diff --git a/examples/distill/SFHunyuan1.5-T2V/distill_dmd.slurm b/examples/distill/SFHunyuan1.5-T2V/distill_dmd.slurm
index d764bba31..8780f00ec 100644
--- a/examples/distill/SFHunyuan1.5-T2V/distill_dmd.slurm
+++ b/examples/distill/SFHunyuan1.5-T2V/distill_dmd.slurm
@@ -47,7 +47,7 @@ FAKE_SCORE_MODEL_PATH="hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2
 # DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
 # DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
 DATA_DIR="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/vidprom_16k_text_embed"
-DATA_DIR_2="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/fv-ode-preprocessing-16k-hy15-121"
+DATA_DIR_2="/mnt/weka/home/hao.zhang/wl/sharefs/wl/release/fv-ode-preprocessing-16k-hy15-distilled-121"
 # DATA_DIR=data/crush-smol_processed_t2v/combined_parquet_dataset
 # DATA_DIR="/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/"
 # VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
@@ -64,16 +64,16 @@ training_args=(
   --train_batch_size 1
   --train_sp_batch_size 1
   --gradient_accumulation_steps 1
-  --num_latent_t 31
+  --num_latent_t 21
   --num_height 480
   --num_width 848
   --enable_gradient_checkpointing_type "full"
   --simulate_generator_forward
   --log_visualization
-  --num_frames 121
+  --num_frames 81
   --num_frame_per_block 3  # Frame generation block size for self-forcing
   # --resume-from-checkpoint "/mnt/weka/home/hao.zhang/wei/SFhy1.5_distill_1333/checkpoint-300"
-  --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_ode_init_1333_new/checkpoint-2400/transformer/diffusion_pytorch_model.safetensors
+  --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wei/hy15_distilled_ode_init_3333/checkpoint-2400/transformer/diffusion_pytorch_model.safetensors
 )
 
 parallel_args=(
@@ -101,14 +101,14 @@ validation_args=(
   --log_validation
   --validation_dataset_file "$VALIDATION_DATASET_FILE"
   --validation_steps 50
-  --validation_sampling_steps "4"
+  --validation_sampling_steps "8"
   --validation_guidance_scale "1.0" # not used for dmd inference
   --text-encoder-cpu-offload
   # --vae_cpu_offload True
 )
 
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 1e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 100
   --weight_only_checkpointing_steps 500
@@ -122,7 +122,7 @@ miscellaneous_args=(
   --checkpoints_total_limit 3
   --training_cfg_rate 0.0
   --dit_precision "fp32"
-  --flow_shift 5
+  --flow_shift 7
   --seed 1000
   --use_ema True
   --ema_decay 0.99
@@ -131,13 +131,13 @@ miscellaneous_args=(
 
 dmd_args=(
   --warp_denoising_step
-  # --dmd_denoising_steps '1000,875,750,625,500,375,250,125'
-  --dmd_denoising_steps '1000,760,520,280'
+  --dmd_denoising_steps '1000,875,750,625,500,375,250,125'
+  # --dmd_denoising_steps '1000,760,520,280'
   --min_timestep_ratio 0.02
   --max_timestep_ratio 0.98
   --dfake_gen_update_ratio 5
   --real_score_guidance_scale 3.5
-  --fake_score_learning_rate 5e-6
+  --fake_score_learning_rate 2e-7
   --fake_score_betas '0.0,0.999'
 )
 
diff --git a/fastvideo/configs/pipelines/hunyuan15.py b/fastvideo/configs/pipelines/hunyuan15.py
index 312986438..2d26057d0 100644
--- a/fastvideo/configs/pipelines/hunyuan15.py
+++ b/fastvideo/configs/pipelines/hunyuan15.py
@@ -143,9 +143,10 @@ class Hunyuan15T2V720PConfig(Hunyuan15T2V480PConfig):
 
 @dataclass
 class SelfForcingHunyuan15T2V480PConfig(Hunyuan15T2V480PConfig):
+    flow_shift: int = 7
     is_causal: bool = True
-    dmd_denoising_steps: list[int] | None = field(
-        default_factory=lambda: [1000, 760, 520, 280])
     # dmd_denoising_steps: list[int] | None = field(
-    #     default_factory=lambda: [1000, 875, 750, 625, 500, 375, 250, 125])
+    #     default_factory=lambda: [1000, 760, 520, 280])
+    dmd_denoising_steps: list[int] | None = field(
+        default_factory=lambda: [1000, 875, 750, 625, 500, 375, 250, 125])
     warp_denoising_step: bool = True
\ No newline at end of file
diff --git a/fastvideo/pipelines/stages/causal_denoising.py b/fastvideo/pipelines/stages/causal_denoising.py
index 7886e4627..d00a5a374 100644
--- a/fastvideo/pipelines/stages/causal_denoising.py
+++ b/fastvideo/pipelines/stages/causal_denoising.py
@@ -178,7 +178,7 @@ class CausalDMDDenosingStage(DenoisingStage):
 
         # For now hardcode the first block to be 1 frame assuming the model is Wan2.2-MoE
         # if boundary_timestep is not None:
-        block_sizes[0] = 1
+        # block_sizes[0] = 1
 
         # Initialize txt kv cache
         with torch.autocast(device_type="cuda",
diff --git a/fastvideo/training/distillation_pipeline.py b/fastvideo/training/distillation_pipeline.py
index 8a9c255c9..01192d66d 100644
--- a/fastvideo/training/distillation_pipeline.py
+++ b/fastvideo/training/distillation_pipeline.py
@@ -147,7 +147,7 @@ class DistillationPipeline(TrainingPipeline):
         self.vae.requires_grad_(False)
 
         self.timestep_shift = self.training_args.pipeline_config.flow_shift
-        assert self.timestep_shift == 5.0, "flow_shift must be 5.0"
+        assert self.timestep_shift == 7.0, "flow_shift must be 7.0"
         # self.noise_scheduler.set_timesteps(sigmas=list(np.linspace(1.0, 0.0, 50 + 1)[:-1]), device=get_local_torch_device())
         self.noise_scheduler.set_timesteps(num_inference_steps=1000, extra_one_step=True, device=get_local_torch_device())
 
diff --git a/fastvideo/training/self_forcing_distillation_pipeline.py b/fastvideo/training/self_forcing_distillation_pipeline.py
index 6a9d835dc..87d238f7f 100644
--- a/fastvideo/training/self_forcing_distillation_pipeline.py
+++ b/fastvideo/training/self_forcing_distillation_pipeline.py
@@ -153,19 +153,19 @@ class SelfForcingDistillationPipeline(DistillationPipeline):
     def _prepare_gt_inputs(self, training_batch: TrainingBatch) -> TrainingBatch:
         trajectory_latents = training_batch.trajectory_latents
 
-        # assert trajectory_latents.shape[1] == len(self.denoising_step_list) + 1
+        assert trajectory_latents.shape[1] == len(self.denoising_step_list) + 1
         
         # start_timestep_index = torch.randint(0, len(self.denoising_step_list), (1, ), device=self.device)
         # if dist.is_initialized():
         #     dist.broadcast(start_timestep_index, src=0)
         # start_timestep_index = start_timestep_index.cpu().item()
-        # training_batch.trajectory_latents = trajectory_latents[:, :, :21]
-        _cached_closest_idx_per_dmd = torch.tensor(
-            [0, 12, 24, 36, 50], dtype=torch.long).cpu()
-        training_batch.trajectory_latents = torch.index_select(
-            trajectory_latents,
-            dim=1,
-            index=_cached_closest_idx_per_dmd.to(trajectory_latents.device))
+        training_batch.trajectory_latents = trajectory_latents[:, :, :21]
+        # _cached_closest_idx_per_dmd = torch.tensor(
+        #     [0, 12, 24, 36, 50], dtype=torch.long).cpu()
+        # training_batch.trajectory_latents = torch.index_select(
+        #     trajectory_latents,
+        #     dim=1,
+        #     index=_cached_closest_idx_per_dmd.to(trajectory_latents.device))
         del trajectory_latents
         # training_batch.start_timestep_index = start_timestep_index
         return training_batch
@@ -603,7 +603,7 @@ class SelfForcingDistillationPipeline(DistillationPipeline):
         # assert self.crossattn_cache is not None
         self._cleanup_simulation_caches(kv_cache1, crossattn_cache)
 
-        assert pred_image_or_video.shape[1] == 31, "pred_image_or_video must have 31 frames"
+        assert pred_image_or_video.shape[1] == 21, "pred_image_or_video must have 21 frames"
 
         return final_output if gradient_mask is not None else pred_image_or_video