sana_video_pretrained

Sana-Video v2 5B video-only backbone, SFT checkpoint at 2,076k steps, EMA weights -- the file sana_v2_5b_sft_2076k_ema.pth that the RoboDojo ARX-X5 abs_sanavideo policy SFT lines start from (model.load_from; runs R14/R15 = sft_robodojo_arxx5_unified_jointonly_f25_320px_4node_abs_sanavideo*, and the pending ..._alw1_v2 relaunch). Uploaded 2026-09-14 from the SFT line's model vault, bytes unchanged.

File

file	bytes	sha256
sana_v2_5b_sft_2076k_ema.pth	17871918206	5bcb59d42b395a80e2e1ebcd4b684271ba10ccf06d64637b64165dac55057c04

• One PyTorch .pth; the outer key is state_dict_ema (not state_dict). Verified at upload (2026-09-14, memory-mapped load): 797 tensors, 4,468,248,160 parameters (796 tensors float32, 1 bfloat16), the same 797 keys / 4.47 B the vault recorded on 2026-07-01; size and sha256 above taken at upload.
• Architecture it was trained as: the Sana-Video QwenNext SubAttnRes V2 SelfFlow world model, 5B, patch 1, depth 36 (SanaRWMVideoQwenNextSubAttnResV2SelfFlowWorldModel*_5B_P1_D36 family): GatedDeltaNet linear attention with GatedSoftmaxAttention every fourth block (softmax_ratio 0.25, head dims 128 / 256), SwiGLU (mlp_ratio 4), wan_rope, qk_norm + cross_norm. It carries the three attn_res projections (attn / mlp / final_proj) and no time projection, so load it with use_time_conditioning: false and use_attn_res: true.
• Video-only: it has no robot modules. The eight policy tensors state_embed.proj.{weight,bias}, action_embed.proj.{weight,bias}, action_head.linear.{weight,bias}, action_head.scale_shift_table and plucker_embed.weight are absent. The file does carry a pos_embed for its own training resolution; the Sana-RWM loader force-drops it and re-creates it for the target resolution, which is the expected Missing keys: ['pos_embed'] warning. Loading it into the policy class through the SFT trainer needs the load-guard waiver SANA_SFT_ALLOWED_MISSING_KEYS=state_embed.proj.weight,state_embed.proj.bias,action_embed.proj.weight,action_embed.proj.bias,action_head.linear.weight,action_head.linear.bias,action_head.scale_shift_table,plucker_embed.weight; those modules then train from their zero init (see the abs_sanavideo launcher).
• Provenance: staged 2026-07-01 into $STORAGE/models/Sana/ from users/haozhu/projects/Sana/output/pretrained_models/sana_v2_5b_sft_2076k_ema.pth with a byte-size check; the vault README records the 797-key / 4.47 B-parameter verification.

Usage (Sana-RWM repo)

model:
  model: SanaRWMVideoQwenNextSubAttnResV2SelfFlowWorldModelCameraConditionMultiViewPolicy_5B_P1_D36
  load_from: <local path>/sana_v2_5b_sft_2076k_ema.pth   # outer key state_dict_ema
  use_time_conditioning: false
  use_attn_res: true

Descendants on the Hub: logits/sana_rwm_policy_robodojo_320px_jointonly_abs_sanavideo_tail_wd1e-2_epoch8_step55000 (R15, step 55000), which carries all robot modules and loads with only the pos_embed warning.