tests

Co-authored-by: kevin314 <kevin.lin.cs1@gmail.com>

.buildkite/pipeline.yml

@@ -176,3 +176,26 @@ steps:
                   - TEST_TYPE=precision_vsa
                 agents:
                   queue: "default"
+            - path:
+                - "csrc/attn/vmoba_attn/**"
+                - "pyproject.toml"
+                - "docker/Dockerfile.python3.12"
+              config:
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                label: "Precision Tests VMoBA"
+                env:
+                  - TEST_TYPE=precision_vmoba
+                agents:
+                  queue: "default"
+            - path:
+                - "csrc/attn/vmoba_attn/vmoba/**"
+                - "fastvideo/attention/backends/vmoba.py"
+                - "pyproject.toml"
+                - "docker/Dockerfile.python3.12"
+              config:
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                label: "Inference Tests VMoBA"
+                env:  
+                  - TEST_TYPE=inference_vmoba
+                agents:
+                  queue: "default"

.buildkite/scripts/pr_test.sh

@@ -109,6 +109,15 @@ case "$TEST_TYPE" in
         log "Running distillation DMD tests..."
         MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_distill_dmd_tests"
         ;;
+        # run_inference_tests_vmoba
+    "inference_vmoba")
+        log "Running V-MoBA inference tests..."
+        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_tests_vmoba"
+        ;;
+    "precision_vmoba")
+        log "Running V-MoBA precision tests..."
+        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_precision_tests_vmoba"
+        ;;
     *)
         log "Error: Unknown test type: $TEST_TYPE"
         exit 1

.github/workflows/pr-test.yml

@@ -235,7 +235,7 @@ jobs:
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
       RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
-  
+
   training-test:
     needs: change-filter
     if: >-
@@ -373,4 +373,4 @@ jobs:
           JOB_IDS: '["encoder-test", "vae-test", "transformer-test", "ssim-test-py3.10", "ssim-test-py3.11", "ssim-test-py3.12", "training-test", "training-test-VSA", "inference-test-STA", "precision-test-STA", "precision-test-VSA"]'
           RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
           GITHUB_RUN_ID: ${{ github.run_id }}
-        run: python .github/scripts/runpod_cleanup.py
+        run: python .github/scripts/runpod_cleanup.py

README.md

@@ -7,7 +7,7 @@
 FastVideo features an end-to-end unified pipeline for accelerating diffusion models, starting from data preprocessing to model training, finetuning, distillation, and inference. FastVideo is designed to be modular and extensible, allowing users to easily add new optimizations and techniques. Whether it is training-free optimizations or post-training optimizations, FastVideo has you covered.
 
 <p align="center">
-    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/rG0QpZdw" target="_blank"> <b> WeChat </b> </a> |
+    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/S7HLCSTh" target="_blank"> <b> WeChat </b> </a> |
 </p>
 
 <div align="center">

csrc/attn/vmoba_attn/README.md

@@ -0,0 +1,32 @@
+# Attention Kernel Used in FastVideo
+
+## VMoBA: Mixture-of-Block Attention for Video Diffusion Models (VMoBA)
+
+### Installation
+Please ensure that you have installed FlashAttention version **2.7.1 or higher**, as some interfaces have changed in recent releases.
+
+### Usage
+
+You can use `moba_attn_varlen` in the following ways:
+
+**Install from source:**
+```bash
+python setup.py install
+```
+
+**Import after installation:**
+```python
+from vmoba import moba_attn_varlen
+```
+
+**Or import directly from the project root:**
+```python
+from csrc.attn.vmoba_attn.vmoba import moba_attn_varlen
+```
+
+### Verify if you have successfully installed
+
+```bash
+python csrc/attn/vmoba_attn/vmoba/vmoba.py
+```
+

csrc/attn/vmoba_attn/setup.py

@@ -0,0 +1,24 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from setuptools import find_packages, setup
+
+PACKAGE_NAME = "vmoba"
+VERSION = "0.0.0"
+AUTHOR = "JianzongWu"
+DESCRIPTION = "VMoBA: Mixture-of-Block Attention for Video Diffusion Models"
+URL = "https://github.com/KwaiVGI/VMoBA"
+
+setup(
+    name=PACKAGE_NAME,
+    version=VERSION,
+    author=AUTHOR,
+    description=DESCRIPTION,
+    url=URL,
+    packages=find_packages(),
+    classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: Apache Software License",
+    ],
+    python_requires='>=3.12',
+    install_requires=[]
+)

csrc/attn/vmoba_attn/tests/test_vmoba_attn.py

@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import pytest
+import random
+from csrc.attn.vmoba_attn.vmoba  import moba_attn_varlen
+
+def generate_test_data(batch_size, total_seqlen, num_heads, head_dim, dtype, device="cuda"):
+    """
+    Generates random data for testing the variable-length attention function.
+    """
+    torch.manual_seed(42)
+    random.seed(42)
+    torch.cuda.manual_seed_all(42)
+
+    # Generate sequence lengths for each item in the batch
+    if batch_size > 1:
+        # Ensure sequence lengths are reasonably distributed
+        avg_seqlen = total_seqlen // batch_size
+        seqlens = [random.randint(avg_seqlen // 2, avg_seqlen + avg_seqlen // 2) for _ in range(batch_size - 1)]
+        remaining_len = total_seqlen - sum(seqlens)
+        if remaining_len > 0:
+            seqlens.append(remaining_len)
+        else: # Adjust if sum exceeds total_seqlen
+            seqlens.append(avg_seqlen)
+            current_sum = sum(seqlens)
+            seqlens[-1] -= (current_sum - total_seqlen)
+        # Ensure all lengths are positive
+        seqlens = [max(1, s) for s in seqlens]
+        # Final adjustment to match total_seqlen
+        seqlens[-1] += total_seqlen - sum(seqlens)
+
+    else:
+        seqlens = [total_seqlen]
+
+    cu_seqlens = torch.tensor([0] + list(torch.cumsum(torch.tensor(seqlens), 0)), device=device, dtype=torch.int32)
+    max_seqlen = max(seqlens) if seqlens else 0
+
+    q = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+    k = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+    v = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+
+    return q, k, v, cu_seqlens, max_seqlen
+
+
+@pytest.mark.parametrize("batch_size", [1, 2])
+@pytest.mark.parametrize("total_seqlen", [512, 1024])
+@pytest.mark.parametrize("num_heads", [8])
+@pytest.mark.parametrize("head_dim", [64])
+@pytest.mark.parametrize("moba_chunk_size", [64])
+@pytest.mark.parametrize("moba_topk", [2, 4])
+@pytest.mark.parametrize("select_mode", ["topk", "threshold"])
+@pytest.mark.parametrize("threshold_type", ["query_head", "head_global", "overall"])
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
+def test_moba_attn_varlen_forward(
+    batch_size, total_seqlen, num_heads, head_dim, moba_chunk_size, moba_topk, select_mode, threshold_type, dtype
+):
+    """
+    Tests the forward pass of moba_attn_varlen for basic correctness.
+    It checks output shape, dtype, and for the presence of NaNs/Infs.
+    """
+    if dtype == torch.float32:
+        pytest.skip("float32 is not supported in flash attention")
+
+    q, k, v, cu_seqlens, max_seqlen = generate_test_data(
+        batch_size, total_seqlen, num_heads, head_dim, dtype
+    )
+
+    # Ensure chunk size is not larger than the smallest sequence length
+    min_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).min().item()
+    if moba_chunk_size > min_seqlen:
+        pytest.skip("moba_chunk_size is larger than the minimum sequence length in the batch")
+
+    try:
+        output = moba_attn_varlen(
+            q=q,
+            k=k,
+            v=v,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            moba_chunk_size=moba_chunk_size,
+            moba_topk=moba_topk,
+            select_mode=select_mode,
+            threshold_type=threshold_type,
+            simsum_threshold=0.5, # A reasonable default for threshold mode
+        )
+    except Exception as e:
+        pytest.fail(f"moba_attn_varlen forward pass failed with exception: {e}")
+
+    # 1. Check output shape
+    assert output.shape == q.shape, f"Expected output shape {q.shape}, but got {output.shape}"
+
+    # 2. Check output dtype
+    assert output.dtype == q.dtype, f"Expected output dtype {q.dtype}, but got {output.dtype}"
+
+    # 3. Check for NaNs or Infs in the output
+    assert torch.all(torch.isfinite(output)), "Output contains NaN or Inf values"

csrc/attn/vmoba_attn/vmoba/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-License-Identifier: Apache-2.0
+from .vmoba import moba_attn_varlen, process_moba_input, process_moba_output

csrc/attn/vmoba_attn/vmoba/vmoba.py

@@ -0,0 +1,860 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapt from https://github.com/KwaiVGI/VMoBA/blob/main/src/vmoba.py
+
+import random
+import time
+import os
+import torch
+from typing import Tuple
+from flash_attn import flash_attn_varlen_func  # Use the new flash attention function
+from flash_attn.flash_attn_interface import _flash_attn_varlen_forward, _flash_attn_varlen_backward
+from functools import lru_cache
+from einops import rearrange
+
+
+@lru_cache(maxsize=16)
+def calc_chunks(cu_seqlen, moba_chunk_size):
+    """
+    Calculate chunk boundaries.
+
+    For vision tasks we include all chunks (even the last one which might be shorter)
+    so that every chunk can be selected.
+    """
+    batch_sizes = cu_seqlen[1:] - cu_seqlen[:-1]
+    batch_num_chunk = (batch_sizes + (moba_chunk_size - 1)) // moba_chunk_size
+    cu_num_chunk = torch.ones(
+        batch_num_chunk.numel() + 1,
+        device=cu_seqlen.device,
+        dtype=batch_num_chunk.dtype,
+    )
+    cu_num_chunk[1:] = batch_num_chunk.cumsum(dim=0)
+    num_chunk = cu_num_chunk[-1]
+    chunk_sizes = torch.full(
+        (num_chunk + 1,), moba_chunk_size, dtype=torch.int32, device=cu_seqlen.device
+    )
+    chunk_sizes[0] = 0
+    batch_last_chunk_size = batch_sizes - (batch_num_chunk - 1) * moba_chunk_size
+    chunk_sizes[cu_num_chunk[1:]] = batch_last_chunk_size
+    cu_chunk = chunk_sizes.cumsum(dim=-1, dtype=torch.int32)
+    chunk_to_batch = torch.zeros(
+        (num_chunk,), dtype=torch.int32, device=cu_seqlen.device
+    )
+    chunk_to_batch[cu_num_chunk[1:-1]] = 1
+    chunk_to_batch = chunk_to_batch.cumsum(dim=0, dtype=torch.int32)
+
+    # Do not filter out any chunk
+    filtered_chunk_indices = torch.arange(
+        num_chunk, device=cu_seqlen.device, dtype=torch.int32
+    )
+    num_filtered_chunk = num_chunk
+
+    return cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch
+
+
+# --- Threshold Selection Helper Functions ---
+
+def _select_threshold_query_head(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects chunks for each <query, head> pair based on threshold.
+    Normalization and sorting happen along the chunk dimension (dim=0).
+    """
+    C, H, S = gate.shape
+    eps = 1e-6
+
+    # LSE‐style normalization per <head, query> (across chunks)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    row_min = gate_min_val.amin(dim=0)        # (H, S)
+    row_max = gate_masked.amax(dim=0)         # (H, S)
+    denom = row_max - row_min
+    denom = torch.where(denom <= eps, torch.ones_like(denom), denom) # avoid divide‑by‑zero
+
+    gate_norm = (gate - row_min.unsqueeze(0)) / denom.unsqueeze(0)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) pull out the self‐chunk’s normalized weight for each <head,seq>
+    self_norm = (gate_norm * gate_self_chunk_mask).sum(dim=0)  # (H, S)
+
+    # 2) compute how much more normalized weight we need beyond self
+    total_norm_sum = gate_norm.sum(dim=0) # (H, S)
+    remain_ratio = simsum_threshold - self_norm / (total_norm_sum + eps)  # (H, S)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # if already ≥ thresh, no extra needed
+
+    # 3) zero out the self‐chunk in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0
+
+    # 4) sort the other chunks by descending norm, per <head,seq>
+    sorted_norm, sorted_idx = torch.sort(others_norm, descending=True, dim=0) # (C, H, S)
+
+    # 5) cumulative‑sum the sorted norms per <head,seq>
+    cumsum_others = sorted_norm.cumsum(dim=0) # (C, H, S)
+
+    # 6) for each <head,seq>, find the smallest k where cumsum_ratio ≥ remain_ratio
+    ratio = cumsum_others / (total_norm_sum.unsqueeze(0) + eps) # (C, H, S)
+    cond = ratio >= remain_ratio.unsqueeze(0) # (C, H, S) boolean mask
+    any_cond = cond.any(dim=0) # (H, S)
+    # Find the index of the first True value along dim 0. If none, use C-1.
+    cutoff = torch.where(any_cond, cond.float().argmax(dim=0), torch.full_like(any_cond, fill_value=C - 1)) # (H, S)
+
+    # 7) build a mask in sorted order up to that cutoff
+    idx_range = torch.arange(C, device=gate.device).view(-1, 1, 1) # (C, 1, 1)
+    sorted_mask = idx_range <= cutoff.unsqueeze(0) # (C, H, S)
+
+    # 8) scatter it back to original chunk order
+    others_mask = torch.zeros_like(gate, dtype=torch.bool)
+    others_mask.scatter_(0, sorted_idx, sorted_mask)
+
+    # 9) finally, include every self‐chunk plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_block(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <query, head> pairs for each block based on threshold.
+    Normalization and sorting happen across the head and sequence dimensions (dim=1, 2).
+    """
+    C, H, S = gate.shape
+    HS = H * S
+    eps = 1e-6
+
+    # LSE‐style normalization per block (across heads and queries)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    block_max = gate_masked.amax(dim=(1, 2), keepdim=True) # (C, 1, 1)
+    block_min = gate_min_val.amin(dim=(1, 2), keepdim=True) # (C, 1, 1)
+    block_denom = block_max - block_min
+    block_denom = torch.where(block_denom <= eps, torch.ones_like(block_denom), block_denom) # (C, 1, 1)
+
+    gate_norm = (gate - block_min) / block_denom # (C, H, S)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) identify normalized weights of entries that *are* self-chunks (from query perspective)
+    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
+    # Sum these weights *per block*
+    self_norm_sum_per_block = self_norm_entries.sum(dim=(1, 2)) # (C,)
+
+    # 2) compute how much more normalized weight each block needs beyond its self-chunk contributions
+    total_norm_sum_per_block = gate_norm.sum(dim=(1, 2)) # (C,)
+    remain_ratio = simsum_threshold - self_norm_sum_per_block / (total_norm_sum_per_block + eps) # (C,)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # (C,)
+
+    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
+
+    # 4) sort the other <head, seq> pairs by descending norm, per block
+    others_flat = others_norm.contiguous().view(C, HS) # (C, H*S)
+    sorted_others_flat, sorted_indices_flat = torch.sort(others_flat, dim=1, descending=True) # (C, H*S)
+
+    # 5) cumulative‑sum the sorted norms per block
+    cumsum_others_flat = sorted_others_flat.cumsum(dim=1) # (C, H*S)
+
+    # 6) for each block, find the smallest k where cumsum_ratio ≥ remain_ratio
+    ratio_flat = cumsum_others_flat / (total_norm_sum_per_block.unsqueeze(1) + eps) # (C, H*S)
+    cond_flat = ratio_flat >= remain_ratio.unsqueeze(1) # (C, H*S) boolean mask
+    any_cond = cond_flat.any(dim=1) # (C,)
+    # Find the index of the first True value along dim 1. If none, use HS-1.
+    cutoff_flat = torch.where(any_cond, cond_flat.float().argmax(dim=1), torch.full_like(any_cond, fill_value=HS - 1)) # (C,)
+
+    # 7) build a mask in sorted order up to that cutoff per block
+    idx_range_flat = torch.arange(HS, device=gate.device).unsqueeze(0) # (1, H*S)
+    sorted_mask_flat = idx_range_flat <= cutoff_flat.unsqueeze(1) # (C, H*S)
+
+    # 8) scatter it back to original <head, seq> order per block
+    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C, H*S)
+    others_mask_flat.scatter_(1, sorted_indices_flat, sorted_mask_flat)
+    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
+
+    # 9) finally, include every self‐chunk entry plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_overall(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <chunk, query, head> triplets globally based on threshold.
+    Normalization and sorting happen across all valid entries.
+    """
+    C, H, S = gate.shape
+    CHS = C * H * S
+    eps = 1e-6
+
+    # LSE‐style normalization globally across all valid entries
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    overall_max = gate_masked.max() # scalar
+    overall_min = gate_min_val.min() # scalar
+    overall_denom = overall_max - overall_min
+    overall_denom = torch.where(overall_denom <= eps, torch.tensor(1.0, device=gate.device, dtype=gate.dtype), overall_denom)
+
+    gate_norm = (gate - overall_min) / overall_denom # (C, H, S)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) identify normalized weights of entries that *are* self-chunks
+    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
+    # Sum these weights globally
+    self_norm_sum_overall = self_norm_entries.sum() # scalar
+
+    # 2) compute how much more normalized weight is needed globally beyond self-chunk contributions
+    total_norm_sum_overall = gate_norm.sum() # scalar
+    remain_ratio = simsum_threshold - self_norm_sum_overall / (total_norm_sum_overall + eps) # scalar
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # scalar
+
+    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
+
+    # 4) sort all other entries by descending norm, globally
+    others_flat = others_norm.flatten() # (C*H*S,)
+    valid_others_mask_flat = valid_gate_mask.flatten() & ~gate_self_chunk_mask.flatten() # Mask for valid, non-self entries
+    
+    # Only sort the valid 'other' entries
+    valid_others_indices = torch.where(valid_others_mask_flat)[0]
+    valid_others_values = others_flat[valid_others_indices]
+
+    sorted_others_values, sort_perm = torch.sort(valid_others_values, descending=True) # (N_valid_others,)
+    sorted_original_indices = valid_others_indices[sort_perm] # Original indices in C*H*S space, sorted by value
+
+    # 5) cumulative‑sum the sorted valid 'other' norms globally
+    cumsum_others_values = sorted_others_values.cumsum(dim=0) # (N_valid_others,)
+
+    # 6) find the smallest k where cumsum_ratio ≥ remain_ratio globally
+    ratio_values = cumsum_others_values / (total_norm_sum_overall + eps) # (N_valid_others,)
+    cond_values = ratio_values >= remain_ratio # (N_valid_others,) boolean mask
+    any_cond = cond_values.any() # scalar
+    
+    # Find the index of the first True value in the *sorted* list. If none, use all valid others.
+    cutoff_idx_in_sorted = torch.where(
+        any_cond, 
+        cond_values.float().argmax(dim=0), 
+        torch.tensor(len(sorted_others_values) - 1, device=gate.device, dtype=torch.long)
+    )
+
+    # 7) build a mask selecting the top-k others based on the cutoff
+    # Select the original indices corresponding to the top entries in the sorted list
+    selected_other_indices = sorted_original_indices[:cutoff_idx_in_sorted + 1]
+
+    # 8) create the mask in the original flat shape
+    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C*H*S,)
+    if selected_other_indices.numel() > 0: # Check if any 'other' indices were selected
+        others_mask_flat[selected_other_indices] = True
+    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
+
+    # 9) finally, include every self‐chunk entry plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_head_global(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <chunk, query> globally for each head based on threshold.
+    """
+    C, H, S = gate.shape
+    eps = 1e-6
+
+    # 1) LSE‐style normalization per head (across chunks and sequence dims)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf)
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf)
+
+    max_per_head = gate_masked.amax(dim=(0, 2), keepdim=True)  # (1, H, 1)
+    min_per_head = gate_min_val.amin(dim=(0, 2), keepdim=True)  # (1, H, 1)
+    denom = max_per_head - min_per_head
+    denom = torch.where(denom <= eps, torch.ones_like(denom), denom)
+
+    gate_norm = (gate - min_per_head) / denom
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0)  # (C, H, S)
+
+    # 2) sum normalized self‐chunk contributions per head
+    self_norm_sum = (gate_norm * gate_self_chunk_mask).sum(dim=(0, 2))  # (H,)
+
+    # 3) total normalized sum per head
+    total_norm_sum = gate_norm.sum(dim=(0, 2))  # (H,)
+
+    # 4) how much more normalized weight needed per head
+    remain_ratio = simsum_threshold - self_norm_sum / (total_norm_sum + eps)  # (H,)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0)
+
+    # 5) zero out self‐chunk entries to focus on "others"
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0  # (C, H, S)
+
+    # 6) flatten chunk and sequence dims, per head
+    CS = C * S
+    others_flat = others_norm.permute(1, 0, 2).reshape(H, CS)  # (H, C*S)
+    valid_flat  = (valid_gate_mask & ~gate_self_chunk_mask) \
+                   .permute(1, 0, 2).reshape(H, CS)           # (H, C*S)
+
+    # 7) vectorized selection of “others” per head
+    masked_flat = torch.where(valid_flat, others_flat, torch.zeros_like(others_flat))
+    sorted_vals, sorted_idx = torch.sort(masked_flat, dim=1, descending=True)  # (H, C*S)
+
+    cumsum_vals = sorted_vals.cumsum(dim=1)                                    # (H, C*S)
+    ratio_vals  = cumsum_vals / (total_norm_sum.unsqueeze(1) + eps)            # (H, C*S)
+    cond        = ratio_vals >= remain_ratio.unsqueeze(1)                      # (H, C*S)
+
+    has_cutoff = cond.any(dim=1)                                               # (H,)
+    default    = torch.full((H,), CS - 1, device=gate.device, dtype=torch.long)
+    cutoff     = torch.where(has_cutoff, cond.float().argmax(dim=1), default)  # (H,)
+
+    idx_range   = torch.arange(CS, device=gate.device).unsqueeze(0)            # (1, C*S)
+    sorted_mask = idx_range <= cutoff.unsqueeze(1)                             # (H, C*S)
+
+    selected_flat = torch.zeros_like(valid_flat)                               # (H, C*S)
+    selected_flat.scatter_(1, sorted_idx, sorted_mask)                         # (H, C*S)
+
+    # 8) reshape selection mask back to (C, H, S)
+    others_mask = selected_flat.reshape(H, C, S).permute(1, 0, 2)               # (C, H, S)
+
+    # 9) include self‐chunks plus selected others, and obey valid mask
+    final_gate_mask = valid_gate_mask & (gate_self_chunk_mask | others_mask)
+
+    return final_gate_mask
+
+
+class MixedAttention(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        q,
+        k,
+        v,
+        self_attn_cu_seqlen,
+        moba_q,
+        moba_kv,
+        moba_cu_seqlen_q,
+        moba_cu_seqlen_kv,
+        max_seqlen,
+        moba_chunk_size,
+        moba_q_sh_indices,
+    ):
+        ctx.max_seqlen = max_seqlen
+        ctx.moba_chunk_size = moba_chunk_size
+        ctx.softmax_scale = softmax_scale = q.shape[-1] ** (-0.5)
+
+        # Non-causal self-attention branch
+        # return out, softmax_lse, S_dmask, rng_state
+        self_attn_out_sh, self_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
+            q=q,
+            k=k,
+            v=v,
+            cu_seqlens_q=self_attn_cu_seqlen,
+            cu_seqlens_k=self_attn_cu_seqlen,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+        )
+        # MOBA attention branch (non-causal)
+        moba_attn_out, moba_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
+            q=moba_q,
+            k=moba_kv[:, 0],
+            v=moba_kv[:, 1],
+            cu_seqlens_q=moba_cu_seqlen_q,
+            cu_seqlens_k=moba_cu_seqlen_kv,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=moba_chunk_size,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+        )
+
+        self_attn_lse_sh = self_attn_lse_hs.t().contiguous()
+        moba_attn_lse = moba_attn_lse_hs.t().contiguous()
+
+        output = torch.zeros((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
+        output_2d = output.view(-1, q.shape[2])
+
+        max_lse_1d = self_attn_lse_sh.view(-1)
+        max_lse_1d = max_lse_1d.index_reduce(
+            0, moba_q_sh_indices, moba_attn_lse.view(-1), "amax"
+        )
+        self_attn_lse_sh = self_attn_lse_sh - max_lse_1d.view_as(self_attn_lse_sh)
+        moba_attn_lse = (
+            moba_attn_lse.view(-1)
+            .sub(max_lse_1d.index_select(0, moba_q_sh_indices))
+            .reshape_as(moba_attn_lse)
+        )
+
+        mixed_attn_se_sh = self_attn_lse_sh.exp()
+        moba_attn_se = moba_attn_lse.exp()
+
+        mixed_attn_se_sh.view(-1).index_add_(
+            0, moba_q_sh_indices, moba_attn_se.view(-1)
+        )
+        mixed_attn_lse_sh = mixed_attn_se_sh.log()
+
+        # Combine self-attention output
+        factor = (self_attn_lse_sh - mixed_attn_lse_sh).exp()  # [S, H]
+        self_attn_out_sh = self_attn_out_sh * factor.unsqueeze(-1)
+        output_2d += self_attn_out_sh.reshape_as(output_2d)
+
+        # Combine MOBA attention output
+        mixed_attn_lse = (
+            mixed_attn_lse_sh.view(-1)
+            .index_select(0, moba_q_sh_indices)
+            .view_as(moba_attn_lse)
+        )
+        factor = (moba_attn_lse - mixed_attn_lse).exp()  # [S, H]
+        moba_attn_out = moba_attn_out * factor.unsqueeze(-1)
+        raw_attn_out = moba_attn_out.view(-1, moba_attn_out.shape[-1])
+        output_2d.index_add_(0, moba_q_sh_indices, raw_attn_out)
+        output = output.to(q.dtype)
+        mixed_attn_lse_sh = mixed_attn_lse_sh + max_lse_1d.view_as(mixed_attn_se_sh)
+        ctx.save_for_backward(
+            output,
+            mixed_attn_lse_sh,
+            q,
+            k,
+            v,
+            self_attn_cu_seqlen,
+            moba_q,
+            moba_kv,
+            moba_cu_seqlen_q,
+            moba_cu_seqlen_kv,
+            moba_q_sh_indices,
+        )
+
+        return output
+
+    @staticmethod
+    def backward(ctx, d_output):
+
+        max_seqlen = ctx.max_seqlen
+        moba_chunk_size = ctx.moba_chunk_size
+        softmax_scale = ctx.softmax_scale
+
+        (
+            output,
+            mixed_attn_vlse_sh,
+            q,
+            k,
+            v,
+            self_attn_cu_seqlen,
+            moba_q,
+            moba_kv,
+            moba_cu_seqlen_q,
+            moba_cu_seqlen_kv,
+            moba_q_sh_indices,
+        ) = ctx.saved_tensors
+
+        d_output = d_output.contiguous()
+
+        dq = torch.empty_like(q)
+        dk = torch.empty_like(k)
+        dv = torch.empty_like(v)
+        _ = _flash_attn_varlen_backward(
+            dout=d_output,
+            q=q,
+            k=k,
+            v=v,
+            out=output,
+            softmax_lse=mixed_attn_vlse_sh.t().contiguous(),
+            dq=dq,
+            dk=dk,
+            dv=dv,
+            cu_seqlens_q=self_attn_cu_seqlen,
+            cu_seqlens_k=self_attn_cu_seqlen,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=True,
+            window_size_left=-1,
+            window_size_right=-1
+        )
+
+        headdim = q.shape[-1]
+        d_moba_output = (
+            d_output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
+        )
+        moba_output = (
+            output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
+        )
+
+        mixed_attn_vlse = (
+            mixed_attn_vlse_sh.view(-1).index_select(0, moba_q_sh_indices).view(1, -1)
+        )
+
+        dmq = torch.empty_like(moba_q)
+        dmkv = torch.empty_like(moba_kv)
+        _ = _flash_attn_varlen_backward(
+            dout=d_moba_output,
+            q=moba_q,
+            k=moba_kv[:, 0],
+            v=moba_kv[:, 1],
+            out=moba_output,
+            softmax_lse=mixed_attn_vlse,
+            dq=dmq,
+            dk=dmkv[:,0],
+            dv=dmkv[:,1],
+            cu_seqlens_q=moba_cu_seqlen_q,
+            cu_seqlens_k=moba_cu_seqlen_kv,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=moba_chunk_size,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=True,
+            window_size_left=-1,
+            window_size_right=-1
+        )
+
+        return dq, dk, dv, None, dmq, dmkv, None, None, None, None, None
+
+
+def moba_attn_varlen(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens: torch.Tensor,
+    max_seqlen: int,
+    moba_chunk_size: int,
+    moba_topk: int,
+    select_mode: str = 'threshold', # "topk" or "threshold"
+    simsum_threshold: float = 0.25,
+    threshold_type: str = 'query_head',
+) -> torch.Tensor:
+    """
+    Accelerated MOBA attention for vision tasks with proper LSE normalization.
+
+    This version:
+      - Splits KV into chunks.
+      - For each query head, selects the top-k relevant KV chunks (including the self chunk)
+        by amplifying the diagonal (self-chunk) logits.
+      - Aggregates the attention outputs from the selected chunks using a log-sum-exp
+        reduction so that attending to each query over the selected chunks is equivalent
+        to the original algorithm.
+    """
+    # Stack keys and values.
+    kv = torch.stack((k, v), dim=1)
+    seqlen, num_head, head_dim = q.shape
+
+    # Compute chunk boundaries.
+    cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch = calc_chunks(
+        cu_seqlens, moba_chunk_size
+    )
+
+    self_attn_cu_seqlen = cu_chunk
+
+    # Update top-k selection to include the self chunk.
+    moba_topk = min(moba_topk, num_filtered_chunk)
+
+    # --- Build filtered KV from chunks ---
+    chunk_starts = cu_chunk[filtered_chunk_indices]          # [num_filtered_chunk]
+    chunk_ends = cu_chunk[filtered_chunk_indices + 1]          # [num_filtered_chunk]
+    chunk_lengths = chunk_ends - chunk_starts                  # [num_filtered_chunk]
+    max_chunk_len = int(chunk_lengths.max().item())
+
+    range_tensor = torch.arange(max_chunk_len, device=kv.device, dtype=chunk_starts.dtype).unsqueeze(0)
+    indices = chunk_starts.unsqueeze(1) + range_tensor
+    indices = torch.clamp(indices, max=kv.shape[0] - 1)
+    valid_mask = range_tensor < chunk_lengths.unsqueeze(1)
+    gathered = kv[indices.view(-1)].view(num_filtered_chunk, max_chunk_len, *kv.shape[1:])
+    gathered = gathered * valid_mask.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).type_as(gathered)
+
+    # Compute key_gate_weight over valid tokens.
+    key_values = gathered[:, :, 0].float()  # [num_filtered_chunk, max_chunk_len, num_head, head_dim]
+    valid_mask_exp = valid_mask.unsqueeze(-1).unsqueeze(-1)
+    key_sum = (key_values * valid_mask_exp).sum(dim=1)
+    divisor = valid_mask.sum(dim=1).unsqueeze(-1).unsqueeze(-1)
+    key_gate_weight = key_sum / divisor  # [num_filtered_chunk, num_head, head_dim]
+
+    # Compute gate logits between key_gate_weight and queries.
+    q_float = q.float()
+    # gate = torch.einsum("nhd,shd->nhs", key_gate_weight, q_float)  # [num_filtered_chunk, num_head, seqlen]
+    gate = torch.bmm(key_gate_weight.permute(1, 0, 2), q_float.permute(1, 0, 2).transpose(1, 2)).permute(1, 0, 2)
+
+    # Amplify the diagonal (self chunk) contributions.
+    gate_seq_idx = torch.arange(seqlen, device=q.device, dtype=torch.int32).unsqueeze(0).expand(num_filtered_chunk, seqlen)
+    chunk_start = cu_chunk[filtered_chunk_indices]  # [num_filtered_chunk]
+    chunk_end = cu_chunk[filtered_chunk_indices + 1]  # [num_filtered_chunk]
+    gate_self_chunk_mask = ((gate_seq_idx >= chunk_start.unsqueeze(1)) &
+                            (gate_seq_idx < chunk_end.unsqueeze(1))).unsqueeze(1).expand(-1, num_head, -1)
+    amplification_factor = 1e9  # Example factor; adjust as needed.
+    origin_gate = gate.clone()
+    gate = gate.clone()
+    if select_mode == "topk":
+        gate[gate_self_chunk_mask] += amplification_factor
+
+    # Exclude positions that are outside the valid batch boundaries.
+    batch_starts = cu_seqlens[chunk_to_batch[filtered_chunk_indices]]
+    batch_ends = cu_seqlens[chunk_to_batch[filtered_chunk_indices] + 1]
+    gate_batch_start_mask = gate_seq_idx < batch_starts.unsqueeze(1)
+    gate_batch_end_mask = gate_seq_idx >= batch_ends.unsqueeze(1)
+    gate_inf_mask = gate_batch_start_mask | gate_batch_end_mask
+    gate.masked_fill_(gate_inf_mask.unsqueeze(1), -float("inf"))
+
+    if select_mode == 'topk':
+        # We amplify self‐chunk in gate already, so self entries will rank highest.
+        valid_gate_mask = gate != -float("inf")
+        if threshold_type == 'query_head':
+            # === per‐<head,seq> top-k across chunks (original behavior) ===
+            # gate: (C, H, S)
+            _, gate_topk_idx = torch.topk(gate, k=moba_topk, dim=0, largest=True, sorted=False)
+            gate_idx_mask = torch.zeros_like(gate, dtype=torch.bool)
+            gate_idx_mask.scatter_(0, gate_topk_idx, True)
+            gate_mask = valid_gate_mask & gate_idx_mask
+        elif threshold_type == 'overall':
+            # === global top-k across all (chunk, head, seq) entries ===
+            C, H, S = gate.shape
+            flat_gate = gate.flatten()
+            flat_mask = valid_gate_mask.flatten()
+            flat_gate_masked = torch.where(flat_mask, flat_gate, -float("inf"))
+            # pick topk global entries
+            vals, idx = torch.topk(flat_gate_masked, k=moba_topk * H * S, largest=True, sorted=False)
+            others_mask_flat = torch.zeros_like(flat_mask, dtype=torch.bool)
+            others_mask_flat[idx] = True
+            gate_mask = (valid_gate_mask.flatten() & others_mask_flat).view(gate.shape)
+        elif threshold_type == 'head_global':
+            # per-head top-k across all chunks and sequence positions
+            C, H, S = gate.shape
+            CS = C * S
+            flat_gate  = gate.permute(1, 0, 2).reshape(H, CS)
+            flat_valid = valid_gate_mask.permute(1, 0, 2).reshape(H, CS)
+            flat_gate_masked = torch.where(flat_valid, flat_gate, torch.full_like(flat_gate, -float('inf')))
+            # pick top-k indices per head
+            _, topk_idx = torch.topk(flat_gate_masked, k=moba_topk * S, dim=1, largest=True, sorted=False)
+            gate_idx_flat = torch.zeros_like(flat_valid, dtype=torch.bool)
+            gate_idx_flat.scatter_(1, topk_idx, True)
+            gate_mask = gate_idx_flat.reshape(H, C, S).permute(1, 0, 2)
+        else:
+            raise ValueError(
+                f"Invalid threshold_type for topk: {threshold_type}. "
+                "Choose 'query_head', 'block', or 'overall'."
+            )
+    elif select_mode == 'threshold':
+        # Delegate to the specific thresholding function
+        valid_gate_mask = gate != -float("inf")  # (num_chunk, num_head, seqlen)
+        if threshold_type == 'query_head':
+            gate_mask = _select_threshold_query_head(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'block':
+            gate_mask = _select_threshold_block(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'overall':
+            gate_mask = _select_threshold_overall(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'head_global':
+            gate_mask = _select_threshold_head_global(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        else:
+            raise ValueError(f"Invalid threshold_type: {threshold_type}. Choose 'query_head', 'block', or 'overall'.")
+    else:
+        raise ValueError(f"Invalid select_mode: {select_mode}. Choose 'topk' or 'threshold'.")
+    
+    # eliminate self_chunk in MoBA branch
+    gate_mask = gate_mask & ~gate_self_chunk_mask
+    # if gate_mask is all false, perform flash_attn instead
+    if gate_mask.sum() == 0:
+        return flash_attn_varlen_func(
+            q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen, causal=False
+        )
+
+    # Determine which query positions are selected.
+    # nonzero_indices has shape [N, 3] where each row is [chunk_index, head_index, seq_index].
+    moba_q_indices = gate_mask.reshape(gate_mask.shape[0], -1).nonzero(as_tuple=True)[-1] # [(h s k)]
+    moba_q_sh_indices = (moba_q_indices % seqlen) * num_head + (moba_q_indices // seqlen)
+    moba_q = rearrange(q, "s h d -> (h s) d").index_select(0, moba_q_indices).unsqueeze(1)
+
+    # Build cumulative sequence lengths for the selected queries.
+    moba_seqlen_q = gate_mask.sum(dim=-1).flatten()
+    q_zero_mask = moba_seqlen_q == 0
+    valid_expert_mask = ~q_zero_mask
+    if q_zero_mask.sum() > 0:
+        moba_seqlen_q = moba_seqlen_q[valid_expert_mask]
+    moba_cu_seqlen_q = torch.cat(
+        (
+            torch.tensor([0], device=q.device, dtype=moba_seqlen_q.dtype),
+            moba_seqlen_q.cumsum(dim=0),
+        ),
+        dim=0,
+    ).to(torch.int32)
+
+    # Rearrange gathered KV for the MOBA branch.
+    experts_tensor = rearrange(gathered, "nc cl two h d -> (nc h) cl two d")
+    valid_expert_lengths = chunk_lengths.unsqueeze(1).expand(num_filtered_chunk, num_head).reshape(-1).to(torch.int32)
+    if q_zero_mask.sum() > 0:
+        experts_tensor = experts_tensor[valid_expert_mask]
+        valid_expert_lengths = valid_expert_lengths[valid_expert_mask]
+
+    seq_range = torch.arange(experts_tensor.shape[1], device=experts_tensor.device).unsqueeze(0)
+    mask = seq_range < valid_expert_lengths.unsqueeze(1)
+    moba_kv = experts_tensor[mask]  # Shape: ((nc h cl_valid) two d)
+    moba_kv = moba_kv.unsqueeze(2)   # Shape: ((nc h cl_valid) two 1 d)
+
+    moba_cu_seqlen_kv = torch.cat(
+        [torch.zeros(1, device=experts_tensor.device, dtype=torch.int32),
+         valid_expert_lengths.cumsum(dim=0)],
+        dim=0,
+    ).to(torch.int32)
+
+    assert (
+        moba_cu_seqlen_kv.shape == moba_cu_seqlen_q.shape
+    ), f"Mismatch between moba_cu_seqlen_kv.shape and moba_cu_seqlen_q.shape: {moba_cu_seqlen_kv.shape} vs {moba_cu_seqlen_q.shape}"
+
+    return MixedAttention.apply(
+        q,
+        k,
+        v,
+        self_attn_cu_seqlen,
+        moba_q,
+        moba_kv,
+        moba_cu_seqlen_q,
+        moba_cu_seqlen_kv,
+        max_seqlen,
+        moba_chunk_size,
+        moba_q_sh_indices,
+    )
+
+
+def process_moba_input(
+    x,
+    patch_resolution,
+    chunk_size,
+):
+    """
+    Process inputs for the attention function.
+
+    Args:
+        x (torch.Tensor): Input tensor with shape [batch_size, num_patches, num_heads, head_dim].
+        patch_resolution (tuple): Tuple containing the patch resolution (t, h, w).
+        chunk_size (int): Size of the chunk. (maybe tuple or int, according to chunk type)
+
+    Returns:
+        torch.Tensor: Processed input tensor.
+    """
+    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
+        moba_chunk_size = int(chunk_size * patch_resolution[1] * patch_resolution[2])
+    else:
+        assert isinstance(chunk_size, (Tuple, list)), f"chunk_size should be a tuple, list, or int, now it is: {type(chunk_size)}"
+        if len(chunk_size) == 2:
+            assert patch_resolution[1] % chunk_size[0] == 0 and patch_resolution[2] % chunk_size[1] == 0, f"spatial patch_resolution {patch_resolution[1:]} should be divisible by 2d chunk_size {chunk_size}"
+            nch, ncw = patch_resolution[1] // chunk_size[0], patch_resolution[2] // chunk_size[1]
+            x = rearrange(x, "b (t nch ch ncw cw) n d -> b (nch ncw t ch cw) n d", t=patch_resolution[0], nch=nch, ncw=ncw, ch=chunk_size[0], cw=chunk_size[1])
+            moba_chunk_size = patch_resolution[0] * chunk_size[0] * chunk_size[1]
+        elif len(chunk_size) == 3:
+            assert patch_resolution[0] % chunk_size[0] == 0 and patch_resolution[1] % chunk_size[1] == 0 and patch_resolution[2] % chunk_size[2] == 0, f"patch_resolution {patch_resolution} should be divisible by 3d chunk_size {chunk_size}"
+            nct, nch, ncw = patch_resolution[0] // chunk_size[0], patch_resolution[1] // chunk_size[1], patch_resolution[2] // chunk_size[2]
+            x = rearrange(x, "b (nct ct nch ch ncw cw) n d -> b (nct nch ncw ct ch cw) n d", nct=nct, nch=nch, ncw=ncw, ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
+            moba_chunk_size = chunk_size[0] * chunk_size[1] * chunk_size[2]
+        else:
+            raise ValueError(f"chunk_size should be a int, or a tuple of length 2 or 3, now it is: {len(chunk_size)}")
+        
+    return x, moba_chunk_size
+
+
+def process_moba_output(
+    x,
+    patch_resolution,
+    chunk_size,
+):
+    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
+        pass
+    elif len(chunk_size) == 2:
+        x = rearrange(x, "b (nch ncw t ch cw) n d -> b (t nch ch ncw cw) n d", nch=patch_resolution[1] // chunk_size[0], ncw=patch_resolution[2] // chunk_size[1], t=patch_resolution[0], ch=chunk_size[0], cw=chunk_size[1])
+    elif len(chunk_size) == 3:
+        x = rearrange(x, "b (nct nch ncw ct ch cw) n d -> b (nct ct nch ch ncw cw) n d", nct=patch_resolution[0] // chunk_size[0], nch=patch_resolution[1] // chunk_size[1], ncw=patch_resolution[2] // chunk_size[2], ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
+
+    return x
+
+
+# TEST
+def generate_data(batch_size, seqlen, num_head, head_dim, dtype):
+    random.seed(0)
+    torch.manual_seed(0)
+    torch.cuda.manual_seed(0)
+    device = torch.cuda.current_device()
+
+    q = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    k = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    v = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    print(f"q.shape: {q.shape}, k.shape: {k.shape}, v.shape: {v.shape}")
+    cu_seqlens = torch.arange(0, q.shape[0] * q.shape[1] + 1, q.shape[1], dtype=torch.int32, device='cuda')
+    max_seqlen = q.shape[1]
+    q = rearrange(q, "b s ... -> (b s) ...")
+    k = rearrange(k, "b s ... -> (b s) ...")
+    v = rearrange(v, "b s ... -> (b s) ...")
+
+    return q, k, v, cu_seqlens, max_seqlen
+
+
+def test_attn_varlen_moba_speed(batch, head, seqlen, head_dim, moba_chunk_size, moba_topk, dtype=torch.bfloat16, select_mode='threshold', simsum_threshold=0.25, threshold_type='query_head'):
+    """Speed test comparing flash_attn vs moba_attention"""
+    # Get data
+    q, k, v, cu_seqlen, max_seqlen = generate_data(batch, seqlen, head, head_dim, dtype)
+    print(f"batch:{batch} head:{head} seqlen:{seqlen} chunk:{moba_chunk_size} topk:{moba_topk} select_mode: {select_mode} simsum_threshold:{simsum_threshold}")
+    vo_grad = torch.randn_like(q)
+    
+    # Warmup
+    warmup_iters = 3
+    perf_test_iters = 10
+
+    # Warmup
+    for _ in range(warmup_iters):
+        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
+        torch.autograd.backward(o, vo_grad)
+    
+    torch.cuda.synchronize()
+    start_flash = time.perf_counter()
+    for _ in range(perf_test_iters):
+        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
+        torch.autograd.backward(o, vo_grad)
+        
+    torch.cuda.synchronize()
+    time_flash = (time.perf_counter() - start_flash) / perf_test_iters * 1000
+
+    # Warmup
+    for _ in range(warmup_iters):
+        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
+        torch.autograd.backward(om, vo_grad)
+
+    torch.cuda.synchronize()
+    start_moba = time.perf_counter()
+    for _ in range(perf_test_iters):
+        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
+        torch.autograd.backward(om, vo_grad)
+    
+    torch.cuda.synchronize()
+    time_moba = (time.perf_counter() - start_moba) / perf_test_iters * 1000
+    
+    print(f"Flash: {time_flash:.2f}ms, MoBA: {time_moba:.2f}ms")
+    print(f"Speedup:  {time_flash / time_moba:.2f}x")
+
+
+
+if __name__ == "__main__":
+    """
+    CUDA_VISIBLE_DEVICES=1 \
+    python -u csrc/attn/vmoba_attn/vmoba/vmoba.py
+    """
+    test_attn_varlen_moba_speed(batch=1, head=12, seqlen=32760, head_dim=128, moba_chunk_size=32760 // 3 // 6 // 4, moba_topk=3, select_mode='threshold', simsum_threshold=0.3, threshold_type='query_head')

examples/datasets/crush_smol/download_dataset.sh

@@ -1,3 +0,0 @@
-#!/bin/bash
-
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"

examples/datasets/crush_smol/validation.json

@@ -1,76 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-034.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-027.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "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.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-030.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-056.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-059.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.",
-      "image_path": null,
-      "video_path": "validation_dataset/EJqsC21GSBY-Scene-059.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/GBSfpTcKegk-Scene-003.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}

examples/datasets/crush_smol/validation_single.json

@@ -1,13 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/distill_dmd_VSA_t2v_5B.sh

@@ -98,6 +98,7 @@ dmd_args=(
 torchrun \
 --nnodes 1 \
 --nproc_per_node $NUM_GPUS \
+--master_port $MASTER_PORT \
     fastvideo/training/wan_distillation_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/distill_dmd_VSA_t2v_5B_lora.sh

@@ -0,0 +1,112 @@
+#!/bin/bash
+
+# Basic Info
+export WANDB_MODE="online"
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export MASTER_PORT=29501
+export TOKENIZERS_PARALLELISM=false
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+# Configs
+NUM_GPUS=1
+MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name wan_t2v_distill_dmd_VSA
+  --output_dir="checkpoints/wan_t2v_finetune"
+  --max_train_steps=4000
+  --train_batch_size=1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps=1
+  --num_latent_t 31
+  --num_height 704
+  --num_width 1280
+  --num_frames 121
+  --enable_gradient_checkpointing_type "full"
+  --training_state_checkpointing_steps=500
+  --weight_only_checkpointing_steps=500
+  --lora_rank 32
+  --lora_training True
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus 1
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 200
+  --validation_sampling_steps "3"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate=1e-4
+  --mixed_precision="bf16"
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --ema_start_step 0
+  --flow_shift 8
+  --seed 1000
+)
+
+# DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '1000,757,522'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --generator_update_interval 5
+  --real_score_guidance_scale 3.5
+  --VSA_sparsity 0.8
+)
+
+torchrun \
+--nnodes 1 \
+--nproc_per_node $NUM_GPUS \
+--master_port $MASTER_PORT \
+    fastvideo/training/wan_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"

examples/inference/basic/basic.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-from fastvideo.configs.sample import SamplingParam
+# from fastvideo.configs.sample import SamplingParam
 
 OUTPUT_PATH = "video_samples"
 def main():
@@ -17,29 +17,30 @@ def main():
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,
         pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
-        ti2v_task=True,
         # image_encoder_cpu_offload=False,
     )
 
-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
     # sampling_param.num_frames = 45
-    sampling_param.image_path = "test.jpg"
+    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
     # Generate videos with the same simple API, regardless of GPU count
     prompt = (
-        "A girl is packing a suitcase when stuff suddently starts flying around the room."
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
     )
     video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True)
     # video = generator.generate_video(prompt, sampling_param=sampling_param, output_path="wan_t2v_videos/")
 
     # Generate another video with a different prompt, without reloading the
     # model!
-    # prompt2 = (
-    #     "A majestic lion strides across the golden savanna, its powerful frame "
-    #     "glistening under the warm afternoon sun. The tall grass ripples gently in "
-    #     "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
-    #     "embodying the raw energy of the wild. Low angle, steady tracking shot, "
-    #     "cinematic.")
-    # video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)
+    prompt2 = (
+        "A majestic lion strides across the golden savanna, its powerful frame "
+        "glistening under the warm afternoon sun. The tall grass ripples gently in "
+        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+        "cinematic.")
+    video2 = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True)
 
 
 if __name__ == "__main__":

examples/inference/basic/basic_self_forcing_causal.py

@@ -19,15 +19,13 @@ def main():
     )
 
     sampling_param = SamplingParam.from_pretrained(model_name)
-    sampling_param.num_frames = 81
 
-    prompts = [
-        "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
-        "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
-    ]
-
-    for prompt in prompts:
-        video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
+    prompt = (
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+    )
+    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
 
 if __name__ == "__main__":
     main()

examples/training/consistency_finetune/causal_ode_init/crush_smol_prompts.txt

@@ -0,0 +1,47 @@
+A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.
+The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.
+The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.
+A red toy car is being crushed by a large hydraulic press, which is flattening objects as if they were under a hydraulic press.
+A large, cylindrical object is seen pressing down on a small orange ball, causing it to flatten as if it were under a hydraulic press. The background features a green wall with yellow and red warning signs.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is shown compressing a wooden object, which shatters into small pieces. The background features a green wall with a yellow sign displaying a lightning bolt.
+A large metal cylinder is seen descending, flattening objects as if they were under a hydraulic press. The cylinder compresses a stack of matches and boxes, causing them to crumble into small pieces. The scene is set against a green background with yellow and red signs.
+A large metal press is shown compressing a pile of colorful macarons, flattening them as if they were under a hydraulic press. The press moves down, crushing the macarons into a pile of crumbs and squishing the colorful filling out.
+The video shows a metal press flattening objects as if they were under a hydraulic press. The press is pressing down on a pile of colorful gummy candies, squishing them into a pile of squiggly shapes. The press is made of metal and has a large base, and the gummy candies are of various colors, including red, green, and orange. The background is a green wall, and the press is placed on a metal surface.
+A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
+The video shows a stack of colorful sponges being flattened as if they were under a hydraulic press. The sponges, which are pink, white, blue, and green, are compressed into a smaller size, demonstrating the press's power. The background features a green wall with a yellow and red sign, adding context to the setting.
+A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, leaving a pile of debris around it.
+A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press.
+The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.
+The video shows a close-up of a metal cylinder pressing down on a yellow object, which is being flattened as if it were under a hydraulic press. The cylinder is positioned above the object, and the force is causing the object to compress and spread out, creating a visible deformation. The background is blurred, focusing attention on the action of the cylinder and the object being flattened.
+A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.
+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 scene shows a metal press with a yellow and black striped pattern, holding a container filled with chocolate. A metal cylinder is descending, flattening the chocolate as if it were under a hydraulic press. The background is a green wall, and the press is mounted on a sturdy metal frame.
+The video shows a colorful sponge being flattened as if it were under a hydraulic press, with the sponge being compressed and eventually flattened into a thin layer.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is pressing down on a stack of wooden blocks, causing them to crumble and break apart. The press is black and yellow striped, and the wooden blocks are small and rectangular. The background is green, and the press is sitting on a metal table.
+A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
+The video shows a stack of colorful sponges being flattened by a large, cylindrical object, which appears to be a hydraulic press. The sponges, which are pink, blue, white, and green, are compressed into a single layer, demonstrating the press's powerful force. The background features a green wall with a yellow and red sign, adding context to the industrial setting.
+A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, demonstrating the immense pressure applied by the cylinder.
+A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press. The popcorn is crushed and scattered around the base of the cylinder, creating a satisfying visual effect.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is composed of a large, cylindrical metal cylinder with yellow and black stripes, and a metal base. The objects being flattened are two cylindrical blocks of cotton candy, one pink and one blue. The press is positioned on a metal table, and the background features a green wall with a yellow and red sign.
+The video shows a large orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.
+The video shows a cylindrical object being pressed down onto a flat surface, causing the objects beneath it to be flattened as if they were under a hydraulic press. The objects being flattened appear to be yellow and are being crushed into a pile of debris. The background is a greenish-gray color, and the surface on which the objects are being flattened is metallic and shiny.
+A green and blue object with a spiky texture is being flattened by a large, cylindrical metal press, demonstrating its resilience and durability.
+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.
+A large metal cylinder is seen pressing down on a pile of colorful jelly beans, flattening them as if they were under a hydraulic press.
+The video shows a machine with a yellow and black striped cylinder pressing down on a stack of colorful sponges, flattening them as if they were under a hydraulic press. The machine is situated in a green-walled room with warning signs in the background.
+The video shows a machine with a yellow and black striped cylinder, which is pressing down on two colorful objects, flattening them as if they were under a hydraulic press. The machine appears to be in a workshop or industrial setting, with a green wall in the background. The objects being flattened are green and orange, and the machine is covered in dirt and grime, indicating it has been used frequently.
+The video shows a large, industrial press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing a pile of pink objects into a pile of crumbs. The press is large and metallic, with a yellow and black striped pattern on its side. The background is a green wall with a yellow warning sign.
+The video shows a pink, sparkly ball being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
+A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and segments.
+The video shows a machine with a yellow and black striped cylinder, which is flattening objects as if they were under a hydraulic press. The machine is pressing down on two colorful objects, causing them to compress and flatten. The background is a green wall, and the machine appears to be in a workshop or industrial setting.
+The video shows a large, yellow and black striped cylinder flattening objects as if they were under a hydraulic press. The objects being flattened are pink and are being crushed into small pieces. The background is a green wall with a yellow sign.
+The video shows a machine with a yellow and black striped cylinder pressing down on two colorful objects, which are flattened as if they were under a hydraulic press. The machine is positioned on a metal platform, and the background is a green wall.
+A green cube is being compressed by a hydraulic press, which flattens the object as if it were under a hydraulic press. The press is shown in action, with the cube being squeezed into a smaller shape.
+A pink, sparkly ball is being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
+A red cabbage is being crushed by a hydraulic press, which flattens the objects as if they were under a hydraulic press. The press is shown in action, compressing the cabbage into a smaller, more compact form.
+A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and pulp.
+A large metal press is shown compressing a stack of burgers, causing them to be flattened and crushed into a pile of ground meat.
+A pizza is being crushed by a hydraulic press, causing the toppings to spread out and the crust to crumble.
+A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.
+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 Oreo cookies, flattening them as if they were under a hydraulic press.

examples/training/consistency_finetune/causal_ode_init/finetune_ode_init.sh

@@ -3,24 +3,19 @@
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
 export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
 
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-# DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-extended-t2v-1-3b/"
+DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
 VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
 NUM_GPUS=1
-# export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_70k"
+  --output_dir "wan_ode_init_crush_smol"
   --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "fixed_wan_ode_init_70k_6e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
+  --wandb_run_name "wan_ode_init_crush_smol"
   --max_train_steps 6000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -30,7 +25,7 @@ training_args=(
   --num_width 832
   --num_frames 77
   --warp_denoising_step
-  # --enable_gradient_checkpointing_type "full"
+  --enable_gradient_checkpointing_type "full"
 )
 
 # Parallel arguments
@@ -82,7 +77,6 @@ miscellaneous_args=(
   --dit_precision "fp32"
   --num_euler_timesteps 50
   --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
 )
 
 # If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t

examples/training/consistency_finetune/causal_ode_init/preprocess_wan_data_crush_smol.sh

@@ -3,8 +3,8 @@
 GPU_NUM=1 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_TYPE="wan"
-DATA_MERGE_PATH="data/crush-smol_single/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_single/"
+DATA_MERGE_PATH="$(dirname "$0")/crush_smol_prompts.txt"
+OUTPUT_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
 
 torchrun --nproc_per_node=$GPU_NUM \
     fastvideo/pipelines/preprocess/v1_preprocess.py \
@@ -18,7 +18,7 @@ torchrun --nproc_per_node=$GPU_NUM \
     --dataloader_num_workers 0 \
     --output_dir=$OUTPUT_DIR \
     --train_fps 16 \
-    --samples_per_file 1 \
-    --flush_frequency 1 \
+    --samples_per_file 8 \
+    --flush_frequency 8 \
     --video_length_tolerance_range 5 \
     --preprocess_task "ode_trajectory" 

examples/training/consistency_finetune/ode_init/finetune_ode_init_16k_vidprom.slurm

@@ -1,132 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=4e6B8_16kFV_no_warp_ode_vidprom
-#SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=ode_vidprom16k_no_warp/ode_vidprom8b16k_4e-6.out
-#SBATCH --error=ode_vidprom16k_no_warp/ode_vidprom8b16k_4e-6.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate will-fv2
-
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-16k-t2v-1-3b/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
-NUM_GPUS=2
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "vidprom_8b16k_test_no_warp_4e-6"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "vidprom_8b16k_wan_ode_init_4e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6001
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --dmd_denoising_steps "1000,750,500,0"
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 50
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 4e-6
-  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_2k_vidprom.slurm

@@ -1,131 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=ode_vidprom2k
-#SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=ode_vidprom2k_output/ode_vidprom2k.out
-#SBATCH --error=ode_vidprom2k_output/ode_vidprom2k.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate will-fv2
-
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
-NUM_GPUS=8
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_vidprom2k"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "vidprom2k_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6001
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 8
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-6
-  --mixed_precision "bf16"
-  --checkpointing_steps 2000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_crush.slurm

@@ -1,132 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=ode_crush
-#SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=ode_crush_output/ode_crush.out
-#SBATCH --error=ode_crush_output/ode_crush.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate will-fv2
-
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
-NUM_GPUS=2
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_warp_2"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "2warp_fixed_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  # --warp_denoising_step
-  --max_train_steps 6001
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-6
-  --mixed_precision "bf16"
-  --checkpointing_steps 2000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_overfit.sh

@@ -1,98 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/weka/home/hao.zhang/wl/FastVideo2/data/crush-smol_processed_t2v_1_3b_ode_init_single"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_crush_smol"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "overfitwan_ode_init_crush_smol"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 2001
-  # --warp_denoising_step
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh

@@ -6,8 +6,8 @@ export TOKENIZERS_PARALLELISM=false
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
 
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_old"
-VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
 NUM_GPUS=4
 # export CUDA_VISIBLE_DEVICES=4,5
 
@@ -52,7 +52,7 @@ dataset_args=(
 validation_args=(
   --log_validation 
   --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 50
+  --validation_steps 200
   --validation_sampling_steps "50" 
   --validation_guidance_scale "6.0"
 )

examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v.sh

@@ -4,7 +4,7 @@ GPU_NUM=1 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_TYPE="wan"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_old/"
+OUTPUT_DIR="data/crush-smol_processed_t2v/"
 
 torchrun --nproc_per_node=$GPU_NUM \
     fastvideo/pipelines/preprocess/v1_preprocess.py \

examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v_new.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-GPU_NUM=1 # 2,4,8
+GPU_NUM=2 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATASET_PATH="data/crush-smol/"
 OUTPUT_DIR="data/crush-smol_processed_t2v/"
@@ -10,10 +10,11 @@ torchrun --nproc_per_node=$GPU_NUM \
     --model_path $MODEL_PATH \
     --mode preprocess \
     --workload_type t2v \
+    --preprocess.video_loader_type torchvision \
     --preprocess.dataset_type merged \
     --preprocess.dataset_path $DATASET_PATH \
     --preprocess.dataset_output_dir $OUTPUT_DIR \
-    --preprocess.preprocess_video_batch_size 8 \
+    --preprocess.preprocess_video_batch_size 2 \
     --preprocess.dataloader_num_workers 0 \
     --preprocess.max_height 480 \
     --preprocess.max_width 832 \

examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json

@@ -28,4 +28,4 @@
       "num_frames": 77
     }
   ]
-}
+}

examples/training/t2v_to_i2v/wan_t2v/finetune_t2v_to_i2v.sh

@@ -1,94 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_old"
-VALIDATION_DATASET_FILE="examples/datasets/crush_smol/validation.json"
-NUM_GPUS=8
-# export CUDA_VISIBLE_DEVICES=4,5
-
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_t2v_i2v_finetune"
-  --output_dir "checkpoints/wan_t2v_i2v_finetune"
-  --max_train_steps 5000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 2
-  --num_latent_t 20
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS 
-  --sp_size 4
-  --tp_size 1
-  --hsdp_replicate_dim 2
-  --hsdp_shard_dim 4
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path $DATA_DIR
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation 
-  --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 50
-  --validation_sampling_steps "50" 
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-  --t2v_as_i2v_task True
-  # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500"
-)
-
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_t2v_i2v_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/t2v_to_i2v/wan_t2v/preprocess_wan_data_t2v.sh

@@ -1,24 +0,0 @@
-#!/bin/bash
-
-GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-MODEL_TYPE="wan"
-DATA_MERGE_PATH="data/crush-smol/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_i2v_1_3b/"
-
-torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
-    --model_path $MODEL_PATH \
-    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 2 \
-    --seed 42 \
-    --max_height 480 \
-    --max_width 832 \
-    --num_frames 77 \
-    --dataloader_num_workers 0 \
-    --output_dir=$OUTPUT_DIR \
-    --train_fps 16 \
-    --samples_per_file 8 \
-    --flush_frequency 8 \
-    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v_ode_trajectory" 

fastvideo/attention/backends/vmoba.py

@@ -0,0 +1,214 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import re
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+from flash_attn.bert_padding import pad_input
+
+from csrc.attn.vmoba_attn.vmoba import (moba_attn_varlen, process_moba_input,
+                                        process_moba_output)
+from fastvideo.attention.backends.abstract import (AttentionBackend,
+                                                   AttentionImpl,
+                                                   AttentionMetadata,
+                                                   AttentionMetadataBuilder)
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class VMOBAAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_name() -> str:
+        return "VMOBA_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["VMOBAAttentionImpl"]:
+        return VMOBAAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["VideoMobaAttentionMetadata"]:
+        return VideoMobaAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["VideoMobaAttentionMetadataBuilder"]:
+        return VideoMobaAttentionMetadataBuilder
+
+
+@dataclass
+class VideoMobaAttentionMetadata(AttentionMetadata):
+    current_timestep: int
+
+    temporal_chunk_size: int
+    temporal_topk: int
+    spatial_chunk_size: tuple[int, int]
+    spatial_topk: int
+    st_chunk_size: tuple[int, int, int]
+    st_topk: int
+
+    moba_select_mode: str
+    moba_threshold: float
+    moba_threshold_type: str
+    patch_resolution: list[int]
+
+    first_full_step: int = 12
+    first_full_layer: int = 0
+    # temporal_layer -> spatial_layer -> st_layer
+    temporal_layer: int = 1
+    spatial_layer: int = 1
+    st_layer: int = 1
+
+
+class VideoMobaAttentionMetadataBuilder(AttentionMetadataBuilder):
+
+    def __init__(self):
+        pass
+
+    def prepare(self):
+        pass
+
+    def build(  # type: ignore
+        self,
+        current_timestep: int,
+        raw_latent_shape: tuple[int, int, int],
+        patch_size: tuple[int, int, int],
+        temporal_chunk_size: int,
+        temporal_topk: int,
+        spatial_chunk_size: tuple[int, int],
+        spatial_topk: int,
+        st_chunk_size: tuple[int, int, int],
+        st_topk: int,
+        moba_select_mode: str = 'threshold',
+        moba_threshold: float = 0.25,
+        moba_threshold_type: str = 'query_head',
+        device: torch.device = None,
+        first_full_layer: int = 0,
+        first_full_step: int = 12,
+        temporal_layer: int = 1,
+        spatial_layer: int = 1,
+        st_layer: int = 1,
+        **kwargs,
+    ) -> VideoMobaAttentionMetadata:
+        if device is None:
+            device = torch.device("cpu")
+        assert raw_latent_shape[0] % patch_size[0] == 0 and raw_latent_shape[
+            1] % patch_size[1] == 0 and raw_latent_shape[2] % patch_size[
+                2] == 0, f"spatial patch_resolution {raw_latent_shape} should be divisible by patch_size {patch_size}"
+        patch_resolution = [
+            t // pt for t, pt in zip(raw_latent_shape, patch_size, strict=False)
+        ]
+
+        return VideoMobaAttentionMetadata(
+            current_timestep=current_timestep,
+            temporal_chunk_size=temporal_chunk_size,
+            temporal_topk=temporal_topk,
+            spatial_chunk_size=spatial_chunk_size,
+            spatial_topk=spatial_topk,
+            st_chunk_size=st_chunk_size,
+            st_topk=st_topk,
+            moba_select_mode=moba_select_mode,
+            moba_threshold=moba_threshold,
+            moba_threshold_type=moba_threshold_type,
+            patch_resolution=patch_resolution,
+            first_full_layer=first_full_layer,
+            first_full_step=first_full_step,
+            temporal_layer=temporal_layer,
+            spatial_layer=spatial_layer,
+            st_layer=st_layer,
+        )
+
+
+class VMOBAAttentionImpl(AttentionImpl):
+
+    def __init__(self,
+                 num_heads,
+                 head_size,
+                 softmax_scale,
+                 causal=False,
+                 num_kv_heads=None,
+                 prefix="",
+                 **extra_impl_args) -> None:
+        self.prefix = prefix
+        self.layer_idx = self._get_layer_idx(prefix)
+
+    def _get_layer_idx(self, prefix: str) -> int | None:
+        match = re.search(r"blocks\.(\d+)", prefix)
+        if not match:
+            raise ValueError(f"Invalid prefix: {prefix}")
+        return int(match.group(1))
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        """
+        query: [B, L, H, D]
+        key:   [B, L, H, D]
+        value: [B, L, H, D]
+        attn_metadata: AttentionMetadata
+        """
+        batch_size, sequence_length, num_heads, head_dim = query.shape
+
+        # select chunk type according to layer idx:
+        loop_layer_num = attn_metadata.temporal_layer + attn_metadata.spatial_layer + attn_metadata.st_layer
+        moba_layer = self.layer_idx - attn_metadata.first_full_layer
+        if moba_layer % loop_layer_num < attn_metadata.temporal_layer:
+            moba_chunk_size = attn_metadata.temporal_chunk_size
+            moba_topk = attn_metadata.temporal_topk
+        elif moba_layer % loop_layer_num < attn_metadata.temporal_layer + attn_metadata.spatial_layer:
+            moba_chunk_size = attn_metadata.spatial_chunk_size
+            moba_topk = attn_metadata.spatial_topk
+        elif moba_layer % loop_layer_num < attn_metadata.temporal_layer + attn_metadata.spatial_layer + attn_metadata.st_layer:
+            moba_chunk_size = attn_metadata.st_chunk_size
+            moba_topk = attn_metadata.st_topk
+
+        # torch.distributed.breakpoint()
+        query, chunk_size = process_moba_input(query,
+                                               attn_metadata.patch_resolution,
+                                               moba_chunk_size)
+        key, chunk_size = process_moba_input(key,
+                                             attn_metadata.patch_resolution,
+                                             moba_chunk_size)
+        value, chunk_size = process_moba_input(value,
+                                               attn_metadata.patch_resolution,
+                                               moba_chunk_size)
+        max_seqlen = query.shape[1]
+        indices_q = torch.arange(0,
+                                 query.shape[0] * query.shape[1],
+                                 device=query.device)
+        cu_seqlens = torch.arange(0,
+                                  query.shape[0] * query.shape[1] + 1,
+                                  query.shape[1],
+                                  dtype=torch.int32,
+                                  device=query.device)
+        query = rearrange(query, "b s ... -> (b s) ...")
+        key = rearrange(key, "b s ... -> (b s) ...")
+        value = rearrange(value, "b s ... -> (b s) ...")
+
+        # current_timestep=attn_metadata.current_timestep
+        hidden_states = moba_attn_varlen(
+            query,
+            key,
+            value,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            moba_chunk_size=chunk_size,
+            moba_topk=moba_topk,
+            select_mode=attn_metadata.moba_select_mode,
+            simsum_threshold=attn_metadata.moba_threshold,
+            threshold_type=attn_metadata.moba_threshold_type,
+        )
+        hidden_states = pad_input(hidden_states, indices_q, batch_size,
+                                  sequence_length)
+        hidden_states = process_moba_output(hidden_states,
+                                            attn_metadata.patch_resolution,
+                                            moba_chunk_size)
+
+        return hidden_states

fastvideo/configs/backend/vmoba/wan_1.3B_77_448_832.json

@@ -0,0 +1,16 @@
+{
+    "temporal_chunk_size": 2,
+    "temporal_topk": 2,
+    "spatial_chunk_size": [4, 13],
+    "spatial_topk": 6,
+    "st_chunk_size": [4, 4, 13],
+    "st_topk": 18,
+    "moba_select_mode": "topk",
+    "moba_threshold": 0.25,
+    "moba_threshold_type": "query_head",
+    "first_full_layer": 0,
+    "first_full_step": 12,
+    "temporal_layer": 1,
+    "spatial_layer": 1,
+    "st_layer": 1
+}

fastvideo/configs/backend/vmoba/wan_1.3B_77_480_832.json

@@ -0,0 +1,16 @@
+{
+    "temporal_chunk_size": 2,
+    "temporal_topk": 3,
+    "spatial_chunk_size": [3, 4],
+    "spatial_topk": 20,
+    "st_chunk_size": [4, 6, 4],
+    "st_topk": 15,
+    "moba_select_mode": "threshold",
+    "moba_threshold": 0.25,
+    "moba_threshold_type": "query_head",
+    "first_full_layer": 0,
+    "first_full_step": 12,
+    "temporal_layer": 1,
+    "spatial_layer": 1,
+    "st_layer": 1
+}

fastvideo/configs/configs.py

@@ -32,6 +32,29 @@ class DatasetType(str, Enum):
         return [dataset_type.value for dataset_type in cls]
 
 
+class VideoLoaderType(str, Enum):
+    """
+    Enumeration for different video loaders.
+    """
+    TORCHCODEC = "torchcodec"
+    TORCHVISION = "torchvision"
+
+    @classmethod
+    def from_string(cls, value: str) -> "VideoLoaderType":
+        """Convert string to VideoLoader enum."""
+        try:
+            return cls(value.lower())
+        except ValueError:
+            raise ValueError(
+                f"Invalid video loader: {value}. Must be one of: {', '.join([m.value for m in cls])}"
+            ) from None
+
+    @classmethod
+    def choices(cls) -> list[str]:
+        """Get all available choices as strings for argparse."""
+        return [video_loader.value for video_loader in cls]
+
+
 @dataclasses.dataclass
 class PreprocessConfig:
     """Configuration for preprocessing operations."""
@@ -51,6 +74,7 @@ class PreprocessConfig:
     flush_frequency: int = 256
 
     # Video processing parameters
+    video_loader_type: VideoLoaderType = VideoLoaderType.TORCHCODEC
     max_height: int = 480
     max_width: int = 848
     num_frames: int = 163
@@ -120,6 +144,12 @@ class PreprocessConfig:
                                      help="How often to save to parquet files")
 
         # Video processing parameters
+        preprocess_args.add_argument(
+            f"--{prefix_with_dot}video-loader-type",
+            type=str,
+            choices=VideoLoaderType.choices(),
+            default=PreprocessConfig.video_loader_type.value,
+            help="Type of the video loader")
         preprocess_args.add_argument(f"--{prefix_with_dot}max-height",
                                      type=int,
                                      default=PreprocessConfig.max_height,
@@ -174,6 +204,10 @@ class PreprocessConfig:
         if 'dataset_type' in kwargs and isinstance(kwargs['dataset_type'], str):
             kwargs['dataset_type'] = DatasetType.from_string(
                 kwargs['dataset_type'])
+        if 'video_loader_type' in kwargs and isinstance(
+                kwargs['video_loader_type'], str):
+            kwargs['video_loader_type'] = VideoLoaderType.from_string(
+                kwargs['video_loader_type'])
 
         preprocess_config = cls()
         if not update_config_from_args(

fastvideo/configs/models/dits/base.py

@@ -15,9 +15,13 @@ class DiTArchConfig(ArchConfig):
     reverse_param_names_mapping: dict = field(default_factory=dict)
     lora_param_names_mapping: dict = field(default_factory=dict)
     _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
-        AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
-        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
-        AttentionBackendEnum.VIDEO_SPARSE_ATTN)
+        AttentionBackendEnum.SLIDING_TILE_ATTN,
+        AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN,
+        AttentionBackendEnum.VMOBA_ATTN,
+    )
 
     hidden_size: int = 0
     num_attention_heads: int = 0

fastvideo/configs/pipelines/__init__.py

@@ -4,12 +4,13 @@ from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.registry import (
     get_pipeline_config_cls_from_name)
 from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
-from fastvideo.configs.pipelines.wan import (WanI2V480PConfig, WanI2V720PConfig,
+from fastvideo.configs.pipelines.wan import (SelfForcingWanT2V480PConfig,
+                                             WanI2V480PConfig, WanI2V720PConfig,
                                              WanT2V480PConfig, WanT2V720PConfig)
 
 __all__ = [
     "HunyuanConfig", "FastHunyuanConfig", "PipelineConfig",
     "SlidingTileAttnConfig", "WanT2V480PConfig", "WanI2V480PConfig",
     "WanT2V720PConfig", "WanI2V720PConfig", "StepVideoT2VConfig",
-    "get_pipeline_config_cls_from_name"
+    "SelfForcingWanT2V480PConfig", "get_pipeline_config_cls_from_name"
 ]

fastvideo/configs/pipelines/base.py

@@ -45,8 +45,6 @@ class PipelineConfig:
     embedded_cfg_scale: float = 6.0
     flow_shift: float | None = None
     disable_autocast: bool = False
-    ti2v_task: bool = False
-    t2v_as_i2v_task: bool = False
 
     # Model configuration
     dit_config: DiTConfig = field(default_factory=DiTConfig)
@@ -87,6 +85,9 @@ class PipelineConfig:
     # DMD parameters
     dmd_denoising_steps: list[int] | None = field(default=None)
 
+    # Wan2.2 TI2V parameters
+    ti2v_task: bool = False
+
     # Compilation
     # enable_torch_compile: bool = False
 
@@ -213,24 +214,6 @@ class PipelineConfig:
             "Comma-separated list of denoising steps (e.g., '1000,757,522')",
         )
 
-        # TI2V task
-        parser.add_argument(
-            f"--{prefix_with_dot}ti2v-task",
-            action=StoreBoolean,
-            dest=f"{prefix_with_dot.replace('-', '_')}ti2v_task",
-            default=PipelineConfig.ti2v_task,
-            help="Enable TI2V",
-        )
-
-        # T2V to I2V task
-        parser.add_argument(
-            f"--{prefix_with_dot}t2v-as-i2v-task",
-            action=StoreBoolean,
-            dest=f"{prefix_with_dot.replace('-', '_')}t2v_as_i2v_task",
-            default=PipelineConfig.t2v_as_i2v_task,
-            help="Enable T2V to I2V task",
-        )
-
         # Add VAE configuration arguments
         from fastvideo.configs.models.vaes.base import VAEConfig
         VAEConfig.add_cli_args(parser, prefix=f"{prefix_with_dot}vae-config")
@@ -262,9 +245,7 @@ class PipelineConfig:
         """
         from fastvideo.configs.pipelines.registry import (
             get_pipeline_config_cls_from_name)
-        logger.info("WTF model_path: %s", model_path)
         pipeline_config_cls = get_pipeline_config_cls_from_name(model_path)
-        logger.info("pipeline_config_cls: %s", pipeline_config_cls)
 
         return cast(PipelineConfig, pipeline_config_cls(model_path=model_path))
 

fastvideo/configs/pipelines/registry.py

@@ -11,9 +11,9 @@ from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
 # isort: off
 from fastvideo.configs.pipelines.wan import (
     FastWan2_1_T2V_480P_Config, FastWan2_2_TI2V_5B_Config,
-    SelfForcingWanT2V480PConfig, Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config,
-    Wan2_2_TI2V_5B_Config, WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig,
-    WanT2V720PConfig, SelfForcingWanT2V480PConfig)
+    Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config, Wan2_2_TI2V_5B_Config,
+    WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig, WanT2V720PConfig,
+    SelfForcingWanT2V480PConfig)
 # isort: on
 from fastvideo.logger import init_logger
 from fastvideo.utils import (maybe_download_model_index,
@@ -48,8 +48,8 @@ PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
     "wanpipeline": lambda id: "wanpipeline" in id.lower(),
     "wanimagetovideo": lambda id: "wanimagetovideo" in id.lower(),
     "wandmdpipeline": lambda id: "wandmdpipeline" in id.lower(),
-    "stepvideo": lambda id: "stepvideo" in id.lower(),
     "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
+    "stepvideo": lambda id: "stepvideo" in id.lower(),
     # Add other pipeline architecture detectors
 }
 
@@ -61,8 +61,8 @@ PIPELINE_FALLBACK_CONFIG: dict[str, type[PipelineConfig]] = {
     WanT2V480PConfig,  # Base Wan config as fallback for any Wan variant
     "wanimagetovideo": WanI2V480PConfig,
     "wandmdpipeline": FastWan2_1_T2V_480P_Config,
-    "stepvideo": StepVideoT2VConfig,
     "wancausaldmdpipeline": SelfForcingWanT2V480PConfig,
+    "stepvideo": StepVideoT2VConfig
     # Other fallbacks by architecture
 }
 

fastvideo/configs/pipelines/wan.py

@@ -12,13 +12,13 @@ from fastvideo.configs.models.vaes import WanVAEConfig
 from fastvideo.configs.pipelines.base import PipelineConfig
 
 
-def t5_postprocess_text(outputs: BaseEncoderOutput) -> torch.tensor:
-    mask: torch.tensor = outputs.attention_mask
-    hidden_state: torch.tensor = outputs.last_hidden_state
+def t5_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
+    mask: torch.Tensor = outputs.attention_mask
+    hidden_state: torch.Tensor = outputs.last_hidden_state
     seq_lens = mask.gt(0).sum(dim=1).long()
     assert torch.isnan(hidden_state).sum() == 0
     prompt_embeds = [u[:v] for u, v in zip(hidden_state, seq_lens, strict=True)]
-    prompt_embeds_tensor: torch.tensor = torch.stack([
+    prompt_embeds_tensor: torch.Tensor = torch.stack([
         torch.cat([u, u.new_zeros(512 - u.size(0), u.size(1))])
         for u in prompt_embeds
     ],
@@ -39,12 +39,12 @@ class WanT2V480PConfig(PipelineConfig):
     vae_sp: bool = False
 
     # Denoising stage
-    flow_shift: float | None = 8.0
+    flow_shift: float | None = 3.0
 
     # Text encoding stage
     text_encoder_configs: tuple[EncoderConfig, ...] = field(
         default_factory=lambda: (T5Config(), ))
-    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.tensor],
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
                                   ...] = field(default_factory=lambda:
                                                (t5_postprocess_text, ))
 
@@ -68,7 +68,7 @@ class WanT2V720PConfig(WanT2V480PConfig):
     # WanConfig-specific parameters with defaults
 
     # Denoising stage
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
 
 
 @dataclass
@@ -94,7 +94,7 @@ class WanI2V720PConfig(WanI2V480PConfig):
     # WanConfig-specific parameters with defaults
 
     # Denoising stage
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
 
 
 @dataclass
@@ -104,7 +104,7 @@ class FastWan2_1_T2V_480P_Config(WanT2V480PConfig):
     # WanConfig-specific parameters with defaults
 
     # Denoising stage
-    flow_shift: int = 8
+    flow_shift: float | None = 8.0
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 757, 522])
 
@@ -115,7 +115,7 @@ class FastWan2_1_T2V_480P_Config(WanT2V480PConfig):
 
 @dataclass
 class Wan2_2_TI2V_5B_Config(WanT2V480PConfig):
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
     ti2v_task: bool = True
 
     def __post_init__(self) -> None:
@@ -125,7 +125,7 @@ class Wan2_2_TI2V_5B_Config(WanT2V480PConfig):
 
 @dataclass
 class FastWan2_2_TI2V_5B_Config(Wan2_2_TI2V_5B_Config):
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 757, 522])
 
@@ -149,3 +149,4 @@ class SelfForcingWanT2V480PConfig(WanT2V480PConfig):
     flow_shift: float | None = 5.0
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 750, 500, 250])
+    warp_denoising_step: bool = True

fastvideo/configs/sample/base.py

@@ -193,6 +193,13 @@ class SamplingParam:
             default=SamplingParam.image_path,
             help="Path to input image for image-to-video generation",
         )
+        parser.add_argument(
+            "--moba-config-path",
+            type=str,
+            default=None,
+            help=
+            "Path to a JSON file containing V-MoBA specific configurations.",
+        )
         parser.add_argument(
             "--return-trajectory-latents",
             action="store_true",

fastvideo/dataset/__init__.py

@@ -4,7 +4,7 @@ from torchvision.transforms import Lambda
 
 from fastvideo.dataset.parquet_dataset_map_style import (
     build_parquet_map_style_dataloader)
-from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset
+from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset, TextDataset
 from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255,
                                          TemporalRandomCrop)
 from fastvideo.dataset.validation_dataset import ValidationDataset
@@ -37,9 +37,15 @@ def getdataset(args) -> VideoCaptionMergedDataset:
                                      temporal_sample=temporal_sample,
                                      transform_topcrop=transform_topcrop,
                                      seed=args.seed)
+                                    
+
+def gettextdataset(args) -> TextDataset:
+    return TextDataset(data_merge_path=args.data_merge_path,
+                       args=args,
+                       seed=args.seed)
 
 
 __all__ = [
     "build_parquet_map_style_dataloader", "ValidationDataset",
-    "VideoCaptionMergedDataset"
+    "VideoCaptionMergedDataset", "TextDataset"
 ]

fastvideo/dataset/dataloader/parquet_io.py

@@ -0,0 +1,264 @@
+"""
+Utilities for converting preprocessing records (dicts) into Arrow tables and
+writing Parquet datasets in fixed-size chunks.
+
+This module centralizes table construction and Parquet file writing so
+pipelines only need to define their PyArrow schema and produce per-sample
+record dictionaries.
+
+Key APIs:
+- records_to_table(records, schema): Safely convert a list of dictionaries into
+  a pa.Table, casting to the provided schema.
+- ParquetDatasetWriter: Buffer tables and flush to a directory as multiple
+  Parquet files with a fixed number of rows per file. Uses temporary files and
+  atomic rename to avoid partially written outputs.
+"""
+
+from __future__ import annotations
+
+import multiprocessing
+import os
+from concurrent.futures import ProcessPoolExecutor
+from typing import Any
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+
+def records_to_table(records: list[dict[str, Any]], schema: pa.Schema) -> pa.Table:
+    """Build a PyArrow table from Python record dicts using an explicit schema.
+
+    Arrow will cast values to the target schema when possible (e.g., promoting
+    Python ints/floats to pa.int64/pa.float64), eliminating hand-written per-
+    field array construction.
+
+    Args:
+        records: List of dictionaries, each representing one row. Keys must
+            match schema field names.
+        schema: Target PyArrow schema. Controls field names and types.
+
+    Returns:
+        pa.Table: In-memory table matching the provided schema. If ``records``
+        is empty, returns an empty table with the given schema.
+    """
+    if not records:
+        return pa.table({}, schema=schema)
+    return pa.Table.from_pylist(records, schema=schema)
+
+
+class ParquetDatasetWriter:
+    """Accumulate tables and flush them to a Parquet directory in fixed-size chunks.
+
+    Behavior:
+    - Writes files under worker-specific subdirectories for parallelism.
+    - Uses temporary files and atomic rename to avoid partial files being left
+      behind on failure.
+    - Only full chunks of ``samples_per_file`` rows are written on each flush;
+      any remainder rows are re-buffered for the next flush.
+
+    Note:
+    - Instances are not meant to be shared across processes. Create one writer
+      per process if using multiprocessing.
+    """
+
+    def __init__(self, out_dir: str, samples_per_file: int, compression: str = "zstd") -> None:
+        """Initialize the dataset writer.
+
+        Args:
+            out_dir: Output directory where Parquet files will be written.
+            samples_per_file: Fixed number of rows per Parquet file.
+            compression: Compression codec passed to ``pyarrow.parquet.write_table``
+                (e.g., ``"zstd"``, ``"snappy"``, ``"gzip"``).
+        """
+        self.out_dir = out_dir
+        self.samples_per_file = max(int(samples_per_file), 1)
+        self.compression = compression
+        os.makedirs(self.out_dir, exist_ok=True)
+        self._tables: list[pa.Table] = []
+
+    def append_table(self, table: pa.Table) -> None:
+        """Append a non-empty table to the internal buffer.
+
+        Args:
+            table: A ``pa.Table`` to buffer. Empty or ``None`` tables are ignored.
+        """
+        if table is None or len(table) == 0:
+            return
+        self._tables.append(table)
+
+    def _combine(self) -> pa.Table | None:
+        """Combine all buffered tables into a single table, if any.
+
+        Returns:
+            A concatenated table, a single table if only one was buffered, or
+            ``None`` if no tables are buffered.
+        """
+        if not self._tables:
+            return None
+        if len(self._tables) == 1:
+            return self._tables[0]
+        return pa.concat_tables(self._tables, promote_options='none')
+
+    def flush(self, num_workers: int | None = None, write_remainder: bool = False) -> int:
+        """Write accumulated tables to disk and clear the written portion.
+
+        Only complete chunks of size ``samples_per_file`` are written. Any
+        remainder rows are kept buffered for the next flush.
+
+        Args:
+            num_workers: Optional override for the number of parallel workers
+                used to write chunks. Defaults to ``min(cpu_count, chunks)``.
+            write_remainder: If True, also write any leftover rows (< samples_per_file)
+                as a final small Parquet file (useful for the last flush at the
+                end of preprocessing).
+
+        Returns:
+            int: Number of rows successfully written in this flush call.
+        """
+        combined = self._combine()
+        self._tables = []
+        if combined is None or len(combined) == 0:
+            return 0
+
+        num_samples = len(combined)
+        total_chunks = num_samples // self.samples_per_file
+        if total_chunks == 0:
+            if not write_remainder:
+                # Not enough to form a full chunk; keep buffered for next round
+                # Re-buffer and return 0 written
+                self._tables = [combined]
+                return 0
+            # Last flush: write the small remainder as a final file in worker_0
+            worker_dir = os.path.join(self.out_dir, "worker_0")
+            os.makedirs(worker_dir, exist_ok=True)
+            # Determine next index
+            num_parquets = 0
+            for _, _, files in os.walk(worker_dir):
+                for file in files:
+                    if file.endswith('.parquet'):
+                        num_parquets += 1
+            chunk_path = os.path.join(worker_dir, f"data_chunk_{num_parquets}.parquet")
+            temp_path = chunk_path + '.tmp'
+            pq.write_table(combined, temp_path, compression=self.compression)
+            if os.path.exists(chunk_path):
+                os.remove(chunk_path)
+            os.rename(temp_path, chunk_path)
+            return num_samples
+
+        # Only write full chunks; keep remainder for next flush
+        written_rows = total_chunks * self.samples_per_file
+        remainder = num_samples - written_rows
+
+        table_to_write = combined.slice(0, written_rows)
+        remainder_table = combined.slice(written_rows, remainder) if remainder > 0 else None
+        if remainder_table is not None and len(remainder_table) > 0:
+            if write_remainder:
+                # Write the remainder as a final small file (worker_0)
+                worker_dir = os.path.join(self.out_dir, "worker_0")
+                os.makedirs(worker_dir, exist_ok=True)
+                num_parquets = 0
+                for _, _, files in os.walk(worker_dir):
+                    for file in files:
+                        if file.endswith('.parquet'):
+                            num_parquets += 1
+                remainder_path = os.path.join(worker_dir,
+                                              f"data_chunk_{num_parquets}.parquet")
+                temp_path = remainder_path + '.tmp'
+                pq.write_table(remainder_table,
+                               temp_path,
+                               compression=self.compression)
+                if os.path.exists(remainder_path):
+                    os.remove(remainder_path)
+                os.rename(temp_path, remainder_path)
+            else:
+                self._tables = [remainder_table]
+
+        # Parallel write by chunk ranges
+        if num_workers is None:
+            num_workers = min(multiprocessing.cpu_count(), max(total_chunks, 1))
+        num_workers = max(int(num_workers), 1)
+        chunks_per_worker = (total_chunks + num_workers - 1) // num_workers
+
+        work_ranges: list[tuple[int, int, pa.Table, int, str, int, str]] = []
+        for worker_id in range(num_workers):
+            start_chunk = worker_id * chunks_per_worker
+            end_chunk = min((worker_id + 1) * chunks_per_worker, total_chunks)
+            if start_chunk < end_chunk:
+                work_ranges.append(
+                    (
+                        start_chunk,
+                        end_chunk,
+                        table_to_write,
+                        worker_id,
+                        self.out_dir,
+                        self.samples_per_file,
+                        self.compression,
+                    )
+                )
+
+        written_total = 0
+        if len(work_ranges) == 1:
+            written_total += _process_chunk_range(work_ranges[0])
+            return written_total
+
+        with ProcessPoolExecutor(max_workers=num_workers) as executor:
+            futures = [executor.submit(_process_chunk_range, args) for args in work_ranges]
+            for f in futures:
+                written_total += f.result()
+        return written_total + (len(remainder_table) if write_remainder and remainder_table is not None else 0)
+
+
+def _process_chunk_range(args: Any) -> int:
+    """Worker function to write a contiguous range of chunk files.
+
+    Args:
+        args: Tuple containing
+            - start_chunk (int): inclusive start chunk index
+            - end_chunk (int): exclusive end chunk index
+            - table (pa.Table): concatenated table containing all rows to write
+            - worker_id (int): numeric worker identifier
+            - output_dir (str): base output directory
+            - samples_per_file (int): rows per chunk file
+            - compression (str): compression codec for Parquet
+
+    Returns:
+        int: Total number of rows written by this worker.
+    """
+    start_chunk, end_chunk, table, worker_id, output_dir, samples_per_file, compression = args
+    total_written = 0
+    num_samples = len(table)
+
+    worker_dir = os.path.join(output_dir, f"worker_{worker_id}")
+    os.makedirs(worker_dir, exist_ok=True)
+
+    # Offset to continue numbering if files exist
+    num_parquets = 0
+    for root, _, files in os.walk(worker_dir):
+        for file in files:
+            if file.endswith('.parquet'):
+                num_parquets += 1
+
+    for i in range(start_chunk, end_chunk):
+        start_sample = i * samples_per_file
+        end_sample = min((i + 1) * samples_per_file, num_samples)
+        if end_sample <= start_sample:
+            continue
+        chunk = table.slice(start_sample, end_sample - start_sample)
+
+        chunk_path = os.path.join(worker_dir, f"data_chunk_{i + num_parquets}.parquet")
+        temp_path = chunk_path + '.tmp'
+        try:
+            pq.write_table(chunk, temp_path, compression=compression)
+            if os.path.exists(chunk_path):
+                os.remove(chunk_path)
+            os.rename(temp_path, chunk_path)
+            total_written += len(chunk)
+        except Exception:
+            if os.path.exists(temp_path):
+                os.remove(temp_path)
+            raise
+
+    return total_written
+
+
+

fastvideo/dataset/dataloader/schema.py

@@ -50,6 +50,7 @@ pyarrow_schema_i2v = pa.schema([
     pa.field("fps", pa.float64()),
 ])
 
+
 pyarrow_schema_t2v = pa.schema([
     pa.field("id", pa.string()),
     # --- Image/Video VAE latents ---
@@ -79,15 +80,9 @@ pyarrow_schema_t2v = pa.schema([
     pa.field("fps", pa.float64()),
 ])
 
-pyarrow_schema_ode_trajectory = pa.schema([
+
+pyarrow_schema_ode_trajectory_text_only = pa.schema([
     pa.field("id", pa.string()),
-    # --- Image/Video VAE latents ---
-    # Tensors are stored as raw bytes with shape and dtype info for loading
-    pa.field("vae_latent_bytes", pa.binary()),
-    # e.g., [C, T, H, W] or [C, H, W]
-    pa.field("vae_latent_shape", pa.list_(pa.int64())),
-    # e.g., 'float32'
-    pa.field("vae_latent_dtype", pa.string()),
     # --- Text encoder output tensor ---
     # Tensors are stored as raw bytes with shape and dtype info for loading
     pa.field("text_embedding_bytes", pa.binary()),
@@ -95,10 +90,6 @@ pyarrow_schema_ode_trajectory = pa.schema([
     pa.field("text_embedding_shape", pa.list_(pa.int64())),
     # e.g., 'bfloat16' or 'float32'
     pa.field("text_embedding_dtype", pa.string()),
-    # I2V
-    pa.field("image_condition_latents_bytes", pa.binary()),
-    pa.field("image_condition_latents_shape", pa.list_(pa.int64())),
-    pa.field("image_condition_latents_dtype", pa.string()),
     # --- ODE Trajectory ---
     pa.field("trajectory_latents_bytes", pa.binary()),
     pa.field("trajectory_latents_shape", pa.list_(pa.int64())),
@@ -109,12 +100,5 @@ pyarrow_schema_ode_trajectory = pa.schema([
     # --- Metadata ---
     pa.field("file_name", pa.string()),
     pa.field("caption", pa.string()),
-    pa.field("media_type", pa.string()),  # 'image' or 'video'
-    pa.field("width", pa.int64()),
-    pa.field("height", pa.int64()),
-    # -- Video-specific (can be null/default for images) ---
-    # Number of frames processed (e.g., 1 for image, N for video)
-    pa.field("num_frames", pa.int64()),
-    pa.field("duration_sec", pa.float64()),
-    pa.field("fps", pa.float64()),
+    pa.field("media_type", pa.string()),  # Always 'text' for text-only
 ])

fastvideo/dataset/preprocessing_datasets.py

@@ -628,3 +628,134 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
     def load_state_dict(self, state_dict: dict[str, Any]) -> None:
         """Load state dict from checkpoint."""
         self.processed_batches = state_dict["processed_batches"]
+
+
+class TextDataset(torch.utils.data.IterableDataset,
+                  torch.distributed.checkpoint.stateful.Stateful):
+    """
+    Text-only dataset for processing prompts from a simple text file.
+    
+    Assumes that data_merge_path is a text file with one prompt per line:
+    A cat playing with a ball
+    A dog running in the park
+    A person cooking dinner
+    ...
+    
+    This dataset processes text data through text encoding stages only.
+    """
+
+    def __init__(self,
+                 data_merge_path: str,
+                 args,
+                 start_idx: int = 0,
+                 seed: int = 42):
+        self.data_merge_path = data_merge_path
+        self.start_idx = start_idx
+        self.args = args
+        self.seed = seed
+
+        # Initialize tokenizer
+        tokenizer_path = os.path.join(args.model_path, "tokenizer")
+        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path,
+                                                  cache_dir=args.cache_dir)
+
+        # Initialize text encoding stage
+        self.text_encoding_stage = TextEncodingStage(
+            tokenizer=tokenizer,
+            text_max_length=args.text_max_length,
+            cfg_rate=getattr(args, 'training_cfg_rate', 0.0),
+            seed=self.seed)
+
+        # Process text data
+        self.processed_batches = self._process_text_data()
+
+    def _load_text_data(self) -> list[str]:
+        """Load text prompts from file."""
+        prompts = []
+        with open(self.data_merge_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                line = line.strip()
+                if line:  # Skip empty lines
+                    prompts.append(line)
+        
+        logger.info(f"Loaded {len(prompts)} text prompts from {self.data_merge_path}")
+        return prompts
+
+    def _process_text_data(self) -> list[PreprocessBatch]:
+        """Process the text prompts through text encoding stage."""
+        raw_prompts = self._load_text_data()
+        processed_batches = []
+
+        for idx, prompt in enumerate(raw_prompts):
+            # Create a text-only batch with dummy path
+            batch = PreprocessBatch(
+                path=f"text_prompt_{idx}",
+                cap=[prompt],  # TextEncodingStage expects a list
+                resolution=None,
+                fps=None,
+                duration=None,
+                num_frames=0,
+                sample_frame_index=None,
+                sample_num_frames=0
+            )
+            
+            processed_batches.append(batch)
+
+        logger.info(f"Processed {len(processed_batches)} text batches")
+        return processed_batches
+
+    def __iter__(self):
+        """Iterator for the dataset."""
+        # Set up distributed sampling if needed
+        if torch.distributed.is_available() and torch.distributed.is_initialized():
+            rank = torch.distributed.get_rank()
+            world_size = torch.distributed.get_world_size()
+        else:
+            rank = 0
+            world_size = 1
+
+        # Calculate chunk for this rank
+        total_items = len(self.processed_batches)
+        items_per_rank = math.ceil(total_items / world_size)
+        start_idx = rank * items_per_rank + self.start_idx
+        end_idx = min(start_idx + items_per_rank, total_items)
+
+        # Yield items for this rank
+        for idx in range(start_idx, end_idx):
+            if idx < len(self.processed_batches):
+                yield self._get_item(idx)
+
+    def _get_item(self, idx: int) -> dict:
+        """Get a single processed text item."""
+        batch = self.processed_batches[idx]
+
+        # Apply text encoding stage
+        batch = self.text_encoding_stage.process(batch)
+
+        # Build result dictionary for text-only processing with required schema fields
+        result = {
+            "text": batch.text,
+            "input_ids": batch.input_ids,
+            "cond_mask": batch.cond_mask,
+            "path": batch.path,
+            # Required schema fields for ODE trajectory processing
+            "id": f"text_{idx}",
+            "file_name": batch.path,
+            "caption": batch.text,
+            "media_type": "text",
+            "width": 1,
+            "height": 1,
+            "num_frames": 0,
+            "duration_sec": 0.0,
+            "fps": 0.0,
+        }
+
+        return result
+
+    def state_dict(self) -> dict[str, Any]:
+        """Return state dict for checkpointing."""
+        return {"processed_batches": self.processed_batches}
+
+    def load_state_dict(self, state_dict: dict[str, Any]) -> None:
+        """Load state dict from checkpoint."""
+        self.processed_batches = state_dict["processed_batches"]

fastvideo/dataset/utils.py

@@ -3,9 +3,6 @@ from typing import Any, cast
 
 import numpy as np
 import torch
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
 
 
 def pad(t: torch.Tensor, padding_length: int) -> tuple[torch.Tensor, torch.Tensor]:

fastvideo/fastvideo_args.py

@@ -1,9 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 # Inspired by SGLang: https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/server_args.py
 """The arguments of FastVideo Inference."""
-
 import argparse
 import dataclasses
+import json
 from contextlib import contextmanager
 from dataclasses import field
 from enum import Enum
@@ -139,6 +139,10 @@ class FastVideoArgs:
     # VSA parameters
     VSA_sparsity: float = 0.0  # inference/validation sparsity
 
+    # V-MoBA parameters
+    moba_config_path: str | None = None
+    moba_config: dict[str, Any] = field(default_factory=dict)
+
     # Master port for distributed training/inference
     master_port: int | None = None
 
@@ -154,7 +158,6 @@ class FastVideoArgs:
         "transformer": True,
         "vae": True,
     })
-    override_transformer_cls_name: str | None = None
 
     # # DMD parameters
     # dmd_denoising_steps: List[int] | None = field(default=None)
@@ -167,6 +170,16 @@ class FastVideoArgs:
         return not self.inference_mode
 
     def __post_init__(self):
+        if self.moba_config_path:
+            try:
+                with open(self.moba_config_path) as f:
+                    self.moba_config = json.load(f)
+                logger.info("Loaded V-MoBA config from %s",
+                            self.moba_config_path)
+            except (FileNotFoundError, json.JSONDecodeError) as e:
+                logger.error("Failed to load V-MoBA config from %s: %s",
+                             self.moba_config_path, e)
+                raise
         self.check_fastvideo_args()
 
     @staticmethod
@@ -383,12 +396,6 @@ class FastVideoArgs:
             default=FastVideoArgs.enable_stage_verification,
             help="Enable input/output verification for pipeline stages",
         )
-        parser.add_argument(
-            "--override-transformer-cls-name",
-            type=str,
-            default=FastVideoArgs.override_transformer_cls_name,
-            help="Override transformer cls name",
-        )
         # Add pipeline configuration arguments
         PipelineConfig.add_cli_args(parser)
 
@@ -681,7 +688,6 @@ class TrainingArgs(FastVideoArgs):
     log_visualization: bool = False
     # simulate generator forward to match inference
     simulate_generator_forward: bool = False
-    warp_denoising_step: bool = False
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace) -> "TrainingArgs":
@@ -993,6 +999,15 @@ class TrainingArgs(FastVideoArgs):
         parser.add_argument("--lora-rank", type=int, help="LoRA rank")
         parser.add_argument("--lora-alpha", type=int, help="LoRA alpha")
 
+        # V-MoBA parameters
+        parser.add_argument(
+            "--moba-config-path",
+            type=str,
+            default=None,
+            help=
+            "Path to a JSON file containing V-MoBA specific configurations.",
+        )
+
         # Distillation arguments
         parser.add_argument("--generator-update-interval",
                             type=int,
@@ -1026,10 +1041,6 @@ class TrainingArgs(FastVideoArgs):
             "--simulate-generator-forward",
             action=StoreBoolean,
             help="Whether to simulate generator forward to match inference")
-        parser.add_argument(
-            "--warp-denoising-step",
-            action=StoreBoolean,
-            help="Whether to warp denoising step according to the scheduler time shift")
 
         return parser
 

fastvideo/layers/layernorm.py

@@ -9,9 +9,6 @@ import torch.nn.functional as F
 from fastvideo.layers.custom_op import CustomOp
 from fastvideo.platforms import current_platform
 
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
 
 @CustomOp.register("rms_norm")
 class RMSNorm(CustomOp):
@@ -103,13 +100,16 @@ class ScaleResidual(nn.Module):
     def forward(self, residual: torch.Tensor, x: torch.Tensor,
                 gate: torch.Tensor) -> torch.Tensor:
         """Apply gated residual connection."""
-        # logger.info("x.shape: %s", x.shape)
-        # if isinstance(gate, torch.Tensor):
-            # logger.info("gate.shape: %s", gate.shape)
-
-        num_frames = gate.shape[1]
-        frame_seqlen = x.shape[1] // num_frames
-        return residual + (x.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * gate).flatten(1, 2)
+        # x.shape: [batch_size, seq_len, inner_dim]
+        if gate.dim() == 4:
+            # gate.shape: [batch_size, num_frames, 1, inner_dim]
+            num_frames = gate.shape[1]
+            frame_seqlen = x.shape[1] // num_frames
+            return residual + (x.unflatten(
+                dim=1, sizes=(num_frames, frame_seqlen)) * gate).flatten(1, 2)
+        else:
+            # gate.shape: [batch_size, 1, inner_dim]
+            return residual + x * gate
 
 
 # adapted from Diffusers: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/normalization.py
@@ -168,7 +168,7 @@ class ScaleResidualLayerNormScaleShift(nn.Module):
             raise NotImplementedError(f"Norm type {norm_type} not implemented")
 
     def forward(self, residual: torch.Tensor, x: torch.Tensor,
-                gate: torch.Tensor, shift: torch.Tensor,
+                gate: torch.Tensor | int, shift: torch.Tensor,
                 scale: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Apply gated residual connection, followed by layernorm and 
@@ -180,34 +180,39 @@ class ScaleResidualLayerNormScaleShift(nn.Module):
             - residual value (value after residual connection 
               but before normalization)
         """
+        # x.shape: [batch_size, seq_len, inner_dim]
         # Apply residual connection with gating
-        # logger.info("x.shape: %s", x.shape)
-        if isinstance(gate, int): 
+        if isinstance(gate, int):
             # used by cross-attention, should be 1
             assert gate == 1
-            residual_output = residual + x * gate
+            residual_output = residual + x
         elif isinstance(gate, torch.Tensor):
-            # logger.info("gate.shape: %s", gate.shape)
-            if gate.dim() == 3:
-                # used by bidirectional self attention
-                residual_output = residual + x * gate
-            else:
-                assert gate.dim() == 4 
+            if gate.dim() == 4:
+                # gate.shape: [batch_size, num_frames, 1, inner_dim]
                 num_frames = gate.shape[1]
                 frame_seqlen = x.shape[1] // num_frames
-                residual_output = residual + (x.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * gate).flatten(1, 2)
-                # residual_output = residual + x * gate
+                residual_output = residual + (
+                    x.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                    gate).flatten(1, 2)
+            else:
+                # used by bidirectional self attention
+                # gate.shape: [batch_size, 1, inner_dim]
+                residual_output = residual + x * gate
         else:
             raise ValueError(f"Gate type {type(gate)} not supported")
-        # logger.info("residual_output.shape: %s", residual_output.shape)
+        # residual_output.shape: [batch_size, seq_len, inner_dim]
 
         # Apply normalization
         normalized = self.norm(residual_output)
         # Apply scale and shift
         if isinstance(scale, torch.Tensor) and scale.dim() == 4:
+            # scale.shape: [batch_size, num_frames, 1, inner_dim]
+            # shift.shape: [batch_size, num_frames, 1, inner_dim]
             num_frames = scale.shape[1]
             frame_seqlen = normalized.shape[1] // num_frames
-            modulated = (normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * (1.0 + scale) + shift).flatten(1, 2)
+            modulated = (
+                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                (1.0 + scale) + shift).flatten(1, 2)
         else:
             modulated = normalized * (1.0 + scale) + shift
         return modulated, residual_output
@@ -251,16 +256,24 @@ class LayerNormScaleShift(nn.Module):
     def forward(self, x: torch.Tensor, shift: torch.Tensor,
                 scale: torch.Tensor) -> torch.Tensor:
         """Apply ln followed by scale and shift in a single fused operation."""
+        # x.shape: [batch_size, seq_len, inner_dim]
         normalized = self.norm(x)
+        if self.compute_dtype == torch.float32:
+            normalized = normalized.float()
+
         if scale.dim() == 4:
+            # scale.shape: [batch_size, num_frames, 1, inner_dim]
             num_frames = scale.shape[1]
             frame_seqlen = normalized.shape[1] // num_frames
-            if self.compute_dtype == torch.float32:
-                return (normalized.float().unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * (1.0 + scale) + shift).flatten(1, 2).to(x.dtype)
-            else:
-                return (normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * (1.0 + scale) + shift).flatten(1, 2)
+            output = (
+                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                (1.0 + scale) + shift).flatten(1, 2)
         else:
-            if self.compute_dtype == torch.float32:
-                return (normalized.float() * (1.0 + scale) + shift).to(x.dtype)
-            else:
-                return normalized * (1.0 + scale) + shift
+            # scale.shape: [batch_size, 1, inner_dim]
+            # shift.shape: [batch_size, 1, inner_dim]
+            output = normalized * (1.0 + scale) + shift
+
+        if self.compute_dtype == torch.float32:
+            output = output.to(x.dtype)
+
+        return output

fastvideo/layers/lora/linear.py

@@ -63,7 +63,7 @@ class BaseLayerWithLoRA(nn.Module):
                             device=self.base_layer.weight.device,
                             dtype=self.base_layer.weight.dtype))
             torch.nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
-            torch.nn.init.kaiming_uniform_(self.lora_B, a=math.sqrt(5))
+            torch.nn.init.zeros_(self.lora_B)
         else:
             self.lora_A = None
             self.lora_B = None

fastvideo/models/dits/causal_wanvideo.py

@@ -244,32 +244,26 @@ class CausalWanTransformerBlock(nn.Module):
         current_start: int = 0,
         cache_start: int | None = None,
     ) -> torch.Tensor:
-        # logger.info("temb.shape: %s", temb.shape)
-        num_frames = temb.shape[1]
-        # logger.info("first hidden_states.shape: %s", hidden_states.shape)
-        # logger.info("num_frames: %s", num_frames)
+        # hidden_states.shape: [batch_size, seq_length, inner_dim]
+        # temb.shape: [batch_size, num_frames, 6, inner_dim]
         if hidden_states.dim() == 4:
             hidden_states = hidden_states.squeeze(1)
-        frame_seqlen = hidden_states.shape[1] // temb.shape[1]
-        # logger.info("frame_seqlen: %s", frame_seqlen)
+        num_frames = temb.shape[1]
+        frame_seqlen = hidden_states.shape[1] // num_frames
         bs, seq_length, _ = hidden_states.shape
         orig_dtype = hidden_states.dtype
         # assert orig_dtype != torch.float32
         e = self.scale_shift_table + temb.float()
-        # logger.info("e.shape: %s", e.shape)
+        # e.shape: [batch_size, num_frames, 6, inner_dim]
+        assert e.shape == (bs, num_frames, 6, self.hidden_dim)
         shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
             6, dim=2)
+        # *_msa.shape: [batch_size, num_frames, 1, inner_dim]
         assert shift_msa.dtype == torch.float32
 
         # 1. Self-attention
-        # print(f"hidden_states: {hidden_states.shape}")
-        # print(f"hidden_states: {scale_msa.shape}")
-        # print(f"hidden_states: {shift_msa.shape}")
-        # norm_hidden_states = (self.norm1(hidden_states.float()) *
-        #                       (1 + scale_msa) + shift_msa).to(orig_dtype)
         norm_hidden_states = (self.norm1(hidden_states.float()).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                              (1 + scale_msa) + shift_msa).flatten(1, 2).to(orig_dtype)
-        # logger.info("1 norm_hidden_states.shape: %s", norm_hidden_states.shape)
+                        (1 + scale_msa) + shift_msa).flatten(1, 2).to(orig_dtype)
         query, _ = self.to_q(norm_hidden_states)
         key, _ = self.to_k(norm_hidden_states)
         value, _ = self.to_v(norm_hidden_states)
@@ -291,8 +285,6 @@ class CausalWanTransformerBlock(nn.Module):
         null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
         norm_hidden_states, hidden_states = self.self_attn_residual_norm(
             hidden_states, attn_output, gate_msa, null_shift, null_scale)
-        # logger.info("after self_attn_residual_norm norm_hidden_states.shape: %s", norm_hidden_states.shape)
-        # logger.info("after self_attn_residual_norm hidden_states.shape: %s", hidden_states.shape)
         norm_hidden_states, hidden_states = norm_hidden_states.to(
             orig_dtype), hidden_states.to(orig_dtype)
 
@@ -303,16 +295,12 @@ class CausalWanTransformerBlock(nn.Module):
                                  crossattn_cache=crossattn_cache)
         norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
             hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
-        # logger.info("after cross_attn_residual_norm norm_hidden_states.shape: %s", norm_hidden_states.shape)
-        # logger.info("after cross_attn_residual_norm hidden_states.shape: %s", hidden_states.shape)
         norm_hidden_states, hidden_states = norm_hidden_states.to(
             orig_dtype), hidden_states.to(orig_dtype)
 
         # 3. Feed-forward
         ff_output = self.ffn(norm_hidden_states)
         hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
-        # logger.info("after mlp_residual norm_hidden_states.shape: %s", norm_hidden_states.shape)
-        # logger.info("after mlp_residual hidden_states.shape: %s", hidden_states.shape)
         hidden_states = hidden_states.to(orig_dtype)
 
         return hidden_states
@@ -378,10 +366,8 @@ class CausalWanTransformer3DModel(BaseDiT):
                                             elementwise_affine=False,
                                             dtype=torch.float32,
                                             compute_dtype=torch.float32)
-        # Debug: Log configuration values
-        proj_out_dim = config.out_channels * math.prod(config.patch_size)
-        
-        self.proj_out = nn.Linear(inner_dim, proj_out_dim)
+        self.proj_out = nn.Linear(
+            inner_dim, config.out_channels * math.prod(config.patch_size))
         self.scale_shift_table = nn.Parameter(
             torch.randn(1, 2, inner_dim) / inner_dim**0.5)
 
@@ -470,7 +456,6 @@ class CausalWanTransformer3DModel(BaseDiT):
         This function will be run for num_frame times.
         Process the latent frames one by one (1560 tokens each)
         """
-        # logger.info("forward inference hidden_states.shape: %s", hidden_states.shape)
 
         orig_dtype = hidden_states.dtype
         if not isinstance(encoder_hidden_states, torch.Tensor):
@@ -507,10 +492,9 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         hidden_states = self.patch_embedding(hidden_states)
         hidden_states = hidden_states.flatten(2).transpose(1, 2)
-        # logger.info("forward inference flattened and transposed hidden_states.shape: %s", hidden_states.shape)
 
         temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
-            timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
+                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
         timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
 
         if encoder_hidden_states_image is not None:
@@ -549,11 +533,9 @@ class CausalWanTransformer3DModel(BaseDiT):
                                         **causal_kwargs)
 
         # 5. Output norm, projection & unpatchify
-        # logger.info("===== INFERENCE 5. Output norm, projection & unpatchify")
-        # logger.info("hidden_states.shape: %s", hidden_states.shape)
-        # logger.info("temb.shape: %s", temb.shape)
-        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2,
-                                                                          dim=1)
+        temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)
+        shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2,
+                                                                    dim=2)
         hidden_states = self.norm_out(hidden_states, shift, scale)
         hidden_states = self.proj_out(hidden_states)
 
@@ -575,8 +557,6 @@ class CausalWanTransformer3DModel(BaseDiT):
                 start_frame: int = 0,
                 **kwargs) -> torch.Tensor:
 
-        # logger.info("===== forward train hidden_states.shape: %s", hidden_states.shape)
-        # logger.info("===== forward train timestep.shape: %s", timestep.shape)
         orig_dtype = hidden_states.dtype
         if not isinstance(encoder_hidden_states, torch.Tensor):
             encoder_hidden_states = encoder_hidden_states[0]
@@ -622,13 +602,9 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         hidden_states = self.patch_embedding(hidden_states)
         hidden_states = hidden_states.flatten(2).transpose(1, 2)
-        # logger.info("forward train flattened and transposed hidden_states.shape: %s", hidden_states.shape)
 
         temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
-            timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
-        # logger.info("forward train timestep_proj.shape: %s", timestep_proj.shape)
-        # logger.info("forward train timestep.shape: %s", timestep.shape)
-        # logger.info("forward train temb.shape: %s", temb.shape)
+                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
         timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
 
         if encoder_hidden_states_image is not None:
@@ -649,35 +625,17 @@ class CausalWanTransformer3DModel(BaseDiT):
                     timestep_proj, freqs_cis,
                     block_mask=self.block_mask)
         else:
-            for block_index, block in enumerate(self.blocks):
-                # logger.info("===== TRAIN block %d", block_index)
-                # logger.info("hidden_states.shape: %s", hidden_states.shape)
-                # logger.info("encoder_hidden_states.shape: %s", encoder_hidden_states.shape)
-                # logger.info("timestep_proj.shape: %s", timestep_proj.shape)
-                # logger.info("freqs_cis.shape: %s", freqs_cis.shape)
-                # logger.info("block_mask.shape: %s", self.block_mask.shape)
+            for block in self.blocks:
                 hidden_states = block(hidden_states, encoder_hidden_states,
                                         timestep_proj, freqs_cis,
                                         block_mask=self.block_mask)
 
         # 5. Output norm, projection & unpatchify
-        # logger.info("===== TRAIN 5. Output norm, projection & unpatchify")
-        # logger.info("hidden_states.shape: %s", hidden_states.shape)
-        # logger.info("temb.shape: %s", temb.shape)
-        # shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2,
         temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)
-        # logger.info("WTFWTF train temb.shape: %s", temb.shape)
-        # logger.info("WTFWTF train self.scale_shift_table.shape: %s", self.scale_shift_table.shape)
         shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2,
-                                                                          dim=2)
-        # logger.info("DEBUG scale.shape: %s", scale.shape)
-        # logger.info("DEBUG shift.shape: %s", shift.shape)
+                                                                    dim=2)
         hidden_states = self.norm_out(hidden_states, shift, scale)
         hidden_states = self.proj_out(hidden_states)
-        # logger.info("DEBUG after proj_out hidden_states.shape: %s", hidden_states.shape)
-        # logger.info(f"DEBUG reshape dimensions: batch_size={batch_size}, post_patch_num_frames={post_patch_num_frames}")
-        # logger.info(f"DEBUG reshape dimensions: post_patch_height={post_patch_height}, post_patch_width={post_patch_width}")
-        # logger.info(f"DEBUG patch dimensions: p_t={p_t}, p_h={p_h}, p_w={p_w}")
 
         hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames,
                                               post_patch_height,

fastvideo/models/loader/component_loader.py

@@ -415,10 +415,6 @@ class TransformerLoader(ComponentLoader):
             raise ValueError(
                 "Model config does not contain a _class_name attribute. "
                 "Only diffusers format is supported.")
-        logger.info("transformer cls_name: %s", cls_name)
-        if fastvideo_args.override_transformer_cls_name is not None:
-            cls_name = fastvideo_args.override_transformer_cls_name
-            logger.info("Overriding transformer cls_name to %s", cls_name)
 
         fastvideo_args.model_paths["transformer"] = model_path
 

fastvideo/models/registry.py

@@ -61,6 +61,9 @@ _SCHEDULERS = {
      "FlowMatchEulerDiscreteScheduler"),
     "UniPCMultistepScheduler":
     ("schedulers", "scheduling_unipc_multistep", "UniPCMultistepScheduler"),
+    "SelfForcingFlowMatchScheduler":
+    ("schedulers", "scheduling_self_forcing_flow_match",
+     "SelfForcingFlowMatchScheduler"),
 }
 
 _FAST_VIDEO_MODELS = {

fastvideo/models/schedulers/scheduling_self_forcing_flow_match.py

@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: Apache-2.0
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput
+import torch
+
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.base import BaseScheduler
+
+logger = init_logger(__name__)
+
+
+class SelfForcingFlowMatchSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+    """
+    prev_sample: torch.FloatTensor
+
+class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
+
+    order = 1
+    def __init__(self, num_inference_steps=100, num_train_timesteps=1000, shift=3.0, sigma_max=1.0, sigma_min=0.003 / 1.002, inverse_timesteps=False, extra_one_step=False, reverse_sigmas=False, training=False):
+        self.num_train_timesteps = num_train_timesteps
+        self.shift = shift
+        self.sigma_max = sigma_max
+        self.sigma_min = sigma_min
+        self.inverse_timesteps = inverse_timesteps
+        self.extra_one_step = extra_one_step
+        self.reverse_sigmas = reverse_sigmas
+        self.set_timesteps(num_inference_steps, training=training)
+
+    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0, training=False, return_dict=False, **kwargs):
+        sigma_start = self.sigma_min + \
+            (self.sigma_max - self.sigma_min) * denoising_strength
+        if self.extra_one_step:
+            self.sigmas = torch.linspace(
+                sigma_start, self.sigma_min, num_inference_steps + 1)[:-1]
+        else:
+            self.sigmas = torch.linspace(
+                sigma_start, self.sigma_min, num_inference_steps)
+        if self.inverse_timesteps:
+            self.sigmas = torch.flip(self.sigmas, dims=[0])
+        self.sigmas = self.shift * self.sigmas / \
+            (1 + (self.shift - 1) * self.sigmas)
+        if self.reverse_sigmas:
+            self.sigmas = 1 - self.sigmas
+        self.timesteps = self.sigmas * self.num_train_timesteps
+        if training:
+            x = self.timesteps
+            y = torch.exp(-2 * ((x - num_inference_steps / 2) /
+                          num_inference_steps) ** 2)
+            y_shifted = y - y.min()
+            bsmntw_weighing = y_shifted * \
+                (num_inference_steps / y_shifted.sum())
+            self.linear_timesteps_weights = bsmntw_weighing
+
+    def step(self, model_output: torch.FloatTensor, timestep: torch.FloatTensor, sample: torch.FloatTensor, to_final=False, return_dict=False, **kwargs):
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+        self.sigmas = self.sigmas.to(model_output.device)
+        self.timesteps = self.timesteps.to(model_output.device)
+        timestep_id = torch.argmin(
+            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
+        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
+        if to_final or (timestep_id + 1 >= len(self.timesteps)).any():
+            sigma_ = 1 if (
+                self.inverse_timesteps or self.reverse_sigmas) else 0
+        else:
+            sigma_ = self.sigmas[timestep_id + 1].reshape(-1, 1, 1, 1)
+        prev_sample = sample + model_output * (sigma_ - sigma)
+        if isinstance(prev_sample, torch.Tensor | float) and not return_dict:
+            return (prev_sample, )
+        return SelfForcingFlowMatchSchedulerOutput(prev_sample=prev_sample)
+
+    def add_noise(self, original_samples, noise, timestep):
+        """
+        Diffusion forward corruption process.
+        Input:
+            - clean_latent: the clean latent with shape [B*T, C, H, W]
+            - noise: the noise with shape [B*T, C, H, W]
+            - timestep: the timestep with shape [B*T]
+        Output: the corrupted latent with shape [B*T, C, H, W]
+        """
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+        self.sigmas = self.sigmas.to(noise.device)
+        self.timesteps = self.timesteps.to(noise.device)
+        timestep_id = torch.argmin(
+            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
+        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
+        sample = (1 - sigma) * original_samples + sigma * noise
+        return sample.type_as(noise)
+
+    def training_target(self, sample, noise, timestep):
+        target = noise - sample
+        return target
+
+    def training_weight(self, timestep):
+        """
+        Input:
+            - timestep: the timestep with shape [B*T]
+        Output: the corresponding weighting [B*T]
+        """
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+        self.linear_timesteps_weights = self.linear_timesteps_weights.to(timestep.device)
+        timestep_id = torch.argmin(
+            (self.timesteps.unsqueeze(1) - timestep.unsqueeze(0)).abs(), dim=0)
+        weights = self.linear_timesteps_weights[timestep_id]
+        return weights
+    
+    def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor:
+        return sample
+
+    def set_shift(self, shift: float) -> None:
+        self.shift = shift
+    

fastvideo/models/utils.py

@@ -6,9 +6,6 @@ from typing import Any
 import torch
 
 
-from fastvideo.logger import init_logger
-logger = init_logger(__name__)
-
 # TODO(PY): move it elsewhere
 def auto_attributes(init_func):
     """
@@ -148,8 +145,30 @@ def pred_noise_to_pred_video(pred_noise: torch.Tensor,
                              scheduler: Any) -> torch.Tensor:
     """
     Convert predicted noise to clean latent.
+
+    Args:
+    pred_noise: the predicted noise with shape [B, C, H, W]
+        where B is batch_size or batch_size * num_frames
+    noise_input_latent: the noisy latent with shape [B, C, H, W],
+    timestep: the timestep with shape [1] or [bs * num_frames] or [bs, num_frames]
+    scheduler: the scheduler
+
+    Returns:
+        the predicted video with shape [B, C, H, W]
     """
-    timestep = timestep.expand(noise_input_latent.shape[0])
+    # If timestep is [bs, num_frames]
+    if timestep.ndim == 2:
+        timestep = timestep.flatten(0, 1)
+        assert timestep.numel() == noise_input_latent.shape[0]
+    elif timestep.ndim == 1:
+        # If timestep is [1]
+        if timestep.shape[0] == 1:
+            timestep = timestep.expand(noise_input_latent.shape[0])
+        else:
+            assert timestep.numel() == noise_input_latent.shape[0]
+    else:
+        raise ValueError(f"[pred_noise_to_pred_video] Invalid timestep shape: {timestep.shape}")
+    # timestep shape should be [B]
     dtype = pred_noise.dtype
     device = pred_noise.device
     pred_noise = pred_noise.float().to(device)

fastvideo/pipelines/basic/wan/wan_causal_dmd_pipeline.py

@@ -16,8 +16,7 @@ from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage,
                                         CausalDMDDenosingStage,
                                         InputValidationStage,
                                         LatentPreparationStage,
-                                        TextEncodingStage,
-                                        TimestepPreparationStage)
+                                        TextEncodingStage)
 # isort: on
 
 logger = init_logger(__name__)
@@ -48,10 +47,6 @@ class WanCausalDMDPipeline(LoRAPipeline, ComposedPipelineBase):
         self.add_stage(stage_name="conditioning_stage",
                        stage=ConditioningStage())
 
-        self.add_stage(stage_name="timestep_preparation_stage",
-                       stage=TimestepPreparationStage(
-                           scheduler=self.get_module("scheduler")))
-
         self.add_stage(stage_name="latent_preparation_stage",
                        stage=LatentPreparationStage(
                            scheduler=self.get_module("scheduler"),

fastvideo/pipelines/composed_pipeline_base.py

@@ -40,7 +40,7 @@ class ComposedPipelineBase(ABC):
     _extra_config_module_map: dict[str, str] = {}
     training_args: TrainingArgs | None = None
     fastvideo_args: FastVideoArgs | TrainingArgs | None = None
-    modules: dict[str, torch.nn.Module] = {}
+    modules: dict[str, Any] = {}
     post_init_called: bool = False
 
     # TODO(will): args should support both inference args and training args
@@ -79,7 +79,7 @@ class ComposedPipelineBase(ABC):
             for name, module in self.modules.items():
                 if not isinstance(module, torch.nn.Module):
                     continue
-                if name == "transformer":
+                if "transformer" in name:
                     module.requires_grad_(True)
                 else:
                     module.requires_grad_(False)
@@ -121,25 +121,14 @@ class ComposedPipelineBase(ABC):
                         model_path: str,
                         device: str | None = None,
                         torch_dtype: torch.dtype | None = None,
-                        pipeline_config: PipelineConfig | None = None,
+                        pipeline_config: str | PipelineConfig | None = None,
                         args: argparse.Namespace | None = None,
                         required_config_modules: list[str] | None = None,
                         loaded_modules: dict[str, torch.nn.Module]
                         | None = None,
                         **kwargs) -> "ComposedPipelineBase":
         """
-        Load a pipeline from a pretrained model. 
-        Few different patterns are supported:
-        - Only provide model_path:
-            - This will load the pipeline in inference mode.
-            - The pipeline will be initialized with the default config.
-            - The pipeline will be initialized with the default modules.
-            - The pipeline will be initialized with the default stages.
-            - The pipeline will be initialized with the default stages.
-        - override the default config using pipeline_config or args or kwargs
-        - override the default modules using loaded_modules
-        - override the pipelineconfig
-
+        Load a pipeline from a pretrained model.
         loaded_modules: Optional[Dict[str, torch.nn.Module]] = None,
         If provided, loaded_modules will be used instead of loading from config/pretrained weights.
         """
@@ -147,18 +136,9 @@ class ComposedPipelineBase(ABC):
 
             kwargs['model_path'] = model_path
             fastvideo_args = FastVideoArgs.from_kwargs(**kwargs)
-            if pipeline_config is not None:
-                fastvideo_args.pipeline_config = pipeline_config
-            if fastvideo_args.override_transformer_cls_name is not None:
-                pipeline_config = PipelineConfig.from_pretrained("wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers")
-                fastvideo_args.pipeline_config = pipeline_config
         else:
             assert args is not None, "args must be provided for training mode"
             fastvideo_args = TrainingArgs.from_cli_args(args)
-            if fastvideo_args.override_transformer_cls_name is not None:
-                pipeline_config = PipelineConfig.from_pretrained("wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers")
-                fastvideo_args.pipeline_config = pipeline_config
-                logger.info("in 2 Overriding transformer cls name to %s", fastvideo_args.override_transformer_cls_name)
             # TODO(will): fix this so that its not so ugly
             fastvideo_args.model_path = model_path
             for key, value in kwargs.items():
@@ -169,8 +149,7 @@ class ComposedPipelineBase(ABC):
             # model is loaded with the correct precision. Subsequently we will
             # use FSDP2's MixedPrecisionPolicy to set the precision for the
             # fwd, bwd, and other operations' precision.
-            fastvideo_args.pipeline_config.dit_precision = 'fp32'
-            # assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
+            assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
 
         logger.info("fastvideo_args in from_pretrained: %s", fastvideo_args)
 

fastvideo/pipelines/lora_pipeline.py

@@ -7,7 +7,7 @@ import torch
 import torch.distributed as dist
 import torch.nn as nn
 from safetensors.torch import load_file
-from torch.distributed.device_mesh import init_device_mesh
+from torch.distributed.device_mesh import DeviceMesh, init_device_mesh
 from torch.distributed.tensor import DTensor
 
 from fastvideo.distributed import get_local_torch_device
@@ -32,6 +32,7 @@ class LoRAPipeline(ComposedPipelineBase):
     cur_adapter_name: str = ""
     cur_adapter_path: str = ""
     lora_layers: dict[str, BaseLayerWithLoRA] = {}
+    lora_layers_critic: dict[str, BaseLayerWithLoRA] = {}
     fastvideo_args: FastVideoArgs | TrainingArgs
     exclude_lora_layers: list[str] = []
     device: torch.device = get_local_torch_device()
@@ -81,6 +82,17 @@ class LoRAPipeline(ComposedPipelineBase):
 
     def set_trainable(self) -> None:
 
+        def set_lora_grads(lora_layers: dict[str, BaseLayerWithLoRA],
+                           device_mesh: DeviceMesh):
+            for name, layer in lora_layers.items():
+                layer.lora_A.requires_grad_(True)
+                layer.lora_B.requires_grad_(True)
+                layer.base_layer.requires_grad_(False)
+                layer.lora_A = nn.Parameter(
+                    DTensor.from_local(layer.lora_A, device_mesh=device_mesh))
+                layer.lora_B = nn.Parameter(
+                    DTensor.from_local(layer.lora_B, device_mesh=device_mesh))
+
         is_lora_training = self.training_mode and getattr(
             self.fastvideo_args, "lora_training", False)
         if not is_lora_training:
@@ -88,18 +100,12 @@ class LoRAPipeline(ComposedPipelineBase):
             return
 
         self.modules["transformer"].requires_grad_(False)
+        if "fake_score_transformer" in self.modules:
+            self.modules["fake_score_transformer"].requires_grad_(False)
         device_mesh = init_device_mesh("cuda", (dist.get_world_size(), 1),
                                        mesh_dim_names=["fake", "replicate"])
-        for name, layer in self.lora_layers.items():
-            # Enable grads for lora weights only
-            # Must convert to DTensor for compatibility with other FSDP modules in grad calculation
-            layer.lora_A.requires_grad_(True)
-            layer.lora_B.requires_grad_(True)
-            layer.base_layer.requires_grad_(False)
-            layer.lora_A = nn.Parameter(
-                DTensor.from_local(layer.lora_A, device_mesh=device_mesh))
-            layer.lora_B = nn.Parameter(
-                DTensor.from_local(layer.lora_B, device_mesh=device_mesh))
+        set_lora_grads(self.lora_layers, device_mesh)
+        set_lora_grads(self.lora_layers_critic, device_mesh)
 
     def convert_to_lora_layers(self) -> None:
         """
@@ -131,6 +137,24 @@ class LoRAPipeline(ComposedPipelineBase):
                 converted_count += 1
         logger.info("Converted %d layers to LoRA layers", converted_count)
 
+        if "fake_score_transformer" in self.modules:
+            for name, layer in self.modules[
+                    "fake_score_transformer"].named_modules():
+                if not self.is_target_layer(name):
+                    continue
+                layer = get_lora_layer(layer,
+                                       lora_rank=self.lora_rank,
+                                       lora_alpha=self.lora_alpha,
+                                       training_mode=self.training_mode)
+                if layer is not None:
+                    self.lora_layers_critic[name] = layer
+                    replace_submodule(self.modules["fake_score_transformer"],
+                                      name, layer)
+                    converted_count += 1
+            logger.info(
+                "Converted %d layers to LoRA layers in the critic model",
+                converted_count)
+
     def set_lora_adapter(self,
                          lora_nickname: str,
                          lora_path: str | None = None):  # type: ignore
@@ -224,4 +248,4 @@ class LoRAPipeline(ComposedPipelineBase):
 
     def unmerge_lora_weights(self) -> None:
         for name, layer in self.lora_layers.items():
-            layer.unmerge_lora_weights()
+            layer.unmerge_lora_weights()

fastvideo/pipelines/pipeline_batch_info.py

@@ -150,7 +150,7 @@ class ForwardBatch:
     output: torch.Tensor | None = None
     return_trajectory_latents: bool = False
     return_trajectory_decoded: bool = False
-    trajectory_timesteps: list[int] | None = None
+    trajectory_timesteps: list[torch.Tensor] | None = None
     trajectory_latents: torch.Tensor | None = None
     trajectory_decoded: list[torch.Tensor] | None = None
 
@@ -250,5 +250,5 @@ class TrainingBatch:
 
 @dataclass
 class PreprocessBatch(ForwardBatch):
-    video_loader: list["VideoDecoder"] = field(default_factory=list)
+    video_loader: list["VideoDecoder"] | list[str] = field(default_factory=list)
     video_file_name: list[str] = field(default_factory=list)

fastvideo/pipelines/preprocess/preprocess_pipeline_base.py

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-import multiprocessing
 import os
-from concurrent.futures import ProcessPoolExecutor
 from typing import Any
 
 import numpy as np
@@ -19,6 +17,8 @@ from fastvideo.logger import init_logger
 from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 from fastvideo.pipelines.stages import TextEncodingStage
+from fastvideo.workflow.preprocess.parquet_io import (ParquetDatasetWriter,
+                                                      records_to_table)
 
 logger = init_logger(__name__)
 
@@ -54,10 +54,14 @@ class BasePreprocessPipeline(ComposedPipelineBase):
         """Get additional features specific to the pipeline type. Override in subclasses."""
         return {}
 
-    def get_schema_fields(self) -> list[str]:
-        """Get the schema fields for the pipeline type. Override in subclasses."""
+    def get_pyarrow_schema(self) -> pa.Schema:
+        """Return the PyArrow schema for this pipeline. Must be overridden."""
         raise NotImplementedError
 
+    def get_schema_fields(self) -> list[str]:
+        """Get the schema fields for the pipeline type."""
+        return [f.name for f in self.get_pyarrow_schema()]
+
     def create_record_for_schema(self,
                                  preprocess_batch: PreprocessBatch,
                                  schema: pa.Schema,
@@ -400,166 +404,26 @@ class BasePreprocessPipeline(ComposedPipelineBase):
                 batch_data.append(record)
 
             if batch_data:
-                # Add progress bar for writing to Parquet dataset
                 write_pbar = tqdm(total=1,
                                   desc="Writing to Parquet dataset",
                                   unit="batch")
-                # Convert batch data to PyArrow arrays
-                arrays = []
-                for field in self.get_schema_fields():
-                    if field.endswith('_bytes'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.binary()))
-                    elif field.endswith('_shape'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.list_(pa.int32())))
-                    elif field in ['width', 'height', 'num_frames']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.int32()))
-                    elif field in ['duration_sec', 'fps']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.float32()))
-                    else:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data]))
-
-                table = pa.Table.from_arrays(arrays,
-                                             names=self.get_schema_fields())
+                table = records_to_table(batch_data, self.get_pyarrow_schema())
                 write_pbar.update(1)
                 write_pbar.close()
 
-                # Store the table in a list for later processing
-                if not hasattr(self, 'all_tables'):
-                    self.all_tables = []
-                self.all_tables.append(table)
-
+                if not hasattr(self, 'dataset_writer'):
+                    self.dataset_writer = ParquetDatasetWriter(
+                        out_dir=combined_parquet_dir,
+                        samples_per_file=args.samples_per_file,
+                    )
+                self.dataset_writer.append_table(table)
                 logger.info("Collected batch with %s samples", len(table))
 
             if num_processed_samples >= args.flush_frequency:
-                self._flush_tables(num_processed_samples, args,
-                                   combined_parquet_dir)
+                written = self.dataset_writer.flush()
+                logger.info("Flushed %s samples to parquet", written)
                 num_processed_samples = 0
-                self.all_tables = []
 
-    def _flush_tables(self, num_processed_samples: int, args,
-                      combined_parquet_dir: str):
-        """Flush collected tables to disk."""
-        assert hasattr(self, 'all_tables') and self.all_tables
-        print(f"Combining {len(self.all_tables)} batches...")
-        combined_table = pa.concat_tables(self.all_tables)
-        assert len(combined_table) == num_processed_samples
-        print(f"Total samples collected: {len(combined_table)}")
-
-        # Calculate total number of chunks needed, discarding remainder
-        total_chunks = max(num_processed_samples // args.samples_per_file, 1)
-
-        print(f"Fixed samples per parquet file: {args.samples_per_file}")
-        print(f"Total number of parquet files: {total_chunks}")
-        print(
-            f"Total samples to be processed: {total_chunks * args.samples_per_file} (discarding {num_processed_samples % args.samples_per_file} samples)"
-        )
-
-        # Split work among processes
-        num_workers = int(min(multiprocessing.cpu_count(), total_chunks))
-        chunks_per_worker = (total_chunks + num_workers - 1) // num_workers
-
-        print(f"Using {num_workers} workers to process {total_chunks} chunks")
-        logger.info("Chunks per worker: %s", chunks_per_worker)
-
-        # Prepare work ranges
-        work_ranges = []
-        for i in range(num_workers):
-            start_idx = i * chunks_per_worker
-            end_idx = min((i + 1) * chunks_per_worker, total_chunks)
-            if start_idx < total_chunks:
-                work_ranges.append(
-                    (start_idx, end_idx, combined_table, i,
-                     combined_parquet_dir, args.samples_per_file))
-
-        total_written = 0
-        failed_ranges = []
-        with ProcessPoolExecutor(max_workers=num_workers) as executor:
-            futures = {
-                executor.submit(self.process_chunk_range, work_range):
-                work_range
-                for work_range in work_ranges
-            }
-            for future in tqdm(futures, desc="Processing chunks"):
-                try:
-                    written = future.result()
-                    total_written += written
-                    logger.info("Processed chunk with %s samples", written)
-                except Exception as e:
-                    work_range = futures[future]
-                    failed_ranges.append(work_range)
-                    logger.error("Failed to process range %s-%s: %s",
-                                 work_range[0], work_range[1], str(e))
-
-        # Retry failed ranges sequentially
-        if failed_ranges:
-            logger.warning("Retrying %s failed ranges sequentially",
-                           len(failed_ranges))
-            for work_range in failed_ranges:
-                try:
-                    total_written += self.process_chunk_range(work_range)
-                except Exception as e:
-                    logger.error(
-                        "Failed to process range %s-%s after retry: %s",
-                        work_range[0], work_range[1], str(e))
-
-        logger.info("Total samples written: %s", total_written)
-
-    @staticmethod
-    def process_chunk_range(args: Any) -> int:
-        start_idx, end_idx, table, worker_id, output_dir, samples_per_file = args
-        try:
-            total_written = 0
-            num_samples = len(table)
-
-            # Create worker-specific subdirectory
-            worker_dir = os.path.join(output_dir, f"worker_{worker_id}")
-            os.makedirs(worker_dir, exist_ok=True)
-
-            # Check how many files there are already in the dir, and update i accordingly
-            num_parquets = 0
-            for root, _, files in os.walk(worker_dir):
-                for file in files:
-                    if file.endswith('.parquet'):
-                        num_parquets += 1
-
-            for i in range(start_idx, end_idx):
-                start_sample = i * samples_per_file
-                end_sample = min((i + 1) * samples_per_file, num_samples)
-                chunk = table.slice(start_sample, end_sample - start_sample)
-
-                # Create chunk file in worker's directory
-                chunk_path = os.path.join(
-                    worker_dir, f"data_chunk_{i + num_parquets}.parquet")
-                temp_path = chunk_path + '.tmp'
-
-                try:
-                    # Write to temporary file
-                    pq.write_table(chunk, temp_path, compression='zstd')
-
-                    # Rename temporary file to final file
-                    if os.path.exists(chunk_path):
-                        os.remove(
-                            chunk_path)  # Remove existing file if it exists
-                    os.rename(temp_path, chunk_path)
-
-                    total_written += len(chunk)
-                except Exception as e:
-                    # Clean up temporary file if it exists
-                    if os.path.exists(temp_path):
-                        os.remove(temp_path)
-                    raise e
-
-            return total_written
-        except Exception as e:
-            logger.error("Error processing chunks %s-%s for worker %s: %s",
-                         start_idx, end_idx, worker_id, str(e))
-            raise
+    def _final_flush_if_any(self):
+        if hasattr(self, 'dataset_writer'):
+            self.dataset_writer.flush()

fastvideo/pipelines/preprocess/preprocess_pipeline_i2v.py

@@ -40,9 +40,9 @@ class PreprocessPipeline_I2V(BasePreprocessPipeline):
                            image_processor=self.get_module("image_processor"),
                        ))
 
-    def get_schema_fields(self) -> list[str]:
-        """Get the schema fields for I2V pipeline."""
-        return [f.name for f in pyarrow_schema_i2v]
+    def get_pyarrow_schema(self):
+        """Return the PyArrow schema for I2V pipeline."""
+        return pyarrow_schema_i2v
 
     def get_extra_features(self, valid_data: dict[str, Any],
                            fastvideo_args: FastVideoArgs) -> dict[str, Any]:

fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py

@@ -15,153 +15,33 @@ from typing import Any
 import numpy as np
 import pyarrow as pa
 import torch
-from PIL import Image
 from torch.utils.data import DataLoader
 from torchdata.stateful_dataloader import StatefulDataLoader
 from tqdm import tqdm
 
 from fastvideo.configs.sample import SamplingParam
-from fastvideo.dataset import getdataset
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_ode_trajectory
-from fastvideo.distributed import get_local_torch_device
+from fastvideo.dataset import gettextdataset
+from fastvideo.dataset.dataloader.schema import (
+    pyarrow_schema_ode_trajectory_text_only)
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
+    SelfForcingFlowMatchScheduler)
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 from fastvideo.pipelines.preprocess.preprocess_pipeline_base import (
     BasePreprocessPipeline)
 from fastvideo.pipelines.stages import (DecodingStage, DenoisingStage,
-                                        ImageVAEEncodingStage,
                                         InputValidationStage,
                                         LatentPreparationStage,
                                         TextEncodingStage,
                                         TimestepPreparationStage)
 from fastvideo.utils import save_decoded_latents_as_video, shallow_asdict
+from fastvideo.workflow.preprocess.parquet_io import (ParquetDatasetWriter,
+                                                      records_to_table)
 
 logger = init_logger(__name__)
 
 
-class FlowMatchScheduler:
-
-    order = 1
-
-    def __init__(self,
-                 num_inference_steps=100,
-                 num_train_timesteps=1000,
-                 shift=3.0,
-                 sigma_max=1.0,
-                 sigma_min=0.003 / 1.002,
-                 inverse_timesteps=False,
-                 extra_one_step=False,
-                 reverse_sigmas=False):
-        self.num_train_timesteps = num_train_timesteps
-        self.shift = shift
-        self.sigma_max = sigma_max
-        self.sigma_min = sigma_min
-        self.inverse_timesteps = inverse_timesteps
-        self.extra_one_step = extra_one_step
-        self.reverse_sigmas = reverse_sigmas
-        self.set_timesteps(num_inference_steps)
-
-    def set_timesteps(self,
-                      num_inference_steps=100,
-                      denoising_strength=1.0,
-                      training=False,
-                      device=None):
-        sigma_start = self.sigma_min + \
-            (self.sigma_max - self.sigma_min) * denoising_strength
-        if self.extra_one_step:
-            self.sigmas = torch.linspace(sigma_start, self.sigma_min,
-                                         num_inference_steps + 1)[:-1]
-        else:
-            self.sigmas = torch.linspace(sigma_start, self.sigma_min,
-                                         num_inference_steps)
-        if self.inverse_timesteps:
-            self.sigmas = torch.flip(self.sigmas, dims=[0])
-        self.sigmas = self.shift * self.sigmas / \
-            (1 + (self.shift - 1) * self.sigmas)
-        if self.reverse_sigmas:
-            self.sigmas = 1 - self.sigmas
-        self.timesteps = self.sigmas * self.num_train_timesteps
-        if training:
-            x = self.timesteps
-            y = torch.exp(
-                -2 * ((x - num_inference_steps / 2) / num_inference_steps)**2)
-            y_shifted = y - y.min()
-            bsmntw_weighing = y_shifted * \
-                (num_inference_steps / y_shifted.sum())
-            self.linear_timesteps_weights = bsmntw_weighing
-
-    def step(self,
-             model_output,
-             timestep,
-             sample,
-             to_final=False,
-             return_dict=False,
-             **kwargs):
-        assert return_dict is False
-        assert kwargs == {}
-        self.sigmas = self.sigmas.to(model_output.device)
-        self.timesteps = self.timesteps.to(model_output.device)
-        logger.info('step timestep: %s', timestep)
-        logger.info('step timestep: %s', timestep.shape)
-        # timestep  is [num_frames]
-        # timestep_id = torch.argmin(
-        #     (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
-        # assert timestep.ndim == 1
-        # assert timestep.shape[0] == 1
-        timestep_id = torch.argmin((self.timesteps - timestep).abs(), dim=0)
-        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
-        if to_final or (timestep_id + 1 >= len(self.timesteps)).any():
-            sigma_ = 1 if (self.inverse_timesteps or self.reverse_sigmas) else 0
-        else:
-            sigma_ = self.sigmas[timestep_id + 1].reshape(-1, 1, 1, 1)
-        prev_sample = sample + model_output * (sigma_ - sigma)
-        return (prev_sample, )
-
-    def scale_model_input(self, sample: torch.Tensor, *args,
-                          **kwargs) -> torch.Tensor:
-        """
-        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
-        current timestep.
-
-        Args:
-            sample (`torch.Tensor`):
-                The input sample.
-
-        Returns:
-            `torch.Tensor`:
-                A scaled input sample.
-        """
-        return sample
-
-    def add_noise(self, original_samples, noise, timestep):
-        """
-        Diffusion forward corruption process.
-        Input:
-            - clean_latent: the clean latent with shape [B, C, H, W]
-            - noise: the noise with shape [B, C, H, W]
-            - timestep: the timestep with shape [B]
-        Output: the corrupted latent with shape [B, C, H, W]
-        """
-        self.sigmas = self.sigmas.to(noise.device)
-        self.timesteps = self.timesteps.to(noise.device)
-        timestep_id = torch.argmin(
-            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
-        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
-        sample = (1 - sigma) * original_samples + sigma * noise
-        return sample.type_as(noise)
-
-    def training_target(self, sample, noise, timestep):
-        target = noise - sample
-        return target
-
-    def training_weight(self, timestep):
-        timestep_id = torch.argmin(
-            (self.timesteps - timestep.to(self.timesteps.device)).abs())
-        weights = self.linear_timesteps_weights[timestep_id]
-        return weights
-
-
 class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
     """ODE Trajectory preprocessing pipeline implementation."""
 
@@ -170,28 +50,35 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
     ]
     preprocess_dataloader: StatefulDataLoader
     preprocess_loader_iter: Iterator[dict[str, Any]]
+    pbar: Any
+    num_processed_samples: int
 
-    def get_schema_fields(self):
-        """Get the schema fields for ODE Trajectory pipeline."""
-        return [f.name for f in pyarrow_schema_ode_trajectory]
+    def get_pyarrow_schema(self) -> pa.Schema:
+        """Return the PyArrow schema for ODE Trajectory pipeline."""
+        return pyarrow_schema_ode_trajectory_text_only
 
     def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
         """Set up pipeline stages with proper dependency injection."""
-        fastvideo_args.pipeline_config.flow_shift = 5
         logger.info('WTF flow_shift: %s',
                     fastvideo_args.pipeline_config.flow_shift)
-
         assert fastvideo_args.pipeline_config.flow_shift == 5
         # self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(
         # shift=fastvideo_args.pipeline_config.flow_shift)
-        self.modules["scheduler"] = FlowMatchScheduler(
+        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
             shift=fastvideo_args.pipeline_config.flow_shift,
             sigma_min=0.0,
             extra_one_step=True)
         self.modules["scheduler"].set_timesteps(num_inference_steps=48,
                                                 denoising_strength=1.0)
-        logger.info('WTF scheduler timesteps: %s',
-                    self.modules["scheduler"].timesteps)
+        # logger.info('WTF scheduler timesteps: %s',
+        #             self.modules["scheduler"].timesteps)
+        # scheduler = FlowMatchScheduler(
+        #     shift=8.0, sigma_min=0.0, extra_one_step=True)
+        # device = get_local_torch_device()
+        # # scheduler.num_train_timesteps = 100
+        # scheduler.set_timesteps(num_inference_steps=50, denoising_strength=1.0)
+        # scheduler.sigmas = scheduler.sigmas.to(device)
+        # self.modules["scheduler"] = scheduler
 
         self.add_stage(stage_name="input_validation_stage",
                        stage=InputValidationStage())
@@ -200,9 +87,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                            text_encoders=[self.get_module("text_encoder")],
                            tokenizers=[self.get_module("tokenizer")],
                        ))
-        self.add_stage(stage_name="vae_encoding_stage",
-                       stage=ImageVAEEncodingStage(
-                           vae=self.get_module("vae"), ))
         self.add_stage(stage_name="timestep_preparation_stage",
                        stage=TimestepPreparationStage(
                            scheduler=self.get_module("scheduler")))
@@ -213,56 +97,47 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
         self.add_stage(stage_name="denoising_stage",
                        stage=DenoisingStage(
                            transformer=self.get_module("transformer"),
-                           transformer_2=self.get_module("transformer_2", None),
                            scheduler=self.get_module("scheduler"),
                            pipeline=self,
                        ))
         self.add_stage(stage_name="decoding_stage",
                        stage=DecodingStage(vae=self.get_module("vae")))
 
-    def preprocess_video_and_text_and_trajectory(self,
-                                                 fastvideo_args: FastVideoArgs,
-                                                 args):
+    def preprocess_text_and_trajectory(self, fastvideo_args: FastVideoArgs,
+                                       args):
+        """Preprocess text-only data and generate trajectory information."""
 
         for batch_idx, data in enumerate(self.pbar):
             if data is None:
                 continue
 
             with torch.inference_mode():
-                # Filter out invalid samples (those with all zeros)
+                # For text-only processing, we only need text data
+                # Filter out samples without text
                 valid_indices = []
-                for i, pixel_values in enumerate(data["pixel_values"]):
-                    if not torch.all(
-                            pixel_values == 0):  # Check if all values are zero
+                for i, text in enumerate(data["text"]):
+                    if text and text.strip():  # Check if text is not empty
                         valid_indices.append(i)
                 self.num_processed_samples += len(valid_indices)
 
                 if not valid_indices:
                     continue
 
-                # Create new batch with only valid samples
+                # Create new batch with only valid samples (text-only)
                 valid_data = {
-                    "pixel_values":
-                    torch.stack(
-                        [data["pixel_values"][i] for i in valid_indices]),
                     "text": [data["text"][i] for i in valid_indices],
                     "path": [data["path"][i] for i in valid_indices],
-                    "fps": [data["fps"][i] for i in valid_indices],
-                    "duration": [data["duration"][i] for i in valid_indices],
                 }
 
-                # VAE
-                with torch.autocast("cuda", dtype=torch.float32):
-                    latents = self.get_module("vae").encode(
-                        valid_data["pixel_values"].to(
-                            get_local_torch_device())).mean
-
-                # Get extra features if needed
-                extra_features = self.get_extra_features(
-                    valid_data, fastvideo_args)
+                # Add fps and duration if available in data
+                if "fps" in data:
+                    valid_data["fps"] = [data["fps"][i] for i in valid_indices]
+                if "duration" in data:
+                    valid_data["duration"] = [
+                        data["duration"][i] for i in valid_indices
+                    ]
 
                 batch_captions = valid_data["text"]
-                logger.info(f"===== batch_captions: {batch_captions}")
                 # Encode text using the standalone TextEncodingStage API
                 prompt_embeds_list, prompt_masks_list = self.prompt_encoding_stage.encode_text(
                     batch_captions,
@@ -274,27 +149,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                 prompt_attention_masks = prompt_masks_list[0]
                 assert prompt_embeds.shape[0] == prompt_attention_masks.shape[0]
 
-                # # Get sequence lengths from attention masks (number of 1s)
-                # seq_lens = prompt_attention_mask.sum(dim=1)
-
-                # non_padded_embeds = []
-                # non_padded_masks = []
-
-                # # Process each item in the batch
-                # for i in range(prompt_embeds.size(0)):
-                #     seq_len = seq_lens[i].item()
-                #     # Slice the embeddings and masks to keep only non-padding parts
-                #     non_padded_embeds.append(prompt_embeds[i, :seq_len])
-                #     non_padded_masks.append(prompt_attention_mask[i, :seq_len])
-
-                # Update the tensors with non-padded versions
-                # prompt_embeds = non_padded_embeds
-                # prompt_attention_masks = non_padded_masks
-                # prompt_embeds = prompt_embeds
-
-                # logger.info(f"===== prompt_embeds: {prompt_embeds[0].shape}")
-                # logger.info(f"===== prompt_attention_masks: {prompt_attention_masks[0].shape}")
-
                 sampling_params = SamplingParam.from_pretrained(args.model_path)
 
                 # encode negative prompt for trajectory collection
@@ -315,52 +169,32 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                 trajectory_latents = []
                 trajectory_timesteps = []
                 trajectory_decoded = []
+
                 for i, (prompt_embed, prompt_attention_mask) in enumerate(
-                        zip(prompt_embeds, prompt_attention_masks, strict=False)):
+                        zip(prompt_embeds, prompt_attention_masks,
+                            strict=False)):
                     prompt_embed = prompt_embed.unsqueeze(0)
                     prompt_attention_mask = prompt_attention_mask.unsqueeze(0)
-                    logger.info("what")
-                    logger.info(f"===== prompt_embed: {prompt_embed.shape}")
-                    logger.info(
-                        f"===== prompt_attention_mask: {prompt_attention_mask.shape}"
-                    )
-                    # Collect the trajectory data
-                    batch = ForwardBatch(
-                        **shallow_asdict(sampling_params),
-                        # data_type="video",
-                        # seed=args.seed,
-                        # prompt=batch_captions[i],
-                        # prompt_embeds=[prompt_embed],
-                        # prompt_attention_mask=[prompt_attention_mask],
-                        # height=args.max_height,
-                        # width=args.max_width,
-                        # num_frames=81,
-                        # fps=args.train_fps,
-                        # return_trajectory_latents=True,
-                        # guidance_scale=3.0,
-                        # do_classifier_free_guidance=True,
-                    )
+
+                    # Collect the trajectory data (text-to-video generation)
+                    batch = ForwardBatch(**shallow_asdict(sampling_params), )
                     batch.prompt_embeds = [prompt_embed]
                     batch.prompt_attention_mask = [prompt_attention_mask]
                     batch.negative_prompt_embeds = [negative_prompt_embed]
                     batch.negative_attention_mask = [
                         negative_prompt_attention_mask
                     ]
+                    batch.num_inference_steps = 48
                     batch.return_trajectory_latents = True
                     batch.return_trajectory_decoded = False
                     batch.height = args.max_height
                     batch.width = args.max_width
-                    batch.num_inference_steps = 48
-                    # batch.num_frames = 81
                     batch.fps = args.train_fps
                     batch.guidance_scale = 6.0
                     batch.do_classifier_free_guidance = True
-                    # fastvideo_args.pipeline_config.ti2v_task = True
 
                     result_batch = self.input_validation_stage(
                         batch, fastvideo_args)
-                    # result_batch = self.prompt_encoding_stage(result_batch, fastvideo_args)
-                    # result_batch = self.vae_encoding_stage(result_batch, fastvideo_args)
                     result_batch = self.timestep_preparation_stage(
                         batch, fastvideo_args)
                     result_batch = self.latent_preparation_stage(
@@ -369,205 +203,127 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                                                         fastvideo_args)
                     result_batch = self.decoding_stage(result_batch,
                                                        fastvideo_args)
-                    # trajectory_latents = result_batch.trajectory_latents
+
                     trajectory_latents.append(
                         result_batch.trajectory_latents.cpu())
                     trajectory_timesteps.append(
                         result_batch.trajectory_timesteps.cpu())
                     trajectory_decoded.append(result_batch.trajectory_decoded)
 
-            extra_features["trajectory_latents"] = trajectory_latents
-            extra_features["trajectory_timesteps"] = trajectory_timesteps
-            logger.info(
-                f"===== trajectory_latents: {trajectory_latents[0].shape}")
-            logger.info(
-                f"===== trajectory_latents len: {len(trajectory_latents)}")
-            logger.info(f"===== trajectory_timesteps: {trajectory_timesteps}")
-            logger.info(
-                f"===== trajectory_timesteps len: {len(trajectory_timesteps)}")
+                # Prepare extra features for text-only processing
+                extra_features = {
+                    "trajectory_latents": trajectory_latents,
+                    "trajectory_timesteps": trajectory_timesteps
+                }
 
-            if batch.return_trajectory_decoded:
-                logger.info("===== SAVING TRAJECTORY DECODED")
-                for i, decoded_frames in enumerate(trajectory_decoded):
-                    for j, decoded_frame in enumerate(decoded_frames):
-                        logger.info(
-                            f"===== SAVING TRAJECTORY DECODED {i} for prompt {batch_captions[i]}"
-                        )
-                        save_decoded_latents_as_video(
-                            decoded_frame,
-                            f"decoded_videos/trajectory_decoded_{i}_{j}.mp4",
-                            args.train_fps)
-            # assert False
-            # Prepare batch data for Parquet dataset
-            batch_data = []
+                if batch.return_trajectory_decoded:
+                    for i, decoded_frames in enumerate(trajectory_decoded):
+                        for j, decoded_frame in enumerate(decoded_frames):
+                            save_decoded_latents_as_video(
+                                decoded_frame,
+                                f"decoded_videos/trajectory_decoded_{i}_{j}.mp4",
+                                args.train_fps)
 
-            # Add progress bar for saving outputs
-            save_pbar = tqdm(enumerate(valid_data["path"]),
-                             desc="Saving outputs",
-                             unit="item",
-                             leave=False)
-            for idx, video_path in save_pbar:
-                # Get the corresponding latent and info using video name
-                latent = latents[idx].cpu()
-                video_name = os.path.basename(video_path).split(".")[0]
+                # Prepare batch data for Parquet dataset
+                batch_data = []
 
-                # Convert tensors to numpy arrays
-                vae_latent = latent.cpu().numpy()
-                text_embedding = prompt_embeds[idx].cpu().numpy()
+                # Add progress bar for saving outputs
+                save_pbar = tqdm(enumerate(valid_data["path"]),
+                                 desc="Saving outputs",
+                                 unit="item",
+                                 leave=False)
 
-                # Get extra features for this sample if needed
-                sample_extra_features = {}
-                if extra_features:
-                    for key, value in extra_features.items():
-                        logger.info(f"===== key: {key}")
-                        if isinstance(value, torch.Tensor):
-                            logger.info(f"===== value: {value[idx].shape}")
-                            sample_extra_features[key] = value[idx].cpu().numpy(
-                            )
-                        else:
-                            assert isinstance(value, list)
-                            if isinstance(value[idx], torch.Tensor):
-                                logger.info(
-                                    f"===== value in list: {value[idx].shape}")
+                for idx, video_path in save_pbar:
+                    video_name = os.path.basename(video_path).split(".")[0]
+
+                    # Convert tensors to numpy arrays
+                    text_embedding = prompt_embeds[idx].cpu().numpy()
+
+                    # Get extra features for this sample
+                    sample_extra_features = {}
+                    if extra_features:
+                        for key, value in extra_features.items():
+                            if isinstance(value, torch.Tensor):
                                 sample_extra_features[key] = value[idx].cpu(
-                                ).float().numpy()
+                                ).numpy()
                             else:
-                                logger.info("===== value in list: not tensor")
-                                sample_extra_features[key] = value[idx]
-                            # logger.info(f"===== value: not tensor")
-                            # sample_extra_features[key] = value[idx]
+                                assert isinstance(value, list)
+                                if isinstance(value[idx], torch.Tensor):
+                                    sample_extra_features[key] = value[idx].cpu(
+                                    ).float().numpy()
+                                else:
+                                    sample_extra_features[key] = value[idx]
 
-                # Create record for Parquet dataset
-                record = self.create_record(
-                    video_name=video_name,
-                    vae_latent=vae_latent,
-                    text_embedding=text_embedding,
-                    valid_data=valid_data,
-                    idx=idx,
-                    extra_features=sample_extra_features)
-                batch_data.append(record)
+                    # Create record for Parquet dataset (without VAE latents for text-only)
+                    record = self.create_text_only_record(
+                        args,
+                        video_name=video_name,
+                        text_embedding=text_embedding,
+                        valid_data=valid_data,
+                        idx=idx,
+                        extra_features=sample_extra_features)
+                    batch_data.append(record)
 
-            if batch_data:
-                # Add progress bar for writing to Parquet dataset
-                write_pbar = tqdm(total=1,
-                                  desc="Writing to Parquet dataset",
-                                  unit="batch")
-                # Convert batch data to PyArrow arrays
-                arrays = []
-                for field in self.get_schema_fields():
-                    if field.endswith('_bytes'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.binary()))
-                    elif field.endswith('_shape'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.list_(pa.int32())))
-                    elif field in ['width', 'height', 'num_frames']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.int32()))
-                    elif field in ['duration_sec', 'fps']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.float32()))
-                    else:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data]))
+                if batch_data:
+                    write_pbar = tqdm(total=1,
+                                      desc="Writing to Parquet dataset",
+                                      unit="batch")
+                    table = records_to_table(batch_data,
+                                             self.get_pyarrow_schema())
+                    write_pbar.update(1)
+                    write_pbar.close()
 
-                table = pa.Table.from_arrays(arrays,
-                                             names=self.get_schema_fields())
-                write_pbar.update(1)
-                write_pbar.close()
+                    if not hasattr(self, 'dataset_writer'):
+                        self.dataset_writer = ParquetDatasetWriter(
+                            out_dir=self.combined_parquet_dir,
+                            samples_per_file=args.samples_per_file,
+                        )
+                    self.dataset_writer.append_table(table)
 
-                # Store the table in a list for later processing
-                if not hasattr(self, 'all_tables'):
-                    self.all_tables = []
-                self.all_tables.append(table)
+                    logger.info("Collected batch with %s samples", len(table))
 
-                logger.info("Collected batch with %s samples", len(table))
+                if self.num_processed_samples >= args.flush_frequency:
+                    written = self.dataset_writer.flush()
+                    logger.info("Flushed %s samples to parquet", written)
+                    self.num_processed_samples = 0
 
-            if self.num_processed_samples >= args.flush_frequency:
-                self._flush_tables(self.num_processed_samples, args,
-                                   self.combined_parquet_dir)
-                self.num_processed_samples = 0
-                self.all_tables = []
+        # Final flush for any remaining samples
+        if hasattr(self, 'dataset_writer'):
+            written = self.dataset_writer.flush()
+            if written:
+                logger.info("Final flush wrote %s samples", written)
 
-    def get_extra_features(self, valid_data: dict[str, Any],
-                           fastvideo_args: FastVideoArgs) -> dict[str, Any]:
-
-        # TODO(will): move these to cpu at some point
-        self.get_module("vae").to(get_local_torch_device())
-
-        # generator = torch.Generator(device=get_local_torch_device(), seed=42)
-        generator = torch.Generator("cpu").manual_seed(42)
-
-        features = {}
-        """Get CLIP features from the first frame of each video."""
-        first_frame = valid_data["pixel_values"][:, :, 0, :, :].permute(
-            0, 2, 3, 1)  # (B, C, T, H, W) -> (B, H, W, C)
-        _, _, num_frames, height, width = valid_data["pixel_values"].shape
-        # latent_height = height // self.get_module(
-        #     "vae").spatial_compression_ratio
-        # latent_width = width // self.get_module("vae").spatial_compression_ratio
-
-        unprocessed_images = []
-        pil_images = []
-        # Frame has values between -1 and 1
-        for frame in first_frame:
-            frame = (frame + 1) * 127.5
-            frame_pil = Image.fromarray(frame.cpu().numpy().astype(np.uint8))
-            pil_images.append(frame_pil)
-            # processed_img = self.get_module("image_processor")(
-            #     images=frame_pil, return_tensors="pt")
-            unprocessed_images.append(frame_pil)
-        """Get VAE features from the first frame of each video"""
-        video_conditions = []
-        for frame in unprocessed_images:
-
-            latent = self.vae_encoding_stage.encode_image(
-                frame, height, width, fastvideo_args, generator)
-            video_conditions.append(latent)
-
-        features["image_condition_latents"] = video_conditions
-        features["pil_images"] = pil_images
-        return features
-
-    def create_record(
+    def create_text_only_record(
             self,
+            args,
             video_name: str,
-            vae_latent: np.ndarray,
             text_embedding: np.ndarray,
             valid_data: dict[str, Any],
             idx: int,
             extra_features: dict[str, Any] | None = None) -> dict[str, Any]:
-        """Create a record for the Parquet dataset with CLIP features."""
-        record = super().create_record(video_name=video_name,
-                                       vae_latent=vae_latent,
-                                       text_embedding=text_embedding,
-                                       valid_data=valid_data,
-                                       idx=idx,
-                                       extra_features=extra_features)
+        """Create a record for text-only preprocessing using text-only schema."""
 
-        if extra_features and "image_condition_latents" in extra_features:
-            image_condition_latents = extra_features["image_condition_latents"]
-            record.update({
-                "image_condition_latents_bytes":
-                image_condition_latents.tobytes(),
-                "image_condition_latents_shape":
-                list(image_condition_latents.shape),
-                "image_condition_latents_dtype":
-                str(image_condition_latents.dtype),
-            })
-        else:
-            record.update({
-                "image_condition_latents_bytes": b"",
-                "image_condition_latents_shape": [],
-                "image_condition_latents_dtype": "",
-            })
+        # Create base record using only fields from text-only schema
+        record = {
+            "id": f"text_{video_name}_{idx}",
+            "text_embedding_bytes": text_embedding.tobytes(),
+            "text_embedding_shape": list(text_embedding.shape),
+            "text_embedding_dtype": str(text_embedding.dtype),
+            "file_name": video_name,
+            "caption": valid_data["text"][idx],
+            "media_type": "text",
+        }
 
+        assert extra_features is not None, "extra_features is required"
+        assert "trajectory_latents" in extra_features, "trajectory_latents is required"
+        assert "trajectory_timesteps" in extra_features, "trajectory_timesteps is required"
+
+        # Add trajectory data if available
         if extra_features and "trajectory_latents" in extra_features:
-            trajectory_latents = extra_features["trajectory_latents"]
+            trajectory_latents = extra_features[
+                "trajectory_latents"][idx] if isinstance(
+                    extra_features["trajectory_latents"],
+                    list) else extra_features["trajectory_latents"]
             record.update({
                 "trajectory_latents_bytes":
                 trajectory_latents.tobytes(),
@@ -584,7 +340,10 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
             })
 
         if extra_features and "trajectory_timesteps" in extra_features:
-            trajectory_timesteps = extra_features["trajectory_timesteps"]
+            trajectory_timesteps = extra_features[
+                "trajectory_timesteps"][idx] if isinstance(
+                    extra_features["trajectory_timesteps"],
+                    list) else extra_features["trajectory_timesteps"]
             record.update({
                 "trajectory_timesteps_bytes":
                 trajectory_timesteps.tobytes(),
@@ -600,20 +359,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                 "trajectory_timesteps_dtype": "",
             })
 
-        if extra_features and "pil_image" in extra_features:
-            pil_image = extra_features["pil_image"]
-            record.update({
-                "pil_image_bytes": pil_image.tobytes(),
-                "pil_image_shape": list(pil_image.shape),
-                "pil_image_dtype": str(pil_image.dtype),
-            })
-        else:
-            record.update({
-                "pil_image_bytes": b"",
-                "pil_image_shape": [],
-                "pil_image_dtype": "",
-            })
-
         return record
 
     def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs, args):
@@ -628,7 +373,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
         os.makedirs(self.combined_parquet_dir, exist_ok=True)
 
         # Loading dataset
-        train_dataset = getdataset(args)
+        train_dataset = gettextdataset(args)
 
         self.preprocess_dataloader = DataLoader(
             train_dataset,
@@ -648,7 +393,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
         # Initialize class variables for data sharing
         self.video_data: dict[str, Any] = {}  # Store video metadata and paths
         self.latent_data: dict[str, Any] = {}  # Store latent tensors
-        self.preprocess_video_and_text_and_trajectory(fastvideo_args, args)
+        self.preprocess_text_and_trajectory(fastvideo_args, args)
 
 
 EntryClass = PreprocessPipeline_ODE_Trajectory

fastvideo/pipelines/preprocess/preprocess_pipeline_t2v.py

@@ -15,9 +15,9 @@ class PreprocessPipeline_T2V(BasePreprocessPipeline):
 
     _required_config_modules = ["text_encoder", "tokenizer", "vae"]
 
-    def get_schema_fields(self):
-        """Get the schema fields for T2V pipeline."""
-        return [f.name for f in pyarrow_schema_t2v]
+    def get_pyarrow_schema(self):
+        """Return the PyArrow schema for T2V pipeline."""
+        return pyarrow_schema_t2v
 
 
 EntryClass = PreprocessPipeline_T2V

fastvideo/pipelines/preprocess/preprocess_stages.py

@@ -4,9 +4,11 @@ from typing import cast
 
 import numpy as np
 import torch
+import torchvision
 from einops import rearrange
 from torchvision import transforms
 
+from fastvideo.configs.configs import VideoLoaderType
 from fastvideo.dataset.transform import (CenterCropResizeVideo,
                                          TemporalRandomCrop)
 from fastvideo.fastvideo_args import FastVideoArgs, WorkloadType
@@ -61,7 +63,16 @@ class VideoTransformStage(PipelineStage):
                 else:
                     frame_indices = frame_indices[:self.num_frames]
 
-            video = batch.video_loader[i].get_frames_at(frame_indices).data
+            if fastvideo_args.preprocess_config.video_loader_type == VideoLoaderType.TORCHCODEC:
+                video = batch.video_loader[i].get_frames_at(frame_indices).data
+            elif fastvideo_args.preprocess_config.video_loader_type == VideoLoaderType.TORCHVISION:
+                video, _, _ = torchvision.io.read_video(batch.video_loader[i],
+                                                        output_format="TCHW")
+                video = video[frame_indices]
+            else:
+                raise ValueError(
+                    f"Invalid video loader type: {fastvideo_args.preprocess_config.video_loader_type}"
+                )
             video = self.video_transform(video)
             video_pixel_batch.append(video)
 

fastvideo/pipelines/preprocess/v1_preprocess.py

@@ -47,10 +47,8 @@ def main(args) -> None:
     else:
         raise ValueError(f"Invalid preprocess task: {args.preprocess_task}")
 
-    logger.info(
-        f"Preprocess task: {args.preprocess_task} using {PreprocessPipeline.__name__}"
-    )
-
+    logger.info("Preprocess task: %s using %s", args.preprocess_task,
+                PreprocessPipeline.__name__)
     pipeline = PreprocessPipeline(args.model_path, fastvideo_args)
     pipeline.forward(batch=None, fastvideo_args=fastvideo_args, args=args)
 

fastvideo/pipelines/stages/causal_denoising.py

@@ -4,11 +4,11 @@ from fastvideo.distributed import get_local_torch_device
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.forward_context import set_forward_context
 from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
 from fastvideo.models.utils import pred_noise_to_pred_video
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 from fastvideo.pipelines.stages.denoising import DenoisingStage
+from fastvideo.pipelines.stages.validators import StageValidators as V
+from fastvideo.pipelines.stages.validators import VerificationResult
 
 try:
     from fastvideo.attention.backends.sliding_tile_attn import (
@@ -36,7 +36,6 @@ class CausalDMDDenosingStage(DenoisingStage):
 
     def __init__(self, transformer, scheduler) -> None:
         super().__init__(transformer, scheduler)
-        self.scheduler = FlowMatchEulerDiscreteScheduler(shift=8.0)
         # KV and cross-attention cache state (initialized on first forward)
         self.kv_cache1: list | None = None
         self.crossattn_cache: list | None = None
@@ -71,8 +70,16 @@ class CausalDMDDenosingStage(DenoisingStage):
         # Timesteps for DMD
         timesteps = torch.tensor(
             fastvideo_args.pipeline_config.dmd_denoising_steps,
-            dtype=torch.long,
-            device=get_local_torch_device())
+            dtype=torch.long).cpu()
+
+        if fastvideo_args.pipeline_config.warp_denoising_step:
+            logger.info("Warping timesteps...")
+            scheduler_timesteps = torch.cat((self.scheduler.timesteps.cpu(),
+                                             torch.tensor([0],
+                                                          dtype=torch.float32)))
+            timesteps = scheduler_timesteps[1000 - timesteps]
+        timesteps = timesteps.to(get_local_torch_device())
+        logger.info("Using timesteps: %s", timesteps)
 
         # Image kwargs (kept empty unless caller provides compatible args)
         image_kwargs: dict = {}
@@ -228,10 +235,7 @@ class CausalDMDDenosingStage(DenoisingStage):
                                                        dim=2)
 
                     # Prepare inputs
-                    t_expand = t_cur.expand(latent_model_input.shape[0])
-                    # t_expand = t_cur * torch.ones((latent_model_input.shape[0], 1), device=latent_model_input.device, dtype=torch.long)
-                    # t_expand = t_expand.repeat(1, self.sliding_window_num_frames)
-
+                    t_expand = t_cur.repeat(latent_model_input.shape[0])
 
                     # Attention metadata if needed
                     if (vsa_available and self.attn_backend
@@ -265,7 +269,10 @@ class CausalDMDDenosingStage(DenoisingStage):
                                             attn_metadata=attn_metadata,
                                             forward_batch=batch):
                         # Run transformer; follow DMD stage pattern
-                        t_expanded_noise= t_cur * torch.ones((latent_model_input.shape[0], 1), device=latent_model_input.device, dtype=torch.long)
+                        t_expanded_noise = t_cur * torch.ones(
+                            (latent_model_input.shape[0], 1),
+                            device=latent_model_input.device,
+                            dtype=torch.long)
                         pred_noise_btchw = self.transformer(
                             latent_model_input,
                             prompt_embeds,
@@ -330,7 +337,7 @@ class CausalDMDDenosingStage(DenoisingStage):
                     set_forward_context(current_timestep=0,
                                         attn_metadata=attn_metadata,
                                         forward_batch=batch):
-                    t_expanded_context = t_context * torch.ones((context_bcthw.shape[0], 1), device=context_bcthw.device, dtype=torch.long)
+                    t_expanded_context = t_context.unsqueeze(1)
                     _ = self.transformer(
                         context_bcthw,
                         prompt_embeds,
@@ -412,3 +419,27 @@ class CausalDMDDenosingStage(DenoisingStage):
                 False,
             })
         self.crossattn_cache = crossattn_cache
+
+    def verify_input(self, batch: ForwardBatch,
+                     fastvideo_args: FastVideoArgs) -> VerificationResult:
+        """Verify denoising stage inputs."""
+        result = VerificationResult()
+        result.add_check("latents", batch.latents,
+                         [V.is_tensor, V.with_dims(5)])
+        result.add_check("prompt_embeds", batch.prompt_embeds, V.list_not_empty)
+        result.add_check("image_embeds", batch.image_embeds, V.is_list)
+        result.add_check("image_latent", batch.image_latent,
+                         V.none_or_tensor_with_dims(5))
+        result.add_check("num_inference_steps", batch.num_inference_steps,
+                         V.positive_int)
+        result.add_check("guidance_scale", batch.guidance_scale,
+                         V.positive_float)
+        result.add_check("eta", batch.eta, V.non_negative_float)
+        result.add_check("generator", batch.generator,
+                         V.generator_or_list_generators)
+        result.add_check("do_classifier_free_guidance",
+                         batch.do_classifier_free_guidance, V.bool_value)
+        result.add_check(
+            "negative_prompt_embeds", batch.negative_prompt_embeds, lambda x:
+            not batch.do_classifier_free_guidance or V.list_not_empty(x))
+        return result

fastvideo/pipelines/stages/decoding.py

@@ -53,7 +53,20 @@ class DecodingStage(PipelineStage):
     @torch.no_grad()
     def decode(self, latents: torch.Tensor,
                fastvideo_args: FastVideoArgs) -> torch.Tensor:
-        """Decode latents into pixel space."""
+        """
+        Decode latent representations into pixel space using VAE.
+        
+        Args:
+            latents: Input latent tensor with shape (batch, channels, frames, height_latents, width_latents)
+            fastvideo_args: Configuration containing:
+                - disable_autocast: Whether to disable automatic mixed precision (default: False)
+                - pipeline_config.vae_precision: VAE computation precision ("fp32", "fp16", "bf16")
+                - pipeline_config.vae_tiling: Whether to enable VAE tiling for memory efficiency
+            
+        Returns:
+            Decoded video tensor with shape (batch, channels, frames, height, width), 
+            normalized to [0, 1] range and moved to CPU as float32
+        """
         self.vae = self.vae.to(get_local_torch_device())
         latents = latents.to(get_local_torch_device())
 
@@ -103,12 +116,26 @@ class DecodingStage(PipelineStage):
         """
         Decode latent representations into pixel space.
         
+        This method processes the batch through the VAE decoder, converting latent
+        representations to pixel-space video/images. It also optionally decodes
+        trajectory latents for visualization purposes.
+        
         Args:
-            batch: The current batch information.
-            fastvideo_args: The inference arguments.
+            batch: The current batch containing:
+                - latents: Tensor to decode (batch, channels, frames, height_latents, width_latents)
+                - return_trajectory_decoded (optional): Flag to decode trajectory latents
+                - trajectory_latents (optional): Latents at different timesteps
+                - trajectory_timesteps (optional): Corresponding timesteps
+            fastvideo_args: Configuration containing:
+                - output_type: "latent" to skip decoding, otherwise decode to pixels
+                - vae_cpu_offload: Whether to offload VAE to CPU after decoding
+                - model_loaded: Track VAE loading state
+                - model_paths: Path to VAE model if loading needed
             
         Returns:
-            The batch with decoded outputs.
+            Modified batch with:
+                - output: Decoded frames (batch, channels, frames, height, width) as CPU float32
+                - trajectory_decoded (if requested): List of decoded frames per timestep
         """
         # load vae if not already loaded (used for memory constrained devices)
         pipeline = self.pipeline() if self.pipeline else None
@@ -128,15 +155,13 @@ class DecodingStage(PipelineStage):
         # decode trajectory latents if needed
         if batch.return_trajectory_decoded:
             batch.trajectory_decoded = []
-            logger.info(f"batch.trajectory_latents.shape: {batch.trajectory_latents.shape}")
             assert batch.trajectory_latents is not None, "batch should have trajectory latents"
             for idx in range(batch.trajectory_latents.shape[1]):
-                # bathc.trajectory_latents is [batch_size, timesteps, channels, frames, height, width]
+                # batch.trajectory_latents is [batch_size, timesteps, channels, frames, height, width]
                 cur_latent = batch.trajectory_latents[:, idx, :, :, :, :]
-                logger.info(f"cur_latent.shape: {cur_latent.shape}")
                 cur_timestep = batch.trajectory_timesteps[idx]
-                logger.info(
-                    f"decoding trajectory latent for timestep: {cur_timestep}")
+                logger.info("decoding trajectory latent for timestep: %s",
+                            cur_timestep)
                 decoded_frames = self.decode(cur_latent, fastvideo_args)
                 batch.trajectory_decoded.append(decoded_frames.cpu().float())
 

fastvideo/pipelines/stages/denoising.py

@@ -40,6 +40,13 @@ try:
 except ImportError:
     st_attn_available = False
 
+try:
+    from fastvideo.attention.backends.vmoba import VMOBAAttentionBackend
+    from fastvideo.utils import is_vmoba_available
+    vmoba_attn_available = is_vmoba_available()
+except ImportError:
+    vmoba_attn_available = False
+
 try:
     from fastvideo.attention.backends.video_sparse_attn import (
         VideoSparseAttentionBackend)
@@ -77,6 +84,7 @@ class DenoisingStage(PipelineStage):
             supported_attention_backends=(
                 AttentionBackendEnum.SLIDING_TILE_ATTN,
                 AttentionBackendEnum.VIDEO_SPARSE_ATTN,
+                AttentionBackendEnum.VMOBA_ATTN,
                 AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA
             )  # hack
         )
@@ -132,12 +140,11 @@ class DenoisingStage(PipelineStage):
             latents = latents[:, :, rank_in_sp_group, :, :, :]
             batch.latents = latents
             if batch.image_latent is not None:
-                if not fastvideo_args.pipeline_config.ti2v_task and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    image_latent = rearrange(batch.image_latent,
-                                             "b c (n t) h w -> b c n t h w",
-                                             n=sp_world_size).contiguous()
-                    image_latent = image_latent[:, :, rank_in_sp_group, :, :, :]
-                    batch.image_latent = image_latent
+                image_latent = rearrange(batch.image_latent,
+                                         "b c (n t) h w -> b c n t h w",
+                                         n=sp_world_size).contiguous()
+                image_latent = image_latent[:, :, rank_in_sp_group, :, :, :]
+                batch.image_latent = image_latent
         # Get timesteps and calculate warmup steps
         timesteps = batch.timesteps
         # TODO(will): remove this once we add input/output validation for stages
@@ -150,7 +157,8 @@ class DenoisingStage(PipelineStage):
         # Prepare image latents and embeddings for I2V generation
         image_embeds = batch.image_embeds
         if len(image_embeds) > 0:
-            assert torch.isnan(image_embeds[0]).sum() == 0
+            assert not torch.isnan(
+                image_embeds[0]).any(), "image_embeds contains nan"
             image_embeds = [
                 image_embed.to(target_dtype) for image_embed in image_embeds
             ]
@@ -187,11 +195,13 @@ class DenoisingStage(PipelineStage):
         # Get latents and embeddings
         latents = batch.latents
         prompt_embeds = batch.prompt_embeds
-        assert torch.isnan(prompt_embeds[0]).sum() == 0
+        assert not torch.isnan(
+            prompt_embeds[0]).any(), "prompt_embeds contains nan"
         if batch.do_classifier_free_guidance:
             neg_prompt_embeds = batch.negative_prompt_embeds
             assert neg_prompt_embeds is not None
-            assert torch.isnan(neg_prompt_embeds[0]).sum() == 0
+            assert not torch.isnan(
+                neg_prompt_embeds[0]).any(), "neg_prompt_embeds contains nan"
 
         # (Wan2.2) Calculate timestep to switch from high noise expert to low noise expert
         if fastvideo_args.boundary_ratio is not None:
@@ -237,7 +247,8 @@ class DenoisingStage(PipelineStage):
                                                              patch_size[2])
             seq_len = int(math.ceil(seq_len / sp_world_size)) * sp_world_size
 
-        trajectory_timesteps: list[int] = []
+        # Initialize lists for ODE trajectory
+        trajectory_timesteps: list[torch.Tensor] = []
         trajectory_latents: list[torch.Tensor] = []
 
         # Run denoising loop
@@ -267,24 +278,14 @@ class DenoisingStage(PipelineStage):
 
                 # Expand latents for I2V
                 latent_model_input = latents.to(target_dtype)
-                if batch.image_latent is not None and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
+                if batch.image_latent is not None:
                     assert not fastvideo_args.pipeline_config.ti2v_task, "image latents should not be provided for TI2V task"
                     latent_model_input = torch.cat(
                         [latent_model_input, batch.image_latent],
                         dim=1).to(target_dtype)
-                elif batch.image_latent is not None and fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    assert batch.image_latent is not None, "image latents should be provided for T2V to I2V task"
-                    if rank_in_sp_group == 0:
-                        logger.info("latent_model_input.shape: %s",
-                                    latent_model_input.shape)
-                        latent_model_input = torch.cat([
-                            batch.image_latent,
-                            latent_model_input[:, :, 1:, :, :],
-                        ],
-                                                       dim=2).to(target_dtype)
-                        logger.info("latent_model_input.shape: %s",
-                                    latent_model_input.shape)
 
+                assert not torch.isnan(
+                    latent_model_input).any(), "latent_model_input contains nan"
                 if fastvideo_args.pipeline_config.ti2v_task and batch.pil_image is not None:
                     timestep = torch.stack([t]).to(get_local_torch_device())
                     temp_ts = (mask2[0][0][:, ::2, ::2] * timestep).flatten()
@@ -297,16 +298,8 @@ class DenoisingStage(PipelineStage):
                 else:
                     t_expand = t.repeat(latent_model_input.shape[0])
 
-                assert torch.isnan(latent_model_input).sum() == 0
                 latent_model_input = self.scheduler.scale_model_input(
                     latent_model_input, t)
-                if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    if rank_in_sp_group == 0:
-                        latent_model_input = torch.cat([
-                            batch.image_latent,
-                            latent_model_input[:, :, 1:, :, :],
-                        ],
-                                                       dim=2).to(target_dtype)
 
                 # Prepare inputs for transformer
                 guidance_expand = (
@@ -349,6 +342,31 @@ class DenoisingStage(PipelineStage):
                             assert attn_metadata is not None, "attn_metadata cannot be None"
                         else:
                             attn_metadata = None
+                    elif (vmoba_attn_available
+                          and self.attn_backend == VMOBAAttentionBackend):
+                        self.attn_metadata_builder_cls = self.attn_backend.get_builder_cls(
+                        )
+                        if self.attn_metadata_builder_cls is not None:
+                            self.attn_metadata_builder = self.attn_metadata_builder_cls(
+                            )
+                            # Prepare V-MoBA parameters from config
+                            moba_params = fastvideo_args.moba_config.copy()
+                            moba_params.update({
+                                "current_timestep":
+                                i,
+                                "raw_latent_shape":
+                                batch.raw_latent_shape[2:5],
+                                "patch_size":
+                                fastvideo_args.pipeline_config.dit_config.
+                                patch_size,
+                                "device":
+                                get_local_torch_device(),
+                            })
+                            attn_metadata = self.attn_metadata_builder.build(
+                                **moba_params)
+                            assert attn_metadata is not None, "attn_metadata cannot be None"
+                        else:
+                            attn_metadata = None
                     else:
                         attn_metadata = None
                     # TODO(will): finalize the interface. vLLM uses this to
@@ -416,7 +434,6 @@ class DenoisingStage(PipelineStage):
                 # save trajectory latents if needed
                 if batch.return_trajectory_latents:
                     trajectory_timesteps.append(t)
-                    # trajectory_latents.append(latents.cpu())
                     trajectory_latents.append(latents)
 
                 # Update progress bar
@@ -430,30 +447,25 @@ class DenoisingStage(PipelineStage):
         trajectory_tensor: torch.Tensor | None = None
         if trajectory_latents:
             trajectory_tensor = torch.stack(trajectory_latents, dim=1)
+            trajectory_timesteps_tensor = torch.stack(trajectory_timesteps,
+                                                      dim=0)
         else:
             trajectory_tensor = None
+            trajectory_timesteps_tensor = None
 
+        # Gather results if using sequence parallelism
         if sp_group:
             latents = sequence_model_parallel_all_gather(latents, dim=2)
             if batch.return_trajectory_latents:
-                # logger.info("before stack trajectory_latents.shape: %s", trajectory_latents[0].shape)
-                logger.info("after stack trajectory_latents.shape: %s", trajectory_tensor.shape)
                 trajectory_tensor = trajectory_tensor.to(
                     get_local_torch_device())
                 trajectory_tensor = sequence_model_parallel_all_gather(
                     trajectory_tensor, dim=3)
 
-        if trajectory_tensor is not None:
-            batch.trajectory_timesteps = torch.tensor(trajectory_timesteps).cpu()
+        if trajectory_tensor is not None and trajectory_timesteps_tensor is not None:
+            batch.trajectory_timesteps = trajectory_timesteps_tensor.cpu()
             batch.trajectory_latents = trajectory_tensor.cpu()
 
-        if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            latents = torch.cat([
-                batch.image_latent,
-                latents[:, :, 1:, :, :],
-            ],
-                                dim=2)
-
         # Update batch with final latents
         batch.latents = latents
 
@@ -791,7 +803,8 @@ class DmdDenoisingStage(DenoisingStage):
 
         video_raw_latent_shape = latents.shape
         prompt_embeds = batch.prompt_embeds
-        assert torch.isnan(prompt_embeds[0]).sum() == 0
+        assert not torch.isnan(
+            prompt_embeds[0]).any(), "prompt_embeds contains nan"
         timesteps = torch.tensor(
             fastvideo_args.pipeline_config.dmd_denoising_steps,
             dtype=torch.long,
@@ -830,7 +843,8 @@ class DmdDenoisingStage(DenoisingStage):
                         batch.image_latent.permute(0, 2, 1, 3, 4)
                     ],
                                                    dim=2).to(target_dtype)
-                assert torch.isnan(latent_model_input).sum() == 0
+                assert not torch.isnan(
+                    latent_model_input).any(), "latent_model_input contains nan"
 
                 # Prepare inputs for transformer
                 t_expand = t.repeat(latent_model_input.shape[0])

fastvideo/pipelines/stages/encoding.py

@@ -92,24 +92,6 @@ class EncodingStage(PipelineStage):
                 latents = latents.to(vae_dtype)
             latents = self.vae.encode(latents).mean
 
-        # Apply shifting if needed (reverse of decoding)
-        if (hasattr(self.vae, "shift_factor")
-                and self.vae.shift_factor is not None):
-            if isinstance(self.vae.shift_factor, torch.Tensor):
-                latents -= self.vae.shift_factor.to(latents.device,
-                                                    latents.dtype)
-            else:
-                latents -= self.vae.shift_factor
-
-        # Apply scaling factor
-        if (hasattr(self.vae, "scaling_factor")
-                and self.vae.scaling_factor is not None):
-            if isinstance(self.vae.scaling_factor, torch.Tensor):
-                latents = latents * self.vae.scaling_factor.to(
-                    latents.device, latents.dtype)
-            else:
-                latents = latents * self.vae.scaling_factor
-
         # Update batch with encoded latents
         batch.latents = latents
 

fastvideo/pipelines/stages/image_encoding.py

@@ -105,81 +105,6 @@ class ImageVAEEncodingStage(PipelineStage):
     def __init__(self, vae: ParallelTiledVAE) -> None:
         self.vae: ParallelTiledVAE = vae
 
-    def encode_image(self,
-                     image: PIL.Image.Image,
-                     height: int,
-                     width: int,
-                     fastvideo_args: FastVideoArgs,
-                     generator: torch.Generator | None = None) -> torch.Tensor:
-        """
-        Encode image into latent space.
-        """
-        image = self.preprocess(
-            image,
-            vae_scale_factor=self.vae.spatial_compression_ratio,
-            height=height,
-            width=width).to(get_local_torch_device(), dtype=torch.float32)
-
-        # (B, C, H, W) -> (B, C, 1, H, W)
-        print(f"image.shape: {image.shape}")
-        image = image.unsqueeze(2)
-        print(f"after unsqueeze image.shape: {image.shape}")
-        return self.encode_tensor(image, fastvideo_args, generator)
-
-    def encode_tensor(self,
-                      video_condition: torch.Tensor,
-                      fastvideo_args: FastVideoArgs,
-                      generator: torch.Generator | None = None) -> torch.Tensor:
-        """
-        Encode frames into latent space.
-        """
-        self.vae = self.vae.to(get_local_torch_device())
-        video_condition = video_condition.to(device=get_local_torch_device(),
-                                             dtype=torch.float32)
-
-        # Setup VAE precision
-        vae_dtype = PRECISION_TO_TYPE[
-            fastvideo_args.pipeline_config.vae_precision]
-        vae_autocast_enabled = (
-            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
-
-        # Encode Image
-        with torch.autocast(device_type="cuda",
-                            dtype=vae_dtype,
-                            enabled=vae_autocast_enabled):
-            if fastvideo_args.pipeline_config.vae_tiling:
-                self.vae.enable_tiling()
-            # if fastvideo_args.vae_sp:
-            #     self.vae.enable_parallel()
-            if not vae_autocast_enabled:
-                video_condition = video_condition.to(vae_dtype)
-            encoder_output = self.vae.encode(video_condition)
-
-        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
-            latent_condition = encoder_output.mean
-        else:
-            generator = generator
-            if generator is None:
-                raise ValueError("Generator must be provided")
-            latent_condition = self.retrieve_latents(encoder_output, generator)
-
-        # Apply shifting if needed
-        if (hasattr(self.vae, "shift_factor")
-                and self.vae.shift_factor is not None):
-            if isinstance(self.vae.shift_factor, torch.Tensor):
-                latent_condition -= self.vae.shift_factor.to(
-                    latent_condition.device, latent_condition.dtype)
-            else:
-                latent_condition -= self.vae.shift_factor
-
-        if isinstance(self.vae.scaling_factor, torch.Tensor):
-            latent_condition = latent_condition * self.vae.scaling_factor.to(
-                latent_condition.device, latent_condition.dtype)
-        else:
-            latent_condition = latent_condition * self.vae.scaling_factor
-
-        return latent_condition
-
     def forward(
         self,
         batch: ForwardBatch,
@@ -232,28 +157,57 @@ class ImageVAEEncodingStage(PipelineStage):
         # (B, C, H, W) -> (B, C, 1, H, W)
         image = image.unsqueeze(2)
 
-        if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            # repeat the image self.vae.temporal_compression_ratio times
-            video_condition = image.repeat(1, 1,
-                                           self.vae.temporal_compression_ratio,
-                                           1, 1)
-            # video_condition = image
-            logger.info("video_condition.shape: %s", video_condition.shape)
-        else:
-            video_condition = torch.cat([
-                image,
-                image.new_zeros(image.shape[0], image.shape[1], num_frames - 1,
-                                image.shape[3], image.shape[4])
-            ],
-                                        dim=2)
+        video_condition = torch.cat([
+            image,
+            image.new_zeros(image.shape[0], image.shape[1], num_frames - 1,
+                            image.shape[3], image.shape[4])
+        ],
+                                    dim=2)
+        video_condition = video_condition.to(device=get_local_torch_device(),
+                                             dtype=torch.float32)
 
-        latent_condition = self.encode_tensor(video_condition, fastvideo_args,
-                                              batch.generator)
+        # Setup VAE precision
+        vae_dtype = PRECISION_TO_TYPE[
+            fastvideo_args.pipeline_config.vae_precision]
+        vae_autocast_enabled = (
+            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
+
+        # Encode Image
+        with torch.autocast(device_type="cuda",
+                            dtype=vae_dtype,
+                            enabled=vae_autocast_enabled):
+            if fastvideo_args.pipeline_config.vae_tiling:
+                self.vae.enable_tiling()
+            # if fastvideo_args.vae_sp:
+            #     self.vae.enable_parallel()
+            if not vae_autocast_enabled:
+                video_condition = video_condition.to(vae_dtype)
+            encoder_output = self.vae.encode(video_condition)
+
+        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
+            latent_condition = encoder_output.mean
+        else:
+            generator = batch.generator
+            if generator is None:
+                raise ValueError("Generator must be provided")
+            latent_condition = self.retrieve_latents(encoder_output, generator)
+
+        # Apply shifting if needed
+        if (hasattr(self.vae, "shift_factor")
+                and self.vae.shift_factor is not None):
+            if isinstance(self.vae.shift_factor, torch.Tensor):
+                latent_condition -= self.vae.shift_factor.to(
+                    latent_condition.device, latent_condition.dtype)
+            else:
+                latent_condition -= self.vae.shift_factor
+
+        if isinstance(self.vae.scaling_factor, torch.Tensor):
+            latent_condition = latent_condition * self.vae.scaling_factor.to(
+                latent_condition.device, latent_condition.dtype)
+        else:
+            latent_condition = latent_condition * self.vae.scaling_factor
 
         if fastvideo_args.mode == ExecutionMode.PREPROCESS:
-            batch.image_latent = latent_condition
-        elif fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            logger.info("latent_condition.shape: %s", latent_condition.shape)
             batch.image_latent = latent_condition
         else:
             mask_lat_size = torch.ones(1, 1, num_frames, latent_height,

fastvideo/pipelines/stages/input_validation.py

@@ -35,15 +35,9 @@ class InputValidationStage(PipelineStage):
         """Generate seeds for the inference"""
         seed = batch.seed
         num_videos_per_prompt = batch.num_videos_per_prompt
-        if isinstance(batch.prompt, list):
-            num_prompts = len(batch.prompt)
-        else:
-            num_prompts = 1
-
-        total_num_videos = num_prompts * num_videos_per_prompt
 
         assert seed is not None
-        seeds = [seed + i for i in range(total_num_videos)]
+        seeds = [seed + i for i in range(num_videos_per_prompt)]
         batch.seeds = seeds
         # Peiyuan: using GPU seed will cause A100 and H100 to generate different results...
         batch.generator = [

fastvideo/pipelines/stages/latent_preparation.py

@@ -59,10 +59,6 @@ class LatentPreparationStage(PipelineStage):
 
         # Adjust batch size for number of videos per prompt
         batch_size *= batch.num_videos_per_prompt
-        logger.info(f"===== batch_size: {batch_size}")
-        logger.info(f"===== batch.prompt: {batch.prompt}")
-        logger.info(f"===== batch.prompt_embeds: {batch.prompt_embeds}")
-        logger.info(f"===== batch.prompt_attention_mask: {batch.prompt_attention_mask}")
 
         # Get required parameters
         dtype = batch.prompt_embeds[0].dtype

fastvideo/platforms/__init__.py

@@ -82,8 +82,8 @@ def rocm_platform_plugin() -> str | None:
                 logger.info("ROCm platform is available")
         finally:
             amdsmi.amdsmi_shut_down()
-    except Exception:
-        pass
+    except Exception as e:
+        logger.info("ROCm platform is unavailable: %s", e)
 
     return "fastvideo.platforms.rocm.RocmPlatform" if is_rocm else None
 

fastvideo/platforms/cuda.py

@@ -159,6 +159,20 @@ class CudaPlatformBase(Platform):
                     str(e))
                 raise ImportError(
                     "Video Sparse Attention backend is not installed. ") from e
+        elif selected_backend == AttentionBackendEnum.VMOBA_ATTN:
+            try:
+                from csrc.attn.vmoba_attn.vmoba import (  # noqa: F401
+                    moba_attn_varlen)
+                from fastvideo.attention.backends.vmoba import (  # noqa: F401
+                    VMOBAAttentionBackend)
+                logger.info("Using Video MOBA Attention backend.")
+
+                return "fastvideo.attention.backends.vmoba.VMOBAAttentionBackend"
+            except ImportError as e:
+                logger.error(
+                    "Failed to import Video MoBA Attention backend: %s", str(e))
+                raise ImportError(
+                    "Video MoBA Attention backend is not installed. ") from e
         elif selected_backend == AttentionBackendEnum.TORCH_SDPA:
             logger.info("Using Torch SDPA backend.")
             return "fastvideo.attention.backends.sdpa.SDPABackend"

fastvideo/platforms/interface.py

@@ -19,6 +19,7 @@ class AttentionBackendEnum(enum.Enum):
     TORCH_SDPA = enum.auto()
     SAGE_ATTN = enum.auto()
     VIDEO_SPARSE_ATTN = enum.auto()
+    VMOBA_ATTN = enum.auto()
     NO_ATTENTION = enum.auto()
 
 

fastvideo/tests/dataset/test_parquet_io.py

@@ -0,0 +1,108 @@
+import os
+from pathlib import Path
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+from fastvideo.dataset.dataloader.parquet_io import (
+    ParquetDatasetWriter,
+    records_to_table,
+)
+
+
+def test_records_to_table_types():
+    schema = pa.schema([
+        pa.field("id", pa.string()),
+        pa.field("vae_latent_bytes", pa.binary()),
+        pa.field("vae_latent_shape", pa.list_(pa.int64())),
+        pa.field("duration_sec", pa.float64()),
+        pa.field("width", pa.int64()),
+    ])
+    records = [{
+        "id": "a",
+        "vae_latent_bytes": b"\x00\x01",
+        "vae_latent_shape": [1, 2, 3],
+        "duration_sec": 1.5,
+        "width": 640,
+    }]
+
+    table = records_to_table(records, schema)
+    assert table.schema == schema
+    assert table.num_rows == 1
+    cols = {name: table.column(name).to_pylist()[0] for name in schema.names}
+    assert cols["id"] == "a"
+    assert isinstance(cols["vae_latent_bytes"], (bytes, bytearray))
+    assert cols["vae_latent_shape"] == [1, 2, 3]
+    assert abs(cols["duration_sec"] - 1.5) < 1e-6
+    assert cols["width"] == 640
+
+
+def test_writer_flush_and_remainder(tmp_path: Path):
+    schema = pa.schema([pa.field("id", pa.string())])
+    records = [{"id": str(i)} for i in range(25)]
+    table = records_to_table(records, schema)
+
+    out_dir = tmp_path / "out"
+    writer = ParquetDatasetWriter(str(out_dir), samples_per_file=10)
+    writer.append_table(table)
+    written = writer.flush(num_workers=1)
+    assert written == 20
+
+    files = sorted(out_dir.rglob("*.parquet"))
+    assert len(files) == 2
+    total_rows = sum(pq.read_table(str(f)).num_rows for f in files)
+    assert total_rows == 20
+
+    # Append remainder to complete another chunk
+    extra = records_to_table([{"id": str(i)} for i in range(5)], schema)
+    writer.append_table(extra)
+    written2 = writer.flush(num_workers=1)
+    assert written2 == 10
+    files2 = sorted(out_dir.rglob("*.parquet"))
+    assert len(files2) == 3
+    total_rows2 = sum(pq.read_table(str(f)).num_rows for f in files2)
+    assert total_rows2 == 30
+
+
+def test_writer_flush_write_remainder(tmp_path: Path):
+    schema = pa.schema([pa.field("id", pa.string())])
+    # 25 rows, 10 per file => 2 full files + 1 remainder(5)
+    records = [{"id": str(i)} for i in range(25)]
+    table = records_to_table(records, schema)
+
+    out_dir = tmp_path / "out_last"
+    writer = ParquetDatasetWriter(str(out_dir), samples_per_file=10)
+    writer.append_table(table)
+    # First flush writes 20
+    written1 = writer.flush(num_workers=1)
+    assert written1 == 20
+    # Final flush with remainder
+    written2 = writer.flush(num_workers=1, write_remainder=True)
+    assert written2 == 5
+    files = sorted(out_dir.rglob("*.parquet"))
+    assert len(files) == 3
+    total_rows = sum(pq.read_table(str(f)).num_rows for f in files)
+    assert total_rows == 25
+
+
+def test_writer_parallel_workers(tmp_path: Path):
+    schema = pa.schema([pa.field("id", pa.string())])
+    # 40 rows, 10 per file => 4 files
+    records = [{"id": str(i)} for i in range(40)]
+    table = records_to_table(records, schema)
+
+    out_dir = tmp_path / "out_parallel"
+    writer = ParquetDatasetWriter(str(out_dir), samples_per_file=10)
+    writer.append_table(table)
+    written = writer.flush(num_workers=2)
+    assert written == 40
+
+    # Ensure files exist under worker subdirs
+    worker_dirs = [p for p in out_dir.iterdir() if p.is_dir() and p.name.startswith("worker_")]
+    assert len(worker_dirs) >= 1
+    files = sorted(out_dir.rglob("*.parquet"))
+    assert len(files) == 4
+    total_rows = sum(pq.read_table(str(f)).num_rows for f in files)
+    assert total_rows == 40
+
+

fastvideo/tests/inference/vmoba/test_vmoba_inference.py

@@ -0,0 +1,58 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+import subprocess
+from pathlib import Path
+
+def test_inference_vmoba():
+    """Test FastVideo VMOBA_ATTN inference pipeline"""
+
+    num_gpus = "1"
+    model_base = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+    output_dir = Path("outputs_video/vmoba_1.3B/")
+    moba_config = "fastvideo/configs/backend/vmoba/wan_1.3B_77_480_832.json"
+
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VMOBA_ATTN"
+
+    cmd = [
+        "fastvideo", "generate",
+        "--model-path", model_base,
+        "--sp-size", num_gpus,
+        "--tp-size", "1",
+        "--num-gpus", num_gpus,
+        "--dit-cpu-offload", "False",
+        "--vae-cpu-offload", "False",
+        "--text-encoder-cpu-offload", "True",
+        "--pin-cpu-memory", "False",
+        "--height", "480",
+        "--width", "832",
+        "--num-frames", "77",
+        "--num-inference-steps", "50",
+        "--moba-config-path", moba_config,
+        "--fps", "16",
+        "--guidance-scale", "6.0",
+        "--flow-shift", "8.0",
+        "--prompt", "A majestic lion strides across the golden savanna, its powerful frame glistening under the warm afternoon sun. The tall grass ripples gently in the breeze, enhancing the lion's commanding presence. The tone is vibrant, embodying the raw energy of the wild. Low angle, steady tracking shot, cinematic.",
+        "--negative-prompt", (
+            "Bright tones, overexposed, static, blurred details, subtitles, style, "
+            "works, paintings, images, static, overall gray, worst quality, low quality, "
+            "JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, "
+            "poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, "
+            "still picture, messy background, three legs, many people in the background, walking backwards"
+        ),
+        "--seed", "1024",
+        "--output-path", str(output_dir),
+    ]
+
+    subprocess.run(cmd, check=True)
+
+    assert output_dir.exists(), f"Output directory {output_dir} does not exist"
+
+    video_files = list(output_dir.glob("*.mp4"))
+    assert len(video_files) > 0, "No video files were generated"
+
+    for video_file in video_files:
+        assert video_file.stat().st_size > 0, f"Video file {video_file} is empty"
+
+if __name__ == "__main__":
+    test_inference_vmoba()

fastvideo/tests/modal/pr_test.py

@@ -102,10 +102,18 @@ def run_precision_tests_STA():
 def run_precision_tests_VSA():
     run_test("python csrc/attn/tests/test_vsa.py")
 
+@app.function(gpu="L40S:1", image=image, timeout=900)
+def run_precision_tests_vmoba():
+    run_test("pytest csrc/attn/vmoba_attn/tests/test_vmoba_attn.py")
+
+@app.function(gpu="L40S:1", image=image, timeout=900)
+def run_inference_tests_vmoba():
+    run_test('python fastvideo/tests/inference/vmoba/test_vmoba_inference.py')
+
 @app.function(gpu="L40S:1", image=image, timeout=3600)
 def run_inference_lora_tests():
     run_test("pytest ./fastvideo/tests/inference/lora/test_lora_inference_similarity.py -vs")
 
 @app.function(gpu="L40S:2", image=image, timeout=900)
 def run_distill_dmd_tests():
-    run_test("pytest ./fastvideo/tests/training/distill/test_distill_dmd.py -vs")
+    run_test("pytest ./fastvideo/tests/training/distill/test_distill_dmd.py -vs")

fastvideo/tests/ssim/README.md

@@ -4,7 +4,8 @@ The reference videos in the `*_reference_videos` directory are used as part of a
 
 run `bash update_reference_videos.sh` from inside the `fastvideo/tests/ssim/` directory after running `test_inference_similarity.py` to update reference videos. Note: make sure to update the path to the corresponding device.
 
-all reference videos are were generated on commit `4aeabbc629e0edf91477e80e795e7bb1823c71cb`
+reference videos were generated on commit `4aeabbc629e0edf91477e80e795e7bb1823c71cb`
+causal videos were generated on commit b318063c0a4618f1d5d99ea82ca67a06aad0d19d
 
 ## Generation Details
 
@@ -76,4 +77,4 @@ Wan2.1-I2V-14B-480P-Diffusers: {
 ### Image-to-Video Prompts
 
 1. "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
-    Image path: "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+    Image path: "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"

fastvideo/tests/ssim/test_causal_similarity.py

@@ -0,0 +1,152 @@
+# SPDX-License-Identifier: Apache-2.0
+import json
+import os
+
+import torch
+import pytest
+
+from fastvideo import VideoGenerator
+from fastvideo.logger import init_logger
+from fastvideo.tests.utils import compute_video_ssim_torchvision, write_ssim_results
+from fastvideo.worker.multiproc_executor import MultiprocExecutor
+
+logger = init_logger(__name__)
+
+device_name = torch.cuda.get_device_name()
+device_reference_folder_suffix = '_reference_videos'
+
+if "A40" in device_name:
+    device_reference_folder = "A40" + device_reference_folder_suffix
+elif "L40S" in device_name:
+    device_reference_folder = "L40S" + device_reference_folder_suffix
+
+# Base parameters from the shell script
+
+SF_WAN_T2V_PARAMS = {
+    "num_gpus": 1,
+    "model_path": "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers",
+    "height": 480,
+    "width": 832,
+    "num_frames": 81,
+    "num_inference_steps": 4,
+    "seed": 1024,
+    "sp_size": 1,
+    "tp_size": 1,
+}
+
+
+MODEL_TO_PARAMS = {
+    "SFWan2.1-T2V-1.3B-Diffusers": SF_WAN_T2V_PARAMS,
+}
+
+I2V_MODEL_TO_PARAMS = {
+}
+
+TEST_PROMPTS = [
+    "Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of a bustling street food market. The scene captures a mix of humor and authenticity. Mid-shot framing, vibrant lighting.",
+    # "A lone hiker stands atop a towering cliff, silhouetted against the vast horizon. The rugged landscape stretches endlessly beneath, its earthy tones blending into the soft blues of the sky. The scene captures the spirit of exploration and human resilience. High angle, dynamic framing, with soft natural lighting emphasizing the grandeur of nature."
+]
+
+I2V_TEST_PROMPTS = [
+    "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot.",
+]
+
+I2V_IMAGE_PATHS = [
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg",
+]
+
+@pytest.mark.parametrize("prompt", TEST_PROMPTS)
+@pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN"])
+@pytest.mark.parametrize("model_id", list(MODEL_TO_PARAMS.keys()))
+def test_causal_similarity(prompt, ATTENTION_BACKEND, model_id):
+    """
+    Test that runs inference with different parameters and compares the output
+    to reference videos using SSIM.
+    """
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = ATTENTION_BACKEND
+
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+
+    base_output_dir = os.path.join(script_dir, 'generated_videos', model_id)
+    output_dir = os.path.join(base_output_dir, ATTENTION_BACKEND)
+    output_video_name = f"{prompt[:100]}.mp4"
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    BASE_PARAMS = MODEL_TO_PARAMS[model_id]
+    num_inference_steps = BASE_PARAMS["num_inference_steps"]
+
+    init_kwargs = {
+        "num_gpus": BASE_PARAMS["num_gpus"],
+        "sp_size": BASE_PARAMS["sp_size"],
+        "tp_size": BASE_PARAMS["tp_size"],
+        "dit_cpu_offload": True,
+    }
+    if BASE_PARAMS.get("vae_sp"):
+        init_kwargs["vae_sp"] = True
+        init_kwargs["vae_tiling"] = True
+    #if "text-encoder-precision" in BASE_PARAMS:
+    #    init_kwargs["text_encoder_precisions"] = BASE_PARAMS["text-encoder-precision"]
+
+    generation_kwargs = {
+        "num_inference_steps": num_inference_steps,
+        "output_path": output_dir,
+        "height": BASE_PARAMS["height"],
+        "width": BASE_PARAMS["width"],
+        "num_frames": BASE_PARAMS["num_frames"],
+        "seed": BASE_PARAMS["seed"],
+    }
+    if "neg_prompt" in BASE_PARAMS:
+        generation_kwargs["neg_prompt"] = BASE_PARAMS["neg_prompt"]
+
+    generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs)
+    generator.generate_video(prompt, **generation_kwargs)
+
+    if isinstance(generator.executor, MultiprocExecutor):
+        generator.executor.shutdown()
+
+    assert os.path.exists(
+        output_dir), f"Output video was not generated at {output_dir}"
+
+    reference_folder = os.path.join(script_dir, device_reference_folder, model_id, ATTENTION_BACKEND)
+
+    if not os.path.exists(reference_folder):
+        logger.error("Reference folder missing")
+        raise FileNotFoundError(
+            f"Reference video folder does not exist: {reference_folder}")
+
+    # Find the matching reference video based on the prompt
+    reference_video_name = None
+
+    for filename in os.listdir(reference_folder):
+        if filename.endswith('.mp4') and prompt[:100] in filename:
+            reference_video_name = filename
+            break
+
+    if not reference_video_name:
+        logger.error(f"Reference video not found for prompt: {prompt} with backend: {ATTENTION_BACKEND}")
+        raise FileNotFoundError(f"Reference video missing")
+
+    reference_video_path = os.path.join(reference_folder, reference_video_name)
+    generated_video_path = os.path.join(output_dir, output_video_name)
+
+    logger.info(
+        f"Computing SSIM between {reference_video_path} and {generated_video_path}"
+    )
+    ssim_values = compute_video_ssim_torchvision(reference_video_path,
+                                                 generated_video_path,
+                                                 use_ms_ssim=True)
+
+    mean_ssim = ssim_values[0]
+    logger.info(f"SSIM mean value: {mean_ssim}")
+    logger.info(f"Writing SSIM results to directory: {output_dir}")
+
+    success = write_ssim_results(output_dir, ssim_values, reference_video_path,
+                                 generated_video_path, num_inference_steps,
+                                 prompt)
+
+    if not success:
+        logger.error("Failed to write SSIM results to file")
+
+    min_acceptable_ssim = 0.98
+    assert mean_ssim >= min_acceptable_ssim, f"SSIM value {mean_ssim} is below threshold {min_acceptable_ssim} for {model_id} with backend {ATTENTION_BACKEND}"

fastvideo/tests/ssim/test_inference_similarity.py

@@ -101,7 +101,7 @@ I2V_IMAGE_PATHS = [
 
 
 @pytest.mark.parametrize("prompt", I2V_TEST_PROMPTS)
-@pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN", "TORCH_SDPA"])
+@pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN"])
 @pytest.mark.parametrize("model_id", list(I2V_MODEL_TO_PARAMS.keys()))
 def test_i2v_inference_similarity(prompt, ATTENTION_BACKEND, model_id):
     """

fastvideo/tests/workflow/test_components_parquet_saver.py

@@ -0,0 +1,67 @@
+from pathlib import Path
+
+import numpy as np
+import pyarrow as pa
+import pyarrow.parquet as pq
+import torch
+
+from fastvideo.workflow.preprocess.components import ParquetDatasetSaver
+from fastvideo.pipelines.pipeline_batch_info import PreprocessBatch
+
+
+def _simple_record_creator(batch: PreprocessBatch) -> list[dict]:
+    # batch.latents will be converted to numpy by the saver before this call
+    assert isinstance(batch.latents, np.ndarray)
+    num = len(batch.video_file_name)
+    records = []
+    for i in range(num):
+        arr = batch.latents[i]
+        records.append({
+            "id": batch.video_file_name[i],
+            "data_bytes": arr.tobytes(),
+            "data_shape": list(arr.shape),
+        })
+    return records
+
+
+def test_parquet_dataset_saver_flush_and_last(tmp_path: Path):
+    # Schema for the simple record creator
+    schema = pa.schema([
+        pa.field("id", pa.string()),
+        pa.field("data_bytes", pa.binary()),
+        pa.field("data_shape", pa.list_(pa.int64())),
+    ])
+
+    B = 5
+    # Build a minimal PreprocessBatch
+    batch = PreprocessBatch(
+        data_type="video",
+        latents=torch.randn(B, 2),
+        prompt_embeds=[torch.randn(B, 1, 1)],
+        prompt_attention_mask=[torch.ones(B, 1)],
+    )
+    batch.video_file_name = [f"vid_{i}" for i in range(B)]
+
+    saver = ParquetDatasetSaver(
+        flush_frequency=10,  # higher than B to avoid auto-flush
+        samples_per_file=3,
+        schema=schema,
+        record_creator=_simple_record_creator,
+    )
+
+    out_dir = tmp_path / "saver_out"
+    saver.save_and_write_parquet_batch(batch, str(out_dir))
+    # First flush: should write one full file (3 rows), keep 2 in buffer
+    saver.flush_tables(str(out_dir))
+    files = sorted(out_dir.rglob("*.parquet"))
+    assert len(files) == 1
+    assert pq.read_table(str(files[0])).num_rows == 3
+
+    # Final flush: write remainder 2 rows
+    saver.flush_last(str(out_dir))
+    files2 = sorted(out_dir.rglob("*.parquet"))
+    assert len(files2) == 2
+    total = sum(pq.read_table(str(f)).num_rows for f in files2)
+    assert total == 5
+
+

fastvideo/training/distillation_pipeline.py

@@ -36,8 +36,9 @@ from fastvideo.training.activation_checkpoint import (
     apply_activation_checkpointing)
 from fastvideo.training.training_pipeline import TrainingPipeline
 from fastvideo.training.training_utils import (
-    clip_grad_norm_while_handling_failing_dtensor_cases, get_scheduler,
-    load_distillation_checkpoint, save_distillation_checkpoint, shift_timestep)
+    clip_grad_norm_while_handling_failing_dtensor_cases, count_trainable,
+    get_scheduler, load_distillation_checkpoint, save_distillation_checkpoint,
+    shift_timestep)
 from fastvideo.utils import is_vsa_available, set_random_seed
 
 import wandb  # isort: skip
@@ -68,11 +69,18 @@ class DistillationPipeline(TrainingPipeline):
     current_trainstep: int
     video_latent_shape: tuple[int, ...]
     video_latent_shape_sp: tuple[int, ...]
+    real_score_transformer: torch.nn.Module
+    fake_score_transformer: torch.nn.Module
 
     def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
         raise RuntimeError(
             "create_pipeline_stages should not be called for training pipeline")
 
+    def set_trainable(self) -> None:
+        super().set_trainable()
+        self.modules["real_score_transformer"].requires_grad_(False)
+        self.modules["vae"].requires_grad_(False)
+
     def initialize_training_pipeline(self, training_args: TrainingArgs):
         """Initialize the distillation training pipeline with multiple models."""
         logger.info("Initializing distillation pipeline...")
@@ -81,8 +89,6 @@ class DistillationPipeline(TrainingPipeline):
 
         self.noise_scheduler = self.get_module("scheduler")
         self.vae = self.get_module("vae")
-        self.vae.requires_grad_(False)
-
         self.timestep_shift = self.training_args.pipeline_config.flow_shift
         self.noise_scheduler = FlowMatchEulerDiscreteScheduler(
             shift=self.timestep_shift)
@@ -90,10 +96,7 @@ class DistillationPipeline(TrainingPipeline):
         # self.transformer is the generator model
         self.real_score_transformer = self.get_module("real_score_transformer")
         self.fake_score_transformer = self.get_module("fake_score_transformer")
-
-        self.real_score_transformer.requires_grad_(False)
         self.real_score_transformer.eval()
-        self.fake_score_transformer.requires_grad_(True)
         self.fake_score_transformer.train()
 
         if training_args.enable_gradient_checkpointing_type is not None:
@@ -167,9 +170,7 @@ class DistillationPipeline(TrainingPipeline):
     def _prepare_distillation(self,
                               training_batch: TrainingBatch) -> TrainingBatch:
         """Prepare training environment for distillation."""
-        self.transformer.requires_grad_(True)
         self.transformer.train()
-        self.fake_score_transformer.requires_grad_(True)
         self.fake_score_transformer.train()
 
         return training_batch
@@ -895,6 +896,14 @@ class DistillationPipeline(TrainingPipeline):
         else:
             set_random_seed(seed + self.global_rank)
 
+        # Check trainable params
+        num_trainable_generator = round(
+            count_trainable(self.transformer) / 1e9, 3)
+        num_trainable_critic = round(
+            count_trainable(self.fake_score_transformer) / 1e9, 3)
+        logger.info(
+            "rank: %s: # of trainable params in generator: %sB, # of trainable params in critic: %sB",
+            self.global_rank, num_trainable_generator, num_trainable_critic)
         # Set random seeds for deterministic training
         self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
             self.seed)

fastvideo/training/ode_causal_pipeline.py

@@ -1,392 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import sys
-from copy import deepcopy
-from typing import cast
-
-import torch
-import torch.nn.functional as F
-
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_ode_trajectory
-from fastvideo.distributed import get_local_torch_device
-from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
-from fastvideo.forward_context import set_forward_context
-from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
-from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
-    WanCausalDMDPipeline)
-from fastvideo.training.training_pipeline import TrainingPipeline
-from fastvideo.training.training_utils import (
-    clip_grad_norm_while_handling_failing_dtensor_cases)
-
-logger = init_logger(__name__)
-
-
-class ODEInitTrainingPipeline(TrainingPipeline):
-    """
-    Training pipeline for ODE-init using precomputed denoising trajectories.
-
-    Supervision: predict the next latent in the stored trajectory by
-    - feeding current latent at timestep t into the transformer to predict noise
-    - stepping the scheduler with the predicted noise
-    - minimizing MSE to the stored next latent at timestep t_next
-    """
-
-    _required_config_modules = ["scheduler", "transformer", "vae"]
-
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        # Match the preprocess/generation scheduler for consistent stepping
-        self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(
-            shift=fastvideo_args.pipeline_config.flow_shift)
-
-    def set_schemas(self):
-        self.train_dataset_schema = pyarrow_schema_ode_trajectory
-
-    def initialize_training_pipeline(self, training_args: TrainingArgs):
-        super().initialize_training_pipeline(training_args)
-
-        self.noise_scheduler = self.get_module("scheduler")
-        self.vae = self.get_module("vae")
-        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"
-        self.noise_scheduler = FlowMatchEulerDiscreteScheduler(
-            shift=self.timestep_shift)
-
-        # logger.info(f"ARG dmd_denoising_steps: {training_args.pipeline_config.dmd_denoising_steps}")
-        logger.info(
-            f"ARG dmd_denoising_steps: {self.training_args.pipeline_config.dmd_denoising_steps}"
-        )
-        self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 0],
-                                                dtype=torch.long,
-                                                device=get_local_torch_device())
-        # self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 250], dtype=torch.long, device=get_local_torch_device())
-        if training_args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
-            timesteps = torch.cat((self.noise_scheduler.timesteps.cpu(),
-                                   torch.tensor([0],
-                                                dtype=torch.float32))).cuda()
-            logger.info(f"timesteps: {timesteps}")
-            self.dmd_denoising_steps = timesteps[1000 -
-                                                 self.dmd_denoising_steps]
-            logger.info(
-                f"warped self.dmd_denoising_steps: {self.dmd_denoising_steps}")
-            assert False, "warp_denoising_step must be false"
-        else:
-            # assert False, "warp_denoising_step must be true"
-            logger.info("not warped")
-        self.dmd_denoising_steps = self.dmd_denoising_steps.to(
-            get_local_torch_device())
-
-        logger.info(f"denoising_step_list: {self.dmd_denoising_steps}")
-
-        logger.info(
-            "Initialized ODE-init training pipeline with %s denoising steps",
-            len(self.dmd_denoising_steps))
-        # Cache for nearest trajectory index per DMD step (computed lazily on first batch)
-        self._cached_closest_idx_per_dmd = None
-        self.num_train_timestep = self.noise_scheduler.num_train_timesteps
-        # self.min_timestep = int(self.training_args.min_timestep_ratio *
-        #                         self.num_train_timestep)
-        # self.max_timestep = int(self.training_args.max_timestep_ratio *
-        #                         self.num_train_timestep)
-        # self.real_score_guidance_scale = self.training_args.real_score_guidance_scale
-
-    def initialize_validation_pipeline(self, training_args: TrainingArgs):
-        logger.info("Initializing validation pipeline...")
-        args_copy = deepcopy(training_args)
-        args_copy.inference_mode = True
-        # Warm start validation with current transformer
-        self.validation_pipeline = WanCausalDMDPipeline.from_pretrained(
-            training_args.model_path,
-            args=args_copy,  # type: ignore
-            inference_mode=True,
-            loaded_modules={
-                "transformer": self.get_module("transformer"),
-            },
-            tp_size=training_args.tp_size,
-            sp_size=training_args.sp_size,
-            num_gpus=training_args.num_gpus,
-            pin_cpu_memory=training_args.pin_cpu_memory,
-            dit_cpu_offload=True)
-
-    def _get_next_batch(self, training_batch):  # type: ignore[override]
-        batch = next(self.train_loader_iter, None)  # type: ignore
-        if batch is None:
-            self.current_epoch += 1
-            logger.info("Starting epoch %s", self.current_epoch)
-            self.train_loader_iter = iter(self.train_dataloader)
-            batch = next(self.train_loader_iter)
-
-        # Required fields from parquet (ODE trajectory schema)
-        encoder_hidden_states = batch['text_embedding']
-        encoder_attention_mask = batch['text_attention_mask']
-        infos = batch['info_list']
-
-        # Trajectory tensors may include a leading singleton batch dim per row
-        trajectory_latents = batch['trajectory_latents']
-        if trajectory_latents.dim() == 7:
-            # [B, 1, S, C, T, H, W] -> [B, S, C, T, H, W]
-            trajectory_latents = trajectory_latents[:, 0]
-        elif trajectory_latents.dim() == 6:
-            # already [B, S, C, T, H, W]
-            pass
-        else:
-            raise ValueError(
-                f"Unexpected trajectory_latents dim: {trajectory_latents.dim()}"
-            )
-
-        trajectory_timesteps = batch['trajectory_timesteps']
-        if trajectory_timesteps.dim() == 3:
-            # [B, 1, S] -> [B, S]
-            trajectory_timesteps = trajectory_timesteps[:, 0]
-        elif trajectory_timesteps.dim() == 2:
-            # [B, S]
-            pass
-        else:
-            raise ValueError(
-                f"Unexpected trajectory_timesteps dim: {trajectory_timesteps.dim()}"
-            )
-        # [B, S, C, T, H, W] -> [B, S, T, C, H, W] to match self-forcing
-        trajectory_latents = trajectory_latents.permute(0, 1, 3, 2, 4, 5)
-
-        # Move to device
-        device = get_local_torch_device()
-        training_batch.encoder_hidden_states = encoder_hidden_states.to(
-            device, dtype=torch.bfloat16)
-        training_batch.encoder_attention_mask = encoder_attention_mask.to(
-            device, dtype=torch.bfloat16)
-        training_batch.infos = infos
-
-        return training_batch, trajectory_latents.to(
-            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
-
-    def _get_timestep(self,
-                      min_timestep: int,
-                      max_timestep: int,
-                      batch_size: int,
-                      num_frame: int,
-                      num_frame_per_block: int,
-                      uniform_timestep: bool = False) -> torch.Tensor:
-        if uniform_timestep:
-            timestep = torch.randint(min_timestep,
-                                     max_timestep, [batch_size, 1],
-                                     device=self.device,
-                                     dtype=torch.long).repeat(1, num_frame)
-            return timestep
-        else:
-            timestep = torch.randint(min_timestep,
-                                     max_timestep, [batch_size, num_frame],
-                                     device=self.device,
-                                     dtype=torch.long)
-            # logger.info(f"individual timestep: {timestep}")
-            # make the noise level the same within every block
-            timestep = timestep.reshape(timestep.shape[0], -1,
-                                        num_frame_per_block)
-            timestep[:, :, 1:] = timestep[:, :, 0:1]
-            timestep = timestep.reshape(timestep.shape[0], -1)
-            return timestep
-
-    def _step_predict_next_latent(
-        self, traj_latents: torch.Tensor, traj_timesteps: torch.Tensor,
-        encoder_hidden_states: torch.Tensor,
-        encoder_attention_mask: torch.Tensor
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        device = get_local_torch_device()
-        target_latent = traj_latents[:, -1]
-
-        # logger.info(f"traj_latents: {traj_latents.shape}")
-        # logger.info(f"traj_timesteps: {traj_timesteps.shape}")
-
-        # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
-        B, S, num_frames, num_channels, height, width = traj_latents.shape
-
-        # Lazily cache nearest trajectory index per DMD step based on the (fixed) S timesteps
-        if self._cached_closest_idx_per_dmd is None:
-            # Use the first sample's trajectory timesteps; assumed identical across batches
-            # s_steps = traj_timesteps[0].to(torch.long)  # [S]
-            # dmd = cast(torch.Tensor, self.dmd_denoising_steps).to(s_steps.device)  # [K]
-            # distances_ks: [K, S] = |s_steps - dmd|
-            # distances_ks = (s_steps.unsqueeze(0) - dmd.unsqueeze(1)).abs()
-            # self._cached_closest_idx_per_dmd = distances_ks.argmin(dim=1).to(torch.long).cpu()  # [K]
-            self._cached_closest_idx_per_dmd = torch.tensor(
-                [0, 12, 24, 36], dtype=torch.long).cpu()
-            # logger.info(f"self._cached_closest_idx_per_dmd: {self._cached_closest_idx_per_dmd}")
-            # logger.info(f"corresponding timesteps: {self.noise_scheduler.timesteps[self._cached_closest_idx_per_dmd]}")
-
-        # logger.info(f"traj_latents: {traj_latents.shape}")
-        # Select the K indexes from traj_latents using self._cached_closest_idx_per_dmd
-        # traj_latents: [B, S, C, T, H, W], self._cached_closest_idx_per_dmd: [K]
-        # Output: [B, K, C, T, H, W]
-        relevant_traj_latents = torch.index_select(
-            traj_latents,
-            dim=1,
-            index=self._cached_closest_idx_per_dmd.to(traj_latents.device))
-        # logger.info(f"relevant_traj_latents: {relevant_traj_latents.shape}")
-        # assert relevant_traj_latents.shape[0] == 1
-
-        indexes = self._get_timestep(  # [B, num_frames]
-            0,
-            len(self.dmd_denoising_steps),
-            B,
-            num_frames,
-            3,
-            uniform_timestep=False)
-        # noisy_input = relevant_traj_latents[indexes]
-        # logger.info(f"indexes: {indexes.shape}")
-        noisy_input = torch.gather(
-            relevant_traj_latents,
-            dim=1,
-            index=indexes.reshape(B, 1, num_frames, 1, 1,
-                                  1).expand(-1, -1, -1, num_channels, height,
-                                            width).to(self.device)).squeeze(1)
-        # noisy_input = noisy_input.unsqueeze(0)
-
-        # # Sample a single DMD step for the whole batch and fetch its cached nearest S-index
-        # K = len(self.dmd_denoising_steps)
-        # dmd_idx = torch.randint(0, K, (1,), device=device)
-        # logger.info(f"dmd_idx: {dmd_idx}")
-        # assert self._cached_closest_idx_per_dmd is not None
-        # nearest_s_idx = int(self._cached_closest_idx_per_dmd[int(dmd_idx.item())])
-        # nearest_idx = torch.full((B,), nearest_s_idx, device=device, dtype=torch.long)
-
-        # batch_indices = torch.arange(B, device=device)
-        # noisy_input = traj_latents[batch_indices, nearest_idx]  # [B, C, T, H, W]
-        # target_latent = traj_latents[batch_indices, -1]  # [B, C, T, H, W]
-        # t = traj_timesteps[batch_indices, nearest_idx]  # [B]
-
-        # Scale model input as in inference for consistency with stored trajectories
-        # noisy_input = self.modules["scheduler"].scale_model_input(noisy_input, t)
-        # logger.info(f"indexes: {indexes.shape}")
-        # logger.info(f"indexes: {indexes}")
-        timestep = self.dmd_denoising_steps[indexes]
-        # logger.info(f"timestep: {timestep.shape}")
-        # logger.info(f"timestep: {timestep}")
-
-        # Prepare inputs for transformer
-        input_kwargs = {
-            "hidden_states": noisy_input.permute(0, 2, 1, 3, 4),
-            "encoder_hidden_states": encoder_hidden_states,
-            "timestep": timestep.to(device, dtype=torch.bfloat16),
-            "encoder_attention_mask": encoder_attention_mask,
-            "return_dict": False,
-        }
-        # Predict noise and step the scheduler to obtain next latent
-        with set_forward_context(current_timestep=timestep,
-                                 attn_metadata=None,
-                                 forward_batch=None):
-            noise_pred = self.transformer(**input_kwargs).permute(0, 2, 1, 3, 4)
-            # logger.info(f"noise_pred: {noise_pred.shape}")
-        if isinstance(noise_pred, (tuple, list)):
-            noise_pred = noise_pred[0]
-
-        from fastvideo.models.utils import pred_noise_to_pred_video
-        noise_pred = pred_noise_to_pred_video(
-            pred_noise=noise_pred.flatten(0, 1),
-            noise_input_latent=noisy_input.flatten(0, 1),
-            timestep=timestep.flatten(0, 1),
-            scheduler=self.modules["scheduler"]).unflatten(
-                0, noise_pred.shape[:2])
-
-        # noisy_input = pred_noise_to_pred_video(noise_pred, noisy_input, t, self.modules["scheduler"])
-        # next_latent_pred = self.modules["scheduler"].step(
-        #     noise_pred, t, current_latents, return_dict=False)[0]
-        return noise_pred, target_latent, timestep
-
-    def train_one_step(self, training_batch):  # type: ignore[override]
-        self.transformer.train()
-        self.optimizer.zero_grad()
-        training_batch.total_loss = 0.0
-        args = cast(TrainingArgs, self.training_args)
-
-        # Using cached nearest index per DMD step; computation happens in _step_predict_next_latent
-
-        for _ in range(args.gradient_accumulation_steps):
-            training_batch, traj_latents, traj_timesteps = self._get_next_batch(
-                training_batch)
-            text_embeds = training_batch.encoder_hidden_states
-            text_attention_mask = training_batch.encoder_attention_mask
-            assert traj_latents.shape[0] == 1
-
-            # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
-            B, S = traj_latents.shape[0], traj_latents.shape[1]
-            if S < 2:
-                raise ValueError("Trajectory must contain at least 2 steps")
-
-            # Sample per-sample current step i in [0, S-2]
-
-            # idx = torch.randint(low=0, high=S - 1, size=(B, ),
-            #                     device=traj_latents.device)
-
-            # Gather current latents and next latents
-            # batch_indices = torch.arange(B, device=traj_latents.device)
-            # current_latents = traj_latents[batch_indices, idx]  # [B, C, T,H,W]
-            # current_latent = traj_timesteps[:, -1, :, :, :, :]
-            # target_latents = traj_latents[:, -1, :, :, :, :]
-
-            # Corresponding timesteps t (long) -> cast per sample
-            # t = traj_timesteps[:, -1, :, :, :, :]
-            # if t.dtype != torch.long:
-            #     t = t.long()
-
-            # Forward to predict next latent by stepping scheduler with predicted noise
-            noise_pred, target_latent, t = self._step_predict_next_latent(
-                traj_latents, traj_timesteps, text_embeds, text_attention_mask)
-
-            mask = t != 0
-
-            # Compute loss
-            loss = F.mse_loss(noise_pred[mask],
-                              target_latent[mask],
-                              reduction="mean")
-            loss = loss / args.gradient_accumulation_steps
-
-            with set_forward_context(current_timestep=t,
-                                     attn_metadata=None,
-                                     forward_batch=None):
-                loss.backward()
-            avg_loss = loss.detach().clone()
-            training_batch.total_loss += avg_loss.item()
-
-        # Clip grad and step optimizers
-        grad_norm = clip_grad_norm_while_handling_failing_dtensor_cases(
-            [p for p in self.transformer.parameters() if p.requires_grad],
-            args.max_grad_norm if args.max_grad_norm is not None else 0.0)
-
-        self.optimizer.step()
-        self.lr_scheduler.step()
-
-        if grad_norm is None:
-            grad_value = 0.0
-        else:
-            try:
-                if isinstance(grad_norm, torch.Tensor):
-                    grad_value = float(grad_norm.detach().float().item())
-                else:
-                    grad_value = float(grad_norm)
-            except Exception:
-                grad_value = 0.0
-        training_batch.grad_norm = grad_value
-        return training_batch
-
-
-def main(args) -> None:
-    logger.info("Starting ODE-init training pipeline...")
-    logger.info(f"ARG dmd_denoising_steps: {args.dmd_denoising_steps}")
-    pipeline = ODEInitTrainingPipeline.from_pretrained(
-        args.pretrained_model_name_or_path, args=args)
-    args = pipeline.training_args
-    pipeline.train()
-    logger.info("ODE-init training pipeline done")
-
-
-if __name__ == "__main__":
-    argv = sys.argv
-    from fastvideo.fastvideo_args import TrainingArgs
-    from fastvideo.utils import FlexibleArgumentParser
-    parser = FlexibleArgumentParser()
-    parser = TrainingArgs.add_cli_args(parser)
-    parser = FastVideoArgs.add_cli_args(parser)
-    args = parser.parse_args()
-    args.dit_cpu_offload = False
-    main(args)

fastvideo/training/training_pipeline.py

@@ -22,6 +22,7 @@ from tqdm.auto import tqdm
 import fastvideo.envs as envs
 from fastvideo.attention.backends.video_sparse_attn import (
     VideoSparseAttentionMetadataBuilder)
+from fastvideo.attention.backends.vmoba import VideoMobaAttentionMetadataBuilder
 from fastvideo.configs.sample import SamplingParam
 from fastvideo.dataset import build_parquet_map_style_dataloader
 from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
@@ -38,22 +39,20 @@ from fastvideo.training.activation_checkpoint import (
     apply_activation_checkpointing)
 from fastvideo.training.training_utils import (
     clip_grad_norm_while_handling_failing_dtensor_cases,
-    compute_density_for_timestep_sampling, get_scheduler, get_sigmas,
-    load_checkpoint, normalize_dit_input, save_checkpoint,
+    compute_density_for_timestep_sampling, count_trainable, get_scheduler,
+    get_sigmas, load_checkpoint, normalize_dit_input, save_checkpoint,
     shard_latents_across_sp)
-from fastvideo.utils import is_vsa_available, set_random_seed, shallow_asdict
+from fastvideo.utils import (is_vmoba_available, is_vsa_available,
+                             set_random_seed, shallow_asdict)
 
 import wandb  # isort: skip
 
 vsa_available = is_vsa_available()
+vmoba_available = is_vmoba_available()
 
 logger = init_logger(__name__)
 
 
-def _get_trainable_params(model: torch.nn.Module) -> int:
-    return sum(p.numel() for p in model.parameters() if p.requires_grad)
-
-
 class TrainingPipeline(LoRAPipeline, ABC):
     """
     A pipeline for training a model. All training pipelines should inherit from this class.
@@ -113,6 +112,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 enable_gradient_checkpointing_type)
 
         noise_scheduler = self.modules["scheduler"]
+        # Set grads for proper modules based on the training mode (Distill, LoRA, etc.)
         self.set_trainable()
         params_to_optimize = self.transformer.parameters()
         params_to_optimize = list(
@@ -272,6 +272,20 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 patch_size=patch_size,
                 VSA_sparsity=current_vsa_sparsity,
                 device=get_local_torch_device())
+        elif vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
+            moba_params = self.training_args.moba_config.copy()
+            moba_params.update({
+                "current_timestep":
+                training_batch.timesteps,
+                "raw_latent_shape":
+                training_batch.raw_latent_shape[2:5],
+                "patch_size":
+                self.training_args.pipeline_config.dit_config.patch_size,
+                "device":
+                get_local_torch_device(),
+            })
+            training_batch.attn_metadata = VideoMobaAttentionMetadataBuilder(
+            ).build(**moba_params)
         else:
             training_batch.attn_metadata = None
 
@@ -296,7 +310,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
     def _transformer_forward_and_compute_loss(
             self, training_batch: TrainingBatch) -> TrainingBatch:
-        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN":
+        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN" or vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
             assert training_batch.attn_metadata is not None
         else:
             assert training_batch.attn_metadata is None
@@ -417,7 +431,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                     local_main_process_only=False)
         if not self.post_init_called:
             self.post_init()
-        num_trainable_params = _get_trainable_params(self.transformer)
+        num_trainable_params = count_trainable(self.transformer)
         logger.info("Starting training with %s B trainable parameters",
                     round(num_trainable_params / 1e9, 3))
 
@@ -462,6 +476,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 current_decay_times = min(step // vsa_decay_interval_steps,
                                           vsa_sparsity // vsa_decay_rate)
                 current_vsa_sparsity = current_decay_times * vsa_decay_rate
+            elif vmoba_available:
+                # TODO: add vmoba sparsity scheduling here
+                pass
             else:
                 current_vsa_sparsity = 0.0
 
@@ -506,7 +523,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 self._log_validation(self.transformer, self.training_args, step)
                 gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
                 trainable_params = round(
-                    _get_trainable_params(self.transformer) / 1e9, 3)
+                    count_trainable(self.transformer) / 1e9, 3)
                 logger.info(
                     "GPU memory usage after validation: %s MB, trainable params: %sB",
                     gpu_memory_usage, trainable_params)
@@ -542,7 +559,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
         logger.info("  Total optimization steps = %s",
                     self.training_args.max_train_steps)
         logger.info("  Total training parameters per FSDP shard = %s B",
-                    round(_get_trainable_params(self.transformer) / 1e9, 3))
+                    round(count_trainable(self.transformer) / 1e9, 3))
         # print dtype
         logger.info("  Master weight dtype: %s",
                     self.transformer.parameters().__next__().dtype)
@@ -610,7 +627,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
         validation_dataloader = DataLoader(validation_dataset,
                                            batch_size=None,
                                            num_workers=0)
-
         transformer.eval()
 
         validation_steps = training_args.validation_sampling_steps.split(",")

fastvideo/training/training_utils.py

@@ -346,14 +346,10 @@ def load_checkpoint(transformer,
     """
     if not os.path.exists(checkpoint_path):
         logger.warning("Checkpoint path %s does not exist", checkpoint_path)
-        assert False
         return 0
 
     # Extract step number from checkpoint path
-    try:
-        step = int(os.path.basename(checkpoint_path).split('-')[-1])
-    except:
-        step = 1
+    step = int(os.path.basename(checkpoint_path).split('-')[-1])
 
     if rank == 0:
         logger.info("Loading checkpoint from step %s", step)
@@ -1282,3 +1278,7 @@ def get_scheduler(
                          num_warmup_steps=num_warmup_steps,
                          num_training_steps=num_training_steps,
                          last_epoch=last_epoch)
+
+
+def count_trainable(model: torch.nn.Module) -> int:
+    return sum(p.numel() for p in model.parameters() if p.requires_grad)

fastvideo/training/wan_t2v_i2v_training_pipeline.py

@@ -1,211 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import sys
-from copy import deepcopy
-from typing import Any
-
-import torch
-
-from fastvideo.configs.pipelines.base import PipelineConfig
-from fastvideo.configs.sample import SamplingParam
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
-from fastvideo.distributed import get_local_torch_device
-from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
-from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_unipc_multistep import (
-    FlowUniPCMultistepScheduler)
-from fastvideo.pipelines.basic.wan.wan_i2v_pipeline import (
-    WanImageToVideoPipeline)
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch, TrainingBatch
-from fastvideo.training.training_pipeline import TrainingPipeline
-from fastvideo.utils import is_vsa_available, shallow_asdict
-
-vsa_available = is_vsa_available()
-
-logger = init_logger(__name__)
-
-
-class WanT2VI2VTrainingPipeline(TrainingPipeline):
-    """
-    A training pipeline for Wan.
-    """
-    _required_config_modules = ["scheduler", "transformer", "vae"]
-
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        self.modules["scheduler"] = FlowUniPCMultistepScheduler(
-            shift=fastvideo_args.pipeline_config.flow_shift)
-
-    def create_training_stages(self, training_args: TrainingArgs):
-        """
-        May be used in future refactors.
-        """
-        pass
-
-    def set_schemas(self):
-        self.train_dataset_schema = pyarrow_schema_t2v
-
-    def initialize_validation_pipeline(self, training_args: TrainingArgs):
-        logger.info("Initializing validation pipeline...")
-        args_copy = deepcopy(training_args)
-
-        args_copy.inference_mode = True
-        args_copy.dit_cpu_offload = True
-        # args_copy.pipeline_config.vae_config.load_encoder = False
-        # validation_pipeline = WanImageToVideoValidationPipeline.from_pretrained(
-        pipeline_config = PipelineConfig.from_pretrained(
-            training_args.model_path)
-        pipeline_config.vae_config.load_encoder = True
-        self.validation_pipeline = WanImageToVideoPipeline.from_pretrained(
-            training_args.model_path,
-            args=None,
-            inference_mode=True,
-            pipeline_config=pipeline_config,
-            loaded_modules={
-                "transformer": self.get_module("transformer"),
-            },
-            required_config_modules=[
-                "scheduler", "transformer", "vae", "text_encoder", "tokenizer"
-            ],
-            tp_size=training_args.tp_size,
-            sp_size=training_args.sp_size,
-            num_gpus=training_args.num_gpus,
-            dit_cpu_offload=True,
-        )
-
-    def _get_next_batch(self, training_batch: TrainingBatch) -> TrainingBatch:
-        batch = next(self.train_loader_iter, None)  # type: ignore
-        if batch is None:
-            self.current_epoch += 1
-            logger.info("Starting epoch %s", self.current_epoch)
-            # Reset iterator for next epoch
-            self.train_loader_iter = iter(self.train_dataloader)
-            # Get first batch of new epoch
-            batch = next(self.train_loader_iter)
-
-        latents = batch['vae_latent']
-        latents = latents[:, :, :self.training_args.num_latent_t]
-        encoder_hidden_states = batch['text_embedding']
-        encoder_attention_mask = batch['text_attention_mask']
-        # clip_features = batch['clip_feature']
-        # image_latents = batch['first_frame_latent']
-        # image_latents = image_latents[:, :, :self.training_args.num_latent_t]
-        # pil_image = batch['pil_image']
-        infos = batch['info_list']
-
-        training_batch.latents = latents.to(get_local_torch_device(),
-                                            dtype=torch.bfloat16)
-        training_batch.encoder_hidden_states = encoder_hidden_states.to(
-            get_local_torch_device(), dtype=torch.bfloat16)
-        training_batch.encoder_attention_mask = encoder_attention_mask.to(
-            get_local_torch_device(), dtype=torch.bfloat16)
-        # training_batch.preprocessed_image = pil_image.to(
-        #     get_local_torch_device())
-        # training_batch.image_embeds = clip_features.to(get_local_torch_device())
-        # training_batch.image_latents = image_latents.to(
-        #     get_local_torch_device())
-        training_batch.infos = infos
-
-        return training_batch
-
-    def _prepare_dit_inputs(self,
-                            training_batch: TrainingBatch) -> TrainingBatch:
-        """Override to properly handle I2V concatenation - call parent first, then concatenate image conditioning."""
-
-        # First, call parent method to prepare noise, timesteps, etc. for video latents
-        training_batch = super()._prepare_dit_inputs(training_batch)
-
-        latents = training_batch.latents
-        logger.info("latents.shape: %s", latents.shape)
-        first_frame_latent = latents[:, :, 0, :, :]
-
-        logger.info("first_frame_latent.shape: %s", first_frame_latent.shape)
-        logger.info("training_batch.noisy_model_input.shape: %s",
-                    training_batch.noisy_model_input.shape)
-
-        training_batch.noisy_model_input = torch.cat([
-            first_frame_latent.unsqueeze(2),
-            training_batch.noisy_model_input[:, :, 1:, :, :]
-        ],
-                                                     dim=2)
-
-        return training_batch
-
-    def _build_input_kwargs(self,
-                            training_batch: TrainingBatch) -> TrainingBatch:
-
-        # Image Embeds for conditioning
-        # image_embeds = training_batch.image_embeds
-        # assert torch.isnan(image_embeds).sum() == 0
-        # image_embeds = image_embeds.to(get_local_torch_device(),
-        #                                dtype=torch.bfloat16)
-        # encoder_hidden_states_image = image_embeds
-
-        # NOTE: noisy_model_input already contains concatenated image_latents from _prepare_dit_inputs
-        training_batch.input_kwargs = {
-            "hidden_states":
-            training_batch.noisy_model_input,
-            "encoder_hidden_states":
-            training_batch.encoder_hidden_states,
-            "timestep":
-            training_batch.timesteps.to(get_local_torch_device(),
-                                        dtype=torch.bfloat16),
-            "encoder_attention_mask":
-            training_batch.encoder_attention_mask,
-            # "encoder_hidden_states_image":
-            # encoder_hidden_states_image,
-            "return_dict":
-            False,
-        }
-        return training_batch
-
-    def _prepare_validation_batch(self, sampling_param: SamplingParam,
-                                  training_args: TrainingArgs,
-                                  validation_batch: dict[str, Any],
-                                  num_inference_steps: int) -> ForwardBatch:
-        sampling_param.prompt = validation_batch['prompt']
-        sampling_param.height = training_args.num_height
-        sampling_param.width = training_args.num_width
-        sampling_param.image_path = validation_batch['video_path']
-        sampling_param.num_inference_steps = num_inference_steps
-        sampling_param.data_type = "video"
-        assert self.seed is not None
-        sampling_param.seed = self.seed
-
-        latents_size = [(sampling_param.num_frames - 1) // 4 + 1,
-                        sampling_param.height // 8, sampling_param.width // 8]
-        n_tokens = latents_size[0] * latents_size[1] * latents_size[2]
-        temporal_compression_factor = training_args.pipeline_config.vae_config.arch_config.temporal_compression_ratio
-        num_frames = (training_args.num_latent_t -
-                      1) * temporal_compression_factor + 1
-        sampling_param.num_frames = num_frames
-        batch = ForwardBatch(
-            **shallow_asdict(sampling_param),
-            latents=None,
-            generator=torch.Generator(device="cpu").manual_seed(self.seed),
-            n_tokens=n_tokens,
-            eta=0.0,
-            VSA_sparsity=training_args.VSA_sparsity,
-        )
-
-        return batch
-
-
-def main(args) -> None:
-    logger.info("Starting training pipeline...")
-
-    pipeline = WanT2VI2VTrainingPipeline.from_pretrained(
-        args.pretrained_model_name_or_path, args=args)
-    args = pipeline.training_args
-    pipeline.train()
-    logger.info("Training pipeline done")
-
-
-if __name__ == "__main__":
-    argv = sys.argv
-    from fastvideo.fastvideo_args import TrainingArgs
-    from fastvideo.utils import FlexibleArgumentParser
-    parser = FlexibleArgumentParser()
-    parser = TrainingArgs.add_cli_args(parser)
-    parser = FastVideoArgs.add_cli_args(parser)
-    args = parser.parse_args()
-    args.dit_cpu_offload = False
-    main(args)

fastvideo/utils.py

@@ -2,10 +2,6 @@
 # Adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/utils.py
 
 import argparse
-from einops import rearrange
-import torchvision
-import numpy as np
-import imageio
 import ctypes
 import hashlib
 import importlib
@@ -27,10 +23,14 @@ from typing import Any, TypeVar, cast
 
 import cloudpickle
 import filelock
+import imageio
+import numpy as np
 import torch
+import torchvision
 import yaml
 from diffusers.loaders.lora_base import (
     _best_guess_weight_name)  # watch out for potetential removal from diffusers
+from einops import rearrange
 from huggingface_hub import snapshot_download
 from remote_pdb import RemotePdb
 from torch.distributed.fsdp import MixedPrecisionPolicy
@@ -818,6 +818,17 @@ def is_vsa_available() -> bool:
     return importlib.util.find_spec("vsa") is not None
 
 
+@lru_cache(maxsize=1)
+def is_vmoba_available() -> bool:
+    if importlib.util.find_spec("csrc.attn.vmoba_attn.vmoba") is None:
+        return False
+    try:
+        import flash_attn
+        return flash_attn.__version__ >= "2.7.4"
+    except Exception:
+        return False
+
+
 # adapted from: https://github.com/Wan-Video/Wan2.2/blob/main/wan/utils/utils.py
 def masks_like(tensor,
                zero=False,
@@ -881,7 +892,8 @@ def best_output_size(w, h, dw, dh, expected_area):
         return ow2, oh2
 
 
-def save_decoded_latents_as_video(decoded_latents: list[torch.Tensor], output_path: str, fps: int):
+def save_decoded_latents_as_video(decoded_latents: list[torch.Tensor],
+                                  output_path: str, fps: int):
     # Process outputs
     videos = rearrange(decoded_latents, "b c t h w -> t b c h w")
     frames = []

fastvideo/workflow/preprocess/components.py

@@ -1,19 +1,19 @@
 import dataclasses
 import gc
-import multiprocessing
 import os
 import random
 from collections.abc import Callable
-from concurrent.futures import ProcessPoolExecutor
 from typing import Any
 
 import numpy as np
 import pyarrow as pa
-import pyarrow.parquet as pq
 import torch
 from datasets import Dataset, Video, load_dataset
 
-from fastvideo.configs.configs import DatasetType, PreprocessConfig
+from fastvideo.configs.configs import (DatasetType, PreprocessConfig,
+                                       VideoLoaderType)
+from fastvideo.dataset.dataloader.parquet_io import (ParquetDatasetWriter,
+                                                     records_to_table)
 from fastvideo.distributed.parallel_state import get_world_rank, get_world_size
 from fastvideo.logger import init_logger
 from fastvideo.pipelines.pipeline_batch_info import PreprocessBatch
@@ -154,31 +154,17 @@ class VideoForwardBatchBuilder:
 
 
 class ParquetDatasetSaver:
-    """Component for saving and writing Parquet datasets"""
+    """Component for saving and writing Parquet datasets using shared parquet_io."""
 
-    def __init__(self,
-                 flush_frequency: int,
-                 samples_per_file: int,
-                 schema_fields: list[str],
-                 record_creator: Callable[..., list[dict[str, Any]]],
-                 file_writer_fn: Callable | None = None):
-        """
-        Initialize ParquetDatasetSaver
-        
-        Args:
-            schema_fields: schema fields list
-            record_creator: Function for creating records
-            file_writer_fn: Function for writing records to files, uses default implementation if None
-        """
+    def __init__(self, flush_frequency: int, samples_per_file: int,
+                 schema: pa.Schema,
+                 record_creator: Callable[..., list[dict[str, Any]]]):
         self.flush_frequency = flush_frequency
         self.samples_per_file = samples_per_file
-        self.schema_fields = schema_fields
+        self.schema = schema
         self.create_records_from_batch = record_creator
-        self.file_writer_fn: Callable[
-            [tuple], int] = file_writer_fn or self._default_file_writer_fn
-        self.all_tables: list[pa.Table] = []
         self.num_processed_samples: int = 0
-        self.num_saved_files: int = 0
+        self._writer: ParquetDatasetWriter | None = None
 
     def save_and_write_parquet_batch(
             self,
@@ -226,11 +212,12 @@ class ParquetDatasetSaver:
 
         if batch_data:
             self.num_processed_samples += len(batch_data)
-            # Convert batch data to PyArrow arrays
-            table = self._convert_batch_to_pyarrow_table(batch_data)
-
-            # Store the table in a list for later processing
-            self.all_tables.append(table)
+            table = records_to_table(batch_data, self.schema)
+            if self._writer is None:
+                os.makedirs(output_dir, exist_ok=True)
+                self._writer = ParquetDatasetWriter(
+                    out_dir=output_dir, samples_per_file=self.samples_per_file)
+            self._writer.append_table(table)
             logger.debug("Collected batch with %s samples", len(table))
 
         # If flush is needed
@@ -260,145 +247,35 @@ class ParquetDatasetSaver:
 
     def _convert_batch_to_pyarrow_table(self,
                                         batch_data: list[dict]) -> pa.Table:
-        """Convert batch data to PyArrow table"""
-        arrays = []
+        # Deprecated path, kept for backward compatibility if needed.
+        return records_to_table(batch_data, self.schema)
 
-        for field in self.schema_fields:
-            if field.endswith('_bytes'):
-                arrays.append(
-                    pa.array([record[field] for record in batch_data],
-                             type=pa.binary()))
-            elif field.endswith('_shape'):
-                arrays.append(
-                    pa.array([record[field] for record in batch_data],
-                             type=pa.list_(pa.int32())))
-            elif field in ['width', 'height', 'num_frames']:
-                arrays.append(
-                    pa.array([record[field] for record in batch_data],
-                             type=pa.int32()))
-            elif field in ['duration_sec', 'fps']:
-                arrays.append(
-                    pa.array([record[field] for record in batch_data],
-                             type=pa.float32()))
-            else:
-                arrays.append(pa.array([record[field]
-                                        for record in batch_data]))
+    def flush_tables(self, output_dir: str, write_remainder: bool = False):
+        """Flush buffered records to disk.
 
-        return pa.Table.from_arrays(arrays, names=self.schema_fields)
-
-    def flush_tables(self, output_dir: str):
-        """Flush collected tables to disk"""
-        if not hasattr(self, 'all_tables') or not self.all_tables:
+        Args:
+            output_dir: Directory where parquet files are written. Kept for API
+                symmetry (writer already configured with this path).
+            write_remainder: If True, also write any leftover rows smaller than
+                ``samples_per_file`` as a final small file. Useful for the last flush.
+        """
+        if self._writer is None:
             return
+        _ = self._writer.flush(write_remainder=write_remainder)
+        # Reset processed sample count modulo samples_per_file
+        remainder = self.num_processed_samples % self.samples_per_file
+        self.num_processed_samples = 0 if write_remainder else remainder
 
-        logger.debug("Combining %d batches...", len(self.all_tables))
-        combined_table = pa.concat_tables(self.all_tables)
-        assert len(combined_table) == self.num_processed_samples
-        logger.debug("Total samples collected: %d", len(combined_table))
-
-        # Calculate total number of chunks needed, putting remainder into self.all_tables
-        total_files = max(self.num_processed_samples // self.samples_per_file,
-                          1)
-
-        logger.debug("Fixed samples per parquet file: %d",
-                     self.samples_per_file)
-        logger.debug("Total number of parquet files: %d", total_files)
-        logger.debug(
-            "Total samples to be processed: %d (putting %d samples into self.all_tables)",
-            total_files * self.samples_per_file,
-            self.num_processed_samples % self.samples_per_file)
-
-        # Split work among processes
-        num_workers = int(min(multiprocessing.cpu_count(), total_files))
-        files_per_worker = (total_files + num_workers - 1) // num_workers
-
-        logger.debug("Using %d workers to process %d files", num_workers,
-                     total_files)
-        logger.debug("Files per worker: %s", files_per_worker)
-
-        # Prepare work ranges
-        work_ranges = []
-        for i in range(num_workers):
-            start_idx = i * files_per_worker
-            end_idx = min((i + 1) * files_per_worker, total_files)
-            if start_idx < total_files:
-                work_ranges.append((start_idx, end_idx, combined_table, i,
-                                    output_dir, self.samples_per_file))
-
-        total_written = 0
-        failed_ranges = []
-        with ProcessPoolExecutor(max_workers=num_workers) as executor:
-            futures = {
-                executor.submit(self.file_writer_fn, work_range): work_range
-                for work_range in work_ranges
-            }
-            for future in futures:
-                try:
-                    written = future.result()
-                    total_written += written
-                    logger.info("Processed file with %s samples", written)
-                except Exception as e:
-                    work_range = futures[future]
-                    failed_ranges.append(work_range)
-                    logger.error("Failed to process range %s-%s: %s",
-                                 work_range[0], work_range[1], str(e))
-
-        # Retry failed ranges sequentially
-        if failed_ranges:
-            logger.warning("Retrying %s failed ranges sequentially",
-                           len(failed_ranges))
-            for work_range in failed_ranges:
-                try:
-                    total_written += self.file_writer_fn(work_range)
-                except Exception as e:
-                    logger.error(
-                        "Failed to process range %s-%s after retry: %s",
-                        work_range[0], work_range[1], str(e))
-
-        self.num_saved_files += total_files
-
-        # Clear tables list
-        self.all_tables = []
-        if self.num_processed_samples > self.samples_per_file:
-            saved_samples = total_files * self.samples_per_file
-            self.all_tables.append(combined_table.slice(saved_samples))
-            self.num_processed_samples -= saved_samples
-        else:
-            self.num_processed_samples = 0
-
-        del combined_table
-        gc.collect()
+    def flush_last(self, output_dir: str):
+        """Flush and write any remaining rows (final flush)."""
+        self.flush_tables(output_dir, write_remainder=True)
 
     def clean_up(self) -> None:
         """Clean up all tables"""
-        self.all_tables = []
+        self._writer = None
         self.num_processed_samples = 0
-        self.num_saved_files = 0
         gc.collect()
 
-    def _default_file_writer_fn(self, args_tuple: tuple) -> int:
-        """Default chunk processing implementation"""
-        start_idx, end_idx, combined_table, worker_id, output_dir, samples_per_file = args_tuple
-
-        written_count = 0
-        for file_idx in range(start_idx, end_idx):
-            start_row = file_idx * samples_per_file
-            end_row = min(start_row + samples_per_file, len(combined_table))
-
-            if start_row >= len(combined_table):
-                break
-
-            chunk_table = combined_table.slice(start_row, end_row - start_row)
-
-            # Write to file
-            output_file = os.path.join(
-                output_dir,
-                f"chunk_{file_idx + self.num_saved_files:06d}.parquet")
-            pq.write_table(chunk_table, output_file)
-            written_count += len(chunk_table)
-
-        return written_count
-
 
 def build_dataset(preprocess_config: PreprocessConfig, split: str,
                   validator: Callable[[dict[str, Any]], bool]) -> Dataset:
@@ -431,7 +308,8 @@ def build_dataset(preprocess_config: PreprocessConfig, split: str,
             return item
 
         dataset = dataset.map(add_video_column)
-        dataset = dataset.cast_column("video", Video())
+        if preprocess_config.video_loader_type == VideoLoaderType.TORCHCODEC:
+            dataset = dataset.cast_column("video", Video())
     else:
         raise ValueError(
             f"Invalid dataset type: {preprocess_config.dataset_type}")

fastvideo/workflow/preprocess/preprocess_workflow.py

@@ -85,16 +85,16 @@ class PreprocessWorkflow(WorkflowBase):
         # record creator
         if self.fastvideo_args.workload_type == WorkloadType.I2V:
             record_creator = i2v_record_creator
-            schema_fields = [f.name for f in pyarrow_schema_i2v]
+            schema = pyarrow_schema_i2v
         else:
             record_creator = basic_t2v_record_creator
-            schema_fields = [f.name for f in pyarrow_schema_t2v]
+            schema = pyarrow_schema_t2v
         processed_dataset_saver = ParquetDatasetSaver(
             flush_frequency=self.fastvideo_args.preprocess_config.
             flush_frequency,
             samples_per_file=self.fastvideo_args.preprocess_config.
             samples_per_file,
-            schema_fields=schema_fields,
+            schema=schema,
             record_creator=record_creator,
         )
         self.add_component("processed_dataset_saver", processed_dataset_saver)

scripts/checkpoint_conversion/wan_to_diffusers.py

@@ -135,8 +135,8 @@ _self_forcing_to_diffusers_param_names_mapping: dict = {
     r"blocks.\1.norm2.\2",
 }
 
-state_dict = load_state_dict_from_file("checkpoints/ode_init.pt")
-state_dict = state_dict["generator"]
+state_dict = load_state_dict_from_file("checkpoints/self_forcing_dmd.pt")
+state_dict = state_dict["generator_ema"]
 new_state_dict = OrderedDict()
 for k, v in state_dict.items():
     new_key = k
@@ -156,6 +156,6 @@ del state_dict
 
 save_torch_state_dict(
     new_state_dict,
-    "ode_init_diffusers/",
+    "new2/",
     max_shard_size="10GB"
 )

scripts/inference/v1_inference_wan_1.3B_VMoba.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+num_gpus=1
+export FASTVIDEO_ATTENTION_BACKEND=VMOBA_ATTN
+export MODEL_BASE=FastVideo/Wan2.1-T2V-1.3B-Diffusers
+# export MODEL_BASE=hunyuanvideo-community/HunyuanVideo
+# You can either use --prompt or --prompt-txt, but not both.
+fastvideo generate \
+    --model-path $MODEL_BASE \
+    --sp-size $num_gpus \
+    --tp-size 1 \
+    --num-gpus $num_gpus \
+    --dit-cpu-offload False \
+    --vae-cpu-offload False \
+    --text-encoder-cpu-offload True \
+    --pin-cpu-memory False \
+    --height 480 \
+    --width 832 \
+    --num-frames 77 \
+    --num-inference-steps 50 \
+    --moba-config-path fastvideo/configs/backend/vmoba/wan_1.3B_77_480_832.json \
+    --fps 16 \
+    --guidance-scale 6.0 \
+    --flow-shift 8.0 \
+    --prompt-txt assets/prompt.txt \
+    --negative-prompt "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" \
+    --seed 1024 \
+    --output-path outputs_video/