Add WorldCompass Sana training artifacts
Browse files
rebuttal_data/worldcompass_sana/wc_sana_smoke_f242310_20260801e/train_log.log
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
2026-08-02 01:07:29 - [Sana] - INFO - [FSDP2] block=SanaVideoMSCamCtrlBlock, cp_size=1, reshard_after_forward=True, limit_all_gathers=False, backward_prefetch=backward_pre
|
| 2 |
+
2026-08-02 01:07:33 - [Sana] - INFO - Failed to load null embed from None....'NoneType' object has no attribute 'seek'. You can only torch.load from a file that is seekable. Please pre-load the data into a buffer like io.BytesIO and try to load from it instead.. Ignore the error during inference
|
| 3 |
+
2026-08-02 01:07:33 - [Sana] - INFO - use pe: True, pos embed type: casual_wan_rope, position embed interpolation: 1.0, base size: 22
|
| 4 |
+
2026-08-02 01:07:33 - [Sana] - INFO - attention type: BidirectionalGDNTriton; ffn type: CachedGLUMBConvTemp; self-attn head dim: 112; self-attn qk norm: True; cross-attn type: flash; cross-attn qk norm: True; autocast linear attn: False
|
| 5 |
+
2026-08-02 01:07:33 - [Sana] - INFO - Hybrid attention (softmax_every_n=4):
|
| 6 |
+
attn_type_list = ['BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton', 'BidirectionalGDNTriton']
|
| 7 |
+
camctrl_cls_list = ['CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedSoftmaxUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedSoftmaxUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedSoftmaxUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedSoftmaxUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedChunkCausalGDNUCPESinglePathLiteLA', 'CachedSoftmaxUCPESinglePathLiteLA']
|
| 8 |
+
2026-08-02 01:07:51 - [Sana] - INFO - Failed to load null embed from None....'NoneType' object has no attribute 'seek'. You can only torch.load from a file that is seekable. Please pre-load the data into a buffer like io.BytesIO and try to load from it instead.. Ignore the error during inference
|
| 9 |
+
2026-08-02 01:08:40 - [Sana] - INFO - Loaded /mnt/workspace/fly/reward_dmd_sanawm/models/SANA-WM_streaming/fb4ef609f5da47d5815e850a5ecec9a12f069b57/sana_dit/model.pt: model=SanaMSVideoCamCtrlStreaming_1600M_P1_D20, missing=1, unexpected=0
|
| 10 |
+
2026-08-02 01:08:43 - [Sana] - INFO - FSDP2: removing custom __call__ from SanaMSVideoCamCtrlStreaming
|
| 11 |
+
2026-08-02 01:08:43 - [Sana] - INFO - FSDP2: removing custom __call__ from Sana
|
| 12 |
+
2026-08-02 01:08:43 - [Sana] - INFO - FSDP2: removing custom __call__ from SanaMSVideoCamCtrlStreaming
|
| 13 |
+
2026-08-02 01:08:43 - [Sana] - INFO - FSDP2: removing custom __call__ from Sana
|
| 14 |
+
2026-08-02 01:08:49 - [Sana] - INFO - Constructing dataset ThreeSourceF19ManifestDataset...
|
| 15 |
+
2026-08-02 01:09:08 - [Sana] - INFO - Dataset ThreeSourceF19ManifestDataset constructed: time: 19.18 s, length (use/ori): 112573/112573
|
| 16 |
+
2026-08-02 01:09:16 - [Sana] - INFO - WorldCompass released-code SANA port: profile=released_code_1588e133_sana_native_smoke G=6 top=3 bottom=3 target_optimizer_steps=3 target_outer_steps=1 N=6 schedule=[1000,960,889,727,0] full_dit=true optimizer=Muon(lr=1e-5,wd=1e-4,momentum=0.95) reward_decode=stage1_causal_vae_before_frozen_refiner evaluation_checkpoint=model_ema.pt
|
| 17 |
+
2026-08-02 01:09:16 - [Sana] - INFO - mode=worldcompass start_optimizer_step=0 target_optimizer_steps=3 start_outer_step=0 target_outer_steps=1 optimizer_updates_per_outer_step=3 save_every_optimizer_steps=3 save_enabled=True cp_size=1 dp_size=2 reward_domain=stage1_causal_vae_before_frozen_refiner
|