 0.6052279472351074, "perf/actor_train_time": 3.6236186027526855, "perf/train_time": 4.935431003570557, "perf/log_probs_tflops": 6.105979662126275, "perf/ref_log_probs_tflops": 6.774109431430098, "perf/actor_train_tflops": 3.0595186264586753, "perf/actor_train_tok_per_s": 536.4803013549049, "perf/step_time": 65.27129006385803, "perf/wait_time_ratio": 0.9243858823880762 } } }, { "run": "alltp8-t4-r1c8-v4", "training_tp": 4, "training_cp": 2, "rollout_tp": 1, "rollout_cp": 8, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r1c8-v4", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "errors": [], "expected_workers_per_cp_rank": 1, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 110.08598232269287, 107.015132188797 ], "mean_step_seconds": 108.55055725574493, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.61300778388977, "perf/data_preprocess_time": 0.011569976806640625, "perf/train_wait_time": 101.85474348068237, "perf/ref_log_probs_time": 2.1814770698547363, "perf/log_probs_time": 0.8866682052612305, "perf/actor_train_time": 4.993817567825317, "perf/train_time": 8.231238842010498, "perf/log_probs_tflops": 4.581632571217705, "perf/ref_log_probs_tflops": 1.8622189456974272, "perf/actor_train_tflops": 2.4404503411948233, "perf/actor_train_tok_per_s": 427.7288809591366, "perf/step_time": 110.08598232269287, "perf/wait_time_ratio": 0.9252290012920771 }, "1": { "perf/update_weights_time": 27.525618314743042, "perf/data_preprocess_time": 0.011382579803466797, "perf/train_wait_time": 102.13341474533081, "perf/ref_log_probs_time": 0.5497012138366699, "perf/log_probs_time": 0.5577092170715332, "perf/actor_train_time": 3.6116809844970703, "perf/train_time": 4.8817174434661865, "perf/log_probs_tflops": 6.62622998445802, "perf/ref_log_probs_tflops": 6.7227603719027424, "perf/actor_train_tflops": 3.069631193312, "perf/actor_train_tok_per_s": 538.2535191631007, "perf/step_time": 107.015132188797, "perf/wait_time_ratio": 0.9543829237639606 } } }, { "run": "alltp8-t4-r2c1-v6", "training_tp": 4, "training_cp": 2, "rollout_tp": 2, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r2c1-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 47.21352815628052, 42.65734839439392 ], "mean_step_seconds": 44.93543827533722, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.507572650909424, "perf/data_preprocess_time": 0.011548995971679688, "perf/train_wait_time": 39.06953954696655, "perf/ref_log_probs_time": 2.054391384124756, "perf/log_probs_time": 0.8848764896392822, "perf/actor_train_time": 5.031513929367065, "perf/train_time": 8.143988609313965, "perf/log_probs_tflops": 4.590909552523, "perf/ref_log_probs_tflops": 1.9774167475973532, "perf/actor_train_tflops": 2.4221663615263154, "perf/actor_train_tok_per_s": 424.5243141498559, "perf/step_time": 47.21352815628052, "perf/wait_time_ratio": 0.827507307177792 }, "1": { "perf/update_weights_time": 27.58127188682556, "perf/data_preprocess_time": 0.010405778884887695, "perf/train_wait_time": 37.80257964134216, "perf/ref_log_probs_time": 0.5464980602264404, "perf/log_probs_time": 0.5572197437286377, "perf/actor_train_time": 3.587411403656006, "perf/train_time": 4.854768753051758, "perf/log_probs_tflops": 6.632050601867562, "perf/ref_log_probs_tflops": 6.762164050933269, "perf/actor_train_tflops": 3.0903978838349815, "perf/actor_train_tok_per_s": 541.8949156538971, "perf/step_time": 42.65734839439392, "perf/wait_time_ratio": 0.8861915019150657 } } }, { "run": "alltp8-t4-r2c2-v6", "training_tp": 4, "training_cp": 2, "rollout_tp": 2, "rollout_cp": 2, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r2c2-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 4, 4 ], "errors": [], "expected_workers_per_cp_rank": 4, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 55.01402950286865, 51.28588581085205 ], "mean_step_seconds": 53.14995765686035, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.637248516082764, "perf/data_preprocess_time": 0.011330366134643555, "perf/train_wait_time": 46.739500999450684, "perf/ref_log_probs_time": 2.0313937664031982, "perf/log_probs_time": 0.9097559452056885, "perf/actor_train_time": 5.161199569702148, "perf/train_time": 8.274528503417969, "perf/log_probs_tflops": 4.465360133666977, "perf/ref_log_probs_tflops": 1.9998032859384498, "perf/actor_train_tflops": 2.3613045034736597, "perf/actor_train_tok_per_s": 413.85727700571516, "perf/step_time": 55.01402950286865, "perf/wait_time_ratio": 0.8495923934641708 }, "1": { "perf/update_weights_time": 27.531531810760498, "perf/data_preprocess_time": 0.008964061737060547, "perf/train_wait_time": 46.320478677749634, "perf/ref_log_probs_time": 0.5539965629577637, "perf/log_probs_time": 0.5565049648284912, "perf/actor_train_time": 3.692288875579834, "perf/train_time": 4.965407133102417, "perf/log_probs_tflops": 6.64056885441609, "perf/ref_log_probs_tflops": 6.670636216654188, "perf/actor_train_tflops": 3.0026168005511167, "perf/actor_train_tok_per_s": 526.502683161462, "perf/step_time": 51.28588581085205, "perf/wait_time_ratio": 0.9031818003219174 } } }, { "run": "alltp8-t4-r2c4-v4", "training_tp": 4, "training_cp": 2, "rollout_tp": 2, "rollout_cp": 4, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r2c4-v4", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 2, 2, 2, 2 ], "errors": [], "expected_workers_per_cp_rank": 2, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 66.72004508972168, 62.61780381202698 ], "mean_step_seconds": 64.66892445087433, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.686144590377808, "perf/data_preprocess_time": 0.011733293533325195, "perf/train_wait_time": 58.613070249557495, "perf/ref_log_probs_time": 2.0509254932403564, "perf/log_probs_time": 0.8790655136108398, "perf/actor_train_time": 5.004491090774536, "perf/train_time": 8.106974840164185, "perf/log_probs_tflops": 4.621257308117321, "perf/ref_log_probs_tflops": 1.9807584149093764, "perf/actor_train_tflops": 2.4352453758445622, "perf/actor_train_tok_per_s": 426.8166255581075, "perf/step_time": 66.72004508972168, "perf/wait_time_ratio": 0.8784926654461588 }, "1": { "perf/update_weights_time": 27.696548461914062, "perf/data_preprocess_time": 0.01165008544921875, "perf/train_wait_time": 57.69498658180237, "perf/ref_log_probs_time": 0.5467743873596191, "perf/log_probs_time": 0.5534319877624512, "perf/actor_train_time": 3.6588964462280273, "perf/train_time": 4.922817230224609, "perf/log_probs_tflops": 6.677441164377037, "perf/ref_log_probs_tflops": 6.758746609572671, "perf/actor_train_tflops": 3.0300197814379666, "perf/actor_train_tok_per_s": 531.3077395245986, "perf/step_time": 62.61780381202698, "perf/wait_time_ratio": 0.9213831062328142 } } }, { "run": "alltp8-t4-r4c1-v3", "training_tp": 4, "training_cp": 2, "rollout_tp": 4, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": null, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 44.254595041275024, 40.09118318557739 ], "mean_step_seconds": 42.17288911342621, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.71545124053955, "perf/data_preprocess_time": 0.010592937469482422, "perf/train_wait_time": 36.03388237953186, "perf/ref_log_probs_time": 2.1297523975372314, "perf/log_probs_time": 0.8786730766296387, "perf/actor_train_time": 5.0397021770477295, "perf/train_time": 8.220712661743164, "perf/log_probs_tflops": 4.623321275155332, "perf/ref_log_probs_tflops": 1.9074461114754926, "perf/actor_train_tflops": 2.4182309507827444, "perf/actor_train_tok_per_s": 423.83456898067624, "perf/step_time": 44.254595041275024, "perf/wait_time_ratio": 0.8142404725639013 }, "1": { "perf/update_weights_time": 27.69651222229004, "perf/data_preprocess_time": 0.011267900466918945, "perf/train_wait_time": 35.21714210510254, "perf/ref_log_probs_time": 0.5441257953643799, "perf/log_probs_time": 0.5572082996368408, "perf/actor_train_time": 3.6087944507598877, "perf/train_time": 4.8740410804748535, "perf/log_probs_tflops": 6.632186812681254, "perf/ref_log_probs_tflops": 6.791645550075899, "perf/actor_train_tflops": 3.0720864714169465, "perf/actor_train_tok_per_s": 538.6840471312132, "perf/step_time": 40.09118318557739, "perf/wait_time_ratio": 0.878426110351658 } } }, { "run": "alltp8-t4-r4c2-v3", "training_tp": 4, "training_cp": 2, "rollout_tp": 4, "rollout_cp": 2, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r4c2-v3", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "cp_rank_worker_counts": [ 4, 4 ], "errors": [], "expected_workers_per_cp_rank": 4, "passed": true, "required": true }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 51.99560809135437, 48.06762647628784 ], "mean_step_seconds": 50.031617283821106, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.554563522338867, "perf/data_preprocess_time": 0.01087188720703125, "perf/train_wait_time": 43.832419633865356, "perf/ref_log_probs_time": 2.0368356704711914, "perf/log_probs_time": 0.8850269317626953, "perf/actor_train_time": 5.069813013076782, "perf/train_time": 8.163188457489014, "perf/log_probs_tflops": 4.590129162506954, "perf/ref_log_probs_tflops": 1.9944603229322997, "perf/actor_train_tflops": 2.4038684968911346, "perf/actor_train_tok_per_s": 421.3173137728207, "perf/step_time": 51.99560809135437, "perf/wait_time_ratio": 0.8430023465992245 }, "1": { "perf/update_weights_time": 27.572540283203125, "perf/data_preprocess_time": 0.011375188827514648, "perf/train_wait_time": 43.171791553497314, "perf/ref_log_probs_time": 0.549454927444458, "perf/log_probs_time": 0.5560750961303711, "perf/actor_train_time": 3.6270573139190674, "perf/train_time": 4.895834922790527, "perf/log_probs_tflops": 6.645702284609403, "perf/ref_log_probs_tflops": 6.725773766295986, "perf/actor_train_tflops": 3.056617982781449, "perf/actor_train_tok_per_s": 535.9716794492808, "perf/step_time": 48.06762647628784, "perf/wait_time_ratio": 0.8981469383514145 } } }, { "run": "alltp8-t4-r8c1-v6", "training_tp": 4, "training_cp": 2, "rollout_tp": 8, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t4-r8c1-v6", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 43.97749853134155, 40.307480335235596 ], "mean_step_seconds": 42.142489433288574, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042619496583939, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762202080339193e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985953032970428, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 27.748736143112183, "perf/data_preprocess_time": 0.010041236877441406, "perf/train_wait_time": 35.754355669021606, "perf/ref_log_probs_time": 2.1010773181915283, "perf/log_probs_time": 0.8989040851593018, "perf/actor_train_time": 5.047789812088013, "perf/train_time": 8.223142862319946, "perf/log_probs_tflops": 4.519267401447032, "perf/ref_log_probs_tflops": 1.9334785511770889, "perf/actor_train_tflops": 2.4143564294375386, "perf/actor_train_tok_per_s": 423.1554956755313, "perf/step_time": 43.97749853134155, "perf/wait_time_ratio": 0.8130147658021172 }, "1": { "perf/update_weights_time": 27.515490531921387, "perf/data_preprocess_time": 0.010199785232543945, "perf/train_wait_time": 35.34765553474426, "perf/ref_log_probs_time": 0.6105740070343018, "perf/log_probs_time": 0.5562090873718262, "perf/actor_train_time": 3.6301169395446777, "perf/train_time": 4.959824800491333, "perf/log_probs_tflops": 6.644101329285815, "perf/ref_log_probs_tflops": 6.052516966318201, "perf/actor_train_tflops": 3.054041727838821, "perf/actor_train_tok_per_s": 535.5199384413864, "perf/step_time": 40.307480335235596, "perf/wait_time_ratio": 0.8769502643370243 } } }, { "run": "alltp8-t8-r4c1-v4", "training_tp": 8, "training_cp": 1, "rollout_tp": 4, "rollout_cp": 1, "status": "SUCCEEDED", "validated": true, "cross_topology_equal": true, "comparison": { "name": "alltp8-t8-r4c1-v4", "weights": [ { "version": "weight_v000003", "equal": true, "different_tensors": [] }, { "version": "weight_v000002", "equal": true, "different_tensors": [] }, { "version": "weight_v000001", "equal": true, "different_tensors": [] } ], "tokens_equal": true, "norms_equal": { "0": true, "1": true } }, "logprob_check": { "bitwise_equal": true, "bitwise_mismatch_count": 0, "element_count": 2048, "errors": [], "max_abs_diff": 0.0, "mismatch_count": 0, "mismatch_k3_kl": 0.0, "mismatch_kl": 0.0, "passed": true, "require_exact": true, "sample_count": 16, "torch_equal": true, "train_rollout_logprob_abs_diff": 0.0 }, "pcp_execution": { "errors": [], "passed": true, "required": false }, "gradient_norms": { "0": 0.05702933925260603, "1": 0.0711779018075793 }, "step_seconds": [ 45.99890923500061, 38.620574951171875 ], "mean_step_seconds": 42.30974209308624, "metrics": { "0": { "train/loss": 3.906372876372188e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.11042621731758118, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.003906372934579849, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.05702933925260603, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 0 }, "1": { "train/loss": 2.7762205718318e-05, "train/pg_loss": 0.0, "train/entropy_loss": 0.08985951542854309, "train/pg_clipfrac": 0.0, "train/ppo_kl": 0.0, "train/train_rollout_logprob_abs_diff": 0.0, "train/train_current_rollout_logprob_mismatch_count": 0.0, "train/train_current_rollout_logprob_max_abs_diff": 0.0, "train/train_current_rollout_logprob_numel": 128.0, "train/kl_loss": 0.002776220440864563, "train/ois": 1.0, "train/tis": 1.0, "train/tis_clipfrac": 0.0, "train/tis_abs": 0.0, "train/grad_norm": 0.0711779018075793, "train/lr-pg_0": 5e-07, "train/lr-pg_1": 5e-07, "train/global_batch_size": 8, "train/step": 1 } }, "perf": { "0": { "perf/update_weights_time": 30.987462282180786, "perf/data_preprocess_time": 0.011212825775146484, "perf/train_wait_time": 39.39871048927307, "perf/ref_log_probs_time": 1.5750758647918701, "perf/log_probs_time": 0.6680095195770264, "perf/actor_train_time": 4.260600328445435, "perf/train_time": 6.600198745727539, "perf/log_probs_tflops": 6.081332391281254, "perf/ref_log_probs_tflops": 2.5791696894706733, "perf/actor_train_tflops": 2.860433471287538, "perf/actor_train_tok_per_s": 501.337801093247, "perf/step_time": 45.99890923500061, "perf/wait_time_ratio": 0.8565140161909439 }, "1": { "perf/update_weights_time": 27.36836576461792, "perf/data_preprocess_time": 0.010560989379882812, "perf/train_wait_time": 34.815091371536255, "perf/ref_log_probs_time": 0.4047837257385254, "perf/log_probs_time": 0.43824267387390137, "perf/actor_train_time": 2.877342700958252, "perf/train_time": 3.80548357963562, "perf/log_probs_tflops": 8.432564323554066, "perf/ref_log_probs_tflops": 9.129590202831317, "perf/actor_train_tflops": 3.8530442017253677, "perf/actor_train_tok_per_s": 675.6233796386446, "perf/step_time": 38.620574951171875, "perf/wait_time_ratio": 0.9014648646622453 } } } ] }