Delete config/rl_v6f_t75.yaml

config/rl_v6f_t75.yaml

@@ -1,179 +0,0 @@
-# ===== 跑5 (v6f): 基于跑4(v6e) events.jsonl 实测动态调参 =====
-# 与跑2/3/4 相比只改 4 项, reward preset / group_size / batch / nft_beta /
-# reference_weight / z_clip / seed / shuffle_seed 一律不动。
-#
-#  1) lr 1e-4 -> 2e-4
-#     依据: v6e 全程 grad_norm 0.058-0.086, 而 grad_clip_norm=1.0 -- 比阈值小
-#     一个数量级, 裁剪一次都没触发。翻倍后预期 ~0.12-0.17, 仍远低于 1.0。
-#  2) lr_min_ratio 0.1 -> 0.3
-#     依据: v6e 末段 lr 已衰到 1e-5, 但 batch/good_rate 在 230-263 段达到全程
-#     最高 0.4744, reward_mean 仍在 0.52 高位 -- 训练是被 lr 掐死的, 不是收敛的。
-#     0.3 让 lr 底线抬到 6e-5。
-#  3) epochs 2 -> 3, max_outer_steps / lr_total_steps 263 -> 395
-#     依据: reward_mean 分段 0.4231->0.5233 单调上行未饱和, advantage 未退化。
-#  4) validation.keep_runs 3 -> 4
-#     依据: v6e 曲线真峰值在 step130 (GDT 0.376947), 但 top3 预算把它挤掉了,
-#     峰值权重不存在、无法评测。
-# ============================================================
-data:
-  pool_manifest: artifacts/evaluation_pools/train_final_v4.pt
-  ride_checkpoint: saved_models/checkpoint.h5
-  processed_data: ../data/processed.pt
-  split_file: ../data/das_split.pt
-  condition_noise_scale: 0.0
-  condition_snapshot_enabled: false
-  x0_representation: one_hot_01
-ride_model:
-  node_in_dim:
-  - 15
-  - 4
-  node_h_dim:
-  - 256
-  - 24
-  edge_in_dim:
-  - 131
-  - 3
-  edge_h_dim:
-  - 128
-  - 4
-  num_layers: 5
-  drop_rate: 0.5
-  out_dim: 4
-  time_cond: true
-noise_schedule:
-  schedule: linear
-  continuous_beta_0: 0.1
-  continuous_beta_1: 20.0
-  eps: 0.001
-rollout:
-  group_size: 8
-  sampler:
-    name: x0_renoise
-    n_steps: 50
-    version: v1
-  temperature:
-    initial: 1.0
-    min: 0.2
-    max: 1.5
-    step: 0.1
-  rescue:
-    max_rounds: 1
-    min_unique_sequences: 2
-    min_valid_sequences: 2
-reward:
-  oracle:
-    backend: local_jsonl_worker
-    python: ${RHOFOLD_PYTHON:-python}
-    worker_script: scripts/rhofold_oracle_worker.py
-    ckpt: ${RHOFOLD_CHECKPOINT:-artifacts/models/rhofold/RhoFold_pretrained.pt}
-    output_dir: outputs/rl_v6f/oracle
-    timeout_sec: 600
-    device: cuda:0
-    fake: false
-  metric_cache:
-    dir: ${RIDE_RL_CACHE:-cache/rl_metrics}
-    metric_version: ride_evaluator_v2_geo6
-    alignment_config: c4p_kabsch_centered
-  composer:
-    # === A/B 实验的唯一变量 ===
-    # A 组: balanced_cosine_v1 = 三分量加权算术平均
-    # B 组: geometric_v3       = 六分量加权几何平均, 权重由
-    #       scripts/reward/reward_design_v3.py 在 1224 条真实候选上求解得到
-    #       (gdt .25 / tm .30 / rmsd .30 / scmcc .05 / edefect .05 / composition .05)
-    #       alignment 0.9608 (门槛 0.70), discrimination 1.2061 (门槛 1.20, 唯一紧约束)
-    # good 门槛由 preset 内部的 threshold_position=0.60 推出, 不再从这里读,
-    # 保留字段只为与 A 组配置结构完全对齐。
-    preset: geometric_v3
-    good:
-      gdt_ts: 0.5
-      tm_score: 0.5
-      rmsd: 4.0
-advantage:
-  group_normalize: true
-  z_clip: 2.0
-  scale:
-    kind: ema
-    decay: 0.95
-    epsilon: 1.0e-06
-    min_valid: 2
-trainer:
-  loss:
-    nft_beta: 0.1
-    reference_weight: 0.02
-    k_t: 8
-    weighting: uniform_eps
-    residual_normalization: none
-    divide_by_beta: true
-  optim:
-    lr: 0.0002
-    weight_decay: 0.0
-    grad_clip_norm: 1.0
-    gradient_accumulation_steps: 1
-    amp: false
-    microbatch_size: 1
-    lr_schedule: warmup_cosine
-    lr_warmup_ratio: 0.05
-    lr_min_ratio: 0.3
-    lr_total_steps: 395
-  ema:
-    decay_max: 0.999
-    update_interval: 1
-runtime:
-  seed: 0
-  device: cuda
-  mode: train
-  ranks_per_target: 1
-  # 527 unique targets * 3 passes / 4 per batch = 395 updates (1581 // 4).
-  # engine.py: total_steps = len(targets) // batch_targets, so max_outer_steps
-  # must equal that or the walk clamps it.
-  epochs: 3
-  shuffle_targets: true
-  shuffle_seed: 20260826
-  budget:
-    max_outer_steps: 395
-    max_oracle_calls: 200000
-    max_wall_hours: 72.0
-    checkpoint_every_oracle_calls: 128
-validation:
-  enabled: true
-  # Held-out 153, disjoint from the 527 training pool (verified zero sequence overlap).
-  pool_manifest: artifacts/evaluation_pools/test_final_v3_std.pt
-  config: configs/diffv3_te0.yaml
-  # t75 only has 4 usable cards (PCI 3E:00.0 / 3F:00.0 are disabled -- they drop
-  # the driver under load), so validation shares them with training instead of
-  # getting spare cards of its own. This is safe here because the worker is
-  # fully detached (start_new_session, torchrun env stripped) and never joins the
-  # process group: the earlier stall came from a mismatched barrier on the
-  # critical path, not from co-residency. Memory fits -- RhoFold+ peaks under
-  # 11 GiB and training holds roughly 10 GiB of a 24 GiB 3090.
-  # 4 shards over 4 cards keeps every card busy during the fold phase.
-  devices: "cuda:0,cuda:1,cuda:2,cuda:3"
-  test_freq: 10
-  # 4 instead of 3: v6e 的真峰值在 step130, 按 top-2+latest 的 top3 预算被挤掉。
-  keep_runs: 4
-  before_train: true
-  n_samples: 8
-  timeout_sec: 7200.0
-  usalign_binary: ../third_party/USalign/USalign
-logging:
-  jsonl: outputs/rl_v6f/train.jsonl
-  wandb:
-    enabled: true
-    project: ride-rl
-    entity: gabriel166714-fudan-university
-    name: v6f-t75-geometric-v3-lr2e4-min03-3ep-395step
-    tags:
-    - v6
-    - 4gpu
-    - ddp
-    - batch4
-    - grpo
-    - geometric_v3
-    - train527
-    - test153
-checkpoint:
-  dir: outputs/rl_v6f/checkpoints
-  save_latest: true
-evaluation:
-  deterministic_anchor_temperature: 0.0
-  group_size: 2