Add files using upload-large-folder tool

artifacts/v2/compressor-structure_20260923/plan/20260923_compressor_structure_experiment.md

@@ -0,0 +1,48 @@
+# 20260923_压缩器结构 B／F 对照实验
+
+创建时间：20260923 10:33:01 UTC+08:00  
+最后修订时间：20260923 10:48:59 UTC+08:00
+
+## 设置
+
+复用已有 mean（pooling）B／F 结果，新增 weighted、spectral 各两组。四份配置只改变对应 mean 配置的 compression，其余字段保持一致。
+
+- B：AE＋LM，单次 warm-up 1 epoch＋递归更新 2 epochs。
+- F：AE＋LM，独立前缀压缩 3 epochs。同一轨迹各终点分别从原文压缩，按位置平均 loss。
+- K=512，LM 目标 512，seed=42，学习率 1e-4，global batch=32，每卡 microbatch=8，双卡梯度累积 2，encoder／decoder token 预算 32768／36928。B 沿用原每 1000 steps 验证，F 每 epoch 验证 4 次。
+- 两组均评估递归与独立前缀路径，训练和最终评估记录在同一个 SwanLab run。新增压缩器参数在三个训练阶段均参与优化。
+
+weighted：连续组内共享打分器 Linear(d,64) → GELU → Linear(64,1,bias=False)，组内 softmax 后加权求和；仅最后一层权重置零，初始等价于 mean。第一步输出层得到梯度，随后隐藏层得到梯度。瓶颈宽度固定在实现中为 64，本轮不增加宽度搜索。
+
+spectral：前后两个 d→256→d 残差 MLP，中间沿位置轴执行固定 Fourier 低通重采样；MLP 与 LayerNorm 可训练，频率截断规则固定。
+
+## 运行组织
+
+project：latent-working-memory-v2；group：qwen3-4b_compressor-structure_20260923；job_type：train；study:reconstruction；method 标签由压缩器自动生成。
+
+| 结构 | 组别 | Run 名称 |
+|---|---|---|
+| weighted | B | qwen3-4b_weighted_ae-lm-warmup_20260923 |
+| weighted | F | qwen3-4b_weighted_ae-lm-static_20260923 |
+| spectral | B | qwen3-4b_spectral_ae-lm-warmup_20260923 |
+| spectral | F | qwen3-4b_spectral_ae-lm-static_20260923 |
+
+物理 GPU 限定 4–7，按 4／5、6／7 双卡调度。保留正在 4／5 上运行的旧 E；空闲双卡先执行新 F，随后接续剩余运行。新实验均从基座初始化，不加载 mean checkpoint。
+
+新系列的规范入口仍为 artifacts/v2/compressor-structure_20260923/。旧服务器产物目录实际位于系统盘，仅剩约 41 GiB，无法容纳四组完整 checkpoint；本系列实际存储于 /data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/（数据盘初始空闲约 333 GiB），原目录保留软链接。未移动、删除旧实验产物。
+
+## 验证与启动
+
+实现提交 a2f2969，追加完整策略测试后为 45fe297，均通过 GitHub 同步到服务器主仓库；已有未提交修改保留。Ruff 通过，四份配置与对应 mean 配置逐字段比较通过。服务器原测试集合 99 项通过，新增四种结构／训练策略的完整链路测试 4 项通过，共 103 项；覆盖各阶段新增参数真实更新、checkpoint、CUDA BF16 varlen、递归反向传播、双卡与最终双路径评估。
+
+d=2560 时，weighted 压缩器新增 163968 个参数，spectral 新增 2637312 个参数。
+
+真实 Qwen3-4B、正式 F 配置、GPU 6／7 的各两步短跑均正常退出，loss、梯度、checkpoint 参数有限，weighted 输出层已从零更新。第一步各有 AE 21 条、LM 11 条，监督 tokens 分别 153882／23598，与原 F 配对；第二步各 16／16 条，监督 tokens 93782／33858。weighted 峰值 allocated 58.48 GiB，spectral 58.67 GiB。短跑关闭 SwanLab，存放在 plan/smoke/，正式运行仍从基座开始。
+
+正式队列 tmux 为 lwm-v2-compressors-20260923，scheduler PID 1971762；weighted F 首先在 6／7 启动，torchrun PID 1971841。队列顺序为 weighted F、spectral F、weighted B、spectral B；4／5 待旧 E 退出后加入调度。四组均为 3000 steps，最终评估追加到训练 run。
+
+配置快照、启动命令和验证记录：artifacts/v2/compressor-structure_20260923/plan/、launch.json、plan/smoke-verification.json。
+
+weighted F 的正式 SwanLab run 为 [a7nrdamp](https://swanlab.cn/@percyWeeeeei/latent-working-memory-v2/runs/a7nrdamp)，日志确认名称为 qwen3-4b_weighted_ae-lm-static_20260923。其余运行在实际获得 GPU 并初始化时创建各自 run。
+
+正式 weighted F 已完成 2 次 optimizer 更新，loss 为 9.5929、9.5261，梯度有限，峰值 allocated 58.48 GiB；首两批样本及监督计数与短跑一致。启动核验保存在 plan/startup-verification.json。

artifacts/v2/compressor-structure_20260923/plan/gpu-expansion_20260924.json

@@ -0,0 +1,185 @@
+{
+  "previous_status": {
+    "group": "qwen3-4b_compressor-structure_20260923",
+    "gpu_pairs": [
+      [
+        4,
+        5
+      ],
+      [
+        6,
+        7
+      ]
+    ],
+    "runs": [
+      {
+        "name": "qwen3-4b_weighted_ae-lm-static_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_weighted_ae-lm-static/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-static_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-static_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-static_20260923",
+        "state": "running",
+        "gpus": [
+          6,
+          7
+        ],
+        "pid": 1971841,
+        "log": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/qwen3-4b_weighted_ae-lm-static_20260923.log",
+        "started_at": "2026-09-23T02:44:22.368755+00:00"
+      },
+      {
+        "name": "qwen3-4b_spectral_ae-lm-static_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_spectral_ae-lm-static/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-static_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-static_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-static_20260923",
+        "state": "running",
+        "gpus": [
+          4,
+          5
+        ],
+        "pid": 2025804,
+        "log": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/qwen3-4b_spectral_ae-lm-static_20260923.log",
+        "started_at": "2026-09-23T06:30:11.564258+00:00"
+      },
+      {
+        "name": "qwen3-4b_weighted_ae-lm-warmup_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_weighted_ae-lm-warmup/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-warmup_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923",
+        "state": "queued"
+      },
+      {
+        "name": "qwen3-4b_spectral_ae-lm-warmup_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_spectral_ae-lm-warmup/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-warmup_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-warmup_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-warmup_20260923",
+        "state": "queued"
+      }
+    ],
+    "scheduler_pid": 1971762,
+    "state": "running",
+    "updated_at": "2026-09-24T03:17:06.840057+00:00"
+  },
+  "command": [
+    "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+    "-u",
+    "-m",
+    "latent_working_memory.v2.pretrain.experiment",
+    "--experiments",
+    "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_weighted_ae-lm-static/experiment.json",
+    "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_spectral_ae-lm-static/experiment.json",
+    "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_weighted_ae-lm-warmup/experiment.json",
+    "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_spectral_ae-lm-warmup/experiment.json",
+    "--output-dir",
+    "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923",
+    "--model-path",
+    "/data/bywei/models/Qwen/Qwen3-4B-Instruct-2507",
+    "--gpus",
+    "0,1,2,3,4,5,6,7",
+    "--group",
+    "qwen3-4b_compressor-structure_20260923",
+    "--run-date",
+    "20260923",
+    "--resume",
+    "--adopt-running",
+    "--min-free-gib",
+    "74"
+  ],
+  "cwd": "/data/bywei/projects/latent_working_memory",
+  "git_commit": "c6b2447ffd3cf69d6c46cd5d95169fda53464d68",
+  "tmux": "lwm-v2-compressors-all-20260924",
+  "created_at": "2026-09-24T03:17:11.581190+00:00",
+  "free_gib": {
+    "0": 74.970703125,
+    "1": 74.970703125,
+    "2": 74.970703125,
+    "3": 74.970703125,
+    "4": 7.18359375,
+    "5": 8.072265625,
+    "6": 7.826171875,
+    "7": 8.748046875
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/policies-tests.log

@@ -0,0 +1,8 @@
+....                                                                     [100%]
+=============================== warnings summary ===============================
+.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24
+  /data/bywei/projects/latent_working_memory/.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24: UserWarning: NPU not support router replay for now.
+    from verl.utils.megatron.router_replay_patch import RouterReplay
+
+-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
+4 passed, 24 deselected, 1 warning in 53.69s

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-warmup_20260923/experiment.json

@@ -0,0 +1,25 @@
+{
+  "model": "model.json",
+  "selection": "selection.json",
+  "training": {
+    "objective": "ae_lm",
+    "warmup_epochs": 1,
+    "multiround_epochs": 2,
+    "independent_prefix_epochs": 0,
+    "global_batch_size": 32,
+    "micro_batch_size": 8,
+    "micro_batch_encoder_tokens": 32768,
+    "micro_batch_decoder_tokens": 36928,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "lm_ratio": 0.5,
+    "seed": 42,
+    "save_every": 1000,
+    "checkpoint_limit": 2,
+    "eval_every": 1000,
+    "generation_samples": 8,
+    "ae_prompt": "Reconstruct the text stored in memory:\n",
+    "lm_prompt": "Continue the text stored in memory:\n"
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-warmup_20260923/model.json

@@ -0,0 +1,16 @@
+{
+  "model_name_or_path": "/data/bywei/models/Qwen/Qwen3-4B-Instruct-2507",
+  "revision": null,
+  "encoder_layers": null,
+  "alignment_layers": 1,
+  "lora_rank": 128,
+  "lora_alpha": 32,
+  "lora_dropout": 0.0,
+  "attention_implementation": "sdpa",
+  "compression": "spectral",
+  "feature_layer": "last",
+  "spectral_bottleneck": 256,
+  "gradient_checkpointing": true,
+  "lm_head_chunk_size": 256,
+  "padding_free": true
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-warmup_20260923/selection.json

@@ -0,0 +1,3 @@
+{
+  "dataset_dir": "/data/bywei/projects/latent_working_memory/data/fineweb-reconstruction-k512-doc100k_20260917"
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-static_20260923/experiment.json

@@ -0,0 +1,26 @@
+{
+  "model": "model.json",
+  "selection": "selection.json",
+  "training": {
+    "objective": "ae_lm",
+    "warmup_epochs": 0,
+    "multiround_epochs": 0,
+    "independent_prefix_epochs": 3,
+    "global_batch_size": 32,
+    "micro_batch_size": 8,
+    "micro_batch_encoder_tokens": 32768,
+    "micro_batch_decoder_tokens": 36928,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "lm_ratio": 0.5,
+    "seed": 42,
+    "save_every": 1000,
+    "checkpoint_limit": 2,
+    "eval_every": null,
+    "evals_per_epoch": 4,
+    "generation_samples": 8,
+    "ae_prompt": "Reconstruct the text stored in memory:\n",
+    "lm_prompt": "Continue the text stored in memory:\n"
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-static_20260923/model.json

@@ -0,0 +1,16 @@
+{
+  "model_name_or_path": "/data/bywei/models/Qwen/Qwen3-4B-Instruct-2507",
+  "revision": null,
+  "encoder_layers": null,
+  "alignment_layers": 1,
+  "lora_rank": 128,
+  "lora_alpha": 32,
+  "lora_dropout": 0.0,
+  "attention_implementation": "sdpa",
+  "compression": "weighted",
+  "feature_layer": "last",
+  "spectral_bottleneck": 256,
+  "gradient_checkpointing": true,
+  "lm_head_chunk_size": 256,
+  "padding_free": true
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-static_20260923/selection.json

@@ -0,0 +1,3 @@
+{
+  "dataset_dir": "/data/bywei/projects/latent_working_memory/data/fineweb-reconstruction-k512-doc100k_20260917"
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-warmup_20260923/experiment.json

@@ -0,0 +1,25 @@
+{
+  "model": "model.json",
+  "selection": "selection.json",
+  "training": {
+    "objective": "ae_lm",
+    "warmup_epochs": 1,
+    "multiround_epochs": 2,
+    "independent_prefix_epochs": 0,
+    "global_batch_size": 32,
+    "micro_batch_size": 8,
+    "micro_batch_encoder_tokens": 32768,
+    "micro_batch_decoder_tokens": 36928,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "lm_ratio": 0.5,
+    "seed": 42,
+    "save_every": 1000,
+    "checkpoint_limit": 2,
+    "eval_every": 1000,
+    "generation_samples": 8,
+    "ae_prompt": "Reconstruct the text stored in memory:\n",
+    "lm_prompt": "Continue the text stored in memory:\n"
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-warmup_20260923/model.json

@@ -0,0 +1,16 @@
+{
+  "model_name_or_path": "/data/bywei/models/Qwen/Qwen3-4B-Instruct-2507",
+  "revision": null,
+  "encoder_layers": null,
+  "alignment_layers": 1,
+  "lora_rank": 128,
+  "lora_alpha": 32,
+  "lora_dropout": 0.0,
+  "attention_implementation": "sdpa",
+  "compression": "weighted",
+  "feature_layer": "last",
+  "spectral_bottleneck": 256,
+  "gradient_checkpointing": true,
+  "lm_head_chunk_size": 256,
+  "padding_free": true
+}

artifacts/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-warmup_20260923/selection.json

@@ -0,0 +1,3 @@
+{
+  "dataset_dir": "/data/bywei/projects/latent_working_memory/data/fineweb-reconstruction-k512-doc100k_20260917"
+}

artifacts/v2/compressor-structure_20260923/plan/smoke-verification.json

@@ -0,0 +1,116 @@
+{
+  "git_commit": "45fe297b116357e33f4ae034a3a4a976b1b01f8d",
+  "tests_passed": 103,
+  "gpus": [
+    6,
+    7
+  ],
+  "runs": [
+    {
+      "name": "qwen3-4b_weighted_ae-lm-static_20260923",
+      "steps": [
+        {
+          "step": 1,
+          "stage": "independent_prefix",
+          "epoch": 1,
+          "global_epoch": 1,
+          "loss": 9.592894613742828,
+          "ae": 9.49606597537086,
+          "ae_samples": 21,
+          "ae_tokens": 153882,
+          "lm": 9.777748704679084,
+          "lm_samples": 11,
+          "lm_tokens": 23598,
+          "samples": 32,
+          "microbatches": 4,
+          "max_microbatch_size": 8,
+          "mean_active_microbatch_size": 6.75,
+          "batched_samples": 32,
+          "source_tokens": 90848,
+          "target_tokens": 177480,
+          "grad_norm": 2.427708148956299,
+          "seconds": 52.97523378697224,
+          "peak_memory_bytes": 62790014976
+        },
+        {
+          "step": 2,
+          "stage": "independent_prefix",
+          "epoch": 1,
+          "global_epoch": 1,
+          "loss": 9.530107498168945,
+          "ae": 9.697963373859722,
+          "ae_samples": 16,
+          "ae_tokens": 93782,
+          "lm": 9.362251124779384,
+          "lm_samples": 16,
+          "lm_tokens": 33858,
+          "samples": 32,
+          "microbatches": 4,
+          "max_microbatch_size": 8,
+          "mean_active_microbatch_size": 6.4,
+          "batched_samples": 32,
+          "source_tokens": 78767,
+          "target_tokens": 127640,
+          "grad_norm": 2.2894504070281982,
+          "seconds": 41.040178007038776,
+          "peak_memory_bytes": 62790014976
+        }
+      ],
+      "compression_parameters": 163968,
+      "checkpoint_verified": true
+    },
+    {
+      "name": "qwen3-4b_spectral_ae-lm-static_20260923",
+      "steps": [
+        {
+          "step": 1,
+          "stage": "independent_prefix",
+          "epoch": 1,
+          "global_epoch": 1,
+          "loss": 9.505487084388733,
+          "ae": 9.421910826365153,
+          "ae_samples": 21,
+          "ae_tokens": 153882,
+          "lm": 9.665041351318362,
+          "lm_samples": 11,
+          "lm_tokens": 23598,
+          "samples": 32,
+          "microbatches": 4,
+          "max_microbatch_size": 8,
+          "mean_active_microbatch_size": 6.75,
+          "batched_samples": 32,
+          "source_tokens": 90848,
+          "target_tokens": 177480,
+          "grad_norm": 1.6032721996307373,
+          "seconds": 57.24044513597619,
+          "peak_memory_bytes": 62995403264
+        },
+        {
+          "step": 2,
+          "stage": "independent_prefix",
+          "epoch": 1,
+          "global_epoch": 1,
+          "loss": 9.40975734591484,
+          "ae": 9.567599389950434,
+          "ae_samples": 16,
+          "ae_tokens": 93782,
+          "lm": 9.251915042599043,
+          "lm_samples": 16,
+          "lm_tokens": 33858,
+          "samples": 32,
+          "microbatches": 4,
+          "max_microbatch_size": 8,
+          "mean_active_microbatch_size": 6.4,
+          "batched_samples": 32,
+          "source_tokens": 78767,
+          "target_tokens": 127640,
+          "grad_norm": 1.6850647926330566,
+          "seconds": 44.85529744299129,
+          "peak_memory_bytes": 62995403264
+        }
+      ],
+      "compression_parameters": 2637312,
+      "checkpoint_verified": true
+    }
+  ]
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923-command.json

@@ -0,0 +1,30 @@
+{
+  "command": [
+    "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+    "-u",
+    "-m",
+    "torch.distributed.run",
+    "--standalone",
+    "--nproc_per_node=2",
+    "-m",
+    "latent_working_memory.v2.pretrain.train",
+    "--experiment",
+    "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-static_20260923/experiment.json",
+    "--output-dir",
+    "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923",
+    "--swanlab-project",
+    "latent-working-memory-v2",
+    "--swanlab-mode",
+    "disabled",
+    "--swanlab-group",
+    "qwen3-4b_compressor-structure_20260923",
+    "--swanlab-tag",
+    "study:reconstruction",
+    "--stop-after-steps",
+    "2"
+  ],
+  "gpus": [
+    6,
+    7
+  ]
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923.log

@@ -0,0 +1,26 @@
+/data/bywei/projects/latent_working_memory/.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24: UserWarning: NPU not support router replay for now.
+  from verl.utils.megatron.router_replay_patch import RouterReplay
+/data/bywei/projects/latent_working_memory/.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24: UserWarning: NPU not support router replay for now.
+  from verl.utils.megatron.router_replay_patch import RouterReplay
+rank 0: loading and filtering prepared datasets
+rank 1: loading and filtering prepared datasets
+multiround/train: retained 31976/32000, rejected {'continuation_length': 24}
+multiround/dev: retained 128/128, rejected {}
+multiround/test: retained 128/128, rejected {}
+`torch_dtype` is deprecated! Use `dtype` instead!
+
+
+multiround/train: retained 31976/32000, rejected {'continuation_length': 24}
+multiround/dev: retained 128/128, rejected {}
+multiround/test: retained 128/128, rejected {}
+`torch_dtype` is deprecated! Use `dtype` instead!
+
+
+/data/bywei/projects/latent_working_memory/src/latent_working_memory/v2/pretrain/engine.py:64: FutureWarning: `broadcast_buffers` is deprecated. Use `forward_sync_buffers` instead. IMPORTANT: unlike `broadcast_buffers=False`, `forward_sync_buffers=False` still syncs buffers at init. If you rely on buffers NOT being synced at init (e.g. rank-local buffers), keep using `broadcast_buffers=False` until `init_sync_buffers` is available.
+  self.module = DistributedDataParallel(
+/data/bywei/projects/latent_working_memory/src/latent_working_memory/v2/pretrain/engine.py:64: FutureWarning: `broadcast_buffers` is deprecated. Use `forward_sync_buffers` instead. IMPORTANT: unlike `broadcast_buffers=False`, `forward_sync_buffers=False` still syncs buffers at init. If you rely on buffers NOT being synced at init (e.g. rank-local buffers), keep using `broadcast_buffers=False` until `init_sync_buffers` is available.
+  self.module = DistributedDataParallel(
+[rank1]:[W923 02:42:02.640311451 reducer.cpp:1529] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration,  which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
+[rank0]:[W923 02:42:02.695255892 reducer.cpp:1529] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration,  which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
+{"step": 1, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.505487084388733, "ae": 9.421910826365153, "ae_samples": 21, "ae_tokens": 153882, "lm": 9.665041351318362, "lm_samples": 11, "lm_tokens": 23598, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.75, "batched_samples": 32, "source_tokens": 90848, "target_tokens": 177480, "grad_norm": 1.6032721996307373, "seconds": 57.24044513597619, "peak_memory_bytes": 62995403264}
+{"step": 2, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.40975734591484, "ae": 9.567599389950434, "ae_samples": 16, "ae_tokens": 93782, "lm": 9.251915042599043, "lm_samples": 16, "lm_tokens": 33858, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.4, "batched_samples": 32, "source_tokens": 78767, "target_tokens": 127640, "grad_norm": 1.6850647926330566, "seconds": 44.85529744299129, "peak_memory_bytes": 62995403264}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/data-filtering.json

@@ -0,0 +1,21 @@
+{
+  "multiround": {
+    "train": {
+      "candidates": 32000,
+      "retained": 31976,
+      "rejected": {
+        "continuation_length": 24
+      }
+    },
+    "dev": {
+      "candidates": 128,
+      "retained": 128,
+      "rejected": {}
+    },
+    "test": {
+      "candidates": 128,
+      "retained": 128,
+      "rejected": {}
+    }
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/data-summary.json

@@ -0,0 +1,63 @@
+{
+  "multiround": {
+    "train": {
+      "trajectories": 31976,
+      "documents": 2473,
+      "rounds": {
+        "4": 10605,
+        "3": 10709,
+        "5": 10662
+      },
+      "ratio_bins": {
+        "5": 7910,
+        "7": 5072,
+        "6": 9183,
+        "8": 3475,
+        "4": 6259,
+        "3": 77
+      },
+      "segment_min": 512,
+      "segment_max": 1536,
+      "source_tokens": 84247773
+    },
+    "dev": {
+      "trajectories": 128,
+      "documents": 68,
+      "rounds": {
+        "5": 45,
+        "3": 51,
+        "4": 32
+      },
+      "ratio_bins": {
+        "6": 54,
+        "4": 34,
+        "5": 26,
+        "7": 10,
+        "8": 4
+      },
+      "segment_min": 512,
+      "segment_max": 1476,
+      "source_tokens": 312059
+    },
+    "test": {
+      "trajectories": 128,
+      "documents": 70,
+      "rounds": {
+        "4": 45,
+        "3": 36,
+        "5": 47
+      },
+      "ratio_bins": {
+        "4": 22,
+        "7": 14,
+        "5": 38,
+        "6": 36,
+        "8": 17,
+        "3": 1
+      },
+      "segment_min": 512,
+      "segment_max": 1536,
+      "source_tokens": 336041
+    }
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/epoch-plan.json

@@ -0,0 +1,41 @@
+{
+  "epochs": [
+    {
+      "stage": "independent_prefix",
+      "epoch": 1,
+      "samples": 31976,
+      "steps": 1000,
+      "evaluation_steps": [
+        250,
+        500,
+        750,
+        1000
+      ]
+    },
+    {
+      "stage": "independent_prefix",
+      "epoch": 2,
+      "samples": 31976,
+      "steps": 1000,
+      "evaluation_steps": [
+        1250,
+        1500,
+        1750,
+        2000
+      ]
+    },
+    {
+      "stage": "independent_prefix",
+      "epoch": 3,
+      "samples": 31976,
+      "steps": 1000,
+      "evaluation_steps": [
+        2250,
+        2500,
+        2750,
+        3000
+      ]
+    }
+  ],
+  "total_steps": 3000
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/provenance.json

@@ -0,0 +1,18 @@
+{
+  "created_at": "2026-09-23T02:41:28.196411+00:00",
+  "execution": {
+    "framework": "verl",
+    "backend": "replicated",
+    "world_size": 2
+  },
+  "engine_origin": "codex/verl-v1@e0bfa37",
+  "packages": {
+    "torch": "2.14.0",
+    "transformers": "4.57.6",
+    "peft": "0.20.0",
+    "verl": "0.8.0",
+    "tensordict": "0.10.0"
+  },
+  "git_commit": "45fe297b116357e33f4ae034a3a4a976b1b01f8d",
+  "git_dirty": true
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/run.json

@@ -0,0 +1,232 @@
+{
+  "model": {
+    "model_name_or_path": "/data/bywei/models/Qwen/Qwen3-4B-Instruct-2507",
+    "revision": null,
+    "encoder_layers": null,
+    "alignment_layers": 1,
+    "lora_rank": 128,
+    "lora_alpha": 32,
+    "lora_dropout": 0.0,
+    "attention_implementation": "sdpa",
+    "compression": "spectral",
+    "feature_layer": "last",
+    "spectral_bottleneck": 256,
+    "gradient_checkpointing": true,
+    "lm_head_chunk_size": 256,
+    "padding_free": true
+  },
+  "selection": {
+    "dataset_dir": "/data/bywei/projects/latent_working_memory/data/fineweb-reconstruction-k512-doc100k_20260917"
+  },
+  "data_preparation": {
+    "preparation_id": "036d92f2-0004-461b-8a14-31fff29ad658",
+    "created_at": "2026-09-17T03:29:24.728506+00:00",
+    "length_estimation": "candidate characters = ceil(4 * planned content tokens * content_reserve_ratio) + 4 * continuation_reserve_tokens; write_token_ends are target token positions",
+    "config": {
+      "source_glob": "data/raw/HuggingFaceFW-fineweb/sample-10BT/**/*.parquet",
+      "capacity": 512,
+      "continuation_tokens": 512,
+      "content_reserve_ratio": 1.5,
+      "continuation_reserve_tokens": 768,
+      "max_documents": 100000,
+      "source_seed": 20260907,
+      "seed": 20260916,
+      "split_fractions": [
+        0.9,
+        0.05,
+        0.05
+      ],
+      "warmup": {
+        "train": 32000,
+        "dev": 128,
+        "test": 128
+      },
+      "multiround": {
+        "train": 32000,
+        "dev": 128,
+        "test": 128
+      }
+    },
+    "source_recipe": {
+      "min_document_chars": 64,
+      "near_duplicate_threshold": 0.9,
+      "near_duplicate_min_words": 64
+    },
+    "source_files": [
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/000_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/001_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/002_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/003_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/004_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/005_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/006_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/007_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/008_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/009_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/010_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/011_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/012_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/013_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/014_00000.parquet"
+    ],
+    "source_statistics": {
+      "eligible": 99998,
+      "duplicate": 2
+    },
+    "referenced_documents": 3433,
+    "statistics": {
+      "warmup": {
+        "train": {
+          "trajectories": 32000,
+          "documents": 3178,
+          "compressions": {
+            "1": 32000
+          },
+          "planned_ratio_bins": {
+            "6": 5443,
+            "8": 5352,
+            "7": 5290,
+            "5": 5391,
+            "4": 5245,
+            "3": 5273,
+            "2": 6
+          },
+          "segment_tokens_min": 1024,
+          "segment_tokens_max": 4096,
+          "planned_source_tokens": 82035196
+        },
+        "dev": {
+          "trajectories": 128,
+          "documents": 61,
+          "compressions": {
+            "1": 128
+          },
+          "planned_ratio_bins": {
+            "4": 17,
+            "7": 25,
+            "5": 31,
+            "3": 12,
+            "6": 22,
+            "8": 21
+          },
+          "segment_tokens_min": 1153,
+          "segment_tokens_max": 4082,
+          "planned_source_tokens": 344604
+        },
+        "test": {
+          "trajectories": 128,
+          "documents": 67,
+          "compressions": {
+            "1": 128
+          },
+          "planned_ratio_bins": {
+            "8": 23,
+            "3": 19,
+            "5": 20,
+            "4": 26,
+            "6": 21,
+            "7": 19
+          },
+          "segment_tokens_min": 1040,
+          "segment_tokens_max": 4081,
+          "planned_source_tokens": 330093
+        }
+      },
+      "multiround": {
+        "train": {
+          "trajectories": 32000,
+          "documents": 2474,
+          "compressions": {
+            "4": 10612,
+            "3": 10719,
+            "5": 10669
+          },
+          "planned_ratio_bins": {
+            "5": 7919,
+            "7": 5075,
+            "6": 9187,
+            "8": 3477,
+            "4": 6265,
+            "3": 77
+          },
+          "segment_tokens_min": 512,
+          "segment_tokens_max": 1536,
+          "planned_source_tokens": 84308173
+        },
+        "dev": {
+          "trajectories": 128,
+          "documents": 68,
+          "compressions": {
+            "5": 45,
+            "3": 51,
+            "4": 32
+          },
+          "planned_ratio_bins": {
+            "6": 54,
+            "4": 34,
+            "5": 26,
+            "7": 10,
+            "8": 4
+          },
+          "segment_tokens_min": 512,
+          "segment_tokens_max": 1476,
+          "planned_source_tokens": 312059
+        },
+        "test": {
+          "trajectories": 128,
+          "documents": 70,
+          "compressions": {
+            "4": 45,
+            "3": 36,
+            "5": 47
+          },
+          "planned_ratio_bins": {
+            "4": 22,
+            "7": 14,
+            "5": 38,
+            "6": 36,
+            "8": 17,
+            "3": 1
+          },
+          "segment_tokens_min": 512,
+          "segment_tokens_max": 1536,
+          "planned_source_tokens": 336041
+        }
+      }
+    }
+  },
+  "training": {
+    "objective": "ae_lm",
+    "warmup_epochs": 0,
+    "multiround_epochs": 0,
+    "independent_prefix_epochs": 3,
+    "global_batch_size": 32,
+    "micro_batch_size": 8,
+    "micro_batch_encoder_tokens": 32768,
+    "micro_batch_decoder_tokens": 36928,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "lm_ratio": 0.5,
+    "seed": 42,
+    "save_every": 1000,
+    "checkpoint_limit": 2,
+    "eval_every": null,
+    "evals_per_epoch": 4,
+    "generation_samples": 8,
+    "ae_prompt": "Reconstruct the text stored in memory:\n",
+    "lm_prompt": "Continue the text stored in memory:\n"
+  },
+  "world_size": 2,
+  "precision": "bf16",
+  "resolved_model_revision": null,
+  "resolved_architecture": {
+    "encoder_layers": 36,
+    "decoder_layers": 36,
+    "alignment_layers": 1,
+    "hidden_size": 2560,
+    "encoder_attention": "causal",
+    "shared_backbone": false,
+    "reader_adapter": null
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/tokenizer/added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/tokenizer/chat_template.jinja

@@ -0,0 +1,61 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/tokenizer/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/tokenizer/tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 1010000,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/train.jsonl

@@ -0,0 +1,2 @@
+{"step": 1, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.505487084388733, "ae": 9.421910826365153, "ae_samples": 21, "ae_tokens": 153882, "lm": 9.665041351318362, "lm_samples": 11, "lm_tokens": 23598, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.75, "batched_samples": 32, "source_tokens": 90848, "target_tokens": 177480, "grad_norm": 1.6032721996307373, "seconds": 57.24044513597619, "peak_memory_bytes": 62995403264}
+{"step": 2, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.40975734591484, "ae": 9.567599389950434, "ae_samples": 16, "ae_tokens": 93782, "lm": 9.251915042599043, "lm_samples": 16, "lm_tokens": 33858, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.4, "batched_samples": 32, "source_tokens": 78767, "target_tokens": 127640, "grad_norm": 1.6850647926330566, "seconds": 44.85529744299129, "peak_memory_bytes": 62995403264}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/training-result.json

@@ -0,0 +1,9 @@
+{
+  "complete": false,
+  "completed_steps": 2,
+  "total_steps": 3000,
+  "completed_epochs": 0,
+  "source_tokens": 169615,
+  "target_tokens": 305120,
+  "checkpoint": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_spectral_ae-lm-static_20260923/checkpoints/step-000002.pt"
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923-command.json

@@ -0,0 +1,30 @@
+{
+  "command": [
+    "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+    "-u",
+    "-m",
+    "torch.distributed.run",
+    "--standalone",
+    "--nproc_per_node=2",
+    "-m",
+    "latent_working_memory.v2.pretrain.train",
+    "--experiment",
+    "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-static_20260923/experiment.json",
+    "--output-dir",
+    "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923",
+    "--swanlab-project",
+    "latent-working-memory-v2",
+    "--swanlab-mode",
+    "disabled",
+    "--swanlab-group",
+    "qwen3-4b_compressor-structure_20260923",
+    "--swanlab-tag",
+    "study:reconstruction",
+    "--stop-after-steps",
+    "2"
+  ],
+  "gpus": [
+    6,
+    7
+  ]
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923.log

@@ -0,0 +1,26 @@
+/data/bywei/projects/latent_working_memory/.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24: UserWarning: NPU not support router replay for now.
+  from verl.utils.megatron.router_replay_patch import RouterReplay
+/data/bywei/projects/latent_working_memory/.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24: UserWarning: NPU not support router replay for now.
+  from verl.utils.megatron.router_replay_patch import RouterReplay
+rank 1: loading and filtering prepared datasets
+rank 0: loading and filtering prepared datasets
+multiround/train: retained 31976/32000, rejected {'continuation_length': 24}
+multiround/dev: retained 128/128, rejected {}
+multiround/train: retained 31976/32000, rejected {'continuation_length': 24}
+multiround/dev: retained 128/128, rejected {}
+multiround/test: retained 128/128, rejected {}
+`torch_dtype` is deprecated! Use `dtype` instead!
+
+
+multiround/test: retained 128/128, rejected {}
+`torch_dtype` is deprecated! Use `dtype` instead!
+
+
+/data/bywei/projects/latent_working_memory/src/latent_working_memory/v2/pretrain/engine.py:64: FutureWarning: `broadcast_buffers` is deprecated. Use `forward_sync_buffers` instead. IMPORTANT: unlike `broadcast_buffers=False`, `forward_sync_buffers=False` still syncs buffers at init. If you rely on buffers NOT being synced at init (e.g. rank-local buffers), keep using `broadcast_buffers=False` until `init_sync_buffers` is available.
+  self.module = DistributedDataParallel(
+/data/bywei/projects/latent_working_memory/src/latent_working_memory/v2/pretrain/engine.py:64: FutureWarning: `broadcast_buffers` is deprecated. Use `forward_sync_buffers` instead. IMPORTANT: unlike `broadcast_buffers=False`, `forward_sync_buffers=False` still syncs buffers at init. If you rely on buffers NOT being synced at init (e.g. rank-local buffers), keep using `broadcast_buffers=False` until `init_sync_buffers` is available.
+  self.module = DistributedDataParallel(
+[rank0]:[W923 02:37:54.384399526 reducer.cpp:1529] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration,  which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
+[rank1]:[W923 02:37:56.987891843 reducer.cpp:1529] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration,  which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
+{"step": 1, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.592894613742828, "ae": 9.49606597537086, "ae_samples": 21, "ae_tokens": 153882, "lm": 9.777748704679084, "lm_samples": 11, "lm_tokens": 23598, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.75, "batched_samples": 32, "source_tokens": 90848, "target_tokens": 177480, "grad_norm": 2.427708148956299, "seconds": 52.97523378697224, "peak_memory_bytes": 62790014976}
+{"step": 2, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.530107498168945, "ae": 9.697963373859722, "ae_samples": 16, "ae_tokens": 93782, "lm": 9.362251124779384, "lm_samples": 16, "lm_tokens": 33858, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.4, "batched_samples": 32, "source_tokens": 78767, "target_tokens": 127640, "grad_norm": 2.2894504070281982, "seconds": 41.040178007038776, "peak_memory_bytes": 62790014976}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/data-filtering.json

@@ -0,0 +1,21 @@
+{
+  "multiround": {
+    "train": {
+      "candidates": 32000,
+      "retained": 31976,
+      "rejected": {
+        "continuation_length": 24
+      }
+    },
+    "dev": {
+      "candidates": 128,
+      "retained": 128,
+      "rejected": {}
+    },
+    "test": {
+      "candidates": 128,
+      "retained": 128,
+      "rejected": {}
+    }
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/data-summary.json

@@ -0,0 +1,63 @@
+{
+  "multiround": {
+    "train": {
+      "trajectories": 31976,
+      "documents": 2473,
+      "rounds": {
+        "4": 10605,
+        "3": 10709,
+        "5": 10662
+      },
+      "ratio_bins": {
+        "5": 7910,
+        "7": 5072,
+        "6": 9183,
+        "8": 3475,
+        "4": 6259,
+        "3": 77
+      },
+      "segment_min": 512,
+      "segment_max": 1536,
+      "source_tokens": 84247773
+    },
+    "dev": {
+      "trajectories": 128,
+      "documents": 68,
+      "rounds": {
+        "5": 45,
+        "3": 51,
+        "4": 32
+      },
+      "ratio_bins": {
+        "6": 54,
+        "4": 34,
+        "5": 26,
+        "7": 10,
+        "8": 4
+      },
+      "segment_min": 512,
+      "segment_max": 1476,
+      "source_tokens": 312059
+    },
+    "test": {
+      "trajectories": 128,
+      "documents": 70,
+      "rounds": {
+        "4": 45,
+        "3": 36,
+        "5": 47
+      },
+      "ratio_bins": {
+        "4": 22,
+        "7": 14,
+        "5": 38,
+        "6": 36,
+        "8": 17,
+        "3": 1
+      },
+      "segment_min": 512,
+      "segment_max": 1536,
+      "source_tokens": 336041
+    }
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/epoch-plan.json

@@ -0,0 +1,41 @@
+{
+  "epochs": [
+    {
+      "stage": "independent_prefix",
+      "epoch": 1,
+      "samples": 31976,
+      "steps": 1000,
+      "evaluation_steps": [
+        250,
+        500,
+        750,
+        1000
+      ]
+    },
+    {
+      "stage": "independent_prefix",
+      "epoch": 2,
+      "samples": 31976,
+      "steps": 1000,
+      "evaluation_steps": [
+        1250,
+        1500,
+        1750,
+        2000
+      ]
+    },
+    {
+      "stage": "independent_prefix",
+      "epoch": 3,
+      "samples": 31976,
+      "steps": 1000,
+      "evaluation_steps": [
+        2250,
+        2500,
+        2750,
+        3000
+      ]
+    }
+  ],
+  "total_steps": 3000
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/provenance.json

@@ -0,0 +1,18 @@
+{
+  "created_at": "2026-09-23T02:37:26.250022+00:00",
+  "execution": {
+    "framework": "verl",
+    "backend": "replicated",
+    "world_size": 2
+  },
+  "engine_origin": "codex/verl-v1@e0bfa37",
+  "packages": {
+    "torch": "2.14.0",
+    "transformers": "4.57.6",
+    "peft": "0.20.0",
+    "verl": "0.8.0",
+    "tensordict": "0.10.0"
+  },
+  "git_commit": "45fe297b116357e33f4ae034a3a4a976b1b01f8d",
+  "git_dirty": true
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/run.json

@@ -0,0 +1,232 @@
+{
+  "model": {
+    "model_name_or_path": "/data/bywei/models/Qwen/Qwen3-4B-Instruct-2507",
+    "revision": null,
+    "encoder_layers": null,
+    "alignment_layers": 1,
+    "lora_rank": 128,
+    "lora_alpha": 32,
+    "lora_dropout": 0.0,
+    "attention_implementation": "sdpa",
+    "compression": "weighted",
+    "feature_layer": "last",
+    "spectral_bottleneck": 256,
+    "gradient_checkpointing": true,
+    "lm_head_chunk_size": 256,
+    "padding_free": true
+  },
+  "selection": {
+    "dataset_dir": "/data/bywei/projects/latent_working_memory/data/fineweb-reconstruction-k512-doc100k_20260917"
+  },
+  "data_preparation": {
+    "preparation_id": "036d92f2-0004-461b-8a14-31fff29ad658",
+    "created_at": "2026-09-17T03:29:24.728506+00:00",
+    "length_estimation": "candidate characters = ceil(4 * planned content tokens * content_reserve_ratio) + 4 * continuation_reserve_tokens; write_token_ends are target token positions",
+    "config": {
+      "source_glob": "data/raw/HuggingFaceFW-fineweb/sample-10BT/**/*.parquet",
+      "capacity": 512,
+      "continuation_tokens": 512,
+      "content_reserve_ratio": 1.5,
+      "continuation_reserve_tokens": 768,
+      "max_documents": 100000,
+      "source_seed": 20260907,
+      "seed": 20260916,
+      "split_fractions": [
+        0.9,
+        0.05,
+        0.05
+      ],
+      "warmup": {
+        "train": 32000,
+        "dev": 128,
+        "test": 128
+      },
+      "multiround": {
+        "train": 32000,
+        "dev": 128,
+        "test": 128
+      }
+    },
+    "source_recipe": {
+      "min_document_chars": 64,
+      "near_duplicate_threshold": 0.9,
+      "near_duplicate_min_words": 64
+    },
+    "source_files": [
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/000_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/001_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/002_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/003_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/004_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/005_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/006_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/007_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/008_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/009_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/010_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/011_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/012_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/013_00000.parquet",
+      "../raw/HuggingFaceFW-fineweb/sample-10BT/014_00000.parquet"
+    ],
+    "source_statistics": {
+      "eligible": 99998,
+      "duplicate": 2
+    },
+    "referenced_documents": 3433,
+    "statistics": {
+      "warmup": {
+        "train": {
+          "trajectories": 32000,
+          "documents": 3178,
+          "compressions": {
+            "1": 32000
+          },
+          "planned_ratio_bins": {
+            "6": 5443,
+            "8": 5352,
+            "7": 5290,
+            "5": 5391,
+            "4": 5245,
+            "3": 5273,
+            "2": 6
+          },
+          "segment_tokens_min": 1024,
+          "segment_tokens_max": 4096,
+          "planned_source_tokens": 82035196
+        },
+        "dev": {
+          "trajectories": 128,
+          "documents": 61,
+          "compressions": {
+            "1": 128
+          },
+          "planned_ratio_bins": {
+            "4": 17,
+            "7": 25,
+            "5": 31,
+            "3": 12,
+            "6": 22,
+            "8": 21
+          },
+          "segment_tokens_min": 1153,
+          "segment_tokens_max": 4082,
+          "planned_source_tokens": 344604
+        },
+        "test": {
+          "trajectories": 128,
+          "documents": 67,
+          "compressions": {
+            "1": 128
+          },
+          "planned_ratio_bins": {
+            "8": 23,
+            "3": 19,
+            "5": 20,
+            "4": 26,
+            "6": 21,
+            "7": 19
+          },
+          "segment_tokens_min": 1040,
+          "segment_tokens_max": 4081,
+          "planned_source_tokens": 330093
+        }
+      },
+      "multiround": {
+        "train": {
+          "trajectories": 32000,
+          "documents": 2474,
+          "compressions": {
+            "4": 10612,
+            "3": 10719,
+            "5": 10669
+          },
+          "planned_ratio_bins": {
+            "5": 7919,
+            "7": 5075,
+            "6": 9187,
+            "8": 3477,
+            "4": 6265,
+            "3": 77
+          },
+          "segment_tokens_min": 512,
+          "segment_tokens_max": 1536,
+          "planned_source_tokens": 84308173
+        },
+        "dev": {
+          "trajectories": 128,
+          "documents": 68,
+          "compressions": {
+            "5": 45,
+            "3": 51,
+            "4": 32
+          },
+          "planned_ratio_bins": {
+            "6": 54,
+            "4": 34,
+            "5": 26,
+            "7": 10,
+            "8": 4
+          },
+          "segment_tokens_min": 512,
+          "segment_tokens_max": 1476,
+          "planned_source_tokens": 312059
+        },
+        "test": {
+          "trajectories": 128,
+          "documents": 70,
+          "compressions": {
+            "4": 45,
+            "3": 36,
+            "5": 47
+          },
+          "planned_ratio_bins": {
+            "4": 22,
+            "7": 14,
+            "5": 38,
+            "6": 36,
+            "8": 17,
+            "3": 1
+          },
+          "segment_tokens_min": 512,
+          "segment_tokens_max": 1536,
+          "planned_source_tokens": 336041
+        }
+      }
+    }
+  },
+  "training": {
+    "objective": "ae_lm",
+    "warmup_epochs": 0,
+    "multiround_epochs": 0,
+    "independent_prefix_epochs": 3,
+    "global_batch_size": 32,
+    "micro_batch_size": 8,
+    "micro_batch_encoder_tokens": 32768,
+    "micro_batch_decoder_tokens": 36928,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "lm_ratio": 0.5,
+    "seed": 42,
+    "save_every": 1000,
+    "checkpoint_limit": 2,
+    "eval_every": null,
+    "evals_per_epoch": 4,
+    "generation_samples": 8,
+    "ae_prompt": "Reconstruct the text stored in memory:\n",
+    "lm_prompt": "Continue the text stored in memory:\n"
+  },
+  "world_size": 2,
+  "precision": "bf16",
+  "resolved_model_revision": null,
+  "resolved_architecture": {
+    "encoder_layers": 36,
+    "decoder_layers": 36,
+    "alignment_layers": 1,
+    "hidden_size": 2560,
+    "encoder_attention": "causal",
+    "shared_backbone": false,
+    "reader_adapter": null
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/tokenizer/added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/tokenizer/chat_template.jinja

@@ -0,0 +1,61 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/tokenizer/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/tokenizer/tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 1010000,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/train.jsonl

@@ -0,0 +1,2 @@
+{"step": 1, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.592894613742828, "ae": 9.49606597537086, "ae_samples": 21, "ae_tokens": 153882, "lm": 9.777748704679084, "lm_samples": 11, "lm_tokens": 23598, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.75, "batched_samples": 32, "source_tokens": 90848, "target_tokens": 177480, "grad_norm": 2.427708148956299, "seconds": 52.97523378697224, "peak_memory_bytes": 62790014976}
+{"step": 2, "stage": "independent_prefix", "epoch": 1, "global_epoch": 1, "loss": 9.530107498168945, "ae": 9.697963373859722, "ae_samples": 16, "ae_tokens": 93782, "lm": 9.362251124779384, "lm_samples": 16, "lm_tokens": 33858, "samples": 32, "microbatches": 4, "max_microbatch_size": 8, "mean_active_microbatch_size": 6.4, "batched_samples": 32, "source_tokens": 78767, "target_tokens": 127640, "grad_norm": 2.2894504070281982, "seconds": 41.040178007038776, "peak_memory_bytes": 62790014976}

artifacts/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/training-result.json

@@ -0,0 +1,9 @@
+{
+  "complete": false,
+  "completed_steps": 2,
+  "total_steps": 3000,
+  "completed_epochs": 0,
+  "source_tokens": 169615,
+  "target_tokens": 305120,
+  "checkpoint": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/smoke/qwen3-4b_weighted_ae-lm-static_20260923/checkpoints/step-000002.pt"
+}

artifacts/v2/compressor-structure_20260923/plan/startup-verification.json

@@ -0,0 +1,201 @@
+{
+  "status": {
+    "group": "qwen3-4b_compressor-structure_20260923",
+    "gpu_pairs": [
+      [
+        4,
+        5
+      ],
+      [
+        6,
+        7
+      ]
+    ],
+    "runs": [
+      {
+        "name": "qwen3-4b_weighted_ae-lm-static_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_weighted_ae-lm-static/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-static_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-static_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-static_20260923",
+        "state": "running",
+        "gpus": [
+          6,
+          7
+        ],
+        "pid": 1971841,
+        "log": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/qwen3-4b_weighted_ae-lm-static_20260923.log",
+        "started_at": "2026-09-23T02:44:22.368755+00:00"
+      },
+      {
+        "name": "qwen3-4b_spectral_ae-lm-static_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_spectral_ae-lm-static/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-static_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-static_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-static_20260923",
+        "state": "queued"
+      },
+      {
+        "name": "qwen3-4b_weighted_ae-lm-warmup_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_weighted_ae-lm-warmup/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-warmup_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923",
+        "state": "queued"
+      },
+      {
+        "name": "qwen3-4b_spectral_ae-lm-warmup_20260923",
+        "source_experiment": "/data/bywei/projects/latent_working_memory/configs/v2/pretrain/qwen3-4b_spectral_ae-lm-warmup/experiment.json",
+        "command": [
+          "/data/bywei/projects/latent_working_memory/.venv/bin/python",
+          "-u",
+          "-m",
+          "torch.distributed.run",
+          "--standalone",
+          "--nproc_per_node=2",
+          "-m",
+          "latent_working_memory.v2.pretrain.train",
+          "--experiment",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_spectral_ae-lm-warmup_20260923/experiment.json",
+          "--output-dir",
+          "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-warmup_20260923",
+          "--swanlab-project",
+          "latent-working-memory-v2",
+          "--swanlab-mode",
+          "online",
+          "--swanlab-group",
+          "qwen3-4b_compressor-structure_20260923",
+          "--swanlab-tag",
+          "study:reconstruction"
+        ],
+        "output": "/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_spectral_ae-lm-warmup_20260923",
+        "state": "queued"
+      }
+    ],
+    "scheduler_pid": 1971762,
+    "state": "running",
+    "updated_at": "2026-09-23T02:48:57.044926+00:00"
+  },
+  "first_steps": [
+    {
+      "step": 1,
+      "stage": "independent_prefix",
+      "epoch": 1,
+      "global_epoch": 1,
+      "loss": 9.592894613742828,
+      "ae": 9.49606597537086,
+      "ae_samples": 21,
+      "ae_tokens": 153882,
+      "lm": 9.777748704679084,
+      "lm_samples": 11,
+      "lm_tokens": 23598,
+      "samples": 32,
+      "microbatches": 4,
+      "max_microbatch_size": 8,
+      "mean_active_microbatch_size": 6.75,
+      "batched_samples": 32,
+      "source_tokens": 90848,
+      "target_tokens": 177480,
+      "grad_norm": 2.430598497390747,
+      "seconds": 55.27200335304951,
+      "peak_memory_bytes": 62790014976
+    },
+    {
+      "step": 2,
+      "stage": "independent_prefix",
+      "epoch": 1,
+      "global_epoch": 1,
+      "loss": 9.526120752096176,
+      "ae": 9.69505270421505,
+      "ae_samples": 16,
+      "ae_tokens": 93782,
+      "lm": 9.357188459237417,
+      "lm_samples": 16,
+      "lm_tokens": 33858,
+      "samples": 32,
+      "microbatches": 4,
+      "max_microbatch_size": 8,
+      "mean_active_microbatch_size": 6.4,
+      "batched_samples": 32,
+      "source_tokens": 78767,
+      "target_tokens": 127640,
+      "grad_norm": 2.226365566253662,
+      "seconds": 41.373851919022854,
+      "peak_memory_bytes": 62790014976
+    }
+  ],
+  "swanlab": {
+    "id": "a7nrdamp",
+    "project": "latent-working-memory-v2",
+    "group": "qwen3-4b_compressor-structure_20260923",
+    "tags": [
+      "data:fineweb",
+      "method:v2-weighted",
+      "scope:main",
+      "study:reconstruction"
+    ],
+    "job_type": "train",
+    "mode": "online",
+    "url": "https://swanlab.cn/@percyWeeeeei/latent-working-memory-v2/runs/a7nrdamp"
+  }
+}

artifacts/v2/compressor-structure_20260923/plan/tests.log

@@ -0,0 +1,9 @@
+........................................................................ [ 72%]
+...........................                                              [100%]
+=============================== warnings summary ===============================
+.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24
+  /data/bywei/projects/latent_working_memory/.venv/lib/python3.11/site-packages/verl/workers/engine/mindspeed/transformer_impl.py:24: UserWarning: NPU not support router replay for now.
+    from verl.utils.megatron.router_replay_patch import RouterReplay
+
+-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
+99 passed, 1 warning in 272.46s (0:04:32)

artifacts/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5/debug/debug.log

@@ -0,0 +1 @@
+Diagnostic log bound to file: /data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5/debug/debug.log

artifacts/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5/files/config.yaml

@@ -0,0 +1,225 @@
+data_preparation:
+  desc: ''
+  sort: 2
+  value:
+    config:
+      capacity: 512
+      content_reserve_ratio: 1.5
+      continuation_reserve_tokens: 768
+      continuation_tokens: 512
+      max_documents: 100000
+      multiround:
+        dev: 128
+        test: 128
+        train: 32000
+      seed: 20260916
+      source_glob: data/raw/HuggingFaceFW-fineweb/sample-10BT/**/*.parquet
+      source_seed: 20260907
+      split_fractions:
+      - 0.9
+      - 0.05
+      - 0.05
+      warmup:
+        dev: 128
+        test: 128
+        train: 32000
+    created_at: '2026-09-17T03:29:24.728506+00:00'
+    length_estimation: candidate characters = ceil(4 * planned content tokens * content_reserve_ratio)
+      + 4 * continuation_reserve_tokens; write_token_ends are target token positions
+    preparation_id: 036d92f2-0004-461b-8a14-31fff29ad658
+    referenced_documents: 3433
+    source_files:
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/000_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/001_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/002_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/003_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/004_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/005_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/006_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/007_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/008_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/009_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/010_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/011_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/012_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/013_00000.parquet
+    - ../raw/HuggingFaceFW-fineweb/sample-10BT/014_00000.parquet
+    source_recipe:
+      min_document_chars: 64
+      near_duplicate_min_words: 64
+      near_duplicate_threshold: 0.9
+    source_statistics:
+      duplicate: 2
+      eligible: 99998
+    statistics:
+      multiround:
+        dev:
+          compressions:
+            '3': 51
+            '4': 32
+            '5': 45
+          documents: 68
+          planned_ratio_bins:
+            '4': 34
+            '5': 26
+            '6': 54
+            '7': 10
+            '8': 4
+          planned_source_tokens: 312059
+          segment_tokens_max: 1476
+          segment_tokens_min: 512
+          trajectories: 128
+        test:
+          compressions:
+            '3': 36
+            '4': 45
+            '5': 47
+          documents: 70
+          planned_ratio_bins:
+            '3': 1
+            '4': 22
+            '5': 38
+            '6': 36
+            '7': 14
+            '8': 17
+          planned_source_tokens: 336041
+          segment_tokens_max: 1536
+          segment_tokens_min: 512
+          trajectories: 128
+        train:
+          compressions:
+            '3': 10719
+            '4': 10612
+            '5': 10669
+          documents: 2474
+          planned_ratio_bins:
+            '3': 77
+            '4': 6265
+            '5': 7919
+            '6': 9187
+            '7': 5075
+            '8': 3477
+          planned_source_tokens: 84308173
+          segment_tokens_max: 1536
+          segment_tokens_min: 512
+          trajectories: 32000
+      warmup:
+        dev:
+          compressions:
+            '1': 128
+          documents: 61
+          planned_ratio_bins:
+            '3': 12
+            '4': 17
+            '5': 31
+            '6': 22
+            '7': 25
+            '8': 21
+          planned_source_tokens: 344604
+          segment_tokens_max: 4082
+          segment_tokens_min: 1153
+          trajectories: 128
+        test:
+          compressions:
+            '1': 128
+          documents: 67
+          planned_ratio_bins:
+            '3': 19
+            '4': 26
+            '5': 20
+            '6': 21
+            '7': 19
+            '8': 23
+          planned_source_tokens: 330093
+          segment_tokens_max: 4081
+          segment_tokens_min: 1040
+          trajectories: 128
+        train:
+          compressions:
+            '1': 32000
+          documents: 3178
+          planned_ratio_bins:
+            '2': 6
+            '3': 5273
+            '4': 5245
+            '5': 5391
+            '6': 5443
+            '7': 5290
+            '8': 5352
+          planned_source_tokens: 82035196
+          segment_tokens_max: 4096
+          segment_tokens_min: 1024
+          trajectories: 32000
+model:
+  desc: ''
+  sort: 0
+  value:
+    alignment_layers: 1
+    attention_implementation: sdpa
+    compression: weighted
+    encoder_layers: null
+    feature_layer: last
+    gradient_checkpointing: true
+    lm_head_chunk_size: 256
+    lora_alpha: 32
+    lora_dropout: 0.0
+    lora_rank: 128
+    model_name_or_path: /data/bywei/models/Qwen/Qwen3-4B-Instruct-2507
+    padding_free: true
+    revision: null
+    spectral_bottleneck: 256
+precision:
+  desc: ''
+  sort: 5
+  value: bf16
+resolved_architecture:
+  desc: ''
+  sort: 7
+  value:
+    alignment_layers: 1
+    decoder_layers: 36
+    encoder_attention: causal
+    encoder_layers: 36
+    hidden_size: 2560
+    reader_adapter: null
+    shared_backbone: false
+resolved_model_revision:
+  desc: ''
+  sort: 6
+  value: {}
+selection:
+  desc: ''
+  sort: 1
+  value:
+    dataset_dir: /data/bywei/projects/latent_working_memory/data/fineweb-reconstruction-k512-doc100k_20260917
+training:
+  desc: ''
+  sort: 3
+  value:
+    ae_prompt: 'Reconstruct the text stored in memory:
+
+      '
+    checkpoint_limit: 2
+    eval_every: 1000
+    generation_samples: 8
+    global_batch_size: 32
+    gradient_clip: 1.0
+    independent_prefix_epochs: 0
+    learning_rate: 0.0001
+    lm_prompt: 'Continue the text stored in memory:
+
+      '
+    lm_ratio: 0.5
+    micro_batch_decoder_tokens: 36928
+    micro_batch_encoder_tokens: 32768
+    micro_batch_size: 8
+    multiround_epochs: 2
+    objective: ae_lm
+    save_every: 1000
+    seed: 42
+    warmup_epochs: 1
+    weight_decay: 0.01
+world_size:
+  desc: ''
+  sort: 4
+  value: 2

artifacts/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5/files/requirements.txt

@@ -0,0 +1,157 @@
+absl-py==2.5.0
+accelerate==1.14.0
+aiohappyeyeballs==2.7.1
+aiohttp==3.14.3
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+annotated-types==0.8.0
+antlr4-python3-runtime==4.9.3
+anyio==4.15.1
+attrs==26.1.0
+certifi==2026.7.22
+cffi==2.1.1
+charset-normalizer==3.5.1
+click==8.5.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+colorful==0.5.8
+cryptography==50.0.1
+cuda-bindings==13.3.1
+cuda-pathfinder==1.8.1
+cuda-toolkit==13.0.3.0
+datasets==5.0.1
+dill==0.4.1
+distlib==0.4.3
+filelock==3.32.5
+frozenlist==1.8.0
+fsspec==2026.2.0
+google-api-core==2.37.0
+google-auth==2.58.0
+googleapis-common-protos==1.75.3
+grpcio==1.84.0
+h11==0.16.0
+hf-xet==1.6.0
+httpcore==1.0.9
+httpx==0.28.1
+huggingface_hub==0.36.2
+hydra-core==1.3.7
+idna==3.19
+importlib_metadata==9.0.1
+iniconfig==2.3.0
+Jinja2==3.1.6
+jsonschema==4.26.0
+jsonschema-specifications==2025.9.1
+latent-working-memory==0.1.0
+lxml==6.1.3
+Markdown==3.10.3
+markdown-it-py==4.2.0
+MarkupSafe==3.0.3
+mdurl==0.1.2
+mpmath==1.3.0
+msgpack==1.2.2
+multidict==6.8.0
+multiprocess==0.70.19
+networkx==3.6.1
+numpy==1.26.4
+nvidia-cublas==13.1.1.3
+nvidia-cuda-cupti==13.0.85
+nvidia-cuda-nvrtc==13.0.88
+nvidia-cuda-runtime==13.0.96
+nvidia-cudnn-cu13==9.24.0.43
+nvidia-cufft==12.0.0.61
+nvidia-cufile==1.15.1.6
+nvidia-curand==10.4.0.35
+nvidia-cusolver==12.0.4.66
+nvidia-cusparse==12.6.3.3
+nvidia-cusparselt-cu13==0.8.1
+nvidia-ml-py==13.610.43
+nvidia-nccl-cu13==2.30.7
+nvidia-nvjitlink==13.3.33
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvtx==13.0.85
+omegaconf==2.3.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opentelemetry-api==1.44.0
+opentelemetry-exporter-otlp-proto-common==1.44.0
+opentelemetry-exporter-otlp-proto-http==1.44.0
+opentelemetry-exporter-prometheus==0.65b0
+opentelemetry-proto==1.44.0
+opentelemetry-sdk==1.44.0
+opentelemetry-semantic-conventions==0.65b0
+orjson==3.12.0
+packaging==25.0
+pandas==3.0.5
+peft==0.20.0
+pillow==12.3.0
+pip==24.0
+platformdirs==4.11.7
+pluggy==1.6.0
+portalocker==4.3.0
+prettytable==3.18.0
+prometheus_client==0.26.0
+propcache==0.5.4
+proto-plus==1.28.4
+protobuf==6.33.6
+psutil==7.2.2
+pwinput==1.0.3
+py-spy==0.4.2
+pyarrow==25.0.1
+pyasn1==0.6.4
+pyasn1_modules==0.4.2
+pybind11==3.1.0
+pycparser==3.0
+pydantic==2.13.5
+pydantic_core==2.46.5
+pydantic-settings==2.15.0
+pyecharts==2.1.0
+Pygments==2.21.0
+pylatexenc==2.11
+pyroaring==1.0.3
+pysbd==0.3.4
+pytest==8.4.2
+python-dateutil==2.9.0.post0
+python-discovery==1.6.0
+python-dotenv==1.2.3
+pyvers==0.1.0
+PyYAML==6.0.3
+ray==2.58.0
+referencing==0.37.0
+regex==2026.9.3
+requests==2.34.2
+rich==15.0.0
+rpds-py==2026.6.3
+ruff==0.13.3
+sacrebleu==2.6.0
+safetensors==0.8.0
+setuptools==84.0.0
+simplejson==4.1.2
+six==1.17.0
+smart_open==8.0.1
+swanlab==0.10.0
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.21.0
+tensorboard-data-server==0.7.2
+tensordict==0.10.0
+tokenizers==0.22.2
+torch==2.14.0
+torchdata==0.11.0
+tqdm==4.70.0
+transformers==4.57.6
+triton==3.8.0
+typing_extensions==4.16.0
+typing-inspection==0.4.4
+urllib3==2.7.0
+uv==0.8.22
+verl==0.8.0
+virtualenv==21.7.10
+wandb==0.30.0
+watchdog==6.0.0
+wcwidth==0.8.3
+Werkzeug==3.1.8
+wrapt==2.4.0
+xxhash==4.0.1
+yarl==1.25.1
+zipp==4.1.0

artifacts/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5/files/swanlab-metadata.json

@@ -0,0 +1 @@
+{"_version":2,"hardware":{"apple_silicon":null,"cpu":{"brand":"Intel(R) Xeon(R) Platinum 8352Y CPU @ 2.20GHz","physical_count":64,"logical_count":128},"memory":{"total":2015,"total_unit":"GB"},"accelerators":[{"vendor":"nvidia","devices":[{"index":0,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":1,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":2,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":3,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":4,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":5,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":6,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"},{"index":7,"name":"NVIDIA A800-SXM4-80GB","memory":80,"memory_unit":"GB"}],"version":"580.178.04","cuda_version":"13.0","cann_version":null}]},"runtime":{"os":"Linux-6.8.0-139-generic-x86_64-with-glibc2.39","os_pretty":"Ubuntu 24.04.2 LTS","hostname":"localhost","pid":2270589,"cwd":"/data/bywei/projects/latent_working_memory","python_version":"3.11.16","python_verbose":"3.11.16 (main, Aug 25 2026, 14:00:53) [Clang 22.1.3 ]","python_executable":"/data/bywei/projects/latent_working_memory/.venv/bin/python","command":"/data/bywei/projects/latent_working_memory/.venv/bin/python -u -m latent_working_memory.v2.pretrain.train --experiment /data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/plan/qwen3-4b_weighted_ae-lm-warmup_20260923/experiment.json --output-dir /data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923 --swanlab-project latent-working-memory-v2 --swanlab-mode online --swanlab-group qwen3-4b_compressor-structure_20260923 --swanlab-tag study:reconstruction"},"git":null,"swanlab":{"version":"0.10.0","run_dir":"/data/bywei/experiment-artifacts/latent_working_memory/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5"}}

artifacts/v2/compressor-structure_20260923/train/qwen3-4b_weighted_ae-lm-warmup_20260923/swanlab/run-20260924_032204-7wtghxu5/media/echarts/2999-d30dd303.json

@@ -0,0 +1,178 @@
+{
+    "animation": true,
+    "animationThreshold": 2000,
+    "animationDuration": 1000,
+    "animationEasing": "cubicOut",
+    "animationDelay": 0,
+    "animationDurationUpdate": 300,
+    "animationEasingUpdate": "cubicOut",
+    "animationDelayUpdate": 0,
+    "aria": {
+        "enabled": false
+    },
+    "series": [
+        {
+            "type": "bar",
+            "name": "write=compressed / test=FineWeb",
+            "legendHoverLink": true,
+            "data": [
+                0.0,
+                null
+            ],
+            "realtimeSort": false,
+            "showBackground": false,
+            "stack": "FineWeb",
+            "stackStrategy": "samesign",
+            "cursor": "pointer",
+            "barMinHeight": 0,
+            "barCategoryGap": "20%",
+            "barGap": "30%",
+            "large": false,
+            "largeThreshold": 400,
+            "seriesLayoutBy": "column",
+            "datasetIndex": 0,
+            "clip": true,
+            "zlevel": 0,
+            "z": 2,
+            "label": {
+                "show": false
+            },
+            "itemStyle": {
+                "color": "#2459a6"
+            }
+        },
+        {
+            "type": "bar",
+            "name": "write=one-shot / test=FineWeb",
+            "legendHoverLink": true,
+            "data": [
+                null,
+                0.0
+            ],
+            "realtimeSort": false,
+            "showBackground": false,
+            "stack": "FineWeb",
+            "stackStrategy": "samesign",
+            "cursor": "pointer",
+            "barMinHeight": 0,
+            "barCategoryGap": "20%",
+            "barGap": "30%",
+            "large": false,
+            "largeThreshold": 400,
+            "seriesLayoutBy": "column",
+            "datasetIndex": 0,
+            "clip": true,
+            "zlevel": 0,
+            "z": 2,
+            "label": {
+                "show": false
+            },
+            "itemStyle": {
+                "color": "#28764a"
+            }
+        }
+    ],
+    "legend": [
+        {
+            "data": [
+                "write=compressed / test=FineWeb",
+                "write=one-shot / test=FineWeb"
+            ],
+            "selected": {},
+            "type": "scroll",
+            "top": 0
+        }
+    ],
+    "tooltip": {
+        "trigger": "axis"
+    },
+    "xAxis": [
+        {
+            "show": true,
+            "scale": false,
+            "nameLocation": "end",
+            "nameGap": 15,
+            "nameTruncate": {},
+            "nameMoveOverlap": true,
+            "axisLabel": {
+                "interval": 0,
+                "rotate": 0,
+                "fontSize": 10,
+                "lineHeight": 11
+            },
+            "inverse": false,
+            "offset": 0,
+            "splitNumber": 5,
+            "minInterval": 0,
+            "silent": false,
+            "triggerEvent": false,
+            "splitLine": {
+                "show": true,
+                "lineStyle": {
+                    "show": false
+                }
+            },
+            "animation": true,
+            "animationThreshold": 2000,
+            "animationDuration": 1000,
+            "animationEasing": "cubicOut",
+            "animationDelay": 0,
+            "animationDurationUpdate": 300,
+            "animationEasingUpdate": "cubicOut",
+            "animationDelayUpdate": 0,
+            "data": [
+                "recurrent",
+                "independent"
+            ]
+        }
+    ],
+    "yAxis": [
+        {
+            "show": true,
+            "scale": false,
+            "nameLocation": "end",
+            "nameGap": 15,
+            "nameTruncate": {},
+            "nameMoveOverlap": true,
+            "inverse": false,
+            "offset": 0,
+            "splitNumber": 3,
+            "minInterval": 0.001,
+            "silent": false,
+            "triggerEvent": false,
+            "splitLine": {
+                "show": true,
+                "lineStyle": {
+                    "show": false
+                }
+            },
+            "animation": true,
+            "animationThreshold": 2000,
+            "animationDuration": 1000,
+            "animationEasing": "cubicOut",
+            "animationDelay": 0,
+            "animationDurationUpdate": 300,
+            "animationEasingUpdate": "cubicOut",
+            "animationDelayUpdate": 0
+        }
+    ],
+    "title": [
+        {
+            "show": true,
+            "target": "blank",
+            "subtarget": "blank",
+            "padding": 5,
+            "itemGap": 10,
+            "textAlign": "auto",
+            "textVerticalAlign": "auto",
+            "triggerEvent": false
+        }
+    ],
+    "grid": {
+        "left": "12%",
+        "right": "4%",
+        "top": "20%",
+        "height": "55%",
+        "containLabel": false
+    }
+}