Upload folder using huggingface_hub

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02b88b20fa7a404ff2ebcb1dae8300a4ca50433c19b3d02101b8f5ff86108ece
+size 10798884178

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/checkpoint_.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02b88b20fa7a404ff2ebcb1dae8300a4ca50433c19b3d02101b8f5ff86108ece
+size 10798884178

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/config.yaml

@@ -0,0 +1,271 @@
+run_dir: exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs
+seed: 0
+allow_tf32: true
+timeout: null
+resolution: 256
+amp: bf16
+cfg_scale: 1.0
+evaluate_split: test
+eval_dir_name: null
+num_save_images: 64
+save_all_images: false
+save_image_format: jpg
+save_images_at_all_procs: false
+save_latent_samples: false
+latent_samples_dir: null
+evaluate_dataset: sample_class
+sample_class:
+  name: SampleClass
+  batch_size: 128
+  n_worker: 8
+  drop_last: false
+  seed: 0
+  shuffle: false
+  num_classes: 1000
+  num_samples: 50000
+autoencoder:
+  num_settings: 1
+  name: dc-ae-f32c32-in-1.0-256px
+  scaling_factor: 0.3285
+  latent_channels: null
+autoencoder_dtype: fp32
+eval_autoencoder_setting_list: null
+model: fp8coat_dit
+dit:
+  name: DiT
+  in_channels: 32
+  input_size: 8
+  cfg_channels: null
+  pretrained_path: null
+  pretrained_source: dc-ae
+  train_scheduler: SiTSampler
+  eval_scheduler: ODE_heun2
+  num_inference_steps: 30
+  flow_shift: 3.0
+  reverse_time: false
+  use_cads: false
+  cads_noise_scale: 0.1
+  cads_mixing_factor: 1.0
+  cads_tau_min: 0.2
+  cads_tau_max: 0.9
+  use_guidance_interval: false
+  guidance_t_min: 0.2
+  guidance_t_max: 0.8
+  count_nfe: false
+  patch_size: 1
+  hidden_size: 1152
+  depth: 28
+  num_heads: 16
+  mlp_ratio: 4.0
+  post_norm: false
+  class_dropout_prob: 0.1
+  num_classes: 1000
+  learn_sigma: false
+  unconditional: false
+  use_checkpoint: true
+  adaptive_channel: false
+  adaptive_channel_share_weights: true
+  only_load_backbone: false
+  freeze_backbone: false
+uvit:
+  name: UViT
+  in_channels: 4
+  input_size: 32
+  cfg_channels: null
+  pretrained_path: null
+  pretrained_source: dc-ae
+  train_scheduler: DPM_Solver
+  eval_scheduler: DPM_Solver
+  num_inference_steps: 30
+  flow_shift: 3.0
+  reverse_time: false
+  use_cads: false
+  cads_noise_scale: 0.1
+  cads_mixing_factor: 1.0
+  cads_tau_min: 0.2
+  cads_tau_max: 0.9
+  use_guidance_interval: false
+  guidance_t_min: 0.2
+  guidance_t_max: 0.8
+  count_nfe: false
+  patch_size: 2
+  hidden_size: 1152
+  depth: 28
+  num_heads: 16
+  mlp_ratio: 4.0
+  mlp_time_embed: false
+  qkv_bias: false
+  act_layer: gelu
+  use_checkpoint: true
+  class_dropout_prob: 0.1
+  num_classes: 1000
+  attn_mode: null
+sana_cls:
+  name: SanaCls
+  in_channels: 4
+  input_size: 32
+  cfg_channels: null
+  pretrained_path: null
+  pretrained_source: dc-ae
+  train_scheduler: SanaScheduler
+  eval_scheduler: SanaScheduler
+  num_inference_steps: 250
+  flow_shift: 3.0
+  reverse_time: false
+  use_cads: false
+  cads_noise_scale: 0.1
+  cads_mixing_factor: 1.0
+  cads_tau_min: 0.2
+  cads_tau_max: 0.9
+  use_guidance_interval: false
+  guidance_t_min: 0.2
+  guidance_t_max: 0.8
+  count_nfe: false
+  patch_size: 2
+  hidden_size: 1152
+  depth: 28
+  num_heads: 16
+  mlp_ratio: 4.0
+  post_norm: false
+  class_dropout_prob: 0.1
+  num_classes: 1000
+  unconditional: false
+  use_checkpoint: true
+  only_load_backbone: false
+  freeze_backbone: false
+  learn_sigma: false
+usana_cls:
+  name: USanaCls
+  in_channels: 4
+  input_size: 32
+  cfg_channels: null
+  pretrained_path: null
+  pretrained_source: dc-ae
+  train_scheduler: DPM_Solver
+  eval_scheduler: DPM_Solver
+  num_inference_steps: 30
+  flow_shift: 3.0
+  reverse_time: false
+  use_cads: false
+  cads_noise_scale: 0.1
+  cads_mixing_factor: 1.0
+  cads_tau_min: 0.2
+  cads_tau_max: 0.9
+  use_guidance_interval: false
+  guidance_t_min: 0.2
+  guidance_t_max: 0.8
+  count_nfe: false
+  patch_size: 2
+  hidden_size: 1152
+  depth: 28
+  num_heads: 16
+  mlp_ratio: 4.0
+  mlp_time_embed: false
+  qkv_bias: false
+  act_layer: gelu
+  use_checkpoint: true
+  class_dropout_prob: 0.1
+  num_classes: 1000
+  num_training_steps: 1000
+fp8:
+  name: FP8DiT
+fp8coat:
+  name: FP8COATDiT
+  qchoice: linear
+  symm: true
+  row_blocksize: -1
+  col_blocksize: -1
+  linear_row_blocksize: 1
+  linear_col_blocksize: 16
+  min_blockunit_row: -1
+  min_blockunit_col: -1
+  fabit: NVE2M1
+  fwbit: NVE2M1
+  babit: NVE2M1
+  bwbit: NVE2M1
+  bobit: NVE2M1
+  epsilon: 1.0e-08
+compute_fid: true
+fid:
+  save_path: null
+  ref_path: assets/data/fid/imagenet_train_256.npz
+  precision_recall_ref_path: assets/data/precision_recall/VIRTUAL_imagenet256.npy
+compute_inception_score: true
+inception_score: {}
+compute_cmmd: true
+cmmd:
+  save_path: null
+  ref_path: assets/data/cmmd/VIRTUAL_imagenet256.npy
+verbose: false
+train_dataset: latent_imagenet
+latent_imagenet:
+  name: LatentImageNet
+  batch_size: 128
+  n_worker: 8
+  drop_last: true
+  seed: 0
+  shuffle: true
+  data_dir: assets/data/latent/dc_ae_f32c32_in_1.0_256px/imagenet_256
+latent_mjhq:
+  name: LatentMJHQ
+  batch_size: 32
+  n_worker: 8
+  drop_last: true
+  seed: 0
+  shuffle: true
+  data_dir: assets/data/latent/dc_ae_f32c32/mjhq_1024
+latent_ffhq:
+  name: LatentFFHQ
+  batch_size: 32
+  n_worker: 8
+  drop_last: true
+  seed: 0
+  shuffle: true
+  data_dir: assets/data/latent/dc_ae_f32c32/ffhq_1024
+latent_mapillary_vistas:
+  name: LatentMapillaryVistas
+  batch_size: 32
+  n_worker: 8
+  drop_last: true
+  seed: 0
+  shuffle: true
+  data_dir: assets/data/latent/dc_ae_f32c32/mapillary_vistas_2048
+latent_multiple_channel_imagenet:
+  name: LatentMultipleChannelImageNet
+  batch_size: 32
+  n_worker: 8
+  drop_last: true
+  seed: 0
+  shuffle: true
+  dataset_sample_ratio: null
+  num_channels_list: null
+  data_dirs:
+  - assets/data/latent/dc_ae_f32c32/imagenet_512
+resume: true
+resume_path: null
+resume_schedule: true
+num_epochs: null
+max_steps: 500000
+clip_grad: null
+num_store_images: 64
+save_checkpoint_steps: 1000
+eval_steps: 20000
+save_eval_checkpoint_steps: 100000
+optimizer:
+  name: adamw
+  lr: 0.0001
+  warmup_lr: 0.0
+  weight_decay: 0.0
+  no_wd_keys: []
+  betas:
+  - 0.9
+  - 0.999
+lr_scheduler:
+  name: constant
+  warmup_steps: 1000
+log: true
+wandb_entity: han2024
+wandb_project: dc_ae_diffusion
+ema_decay: 0.9999
+ema_warmup_steps: 2000
+eval_ema: true

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/eval_results.csv

@@ -0,0 +1,77 @@
+,fid,precision,recall,inception_score_mean,inception_score_std,cmmd
+step_100000_autoencoder_setting_0_cfg_1.0,27.82081598705525,0.580020010471344,0.6139999628067017,46.420777211047714,1.0247874181815988,0.6939172744750977
+step_10000_autoencoder_setting_0_cfg_1.0,78.50917779509035,0.3088999986648559,0.4490000009536743,15.570617267902197,0.3332917058620098,1.505136489868164
+step_105000_autoencoder_setting_0_cfg_1.0,27.108763303815124,0.58433997631073,0.6107999682426453,47.7493937125107,1.2485355820824466,0.6833076477050781
+step_110000_autoencoder_setting_0_cfg_1.0,26.572085273128607,0.5875200033187866,0.6103000044822693,48.91806079101078,1.320462264402051,0.6742477416992188
+step_115000_autoencoder_setting_0_cfg_1.0,26.04053809255868,0.5945199728012085,0.6122999787330627,50.06943499322904,1.271234794980543,0.6644725799560547
+step_120000_autoencoder_setting_0_cfg_1.0,25.48026900045761,0.59961998462677,0.6194999814033508,51.20050723062443,1.5078486704815637,0.6552934646606445
+step_125000_autoencoder_setting_0_cfg_1.0,25.05591886633323,0.6003400087356567,0.6232999563217163,52.10802955307636,1.581453473677152,0.649571418762207
+step_130000_autoencoder_setting_0_cfg_1.0,24.5844863096371,0.6046000123023987,0.6189000010490417,52.79456366277475,1.507362587027362,0.6395578384399414
+step_135000_autoencoder_setting_0_cfg_1.0,24.18458923313409,0.6058200001716614,0.620199978351593,53.62105523874786,1.4409782134398583,0.6331205368041992
+step_140000_autoencoder_setting_0_cfg_1.0,23.866081545629527,0.6055200099945068,0.6236000061035156,54.368579244722106,1.4469492288369885,0.6299018859863281
+step_145000_autoencoder_setting_0_cfg_1.0,23.26773026378873,0.6069599986076355,0.6215999722480774,55.51033309373391,1.387295168322456,0.6201267242431641
+step_150000_autoencoder_setting_0_cfg_1.0,22.445572104073108,0.6144399642944336,0.6243999600410461,57.17583559018384,1.5491262770197771,0.6072521209716797
+step_15000_autoencoder_setting_0_cfg_1.0,68.12664722088442,0.3365599811077118,0.5128999948501587,17.673999189426127,0.1972535523911879,1.3135671615600586
+step_155000_autoencoder_setting_0_cfg_1.0,21.6969907496823,0.6131199598312378,0.6244999766349792,58.32637361785574,1.5920632516988755,0.5936622619628906
+step_160000_autoencoder_setting_0_cfg_1.0,21.259075753956097,0.6131199598312378,0.6218000054359436,59.29787913474204,1.77534055094449,0.5881786346435547
+step_165000_autoencoder_setting_0_cfg_1.0,21.028425986100444,0.6096999645233154,0.6220999956130981,60.28639353096686,1.6039359868962844,0.5855560302734375
+step_170000_autoencoder_setting_0_cfg_1.0,20.78157261474053,0.6087200045585632,0.623699963092804,61.02744304547458,1.5003141917637632,0.5818605422973633
+step_175000_autoencoder_setting_0_cfg_1.0,20.412292669880458,0.613319993019104,0.6223999857902527,61.96596631327077,1.7293084120916875,0.577092170715332
+step_180000_autoencoder_setting_0_cfg_1.0,20.27882333039713,0.612280011177063,0.6218999624252319,62.114020147042574,1.8010197629079103,0.5730390548706055
+step_185000_autoencoder_setting_0_cfg_1.0,19.89168937788316,0.615339994430542,0.6232999563217163,63.27708418568627,1.451279607657161,0.5658864974975586
+step_190000_autoencoder_setting_0_cfg_1.0,19.419406988353785,0.6139999628067017,0.6232999563217163,64.56478640219703,1.471862549083792,0.5564689636230469
+step_195000_autoencoder_setting_0_cfg_1.0,18.97561403631505,0.6157199740409851,0.6226999759674072,65.76082161550052,1.3146110081612372,0.5470514297485352
+step_200000_autoencoder_setting_0_cfg_1.0,19.026450618121032,0.6161999702453613,0.6227999925613403,66.43766544317899,1.4349545277425972,0.5500316619873047
+step_20000_autoencoder_setting_0_cfg_1.0,69.80893117916492,0.3557800054550171,0.5060999989509583,17.63329056852647,0.2319949836815072,1.3469457626342771
+step_205000_autoencoder_setting_0_cfg_1.0,18.5947097582615,0.6192600131034851,0.6297999620437622,67.63410420493474,1.536225845054557,0.5457401275634766
+step_210000_autoencoder_setting_0_cfg_1.0,18.233572638968496,0.6248599886894226,0.623199999332428,68.6561121314874,1.5783403374324958,0.538945198059082
+step_215000_autoencoder_setting_0_cfg_1.0,17.797767795291293,0.6285399794578552,0.6223999857902527,70.42323234086862,1.5883405198437344,0.5366802215576172
+step_220000_autoencoder_setting_0_cfg_1.0,17.74635629321216,0.6273399591445923,0.6311999559402466,70.2145545808094,1.5785695221108964,0.5356073379516602
+step_225000_autoencoder_setting_0_cfg_1.0,17.623212029857086,0.6253600120544434,0.6233999729156494,70.10889264234,1.402678692708326,0.5316734313964844
+step_230000_autoencoder_setting_0_cfg_1.0,17.56820183523979,0.6272199749946594,0.6281999945640564,70.24948808293567,1.3719698368922175,0.5235671997070312
+step_235000_autoencoder_setting_0_cfg_1.0,17.140324617321994,0.6325399875640869,0.6326999664306641,71.8477116161524,1.4453070017749934,0.5209445953369141
+step_240000_autoencoder_setting_0_cfg_1.0,16.875031621310654,0.6319199800491333,0.6304000020027161,72.40102219505278,1.572925419822807,0.5216598510742188
+step_245000_autoencoder_setting_0_cfg_1.0,16.603363011407623,0.6376399993896484,0.6248999834060669,73.19131748156653,1.634987063086786,0.5117654800415039
+step_250000_autoencoder_setting_0_cfg_1.0,16.0659137008808,0.644540011882782,0.6232999563217163,75.05675046856344,1.4611675195953364,0.4990100860595703
+step_25000_autoencoder_setting_0_cfg_1.0,69.20439116845444,0.3652600049972534,0.5146999955177307,17.835205670638455,0.1921896922272648,1.3399124145507812
+step_255000_autoencoder_setting_0_cfg_1.0,16.18191134397,0.643779993057251,0.6266999840736389,74.87994735992554,1.5313646958234384,0.5011558532714844
+step_260000_autoencoder_setting_0_cfg_1.0,16.307349111818837,0.6428399682044983,0.6293999552726746,74.55537296593175,1.7462488184454217,0.5011558532714844
+step_265000_autoencoder_setting_0_cfg_1.0,16.351515906573184,0.6398400068283081,0.6279999613761902,73.60775233144288,2.1094810857369937,0.4945993423461914
+step_270000_autoencoder_setting_0_cfg_1.0,15.720775344146716,0.6462000012397766,0.6182999610900879,75.93336593862954,1.9112681424085896,0.4835128784179687
+step_275000_autoencoder_setting_0_cfg_1.0,15.563540840291353,0.6510800123214722,0.6226999759674072,76.73444862274553,2.1060209810007886,0.4808902740478515
+step_280000_autoencoder_setting_0_cfg_1.0,15.702402057568406,0.6554399728775024,0.6204000115394592,76.80205589593925,1.7598216362881405,0.4860162734985351
+step_285000_autoencoder_setting_0_cfg_1.0,15.735781046575084,0.6553199887275696,0.6211000084877014,76.72250513048573,1.9211665275192824,0.4916191101074219
+step_290000_autoencoder_setting_0_cfg_1.0,15.69120289879271,0.6527199745178223,0.6200000047683716,77.60842637374755,2.02590143858928,0.4960298538208008
+step_295000_autoencoder_setting_0_cfg_1.0,15.545950959937784,0.6571399569511414,0.626800000667572,78.00591850878438,1.96422643425752,0.4917383193969726
+step_300000_autoencoder_setting_0_cfg_1.0,14.454307153848449,0.6623199582099915,0.616599977016449,82.6760750956561,2.181097692838078,0.4781484603881836
+step_30000_autoencoder_setting_0_cfg_1.0,71.33487248804539,0.3709399998188019,0.5120999813079834,17.564975662944967,0.2190062162059579,1.3840198516845703
+step_305000_autoencoder_setting_0_cfg_1.0,14.727343043843916,0.660319983959198,0.618399977684021,81.5701813105076,1.9043765039281204,0.4681348800659179
+step_310000_autoencoder_setting_0_cfg_1.0,15.338016119750534,0.6581599712371826,0.6150999665260315,78.50163318136279,1.3744470633165384,0.4767179489135742
+step_315000_autoencoder_setting_0_cfg_1.0,15.670858997474229,0.6570199728012085,0.6189999580383301,77.5421772667251,1.3151557061430104,0.4895925521850586
+step_320000_autoencoder_setting_0_cfg_1.0,15.351098933787853,0.6589800119400024,0.6159999966621399,78.85159621337543,1.8036081361014704,0.4880428314208984
+step_325000_autoencoder_setting_0_cfg_1.0,16.137012142700883,0.6609199643135071,0.6116999983787537,77.20548121816479,1.4635409927233674,0.506281852722168
+step_330000_autoencoder_setting_0_cfg_1.0,16.935113423647522,0.6579399704933167,0.6132999658584595,75.14944105179288,1.6728809851158315,0.5131959915161133
+step_335000_autoencoder_setting_0_cfg_1.0,16.642815521207183,0.6565999984741211,0.6182000041007996,74.89506149880103,1.7185640977473733,0.5041360855102539
+step_340000_autoencoder_setting_0_cfg_1.0,16.329822569891235,0.6578199863433838,0.6171000003814697,76.13033508421296,1.453925495743798,0.5016326904296875
+step_35000_autoencoder_setting_0_cfg_1.0,68.4526505825479,0.3855199813842773,0.523099958896637,18.50040439034826,0.2102358995042317,1.3387203216552734
+step_360000_autoencoder_setting_0_cfg_1.0,18.51766157845032,0.641539990901947,0.6207000017166138,68.6112735921217,1.5894359804544205,0.5263090133666992
+step_380000_autoencoder_setting_0_cfg_1.0,19.047118711635164,0.6372199654579163,0.6157000064849854,67.32176540925762,1.6363590511183173,0.5387067794799805
+step_400000_autoencoder_setting_0_cfg_1.0,20.46042710661544,0.6325799822807312,0.6110000014305115,64.22305233265526,1.8580822376880224,0.5658864974975586
+step_40000_autoencoder_setting_0_cfg_1.0,59.98117677660253,0.4123599827289581,0.5669999718666077,20.98464259395777,0.3452826016560876,1.1975765228271484
+step_420000_autoencoder_setting_0_cfg_1.0,20.048889679310776,0.6316800117492676,0.611299991607666,64.82227407909255,1.662291796280963,0.5481243133544922
+step_440000_autoencoder_setting_0_cfg_1.0,21.6126295301454,0.6124599575996399,0.6274999976158142,62.33693934455978,1.8068080125059365,0.582575798034668
+step_45000_autoencoder_setting_0_cfg_1.0,52.35728384337608,0.4363399744033813,0.5849999785423279,23.625650383680828,0.4867593086153808,1.0684728622436523
+step_460000_autoencoder_setting_0_cfg_1.0,24.502855264928712,0.5979999899864197,0.6233999729156494,57.90086638570746,1.6518872144366603,0.6576776504516602
+step_480000_autoencoder_setting_0_cfg_1.0,22.52686733943784,0.6110399961471558,0.6092999577522278,62.57142330892974,1.7219275950198052,0.6278753280639648
+step_500000_autoencoder_setting_0_cfg_1.0,23.343317025502756,0.6067399978637695,0.6193000078201294,60.4235201455905,1.2737023659678393,0.6389617919921875
+step_50000_autoencoder_setting_0_cfg_1.0,49.921173116398734,0.4559399783611297,0.5971999764442444,25.213578656547632,0.4281732070550088,1.0330677032470703
+step_5000_autoencoder_setting_0_cfg_1.0,90.67715449183368,0.2694199979305267,0.3879999816417694,13.429481826679131,0.210033970069514,1.7157793045043943
+step_55000_autoencoder_setting_0_cfg_1.0,47.73685522428235,0.4755999743938446,0.5899999737739563,26.89910198011636,0.4931733345923623,1.0018348693847656
+step_60000_autoencoder_setting_0_cfg_1.0,44.80032758845715,0.4915199875831604,0.5986999869346619,28.861349883325506,0.6176201506732034,0.9595155715942384
+step_65000_autoencoder_setting_0_cfg_1.0,41.53898757569908,0.5051400065422058,0.6049000024795532,31.154324157233617,0.6160978188506026,0.9080171585083008
+step_70000_autoencoder_setting_0_cfg_1.0,37.99431155890278,0.5188400149345398,0.6146000027656555,33.93576625895251,0.5479828376716884,0.8559226989746094
+step_75000_autoencoder_setting_0_cfg_1.0,35.3211073866222,0.5319399833679199,0.6132999658584595,36.31162383244789,0.8240562594862988,0.8103847503662109
+step_80000_autoencoder_setting_0_cfg_1.0,33.006981752393926,0.5404999852180481,0.6175000071525574,38.86142600259436,0.8896361831768164,0.774383544921875
+step_85000_autoencoder_setting_0_cfg_1.0,30.9545240118008,0.5546599626541138,0.6133999824523926,41.26951645107668,0.9114569673854772,0.7370710372924805
+step_90000_autoencoder_setting_0_cfg_1.0,29.674156243314485,0.5648999810218811,0.6136999726295471,43.06725174936357,0.909306073199211,0.7197856903076172
+step_95000_autoencoder_setting_0_cfg_1.0,28.471735321998835,0.572659969329834,0.6085000038146973,44.97528607860245,0.9621155062307708,0.7002353668212891

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/model.txt

@@ -0,0 +1,56 @@
+FP8COATDiT(
+  (x_embedder): PatchEmbed(
+    (proj): Conv2d(32, 1152, kernel_size=(1, 1), stride=(1, 1))
+    (norm): Identity()
+  )
+  (t_embedder): TimestepEmbedder(
+    (mlp): Sequential(
+      (0): Linear(in_features=256, out_features=1152, bias=True)
+      (1): SiLU()
+      (2): Linear(in_features=1152, out_features=1152, bias=True)
+    )
+  )
+  (y_embedder): LabelEmbedder(
+    (embedding_table): Embedding(1001, 1152)
+  )
+  (blocks): ModuleList(
+    (0-27): 28 x FP8COATDiTBlock(
+      (norm1): LayerNorm((1152,), eps=1e-06, elementwise_affine=False)
+      (attn): FP8COATAttention(
+        (qkv): QAct_Linear(in_features=1152, out_features=3456, bias=True)
+        (q_norm): Identity()
+        (k_norm): Identity()
+        (attn_drop): Dropout(p=0.0, inplace=False)
+        (proj): QAct_Linear(in_features=1152, out_features=1152, bias=True)
+        (proj_drop): Dropout(p=0.0, inplace=False)
+        (qact_q_quantize): QAct_Quantize()
+        (qact_k_quantize): QAct_Quantize()
+      )
+      (norm2): LayerNorm((1152,), eps=1e-06, elementwise_affine=False)
+      (mlp): FP8COATMlp(
+        (fc1): QAct_Linear(in_features=1152, out_features=4608, bias=True)
+        (act): GELU(approximate='tanh')
+        (drop1): Dropout(p=0, inplace=False)
+        (norm): Identity()
+        (fc2): QAct_Linear(in_features=4608, out_features=1152, bias=True)
+        (drop2): Dropout(p=0, inplace=False)
+        (qact_quantize_gelu): QAct_Quantize()
+        (qact_quantize_norm): QAct_Quantize()
+      )
+      (adaLN_modulation): Sequential(
+        (0): SiLU()
+        (1): Linear(in_features=1152, out_features=6912, bias=True)
+      )
+      (qact_quantize_norm1): QAct_Quantize()
+      (qact_quantize_norm2): QAct_Quantize()
+    )
+  )
+  (final_layer): FinalLayer(
+    (norm_final): LayerNorm((1152,), eps=1e-06, elementwise_affine=False)
+    (linear): Linear(in_features=1152, out_features=32, bias=True)
+    (adaLN_modulation): Sequential(
+      (0): SiLU()
+      (1): Linear(in_features=1152, out_features=2304, bias=True)
+    )
+  )
+)

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/slurm.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+#SBATCH -A nvr_elm_llm                 #account
+#SBATCH -p grizzly,polar,polar3,polar4 #partition
+#SBATCH -t 04:00:00                    #wall time limit, hr:min:sec
+#SBATCH -N 1                           #number of nodes
+#SBATCH -J train_diffusion_dit         #job name
+#SBATCH --array=1-30%1
+#SBATCH --output=exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/slurm_out/%A_%a.out
+#SBATCH --gpus-per-node 8
+#SBATCH --exclusive
+
+export LOGLEVEL=INFO
+export PATH="/home/hcai/workspace/anaconda3/envs/efficientvit/bin:$PATH"
+
+cd /lustre/fs12/portfolios/nvr/users/hcai/workspace/code/fp8-dit
+
+read -r -d '' cmd <<EOF
+torchrun --nnodes=1 --nproc_per_node=8 \
+-m efficientvit.diffusioncore.trainer yaml=exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/config.yaml
+EOF
+
+srun bash -c "${cmd}"

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/step_100000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46e93e161bc858fecdac014ca25bd5c06721b2b561375aa3a3dcc8c1f990a349
+size 10798884178

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/step_200000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ef4dbcf4e31575f3e05c8413b7ab0b60e66680d1b8e48b446ae3fa9fc2a93e0
+size 10798884178

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/step_300000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:141dc8d80d0907cc16f2d3786fa475ce7a823d263208636e92ba00eeb05c5d2c
+size 10798884178

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/step_400000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c4c0ae8c74d2d361bd0e4a4a1570ee2cd0d8688a26a07067e5ccbde17cdd431
+size 10798884178

exp_hcai/latent_diffusion/imagenet_256/dc_ae_f32c32_in_1.0_256px/dit_xl_1/SiTSampler_ODE_heun2_30/nvfp4_e2m1_cs/step_500000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11fe39b36b6414f011fa2cabf7e212f1e9f0cbb5ae631a7896b11a0f20cbade0
+size 10798884178