Add HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3 step 304578 weights

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/README.md

@@ -0,0 +1,16 @@
+# HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3
+
+Successful final epoch4 checkpoint, optimizer step304578; 1,232,342,016 parameters. Native weights-only export: model.safetensors contains EMA weights matching in-training reference evaluation; original_model.safetensors contains ordinary weights. Optimizer/RNG/data state stays in the local full checkpoint.
+
+Native HRM epoch3 continuation source and full fourth-epoch K2/K3 curriculum are pinned in training_recipe.json/config.json. Six feedback sources are ['L1', 'L3', 'H1', 'L4', 'L6', 'H2']; evaluation depthK3. Preserve the original cycle-output interpretation, not physical layer numbering. Original W&B: https://wandb.ai/LoopTF-4-CSPs/causal_hrm_text/runs/wzacavn3. No unverified final autoregressive benchmark score is claimed.
+
+Use https://github.com/FFishy-git/loopTF-workspace/tree/24e8310581000309bad97cb807ed5857ab0d0c85 and pinned HRM-Text upstreamaaa948ea674fd84b7bc455c9cfb455ecfefdf914. This custom native model is not a Transformers AutoModel checkpoint. With dependencies installed:
+
+    import json
+    from safetensors.torch import load_file
+    from hrm_text_tf.official.evaluation.checkpoint import build_fbw_model
+    config = json.load(open('config.json'))
+    state = load_file('model.safetensors', device='cpu')
+    model = build_fbw_model(state, config['native_model_config'], '/path/to/pinned/HRM-Text', config['fbw']['num_passes'], tuple(config['fbw']['feedback_sources']))
+
+The builder loads strictly with assign=True, preserving saved tensor dtypes and rebuilding rotary buffers. To use ordinary final weights, select original_model.safetensors. Tokenizer and special-token metadata are included. Full training recovery state is excluded.

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/checkpoint_receipt.json

@@ -0,0 +1,134 @@
+{
+  "epoch": 4,
+  "epoch_step": 76143,
+  "epoch_micro_step": 76143,
+  "accumulation_steps": 1,
+  "reference_replicas": 32,
+  "gradient_reduction": "sum",
+  "gradient_reference_replicas": 32,
+  "gradient_scale": 1.0,
+  "step": 304578,
+  "input_tokens": 59692873068,
+  "response_labels": 25612353450,
+  "examples": 230796395,
+  "debug_partial_epoch": false,
+  "max_grad_norm": 1.0,
+  "continuation": {
+    "mode": "hrm_to_fbw",
+    "checkpoint": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/causal_hrm_text/continuation_sources/ty8nbhn0/full_epoch_3/checkpoints/fsdp2_epoch_3",
+    "preserve_optimizer": true,
+    "preserve_ema": true,
+    "curriculum": [
+      {
+        "start_update": 228436,
+        "mixture": {
+          "2": 1.0
+        }
+      },
+      {
+        "start_update": 274312,
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        }
+      }
+    ],
+    "source_epoch": 3,
+    "source_step": 228435,
+    "source_world_size": 32,
+    "source_recipe_sha256": "8e53dcb59af9678277550dfa2540f3f2e32268c9386fa8f42936e616d8c3307b",
+    "checkpoint_metadata_sha256": "4898bea588e7717f169d4b648be3faee795b2b8f821815301a886fca132543dd",
+    "source_run_name": "HRM-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-H800x32-20260925-002521",
+    "source_wandb_run_id": "ty8nbhn0",
+    "data_epoch_index": 3,
+    "rng_restored": false,
+    "bit_exact_resume": false
+  },
+  "architecture": {
+    "arm": "hrm_fbw",
+    "fbw": {
+      "num_passes": 3,
+      "feedback_sources": [
+        "L1",
+        "L3",
+        "H1",
+        "L4",
+        "L6",
+        "H2"
+      ]
+    },
+    "hrm": {
+      "h_layers": 18,
+      "l_layers": 15,
+      "h_cycles": 2,
+      "l_cycles": 3
+    },
+    "forward_blocks": 378,
+    "response_ce_weights": [
+      0.5,
+      0.25,
+      0.25
+    ],
+    "pass_schedule": null,
+    "feedback": "shifted previous-pass latent, normalized FBW gate, reset at document starts",
+    "inner_bp": "K1: upstream warmup; K>1: preceding passes max(5, scheduled BP), final pass full BP8",
+    "feedback_source_fusion": {
+      "operation": "concatenate_then_bias_free_linear",
+      "initialization": "final_H2_identity",
+      "states": "outputs of complete L/H stack invocations, including stack normalization",
+      "projection_parameters": 14155776,
+      "lm_head_source": "H2"
+    }
+  },
+  "pass_schedule": {
+    "progress": "absolute_optimizer_updates",
+    "stages": [
+      {
+        "start_update": 228436,
+        "mixture": {
+          "2": 1.0
+        }
+      },
+      {
+        "start_update": 274312,
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        }
+      }
+    ],
+    "estimated_total_steps": 304789,
+    "seed": 0,
+    "sampler": "sha256-counter-v1",
+    "step_convention": "completed updates before the next update; first update uses zero",
+    "start_completed_updates": [
+      228435,
+      274311
+    ],
+    "first_update_per_stage": [
+      228436,
+      274312
+    ],
+    "after_horizon": "keep final mixture"
+  },
+  "pass_accounting": {
+    "1": {
+      "updates": 228435,
+      "input_tokens": 44769895151,
+      "response_labels": 19208956003,
+      "examples": 173097293
+    },
+    "2": {
+      "updates": 67046,
+      "input_tokens": 13140065044,
+      "response_labels": 5638502351,
+      "examples": 50810679
+    },
+    "3": {
+      "updates": 9097,
+      "input_tokens": 1782912873,
+      "response_labels": 764895096,
+      "examples": 6888423
+    }
+  }
+}

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/config.json

@@ -0,0 +1,263 @@
+{
+  "architecture": "hrm_fbw",
+  "architecture_receipt": {
+    "arm": "hrm_fbw",
+    "fbw": {
+      "feedback_sources": [
+        "L1",
+        "L3",
+        "H1",
+        "L4",
+        "L6",
+        "H2"
+      ],
+      "num_passes": 3
+    },
+    "feedback": "shifted previous-pass latent, normalized FBW gate, reset at document starts",
+    "feedback_source_fusion": {
+      "initialization": "final_H2_identity",
+      "lm_head_source": "H2",
+      "operation": "concatenate_then_bias_free_linear",
+      "projection_parameters": 14155776,
+      "states": "outputs of complete L/H stack invocations, including stack normalization"
+    },
+    "forward_blocks": 378,
+    "hrm": {
+      "h_cycles": 2,
+      "h_layers": 18,
+      "l_cycles": 3,
+      "l_layers": 15
+    },
+    "inner_bp": "K1: upstream warmup; K>1: preceding passes max(5, scheduled BP), final pass full BP8",
+    "lookahead": null,
+    "pass_schedule": null,
+    "response_ce_weights": [
+      0.5,
+      0.25,
+      0.25
+    ]
+  },
+  "checkpoint_format": "safetensors_native_hrm_fbw_state_dict",
+  "continuation": {
+    "bit_exact_resume": false,
+    "checkpoint": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/causal_hrm_text/continuation_sources/ty8nbhn0/full_epoch_3/checkpoints/fsdp2_epoch_3",
+    "checkpoint_metadata_sha256": "4898bea588e7717f169d4b648be3faee795b2b8f821815301a886fca132543dd",
+    "curriculum": [
+      {
+        "mixture": {
+          "2": 1.0
+        },
+        "start_update": 228436
+      },
+      {
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        },
+        "start_update": 274312
+      }
+    ],
+    "data_epoch_index": 3,
+    "mode": "hrm_to_fbw",
+    "preserve_ema": true,
+    "preserve_optimizer": true,
+    "rng_restored": false,
+    "source_epoch": 3,
+    "source_recipe_sha256": "8e53dcb59af9678277550dfa2540f3f2e32268c9386fa8f42936e616d8c3307b",
+    "source_run_name": "HRM-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-H800x32-20260925-002521",
+    "source_step": 228435,
+    "source_wandb_run_id": "ty8nbhn0",
+    "source_world_size": 32
+  },
+  "dtype": "preserved per tensor from source checkpoint",
+  "epoch": 4,
+  "eval_num_passes": 3,
+  "export_code_revision": "24e8310581000309bad97cb807ed5857ab0d0c85",
+  "fbw": {
+    "feedback_sources": [
+      "L1",
+      "L3",
+      "H1",
+      "L4",
+      "L6",
+      "H2"
+    ],
+    "num_passes": 3
+  },
+  "iteration": 304578,
+  "native_model_config": {
+    "H_cycles": 2,
+    "H_override": {
+      "n_layers": 18
+    },
+    "L_cycles": 3,
+    "attn_type": "causal",
+    "bp_max_steps": 5,
+    "bp_warmup_ratio": 0.2,
+    "expansion": 4,
+    "half_layers": false,
+    "head": "lm_head@LMHead",
+    "hidden_size": 1536,
+    "init_type": "lecun_normal",
+    "max_seq_len": 4096,
+    "n_layers": 15,
+    "name": "baselines.hrm_nocarry_bp_warmup@HierarchicalReasoningModel",
+    "norm_eps": 1e-06,
+    "norm_type": "pre",
+    "num_heads": 12,
+    "pos_emb_type": "rope",
+    "rope_theta": 10000.0,
+    "vocab_size": 65536
+  },
+  "ordinary_weights_file": "original_model.safetensors",
+  "original_training_run_name": "HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun",
+  "parameter_count": 1232342016,
+  "pass_schedule": {
+    "after_horizon": "keep final mixture",
+    "estimated_total_steps": 304789,
+    "first_update_per_stage": [
+      228436,
+      274312
+    ],
+    "progress": "absolute_optimizer_updates",
+    "sampler": "sha256-counter-v1",
+    "seed": 0,
+    "stages": [
+      {
+        "mixture": {
+          "2": 1.0
+        },
+        "start_update": 228436
+      },
+      {
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        },
+        "start_update": 274312
+      }
+    ],
+    "start_completed_updates": [
+      228435,
+      274311
+    ],
+    "step_convention": "completed updates before the next update; first update uses zero"
+  },
+  "pipeline_parallel_size": 1,
+  "run_name": "HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3",
+  "source_checkpoint": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/launchings/20261007-004905_hrm-fbw-epoch3-rerun/HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun/checkpoints/fsdp2_epoch_4",
+  "source_metadata_sha256": "159c5528d402be1f6382ffb16cb6d4578075d3540249e93c5697732ee512b80e",
+  "source_wandb": {
+    "entity": "LoopTF-4-CSPs",
+    "group": "official_recipe",
+    "history_step": 304575,
+    "history_step_offset": 0,
+    "id": "wzacavn3",
+    "name": "HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun",
+    "project": "causal_hrm_text",
+    "tags": [
+      "product",
+      "causal_hrm_text",
+      "official_recipe",
+      "hrm_fbw",
+      "flex_causal",
+      "h800",
+      "resume",
+      "hrm_to_fbw",
+      "six_sources",
+      "epoch4",
+      "progressive_passes",
+      "optimized",
+      "selective_recompute",
+      "model_fsdp",
+      "k2_two_thirds_retention",
+      "rerun_from_epoch3",
+      "progressive_passes"
+    ],
+    "url": "https://wandb.ai/LoopTF-4-CSPs/causal_hrm_text/runs/wzacavn3"
+  },
+  "tensor_parallel_size": 1,
+  "tokenizer_info": {
+    "boq": "<|im_start|>",
+    "condition_mapping": {
+      "cot": "<|object_ref_end|>",
+      "direct": "<|object_ref_start|>",
+      "noisy": "<|quad_start|>",
+      "synth": "<|quad_end|>"
+    },
+    "eoa": "<|box_end|>",
+    "eoq": "<|im_end|>",
+    "tokenizer_path": "/root/dl/runtime/script/loopTF-workspace/data/causal_hrm_text_sources/data_io/trained_tokenizers/bpe/tokenizer.json",
+    "vocab_size": 65536
+  },
+  "trainable_parameter_count": 1232342016,
+  "training": {
+    "elapsed_seconds": 69863.77334650792,
+    "epochs": 4,
+    "examples": 230796395,
+    "final_checkpoint": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/launchings/20261007-004905_hrm-fbw-epoch3-rerun/HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun/checkpoints/fsdp2_epoch_4",
+    "input_tokens": 59692873068,
+    "partial_epoch": false,
+    "pass_accounting": {
+      "1": {
+        "examples": 173097293,
+        "input_tokens": 44769895151,
+        "response_labels": 19208956003,
+        "updates": 228435
+      },
+      "2": {
+        "examples": 50810679,
+        "input_tokens": 13140065044,
+        "response_labels": 5638502351,
+        "updates": 67046
+      },
+      "3": {
+        "examples": 6888423,
+        "input_tokens": 1782912873,
+        "response_labels": 764895096,
+        "updates": 9097
+      }
+    },
+    "pass_schedule": {
+      "after_horizon": "keep final mixture",
+      "estimated_total_steps": 304789,
+      "first_update_per_stage": [
+        228436,
+        274312
+      ],
+      "progress": "absolute_optimizer_updates",
+      "sampler": "sha256-counter-v1",
+      "seed": 0,
+      "stages": [
+        {
+          "mixture": {
+            "2": 1.0
+          },
+          "start_update": 228436
+        },
+        {
+          "mixture": {
+            "2": 0.7,
+            "3": 0.3
+          },
+          "start_update": 274312
+        }
+      ],
+      "start_completed_updates": [
+        228435,
+        274311
+      ],
+      "step_convention": "completed updates before the next update; first update uses zero"
+    },
+    "response_labels": 25612353450,
+    "status": "complete",
+    "step": 304578,
+    "stop_at_unix_time": null
+  },
+  "training_dispatch_revision": "24e8310581000309bad97cb807ed5857ab0d0c85",
+  "training_source_revision": "58f77006338b4316dae792560be944e3d73367ab plus qualified source snapshot/code.diff",
+  "transformers_auto_model_compatible": false,
+  "upstream_revision": "aaa948ea674fd84b7bc455c9cfb455ecfefdf914",
+  "weights": "ema",
+  "world_size": 32
+}

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/launch_config.env

@@ -0,0 +1,5 @@
+RUN_NAME=HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3
+SOURCE_WANDB_ID=wzacavn3
+CHECKPOINT_STEP=304578
+TRAINING_SOURCE_REVISION=58f77006338b4316dae792560be944e3d73367ab plus qualified source snapshot/code.diff
+EFFECTIVE_CONFIG=launch_config.yaml

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/launch_config.yaml

@@ -0,0 +1,3 @@
+base: /root/dl/runtime/script/loopTF-workspace/output/launchings/20261007-004905_hrm-fbw-epoch3-rerun/source_snapshot/configs/causal_hrm_text/official_recipe/hrm_fbw/HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2.yaml
+project_name: causal_hrm_text
+run_name: HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/manifest.json

@@ -0,0 +1,48 @@
+{
+  "checkpoint_iteration": 304578,
+  "files": {
+    "README.md": {
+      "sha256": "95bbc0a4454ba91b8af7aca384c273af829a8a27892e312220b98504d3506dc2",
+      "size": 1812
+    },
+    "checkpoint_receipt.json": {
+      "sha256": "2897b0f4ae5f82f5c1b0be286a1346b4e64288425695ebe4d1e207fd963c292d",
+      "size": 3462
+    },
+    "config.json": {
+      "sha256": "09e8376b2854c15be9a06a3654d8c667780a5c23e7aee4ddb6caf1d5c6627110",
+      "size": 7985
+    },
+    "launch_config.env": {
+      "sha256": "2db6c16aec3cea3f96292f4c9ddfce394ebf29f8538f7480bbbf6699f9b43802",
+      "size": 336
+    },
+    "launch_config.yaml": {
+      "sha256": "1d9f40773888f020d2a33265a24b9d7b6626688ce431eccffb59f61642c79e8b",
+      "size": 425
+    },
+    "model.safetensors": {
+      "sha256": "28691e1f86d814e2f3041c81f65aa940ae0662d1eaa5f4c8e0910feca02a52c7",
+      "size": 4929389352
+    },
+    "original_model.safetensors": {
+      "sha256": "8106e142bcd4ec36cc1e081080228a42c219229bd977ad6cccb5896e91e6f931",
+      "size": 4929389352
+    },
+    "tokenizer/tokenizer.json": {
+      "sha256": "d569c356b9c88d4fb5fc2ab43f0b9f9c445336d3121588f1f2f6e2bdbc6fdc0f",
+      "size": 4732721
+    },
+    "training_completion.json": {
+      "sha256": "afc8794cf05903b48cf4f593a3abca331d2632854fea424a80c0f8284d7cd576",
+      "size": 1659
+    },
+    "training_recipe.json": {
+      "sha256": "81309a65964dce0e1b12931ead82474251cd8860b17a5cf70d23bb2f6cbf971c",
+      "size": 8174
+    }
+  },
+  "manifest_hashes_itself": false,
+  "path_in_repo": "HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578",
+  "repo_id": "Siyuc/loopTF-workspace"
+}

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28691e1f86d814e2f3041c81f65aa940ae0662d1eaa5f4c8e0910feca02a52c7
+size 4929389352

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/original_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8106e142bcd4ec36cc1e081080228a42c219229bd977ad6cccb5896e91e6f931
+size 4929389352

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/training_completion.json

@@ -0,0 +1,63 @@
+{
+  "status": "complete",
+  "partial_epoch": false,
+  "stop_at_unix_time": null,
+  "final_checkpoint": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/launchings/20261007-004905_hrm-fbw-epoch3-rerun/HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun/checkpoints/fsdp2_epoch_4",
+  "epochs": 4,
+  "step": 304578,
+  "input_tokens": 59692873068,
+  "response_labels": 25612353450,
+  "examples": 230796395,
+  "elapsed_seconds": 69863.77334650792,
+  "pass_schedule": {
+    "progress": "absolute_optimizer_updates",
+    "stages": [
+      {
+        "start_update": 228436,
+        "mixture": {
+          "2": 1.0
+        }
+      },
+      {
+        "start_update": 274312,
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        }
+      }
+    ],
+    "estimated_total_steps": 304789,
+    "seed": 0,
+    "sampler": "sha256-counter-v1",
+    "step_convention": "completed updates before the next update; first update uses zero",
+    "start_completed_updates": [
+      228435,
+      274311
+    ],
+    "first_update_per_stage": [
+      228436,
+      274312
+    ],
+    "after_horizon": "keep final mixture"
+  },
+  "pass_accounting": {
+    "1": {
+      "updates": 228435,
+      "input_tokens": 44769895151,
+      "response_labels": 19208956003,
+      "examples": 173097293
+    },
+    "2": {
+      "updates": 67046,
+      "input_tokens": 13140065044,
+      "response_labels": 5638502351,
+      "examples": 50810679
+    },
+    "3": {
+      "updates": 9097,
+      "input_tokens": 1782912873,
+      "response_labels": 764895096,
+      "examples": 6888423
+    }
+  }
+}

HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun--wzacavn3/step_304578/training_recipe.json

@@ -0,0 +1,274 @@
+{
+  "launching_card": "/root/dl/runtime/script/loopTF-workspace/launchings/product/20261007_hrm_fbw_epoch3_rerun/20261007-004905_hrm-fbw-epoch3-rerun.md",
+  "upstream_revision": "aaa948ea674fd84b7bc455c9cfb455ecfefdf914",
+  "recipe": {
+    "arch": {
+      "name": "baselines.hrm_nocarry_bp_warmup@HierarchicalReasoningModel",
+      "head": "lm_head@LMHead",
+      "half_layers": false,
+      "H_cycles": 2,
+      "L_cycles": 3,
+      "H_override": {
+        "n_layers": 18
+      },
+      "bp_warmup_ratio": 0.2,
+      "bp_max_steps": 5,
+      "n_layers": 15,
+      "hidden_size": 1536,
+      "num_heads": 12,
+      "expansion": 4,
+      "norm_type": "pre",
+      "norm_eps": 1e-06,
+      "rope_theta": 10000.0,
+      "pos_emb_type": "rope",
+      "init_type": "lecun_normal",
+      "attn_type": "causal"
+    },
+    "data": {
+      "path": "/root/dl/runtime/script/loopTF-workspace/data/causal_hrm_text_sources/sampled",
+      "target_only": true
+    },
+    "global_batch_size": 196608,
+    "epochs": 4,
+    "lr": 0.00022,
+    "lr_min_ratio": 1.0,
+    "lr_warmup_steps": 2000,
+    "weight_decay": 0.1,
+    "beta1": 0.9,
+    "beta2": 0.95,
+    "ema": 0.9999,
+    "fwd_bwd_dtype": "bfloat16",
+    "project_name": "causal_hrm_text",
+    "run_name": "HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun",
+    "checkpoint_path": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/launchings/20261007-004905_hrm-fbw-epoch3-rerun/HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun/checkpoints",
+    "resume_from": null,
+    "resume_epoch": null,
+    "weights_only_resume_from_ema": false,
+    "seed": 0,
+    "checkpoint_interval": 1,
+    "log_interval": 5
+  },
+  "parameters": 1232342016,
+  "world_size": 32,
+  "local_batch_tokens": 6144,
+  "estimated_total_steps": 304789,
+  "execution": {
+    "stop_at_unix_time": null,
+    "runtime_cache_rank0": "/root/dl/runtime/script/loopTF-workspace/data/cache/hrm_official/runtime/cc557657b87216cad956/rank0",
+    "attention": "flex_triton_segmented_causal",
+    "checkpointing": true,
+    "checkpointing_skip_every": 2,
+    "checkpointing_retain_two_thirds_k2": true,
+    "checkpointing_min_passes": 0,
+    "compile_blocks": true,
+    "hrm_fsdp_grouping": "model",
+    "compile_optimizer": false,
+    "compile_vocabulary_loss": false,
+    "split_pseudo_attention": false,
+    "compile_a2_mix": false,
+    "lookahead_fsdp_grouping": "block",
+    "accumulation_steps": 1,
+    "gradient_reduction": "sum",
+    "gradient_reference_replicas": 32,
+    "gradient_scale": 1.0,
+    "virtual_replicas": 32,
+    "micro_batch_tokens": 196608,
+    "redistribution": null
+  },
+  "debug": {
+    "max_steps": 0,
+    "force_bp": 0,
+    "force_passes": null,
+    "benchmark_execution": false,
+    "step_offset": 0
+  },
+  "data_order": "official multipack, no holdout, upstream per-epoch drop_last",
+  "torch": "2.13.0+cu129",
+  "cuda": "12.9",
+  "gpu": "NVIDIA H800",
+  "config_sha256": "1d9f40773888f020d2a33265a24b9d7b6626688ce431eccffb59f61642c79e8b",
+  "resolved_overrides": {
+    "arch": {
+      "half_layers": false,
+      "n_layers": 15,
+      "H_override": {
+        "n_layers": 18
+      },
+      "attn_type": "causal"
+    },
+    "data": {
+      "path": "/root/dl/runtime/script/loopTF-workspace/data/causal_hrm_text_sources/sampled"
+    },
+    "project_name": "causal_hrm_text",
+    "run_name": "HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2__20261007-004905_hrm-fbw-epoch3-rerun",
+    "checkpoint_path": "/root/dl/runtime/script/loopTF-workspace/output/causal_hrm_text/official_recipe/HRM-FBW-K3-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-fromHRM-epoch3-SrcL1-L3-H1-L4-L6-H2"
+  },
+  "continuation": {
+    "mode": "hrm_to_fbw",
+    "checkpoint": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/output/causal_hrm_text/continuation_sources/ty8nbhn0/full_epoch_3/checkpoints/fsdp2_epoch_3",
+    "preserve_optimizer": true,
+    "preserve_ema": true,
+    "curriculum": [
+      {
+        "start_update": 228436,
+        "mixture": {
+          "2": 1.0
+        }
+      },
+      {
+        "start_update": 274312,
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        }
+      }
+    ],
+    "source_epoch": 3,
+    "source_step": 228435,
+    "source_world_size": 32,
+    "source_recipe_sha256": "8e53dcb59af9678277550dfa2540f3f2e32268c9386fa8f42936e616d8c3307b",
+    "checkpoint_metadata_sha256": "4898bea588e7717f169d4b648be3faee795b2b8f821815301a886fca132543dd",
+    "source_run_name": "HRM-H2-L3-H18-L15-causal-AdamATan2-EMA-seed0-H800x32-20260925-002521",
+    "source_wandb_run_id": "ty8nbhn0",
+    "data_epoch_index": 3,
+    "rng_restored": false,
+    "bit_exact_resume": false
+  },
+  "resume": null,
+  "resume_execution": null,
+  "checkpointing_interval": {
+    "save_every": 1000,
+    "keep_step_checkpoints": 3,
+    "save_first_step": true,
+    "verify_first": true
+  },
+  "max_grad_norm": 1.0,
+  "pass_schedule": {
+    "progress": "absolute_optimizer_updates",
+    "stages": [
+      {
+        "start_update": 228436,
+        "mixture": {
+          "2": 1.0
+        }
+      },
+      {
+        "start_update": 274312,
+        "mixture": {
+          "2": 0.7,
+          "3": 0.3
+        }
+      }
+    ],
+    "estimated_total_steps": 304789,
+    "seed": 0,
+    "sampler": "sha256-counter-v1",
+    "step_convention": "completed updates before the next update; first update uses zero",
+    "start_completed_updates": [
+      228435,
+      274311
+    ],
+    "first_update_per_stage": [
+      228436,
+      274312
+    ],
+    "after_horizon": "keep final mixture"
+  },
+  "reference_validation": {
+    "suite": "/apdcephfs_tj31/share_303214315/hunyuan/kenanchen/loopTF-workspace/data/evaluation/hrm_text/reference-v1-seed20261004",
+    "manifest_sha256": "98c9561e9ae9ebf972500e8dea63dde867b356f6e85016dd4a950120cc2635ba",
+    "weights": "ema",
+    "capacity": 4096,
+    "examples": 1280,
+    "teacher_forced": true,
+    "tasks": {
+      "GSM8k": {
+        "examples": 256,
+        "response_tokens": 24971,
+        "packs": 11
+      },
+      "MATH": {
+        "examples": 256,
+        "response_tokens": 56990,
+        "packs": 21
+      },
+      "DROP": {
+        "examples": 128,
+        "response_tokens": 533,
+        "packs": 54
+      },
+      "MMLU": {
+        "examples": 128,
+        "response_tokens": 128,
+        "packs": 22
+      },
+      "ARC": {
+        "examples": 128,
+        "response_tokens": 128,
+        "packs": 128
+      },
+      "HellaSwag": {
+        "examples": 128,
+        "response_tokens": 128,
+        "packs": 64
+      },
+      "Winogrande": {
+        "examples": 128,
+        "response_tokens": 128,
+        "packs": 7
+      },
+      "BoolQ": {
+        "examples": 128,
+        "response_tokens": 128,
+        "packs": 39
+      }
+    },
+    "every": 1000,
+    "first_update": true,
+    "training_horizon_changed": false
+  },
+  "tracking": {
+    "mode": "online",
+    "entity": "LoopTF-4-CSPs",
+    "project": "causal_hrm_text",
+    "group": "official_recipe"
+  },
+  "architecture": {
+    "arm": "hrm_fbw",
+    "fbw": {
+      "num_passes": 3,
+      "feedback_sources": [
+        "L1",
+        "L3",
+        "H1",
+        "L4",
+        "L6",
+        "H2"
+      ]
+    },
+    "hrm": {
+      "h_layers": 18,
+      "l_layers": 15,
+      "h_cycles": 2,
+      "l_cycles": 3
+    },
+    "forward_blocks": 378,
+    "response_ce_weights": [
+      0.5,
+      0.25,
+      0.25
+    ],
+    "pass_schedule": null,
+    "feedback": "shifted previous-pass latent, normalized FBW gate, reset at document starts",
+    "inner_bp": "K1: upstream warmup; K>1: preceding passes max(5, scheduled BP), final pass full BP8",
+    "feedback_source_fusion": {
+      "operation": "concatenate_then_bias_free_linear",
+      "initialization": "final_H2_identity",
+      "states": "outputs of complete L/H stack invocations, including stack normalization",
+      "projection_parameters": 14155776,
+      "lm_head_source": "H2"
+    },
+    "lookahead": null
+  },
+  "objective": "normalized_feedback_pass_response_CE"
+}