nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[evidence]: RL (DiffusionNFT) + VideoAlign + dataloader gates PASS — goldens, ledger, report

Browse Source
This commit is contained in:
SolitaryThinker
2026-07-23 14:27:33 -07:00
parent 4dcdb0d42c
commit 60054536d8
9 changed files with 526 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastvideo2/evidence/goldens/train-rl-main/inner0.npz
BIN
View File
Binary file not shown.
fastvideo2/evidence/goldens/train-rl-main/inner1.npz
BIN
View File
Binary file not shown.
fastvideo2/evidence/goldens/train-rl-main/manifest.json
+20
View File
@@ -0,0 +1,20 @@
{
"fastvideo_commit": "c459a1897899ffcec3be7765534d81000b9bb9c1",
"steps": 2,
"config": "diffusion_nft single-frame: 2 outer x 2 batches x 4 timesteps, pickscore+clipscore, crush-smol prompts, beta .1 kl 1e-4 lr 3e-5",
"probe": {
"param_dtype": "torch.float32",
"decay_type": 1,
"ntt": 1000,
"nft_beta": 0.1,
"kl_beta": 0.0001,
"adv_clip_max": 5.0,
"adv_mode": "all",
"max_grad_norm": 1.0
},
"gate_yaml": "method:\n _target_: fastvideo.train.methods.rl.diffusion_nft.DiffusionNFTMethod\n adv_clip_max: 5\n adv_mode: all\n beta: 0.1\n decay_type: 1\n ema:\n decay: 0.9\n enabled: false\n update_after_step: 0\n validation: true\n kl_beta: 0.0001\n max_grad_norm: 1.0\n num_batches_per_epoch: 2\n num_inner_epochs: 1\n num_video_per_prompt: 2\n reward_fn:\n clipscore: 1.0\n pickscore: 1.0\n sample_train_batch_size: 2\n sampling:\n flow_shift: inherit\n num_steps: 4\n scheduler: flow_match_euler\n trajectory: ode\n terminal_progress: false\n timestep_fraction: 1.0\n train_batch_size: 2\n validation:\n batch_size: 16\n data_path: null\n every_steps: 1000000\n log_samples: true\n num_prompts: 16\n num_steps: 40\n seed: 42\nmodels:\n old:\n _target_: fastvideo.train.models.wan.WanModel\n disable_custom_init_weights: true\n init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers\n trainable: false\n reference:\n _target_: fastvideo.train.models.wan.WanModel\n disable_custom_init_weights: true\n init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers\n trainable: false\n student:\n _target_: fastvideo.train.models.wan.WanModel\n init_from: Wan-AI/Wan2.1-T2V-1.3B-Diffusers\n trainable: true\npipeline:\n flow_shift: 8\ntraining:\n checkpoint:\n checkpoints_total_limit: 3\n output_dir: /tmp/fv2_rl_capture\n training_state_checkpointing_steps: 1000000\n data:\n data_path: /home/willlin/.cache/huggingface/hub/datasets--wlsaidhi--crush-smol_processed_t2v/snapshots/67dd07f2163ad2b3397f8b3d8125b67ca452dc85/combined_parquet_dataset\n dataloader_num_workers: 0\n num_frames: 1\n num_height: 448\n num_latent_t: 1\n num_width: 832\n preprocessed_data_type: t2v\n seed: 42\n train_batch_size: 1\n training_cfg_rate: 0.0\n distributed:\n hsdp_replicate_dim: 1\n hsdp_shard_dim: 1\n num_gpus: 1\n sp_size: 1\n tp_size: 1\n loop:\n gradient_accumulation_steps: 2\n max_train_steps: 2\n model:\n enable_gradient_checkpointing_type: null\n optimizer:\n betas:\n - 0.9\n - 0.999\n learning_rate: 3.0e-05\n lr_scheduler: constant\n lr_warmup_steps: 0\n weight_decay: 0.0001\n tracker:\n project_name: diffusion_nft_wan\n run_name: wan2.1_diffusion_nft_pick_clip\n",
"torch": "2.12.0+cu130",
"gpu": "NVIDIA GB200",
"self_noise_max": 0.04484055485020377,
"self_noise_note": "max |total_loss| spread over 3 identical-seed capture runs (all steps)"
}
fastvideo2/evidence/goldens/train-rl-main/samples0.npz
BIN
View File
Binary file not shown.
fastvideo2/evidence/goldens/train-rl-main/samples1.npz
BIN
View File
Binary file not shown.
fastvideo2/evidence/goldens/train-rl-main/steps.json
+408
View File
@@ -0,0 +1,408 @@
[
{
"items": [
{
"prompts": [
"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."
]
},
{
"prompts": [
"A pizza is being crushed by a hydraulic press, causing the toppings to spread out and the crust to crumble.",
"A pizza is being crushed by a hydraulic press, causing the toppings to spread out and the crust to crumble."
]
}
],
"inner": [
{
"timestep_idx": 0,
"row_idx": [
1,
3
],
"timestep": [
803.7713623046875,
60.20697021484375
],
"losses": {
"total_loss": 5.342502646625078,
"policy_loss": 5.342502646625078,
"unweighted_policy_loss": 1.0685005293250156,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1357421875
}
},
{
"timestep_idx": 1,
"row_idx": [
1,
3
],
"timestep": [
941.8340454101562,
1000.0
],
"losses": {
"total_loss": 13.446108607202637,
"policy_loss": 13.446108607202637,
"unweighted_policy_loss": 2.6892217214405276,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1357421875
}
},
{
"timestep_idx": 2,
"row_idx": [
1,
3
],
"timestep": [
60.20697021484375,
803.7713623046875
],
"losses": {
"total_loss": 4.515896192359375,
"policy_loss": 4.515896192359375,
"unweighted_policy_loss": 0.9031792384718751,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1357421875
}
},
{
"timestep_idx": 3,
"row_idx": [
1,
3
],
"timestep": [
1000.0,
941.8340454101562
],
"losses": {
"total_loss": 14.399518803001467,
"policy_loss": 14.399518803001467,
"unweighted_policy_loss": 2.8799037606002935,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1357421875
}
},
{
"timestep_idx": 0,
"row_idx": [
2,
0
],
"timestep": [
941.8340454101562,
941.8340454101562
],
"losses": {
"total_loss": 13.118073008759543,
"policy_loss": 13.118073008759543,
"unweighted_policy_loss": 2.6236146017519086,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1845703125
}
},
{
"timestep_idx": 1,
"row_idx": [
2,
0
],
"timestep": [
1000.0,
1000.0
],
"losses": {
"total_loss": 18.29075873703107,
"policy_loss": 18.29075873703107,
"unweighted_policy_loss": 3.6581517474062135,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1845703125
}
},
{
"timestep_idx": 2,
"row_idx": [
2,
0
],
"timestep": [
60.20697021484375,
60.20697021484375
],
"losses": {
"total_loss": 1.6802980970623111,
"policy_loss": 1.6802980970623111,
"unweighted_policy_loss": 0.33605961941246226,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1845703125
}
},
{
"timestep_idx": 3,
"row_idx": [
2,
0
],
"timestep": [
803.7713623046875,
803.7713623046875
],
"losses": {
"total_loss": 7.296204002432377,
"policy_loss": 7.296204002432377,
"unweighted_policy_loss": 1.4592408004864756,
"kl_div_loss": 0.0,
"old_deviate": 0.0,
"old_kl_div": 0.0,
"x0_norm": 0.1845703125
}
}
],
"rewards": {
"clipscore": [
0.13153909146785736,
0.20024922490119934,
0.14863893389701843,
0.2340940237045288
],
"pickscore": [
0.6811658143997192,
0.7108843326568604,
0.7316142916679382,
0.7339301109313965
],
"avg": [
0.8127049207687378,
0.9111335277557373,
0.8802531957626343,
0.9680241346359253
]
},
"old_decay": 0.001,
"advantages_shape": [
4,
4
]
},
{
"items": [
{
"prompts": [
"The video shows a stack of caramelized sugar cubes being flattened as if they were under a hydraulic press, resulting in a messy pile of broken sugar on the table.",
"The video shows a stack of caramelized sugar cubes being flattened as if they were under a hydraulic press, resulting in a messy pile of broken sugar on the table."
]
},
{
"prompts": [
"The video shows a hydraulic press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing two colorful objects that resemble sandwiches. The press is yellow and black striped, and the objects being flattened are placed on a metal plate. The background is green, and the press is moving down, compressing the objects.",
"The video shows a hydraulic press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing two colorful objects that resemble sandwiches. The press is yellow and black striped, and the objects being flattened are placed on a metal plate. The background is green, and the press is moving down, compressing the objects."
]
}
],
"inner": [
{
"timestep_idx": 0,
"row_idx": [
1,
2
],
"timestep": [
941.8340454101562,
1000.0
],
"losses": {
"total_loss": 12.398794127316185,
"policy_loss": 12.398782623619743,
"unweighted_policy_loss": 2.479756524723949,
"kl_div_loss": 0.11474609375,
"old_deviate": 2.8848648071289062e-05,
"old_kl_div": 0.11474609375,
"x0_norm": 0.232421875
}
},
{
"timestep_idx": 1,
"row_idx": [
1,
2
],
"timestep": [
60.20697021484375,
941.8340454101562
],
"losses": {
"total_loss": 5.758004918792544,
"policy_loss": 5.758003212609587,
"unweighted_policy_loss": 1.1516006425219176,
"kl_div_loss": 0.01708984375,
"old_deviate": 0.00015544891357421875,
"old_kl_div": 0.0169677734375,
"x0_norm": 0.232421875
}
},
{
"timestep_idx": 2,
"row_idx": [
1,
2
],
"timestep": [
1000.0,
803.7713623046875
],
"losses": {
"total_loss": 10.294339807390942,
"policy_loss": 10.29431918418385,
"unweighted_policy_loss": 2.05886383683677,
"kl_div_loss": 0.2060546875,
"old_deviate": 3.147125244140625e-05,
"old_kl_div": 0.205078125,
"x0_norm": 0.232421875
}
},
{
"timestep_idx": 3,
"row_idx": [
1,
2
],
"timestep": [
803.7713623046875,
60.20697021484375
],
"losses": {
"total_loss": 3.6922912036167417,
"policy_loss": 3.69229043248165,
"unweighted_policy_loss": 0.7384580864963299,
"kl_div_loss": 0.007720947265625,
"old_deviate": 0.00016689300537109375,
"old_kl_div": 0.00775146484375,
"x0_norm": 0.232421875
}
},
{
"timestep_idx": 0,
"row_idx": [
0,
3
],
"timestep": [
941.8340454101562,
1000.0
],
"losses": {
"total_loss": 10.91259682812316,
"policy_loss": 10.91258747019393,
"unweighted_policy_loss": 2.1825174940387857,
"kl_div_loss": 0.09375,
"old_deviate": 2.9206275939941406e-05,
"old_kl_div": 0.09326171875,
"x0_norm": 0.296875
}
},
{
"timestep_idx": 1,
"row_idx": [
0,
3
],
"timestep": [
60.20697021484375,
803.7713623046875
],
"losses": {
"total_loss": 3.5539959718750014,
"policy_loss": 3.5539953236744894,
"unweighted_policy_loss": 0.7107990647348978,
"kl_div_loss": 0.0064697265625,
"old_deviate": 0.00017261505126953125,
"old_kl_div": 0.006439208984375,
"x0_norm": 0.296875
}
},
{
"timestep_idx": 2,
"row_idx": [
0,
3
],
"timestep": [
1000.0,
941.8340454101562
],
"losses": {
"total_loss": 10.252210426698156,
"policy_loss": 10.252192783723302,
"unweighted_policy_loss": 2.050438556744661,
"kl_div_loss": 0.1767578125,
"old_deviate": 2.9325485229492188e-05,
"old_kl_div": 0.1767578125,
"x0_norm": 0.296875
}
},
{
"timestep_idx": 3,
"row_idx": [
0,
3
],
"timestep": [
803.7713623046875,
60.20697021484375
],
"losses": {
"total_loss": 4.397589007283385,
"policy_loss": 4.397588385159906,
"unweighted_policy_loss": 0.8795176770319812,
"kl_div_loss": 0.0062255859375,
"old_deviate": 0.00019168853759765625,
"old_kl_div": 0.0062255859375,
"x0_norm": 0.296875
}
}
],
"rewards": {
"clipscore": [
0.20155902206897736,
0.1521805226802826,
0.18642529845237732,
0.19080127775669098
],
"pickscore": [
0.7295283079147339,
0.7310761213302612,
0.7291601896286011,
0.7219043374061584
],
"avg": [
0.93108731508255,
0.8832566738128662,
0.9155855178833008,
0.9127056002616882
]
},
"old_decay": 0.002,
"advantages_shape": [
4,
4
]
}
]
fastvideo2/evidence/goldens/videoalign-pr1476/scores.json
+23
View File
@@ -0,0 +1,23 @@
{
"scores": {
"MQ": [
0.497402,
0.457805,
-0.017359
],
"VQ": [
-1.186994,
-1.075764,
0.408458
],
"TA": [
-2.897675,
-2.61778,
-1.07544
]
},
"checkpoint": "/home/willlin/.cache/huggingface/hub/models--KwaiVGI--VideoReward/snapshots/4f26600130683e6f1de9f5d463887f28e8ef995c",
"authority": "PR #1476 maint/pr1476-runtime-compat @518aeab0b (vendored byte-identical, see test_videoalign.py)",
"torch": "2.12.0+cu130",
"gpu": "NVIDIA GB200"
}
fastvideo2/evidence/ledger.jsonl
+4
View File
@@ -166,3 +166,7 @@
{"ts": "2026-07-23T17:51:55Z", "gate": "anchor.train-self_forcing-main", "status": "pass", "model_id": "sfwan-t2v-1.3b", "card_digest": "9c9b4cc573606028", "metrics": {"x0.step0": 0.0, "fake.step0": 0.00029414892196655273, "x0.step1": 0.0, "fake.step1": 0.0005563050508499146, "x0.step2": 0.0, "fake.step2": 0.00035434961318969727, "x0.step3": 0.0, "fake.step3": 0.0002406388521194458, "x0.step4": 0.0, "fake.step4": 0.00022238492965698242, "fwd.matches": 0.0}, "tolerances": {"x0-pre-update": 0.0, "later": "band"}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "93153f8", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897"}
{"ts": "2026-07-23T18:23:19Z", "gate": "anchor.serve-identity", "status": "pass", "model_id": "fastwan-qad-fp8-1.3b", "card_digest": "1c8e6f7d1380552d", "metrics": {"identity": 0.0, "ws_steps": 3.0}, "tolerances": {"identity": 0.0}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "b376207", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "served 41c4817d3d0f666b offline 41c4817d3d0f666b"}
{"ts": "2026-07-23T18:27:47Z", "gate": "anchor.dreamverse-runtime", "status": "pass", "model_id": "fastwan-qad-fp8-1.3b", "card_digest": "-", "metrics": {"steps": 3.0, "chunks": 32.0, "ftyp": 1.0}, "tolerances": {}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "fdddc59", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "latents 9671b076aafff016"}
{"ts": "2026-07-23T19:26:25Z", "gate": "anchor.train-rl-main", "status": "fail", "model_id": "wan2.1-t2v-1.3b", "card_digest": "7171d8729ae72ae4", "metrics": {"advantages.step0": 0.0, "loss.s0k0": 0.0034792618441672474, "loss.s0k1": 0.0007733339205699963, "loss.s0k2": 0.00017366389334405596, "loss.s0k3": 0.004427341173300903, "loss.s0k4": 0.021503618076398112, "loss.s0k5": 0.0033195134562689077, "loss.s0k6": 2.1168774413382607e-05, "loss.s0k7": 0.000990179815745229, "advantages.step1": 0.0, "loss.s1k0": 0.05847059518240627, "loss.s1k1": 0.0010533409492632373, "loss.s1k2": 0.0034236295955523133, "loss.s1k3": 0.004349640925024101, "loss.s1k4": 0.047157419533888856, "loss.s1k5": 0.002190229000701205, "loss.s1k6": 0.011363542387446302, "loss.s1k7": 0.005596577394816116}, "tolerances": {"advantages": 0.0, "step0": "exact-or-band", "step1": "main self-noise band"}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "3866db2", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897, band=6.726e-02"}
{"ts": "2026-07-23T19:54:30Z", "gate": "anchor.train-rl-main", "status": "pass", "model_id": "wan2.1-t2v-1.3b", "card_digest": "7171d8729ae72ae4", "metrics": {"advantages.step0": 0.0, "loss.s0k0": 0.0, "loss.s0k1": 0.0, "loss.s0k2": 0.0, "loss.s0k3": 0.0, "loss.s0k4": 0.0, "loss.s0k5": 0.0, "loss.s0k6": 0.0, "loss.s0k7": 0.0, "advantages.step1": 0.0, "loss.s1k0": 0.011695116289581975, "loss.s1k1": 0.0015937395833205414, "loss.s1k2": 0.002726593122572396, "loss.s1k3": 0.0009456473942783816, "loss.s1k4": 0.011560843720619829, "loss.s1k5": 0.0038772091026810784, "loss.s1k6": 0.00849335173029786, "loss.s1k7": 0.0002466132267455734}, "tolerances": {"advantages": 0.0, "step0": "exact-or-band", "step1": "main self-noise band"}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "b01bde8", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897, band=6.726e-02"}
{"ts": "2026-07-23T19:58:25Z", "gate": "anchor.videoalign-pr1476", "status": "pass", "model_id": "wan2.1-t2v-1.3b", "card_digest": "7171d8729ae72ae4", "metrics": {"MQ.v0": 0.497402, "MQ.v1": 0.457805, "MQ.v2": -0.017359, "VQ.v0": -1.186994, "VQ.v1": -1.075764, "VQ.v2": 0.408458, "TA.v0": -2.897675, "TA.v1": -2.61778, "TA.v2": -1.07544}, "tolerances": {"determinism": "exact", "vs_golden": "exact"}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "b01bde8", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "MQ/VQ/TA on 3 fixed media, KwaiVGI/VideoReward, use_norm=True"}
{"ts": "2026-07-23T19:58:44Z", "gate": "anchor.train-data-main", "status": "pass", "model_id": "wan2.1-t2v-1.3b", "card_digest": "7171d8729ae72ae4", "metrics": {"step0.row_match": 1.0, "step1.row_match": 1.0, "step2.row_match": 1.0, "step3.row_match": 1.0, "step4.row_match": 1.0}, "tolerances": {"caption/latents/embeds": "exact"}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "b01bde8", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "loader order vs goldens c459a1897: 5 batches, files=4"}
fastvideo2/evidence/train_numerics_report.md
+71
View File
@@ -186,3 +186,74 @@ Gate transcript: 3 live steps, 32 fMP4 chunks (ISO-BMFF ftyp verified),
Scope note: the original app is LTX2-specific (audio, refine, continuation
state, LoRA stacks) — this port re-bases Dreamverse's ARCHITECTURE onto the
wan family; their Next.js client's message schema maps onto these events.
## Dataloader order — added 2026-07-23
Gate `anchor.train-data-main`: PASS. `fastvideo2/train/data.py` reproduces
main's map-style parquet loader at world size 1: file discovery is
`os.walk(realpath(root))` with each file realpath'd (on an HF snapshot the
sort key is therefore the resolved BLOB path) and one global path sort;
the sampler is `torch.randperm(total, generator=manual_seed(seed))`
truncated to a batch multiple and chunked sequentially; rows decode
`np.frombuffer(..., np.float32)` (main ignores the `_dtype` column); text
embeddings pad/crop to 512 with a 1/0 mask. Checked against the finetune
capture's recorded stream: first 5 batches match on caption, latents hash
(post `num_latent_t` slice, bf16-cast) and embeds hash — all exact.
## DiffusionNFT RL — added 2026-07-23
Gate `anchor.train-rl-main`: PASS. Goldens from main's MERGED modular
trainer (`train/methods/rl/diffusion_nft.py`) at a shrunk single-frame
config (2 outer steps x 2 sample batches x 2 videos/prompt, 4 sampling
steps, timestep_fraction 1.0, pickscore+clipscore rewards, crush-smol
prompts, Wan2.1-T2V-1.3B for student/old/reference).
Capture records the DIRECT inputs of every `_training_timestep_loss` call
(pre-shuffle row indices by byte-match, timestep column, advantages, the
xt noise draw) — the inner loop's cuda_generator sample/timestep
permutations never need replaying. At this config the effective grad accum
(2 x 4 timesteps = 8) equals the calls per outer step, so the optimizer
fires once per outer step and ALL step-0 losses are pre-update:
advantages exact (0.0) both outer steps (group-by-prompt,
(r-mean)/(std_unbiased=False+1e-4), rewards["avg"])
loss.s0k0..k7 BITWISE 0.0 (8/8) at pre-update weights
loss.s1k0..k7 max diff 1.17e-2, within the measured band
(main self-noise 4.48e-2 over 3 identical-seed
capture runs; band = 1.5x). Step-0 losses are
exactly repeatable on main across runs.
Root cause worth keeping (cost one bisection): the modular stack loads
every role through `maybe_load_fsdp_model` with
`MixedPrecisionPolicy(param_dtype=bf16, reduce_dtype=fp32)` over fp32
storage. Probing `p.dtype` shows only the fp32 STORAGE dtype; the FORWARD
runs on bf16-cast params (RMSNorm weights and scale_shift_table included —
confirmed by per-module hash probes: inputs bitwise, `to_q` diverging,
main's `norm_q` emitting bf16). AdamW (eps 1e-8) and the old-policy EMA
operate on the fp32 storage — at world 1 exactly the fp32-master /
bf16-compute chain (`_MasterOpt`) the finetune/DMD2 gates proved. An
NFTStep variant with plain AdamW over fp32 live params + autocast produced
2e-5..2e-2 step-0 errors; the restored master-chain NFTStep is bitwise.
Upstream findings (transformers 5.13, filed in PORT_NOTES): main's
PickScoreScorer crashes (`get_*_features` now returns a ModelOutput whose
`.pooler_output` holds the projected features; unwrap must be scoped to the
PickScore instance — `CLIPModel.forward` consumes the new contract);
RL validation fires at iteration 0 (`0 % every_steps == 0`); `_return_decay`
is a staticmethod (plain-function monkeypatches shift its args); the
VideoAlign runtime needs `accelerate>=1.1` and headless opencv on ARM.
## VideoAlign rewards — added 2026-07-23
Gate `anchor.videoalign-pr1476`: PASS. The reward runtime under
`fastvideo2/rl_rewards/VideoAlign/` is vendored BYTE-IDENTICAL from
PR #1476 (`maint/pr1476-runtime-compat` @518aeab0b) — sha256 of all 8 files
is asserted in `tests/test_videoalign.py`, and pre-commit excludes the
directory so formatters cannot drift it. The wrapper
(`fastvideo2/train/videoalign.py`) differs from the PR's
`rewards/videoalign.py` only by the vendored-root path and the inlined
(verbatim) `media_to_uint8_array`. Code identity => behavior identity; the
cluster gate then proves the stack end-to-end on the real
KwaiVGI/VideoReward checkpoint: MQ/VQ/TA on 3 fixed synthetic media,
scored twice — deterministic (exact), finite, goldens recorded
(`videoalign-pr1476/scores.json`).