diff --git a/benchmarks/results/h20-mcore-all-families/README.md b/benchmarks/results/h20-mcore-all-families/README.md new file mode 100644 index 000000000..9d93d8de6 --- /dev/null +++ b/benchmarks/results/h20-mcore-all-families/README.md @@ -0,0 +1,16 @@ +# ScaleRLT — H20 MCore two-update correctness + +All twelve frozen tiny FP32 CUDA configurations passed on the original H20 using the existing 0z5a environment, source `2ad271d5fbb58f06170a05277b03663bf9d66b24`. Each candidate performs two actual MCore/DDP/AdamW updates against the ordinary replay oracle. Both token-mean and response-mean losses are included. The original full-gradient, parameter, Adam first/second-moment, counter and sample-accounting assertions are unchanged. + +| Family / schedule | Candidate updates | Maximum main-gradient error | Maximum parameter error | Complete update gate | Speed ratio | +| --- | ---: | ---: | ---: | --- | --- | +| Ouro ordinary | 4 | 1.192093e-07 | 5.960464e-08 | PASS | NOT_MEASURED | +| Ouro joint | 4 | 1.192093e-07 | 5.960464e-08 | PASS | NOT_MEASURED | +| Nanbeige ordinary | 4 | 5.960464e-08 | 2.980232e-08 | PASS | NOT_MEASURED | +| Nanbeige joint | 4 | 5.960464e-08 | 2.980232e-08 | PASS | NOT_MEASURED | +| Huginn independent latent | 4 | 7.152557e-07 | 9.302312e-08 | PASS | NOT_MEASURED | +| Huginn mixed latent | 4 | 1.177192e-06 | 2.094894e-07 | PASS | NOT_MEASURED | + +Pytest reports **12 passed, 36 warnings in 22.66s**. This is verification duration, not a baseline/candidate speed comparison. Twenty-four candidate updates and their matched ordinary references use tiny fixed trajectories; they do not establish official checkpoint accuracy, online reward convergence, memory saving or large-scale stability. Existing CPU BF16 failures remain failed. + +The natural controller, test, wrapper and finalizer exits are recorded; all four recorded PIDs are absent in a fresh read-only check and the compute list is empty. The offbox archive is 21,546 bytes with 84 verified payloads plus its manifest, SHA-256 `5cbea08375482e57936c9808a352fa3d12b22aabbb4a04a21723deac117f42ff`. Raw data: [`cuda-gate.tar.gz`](cuda-gate.tar.gz); independent handback: [`independent-handback.json`](independent-handback.json). diff --git a/benchmarks/results/h20-mcore-all-families/artifact.sha256.json b/benchmarks/results/h20-mcore-all-families/artifact.sha256.json new file mode 100644 index 000000000..f6169e93c --- /dev/null +++ b/benchmarks/results/h20-mcore-all-families/artifact.sha256.json @@ -0,0 +1,14 @@ +{ + "README.md": { + "bytes": 1961, + "sha256": "fdf98ac52d107f3e722a0386309b6b38d5fda3933fc1b9a949678b895a46c65c" + }, + "cuda-gate.tar.gz": { + "bytes": 21546, + "sha256": "5cbea08375482e57936c9808a352fa3d12b22aabbb4a04a21723deac117f42ff" + }, + "independent-handback.json": { + "bytes": 12383, + "sha256": "97c3785f3e2194e75195eb1e6aac4aad4b5beae3a56ed0fc3faa1a0de054676b" + } +} diff --git a/benchmarks/results/h20-mcore-all-families/cuda-gate.tar.gz b/benchmarks/results/h20-mcore-all-families/cuda-gate.tar.gz new file mode 100644 index 000000000..f0753cd98 Binary files /dev/null and b/benchmarks/results/h20-mcore-all-families/cuda-gate.tar.gz differ diff --git a/benchmarks/results/h20-mcore-all-families/independent-handback.json b/benchmarks/results/h20-mcore-all-families/independent-handback.json new file mode 100644 index 000000000..d3cedc55a --- /dev/null +++ b/benchmarks/results/h20-mcore-all-families/independent-handback.json @@ -0,0 +1,377 @@ +{ + "state": "VERIFIED_WHOLE_HANDBACK", + "recorded_unix": 1791246104.340163, + "grant_sha256": "cc40350d2a88133d18e351388d866f06d7929dbba83c9a2adf9d3c84f43ed23f", + "packet_sha256": "79ff7ca7e98ab93173fc10bf7f5594582c7a4c849e253ba42f7629dec26452c0", + "source": "2ad271d5fbb58f06170a05277b03663bf9d66b24", + "archive_sha256": "5cbea08375482e57936c9808a352fa3d12b22aabbb4a04a21723deac117f42ff", + "archive_bytes": 21546, + "payload_files_verified": 84, + "regular_members": 85, + "actual_cuda_mcore_two_update_cases": 12, + "case_status": "12PASS", + "natural_test_and_controller_exit_code": 0, + "finalizer_actual_exit_code": 0, + "finalizer_pid": 224479, + "fresh_owned_process_absence": { + "absent_owned_pids": [ + 223243, + 223290, + 224479, + 223242 + ], + "empty_compute": true + }, + "cases": [ + { + "name": "test_actual_mcore_prefix_train_step[ouro-False-none-token_mean]", + "seconds": 7.87, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 7.450580596923828e-08, + "parameter_max_error": 5.8673322200775146e-08, + "delta_norm": 0.03521193191409111, + "loss": { + "loss": 0.4919145584106445, + "rltt_terminal_logprob": -2.533199501037598, + "entropy_loss": 2.5616533279418947 + } + }, + { + "update": 1, + "gradient_max_error": 5.960464477539063e-08, + "parameter_max_error": 5.960464477539063e-08, + "delta_norm": 0.0270307045429945, + "loss": { + "loss": 0.47786765098571776, + "rltt_terminal_logprob": -2.5281244277954102, + "entropy_loss": 2.5631681442260743 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[ouro-False-none-response_mean]", + "seconds": 1.131, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 1.1920928955078125e-07, + "parameter_max_error": 1.6763806343078613e-08, + "delta_norm": 0.03511379286646843, + "loss": { + "loss": 0.6862307071685791, + "rltt_terminal_logprob": -2.523184967041016, + "entropy_loss": 2.5616641998291017 + } + }, + { + "update": 1, + "gradient_max_error": 8.940696716308594e-08, + "parameter_max_error": 2.0489096641540527e-08, + "delta_norm": 0.028335392475128174, + "loss": { + "loss": 0.6705065250396729, + "rltt_terminal_logprob": -2.5039222717285154, + "entropy_loss": 2.562147521972656 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[ouro-True-none-token_mean]", + "seconds": 1.633, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 7.450580596923828e-08, + "parameter_max_error": 5.8673322200775146e-08, + "delta_norm": 0.03521193191409111, + "loss": { + "loss": 0.4919145584106445, + "rltt_terminal_logprob": -2.533199501037598, + "entropy_loss": 2.5616533279418947 + } + }, + { + "update": 1, + "gradient_max_error": 5.960464477539063e-08, + "parameter_max_error": 5.960464477539063e-08, + "delta_norm": 0.0270307045429945, + "loss": { + "loss": 0.47786765098571776, + "rltt_terminal_logprob": -2.5281244277954102, + "entropy_loss": 2.5631681442260743 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[ouro-True-none-response_mean]", + "seconds": 1.636, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 1.1920928955078125e-07, + "parameter_max_error": 1.6763806343078613e-08, + "delta_norm": 0.03511379286646843, + "loss": { + "loss": 0.6862307071685791, + "rltt_terminal_logprob": -2.523184967041016, + "entropy_loss": 2.5616641998291017 + } + }, + { + "update": 1, + "gradient_max_error": 8.940696716308594e-08, + "parameter_max_error": 2.0489096641540527e-08, + "delta_norm": 0.028335392475128174, + "loss": { + "loss": 0.6705065250396729, + "rltt_terminal_logprob": -2.5039222717285154, + "entropy_loss": 2.562147521972656 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[nanbeige-False-none-token_mean]", + "seconds": 1.15, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 2.9802322387695312e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03637052699923515, + "loss": { + "loss": 0.1866389036178589, + "rltt_terminal_logprob": -2.5343578338623045, + "entropy_loss": 2.176030731201172 + } + }, + { + "update": 1, + "gradient_max_error": 2.9802322387695312e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03611896559596062, + "loss": { + "loss": 0.16810529232025145, + "rltt_terminal_logprob": -2.5119512557983397, + "entropy_loss": 2.1729530334472655 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[nanbeige-False-none-response_mean]", + "seconds": 1.115, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 5.960464477539063e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03630533441901207, + "loss": { + "loss": 0.33352444171905515, + "rltt_terminal_logprob": -2.3822282791137694, + "entropy_loss": 2.1632238388061524 + } + }, + { + "update": 1, + "gradient_max_error": 5.960464477539063e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03600601851940155, + "loss": { + "loss": 0.31150450706481936, + "rltt_terminal_logprob": -2.3452426910400392, + "entropy_loss": 2.1597824096679688 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[nanbeige-True-none-token_mean]", + "seconds": 1.117, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 2.9802322387695312e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03637052699923515, + "loss": { + "loss": 0.1866389036178589, + "rltt_terminal_logprob": -2.5343578338623045, + "entropy_loss": 2.176030731201172 + } + }, + { + "update": 1, + "gradient_max_error": 2.9802322387695312e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03611896559596062, + "loss": { + "loss": 0.16810529232025145, + "rltt_terminal_logprob": -2.5119512557983397, + "entropy_loss": 2.1729530334472655 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[nanbeige-True-none-response_mean]", + "seconds": 1.137, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 5.960464477539063e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03630533441901207, + "loss": { + "loss": 0.33352444171905515, + "rltt_terminal_logprob": -2.3822282791137694, + "entropy_loss": 2.1632238388061524 + } + }, + { + "update": 1, + "gradient_max_error": 5.960464477539063e-08, + "parameter_max_error": 2.9802322387695312e-08, + "delta_norm": 0.03600601851940155, + "loss": { + "loss": 0.31150450706481936, + "rltt_terminal_logprob": -2.3452426910400392, + "entropy_loss": 2.1597824096679688 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[huginn-False-independent-token_mean]", + "seconds": 1.132, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 2.384185791015625e-07, + "parameter_max_error": 5.549372872337699e-08, + "delta_norm": 0.04711056873202324, + "loss": { + "loss": -0.029158186912536622, + "rltt_terminal_logprob": -4.236648178100586, + "entropy_loss": 0.6799128532409668 + } + }, + { + "update": 1, + "gradient_max_error": 6.556510925292969e-07, + "parameter_max_error": 9.30231180973351e-08, + "delta_norm": 0.046330101788043976, + "loss": { + "loss": -0.10193557739257812, + "rltt_terminal_logprob": -4.109548187255859, + "entropy_loss": 0.694879150390625 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[huginn-False-independent-response_mean]", + "seconds": 1.13, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 3.5762786865234375e-07, + "parameter_max_error": 4.259163688402623e-08, + "delta_norm": 0.04705020412802696, + "loss": { + "loss": 0.35195763111114503, + "rltt_terminal_logprob": -3.976166534423828, + "entropy_loss": 0.6500929832458496 + } + }, + { + "update": 1, + "gradient_max_error": 7.152557373046875e-07, + "parameter_max_error": 8.133065421134233e-08, + "delta_norm": 0.04666852205991745, + "loss": { + "loss": 0.2437004804611206, + "rltt_terminal_logprob": -3.8089141845703125, + "entropy_loss": 0.6827536582946777 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[huginn-False-mixed-token_mean]", + "seconds": 1.122, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 1.7881393432617188e-07, + "parameter_max_error": 1.7881393432617188e-07, + "delta_norm": 0.04707252234220505, + "loss": { + "loss": 0.02080817222595215, + "rltt_terminal_logprob": -3.8632606506347655, + "entropy_loss": 0.7506943702697754 + } + }, + { + "update": 1, + "gradient_max_error": 1.1771917343139648e-06, + "parameter_max_error": 1.7881393432617188e-07, + "delta_norm": 0.046352267265319824, + "loss": { + "loss": -0.05286041498184204, + "rltt_terminal_logprob": -3.748757553100586, + "entropy_loss": 0.7775139331817627 + } + } + ] + }, + { + "name": "test_actual_mcore_prefix_train_step[huginn-False-mixed-response_mean]", + "seconds": 1.121, + "status": "PASS", + "actual_mcore_two_update": [ + { + "update": 0, + "gradient_max_error": 2.980232238769531e-07, + "parameter_max_error": 1.2610689736902714e-07, + "delta_norm": 0.04701389744877815, + "loss": { + "loss": 0.39359636306762696, + "rltt_terminal_logprob": -3.665010452270508, + "entropy_loss": 0.7090775966644287 + } + }, + { + "update": 1, + "gradient_max_error": 1.0132789611816406e-06, + "parameter_max_error": 2.094893716275692e-07, + "delta_norm": 0.04656941071152687, + "loss": { + "loss": 0.2847589015960693, + "rltt_terminal_logprob": -3.5019454956054688, + "entropy_loss": 0.7516379833221436 + } + } + ] + } + ], + "scope": "Twelve tiny actual CUDA MCore/DDP/Adam cases; Ouro/Nanbeige ordinary/joint plus Huginn independent/mixed latent, not official online RL or speed/convergence evidence", + "logical_whole_returned": true +}