refactor sampling pipeline

Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com>

.buildkite/pipeline.yml

@@ -61,7 +61,7 @@ steps:
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
-                command: "timeout 60m .buildkite/scripts/pr_test.sh"
+                command: "timeout 90m .buildkite/scripts/pr_test.sh"
                 label: "SSIM Tests"
                 env:
                   - TEST_TYPE=ssim

.github/workflows/fastvideo-kernel-publish.yml

@@ -156,8 +156,23 @@ jobs:
           
           # Fix the wheel to be manylinux compliant
           pip install auditwheel
+          # Point auditwheel at torch libs, but do not vendor them into the wheel.
+          TORCH_LIB_DIR=$(python - <<'PY'
+          import os
+          import torch
+
+          print(os.path.join(os.path.dirname(torch.__file__), "lib"))
+          PY
+          )
+          export LD_LIBRARY_PATH="${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}"
           # Target manylinux_2_35 (Ubuntu 22.04 native)
-          auditwheel repair dist/*.whl --plat manylinux_2_35_x86_64 -w fixed_dist
+          auditwheel repair dist/*.whl --plat manylinux_2_35_x86_64 -w fixed_dist \
+            --exclude libtorch_cuda.so \
+            --exclude libtorch_cpu.so \
+            --exclude libtorch.so \
+            --exclude libc10.so \
+            --exclude libc10_cuda.so \
+            --exclude libtorch_python.so
           # Move fixed wheels back to dist for upload consistency
           rm dist/*.whl
           mv fixed_dist/*.whl dist/

.pre-commit-config.yaml

@@ -68,7 +68,7 @@ repos:
     entry: bash
     args:
       - -c
-      - 'git ls-files | grep -v "^fastvideo/tests/ssim/" | grep -v "^fastvideo/tests/inference/lora/L40S_reference_videos/" | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
+      - 'git ls-files | grep -v "^\"*fastvideo/tests/ssim/" | grep -v "^\"*fastvideo/tests/inference/lora/L40S_reference_videos/" | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
     language: system
     always_run: true
     pass_filenames: false

docs/attention/vsa/index.md

@@ -13,11 +13,13 @@ from fastvideo_kernel import video_sparse_attn
 
 # q, k, v: [batch_size, num_heads, seq_len, head_dim]
 # variable_block_sizes: Number of valid tokens per block
+# q_variable_block_sizes: Number of valid tokens per q block (can differ from KV for q/k of different lengths)
 # topk: Number of blocks to attend
 
 output = video_sparse_attn(
     q, k, v, 
-    variable_block_sizes=block_sizes, 
+    block_sizes,
+    block_sizes,
     topk=32
 )
 ```

examples/inference/basic/basic_cosmos2_5_t2w.py

@@ -0,0 +1,42 @@
+from fastvideo import VideoGenerator
+
+
+def main():
+    # Point this to your local diffusers model dir (or replace with a HF model ID).
+    model_path = "KyleShao/Cosmos-Predict2.5-2B-Diffusers"
+
+    generator = VideoGenerator.from_pretrained(
+        model_path,
+        num_gpus=1,
+        use_fsdp_inference=False,  # set True if GPU is out of memory
+        dit_cpu_offload=False,
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+    )
+
+    prompt = (
+       "A high-definition video captures the precision of robotic welding in an industrial setting. The first frame showcases a robotic arm, equipped with a welding torch, positioned over a large metal structure. The welding process is in full swing, with bright sparks and intense light illuminating the scene, creating a vivid display of blue and white hues. A significant amount of smoke billows around the welding area, partially obscuring the view but emphasizing the heat and activity. The background reveals parts of the workshop environment, including a ventilation system and various pieces of machinery, indicating a busy and functional industrial workspace. As the video progresses, the robotic arm maintains its steady position, continuing the welding process and moving to its left. The welding torch consistently emits sparks and light, and the smoke continues to rise, diffusing slightly as it moves upward. The metal surface beneath the torch shows ongoing signs of heating and melting. The scene retains its industrial ambiance, with the welding sparks and smoke dominating the visual field, underscoring the ongoing nature of the welding operation."
+    )
+
+    video = generator.generate_video(
+        prompt,
+        negative_prompt="",
+        height=704,
+        width=1280,
+        num_frames=77,
+        num_inference_steps=35,
+        guidance_scale=7.0,
+        fps=24,
+        output_path="outputs_video/cosmos2_5_t2w.mp4",
+        save_video=True,
+    )
+
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()
+
+
+

examples/training/rl/finetune_t2v_grpo.sh

@@ -0,0 +1,129 @@
+#!/bin/bash
+
+# Change to FastVideo root directory (3 levels up from this script)
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+FASTVIDEO_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
+cd "$FASTVIDEO_ROOT"
+
+# Add FastVideo root to PYTHONPATH so Python can find the fastvideo package
+export PYTHONPATH="$FASTVIDEO_ROOT${PYTHONPATH:+:$PYTHONPATH}"
+
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+RL_DATASET_DIR="data/ocr/"  # Path to RL prompt dataset directory (should contain train.txt and test.txt)
+VALIDATION_DATASET_FILE="$SCRIPT_DIR/validation.json"
+NUM_GPUS=1
+
+# use GPU 3
+export CUDA_VISIBLE_DEVICES=3
+
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_t2v_grpo"
+  --output_dir "checkpoints/wan_t2v_grpo"
+  --max_train_steps 5000
+  --train_batch_size 4
+  # --train_sp_batch_size 4
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 5
+  --num_height 240
+  --num_width 416
+  --num_frames 33
+  --lora_rank 32
+  --lora_training True
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS 
+  --sp_size $NUM_GPUS 
+  --tp_size $NUM_GPUS
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim $NUM_GPUS
+  # --use-fsdp-inference False
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments (for RL prompt dataset)
+dataset_args=(
+  --data_path $RL_DATASET_DIR  # Used as fallback if rl_dataset_path not set
+  --rl_dataset_path $RL_DATASET_DIR  # RL prompt dataset directory
+  --rl_dataset_type "text"  # "text" or "geneval"
+  --rl_num_image_per_prompt 4  # k parameter (number of samples per prompt)
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation True
+  --validation_dataset_file $VALIDATION_DATASET_FILE
+  --validation_steps 5
+  --validation_sampling_steps "50" 
+  --validation_guidance_scale "6.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 5e-5
+  --mixed_precision "bf16"
+  --weight_only_checkpointing_steps 10
+  --training_state_checkpointing_steps 10
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# RL-specific arguments
+rl_args=(
+  --inference_mode False
+  --rl_mode True
+  --rl_algorithm "grpo"
+  --rl_kl_beta 0.004  # KL regularization coefficient
+  --rl_policy_clip_range 0.2  # Policy clipping range for GRPO
+  --rl_kl_reward 0.0  # KL reward coefficient (typically 0)
+  --rl_global_std False  # Use per-prompt std (recommended for GRPO)
+  --rl_per_prompt_stat_tracking True  # Enable per-prompt stat tracking
+  --rl_warmup_steps 0  # Number of warmup steps (SFT before RL)
+  --reward-models "{\"paddle_ocr\": 1.0}" # use video_ocr reward function
+)
+
+# CFG arguments
+cfg_args=(
+  --guidance_scale 1.0 # use guidance_scale > 1.0 to enable CFG
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0  # No CFG during training (CFG used in sampling)
+  --dit_precision "fp32"
+  # --dit_precision "bf16"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+  # --resume_from_checkpoint "checkpoints/wan_t2v_grpo/checkpoint-XXX"
+  --enable-gradient-checkpointing-type "full"
+)
+
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+  --master_port 29501 \
+    "$FASTVIDEO_ROOT/fastvideo/training/wan_rl_training_pipeline.py" \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${rl_args[@]}" \
+    "${miscellaneous_args[@]}"

fastvideo-kernel/CMakeLists.txt

@@ -10,6 +10,8 @@ if(GPU_BACKEND STREQUAL "ROCM")
     enable_language(HIP)
 else()
     enable_language(CUDA)
+    # Ensure CUDA toolkit targets (CUDA::cudart, CUDA::cuda_driver, etc.) are available.
+    find_package(CUDAToolkit REQUIRED)
 endif()
 
 # Import common utils if needed, but we keep it simple for now
@@ -153,6 +155,30 @@ if(BUILD_CXX_KERNELS)
         $<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>
     )
 
+    # Link against Torch libraries to avoid undefined symbols at import time
+    # (e.g., torch::autograd vtables) when loading the extension module.
+    target_link_libraries(fastvideo_kernel_ops PRIVATE ${TORCH_LIBRARIES})
+
+    # Also link against libtorch_python to satisfy Python-binding symbols
+    # (e.g., torch::PyWarningHandler) required by torch/extension.h.
+    execute_process(
+        COMMAND "${Python_EXECUTABLE}" -c "import torch; from pathlib import Path; p=Path(torch.__file__).parent/'lib'; m=sorted(p.glob('libtorch_python*')); print(str(m[0]) if m else '')"
+        OUTPUT_VARIABLE TORCH_PYTHON_LIBRARY_PATH
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+    if(TORCH_PYTHON_LIBRARY_PATH)
+        message(STATUS "TORCH_PYTHON_LIBRARY_PATH: ${TORCH_PYTHON_LIBRARY_PATH}")
+        target_link_libraries(fastvideo_kernel_ops PRIVATE "${TORCH_PYTHON_LIBRARY_PATH}")
+    else()
+        message(WARNING "Could not locate libtorch_python; fastvideo_kernel_ops may fail to import.")
+    endif()
+
+    # Link CUDA runtime + driver explicitly (fixes missing symbols like cuGetErrorString at import time)
+    if(NOT GPU_BACKEND STREQUAL "ROCM")
+        target_link_libraries(fastvideo_kernel_ops PRIVATE CUDA::cudart CUDA::cuda_driver)
+    endif()
+
     # We install it to fastvideo_kernel/_C so we can load it to register the ops
     install(TARGETS fastvideo_kernel_ops LIBRARY DESTINATION fastvideo_kernel/_C)
 endif()

fastvideo-kernel/README.md

@@ -34,7 +34,7 @@ from fastvideo_kernel import sliding_tile_attention, video_sparse_attn, moba_att
 out = sliding_tile_attention(q, k, v, window_sizes, text_len)
 
 # Example: Video Sparse Attention (with Triton fallback)
-out = video_sparse_attn(q, k, v, block_sizes, topk=5)
+out = video_sparse_attn(q, k, v, block_sizes, block_sizes, topk=5)
 
 # Example: VMoBA
 out = moba_attn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, ...)

fastvideo-kernel/csrc/attention/block_sparse_h100.cu

@@ -639,7 +639,8 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
 
     // store kq and vq
 
-    // ensuring all writes are finished
+    // ! the following two line seems unnecessary.
+    // tma::store_async_wait(); // ensure qg is finished
     __syncthreads();
 
     warpgroup::store(kg_smem[0], kg_reg);
@@ -660,145 +661,6 @@ void bwd_attend_ker(const __grid_constant__ bwd_globals<D> g) {
     tma::store_async_wait(); 
 }
 
-
-template<int D>
-void block_sparse_attention_forward_impl(
-    bf16* d_q, bf16* d_k, bf16* d_v, float* d_l, bf16* d_o,
-    int batch, int qo_heads, int kv_heads, int seq_len, int hr,
-    int max_kv_blocks_per_q, 
-    int32_t* q2k_block_sparse_index_ptr,
-    int32_t* q2k_block_sparse_num_ptr,
-    int32_t* block_size_ptr,
-    cudaStream_t stream
-) {
-    using K = fwd_attend_ker_tile_dims<D>;
-    using q_tile    =         st_bf<K::qo_height, K::tile_width>;
-    using k_tile    =         st_bf<K::kv_height, K::tile_width>;
-    using v_tile    =         st_bf<K::kv_height, K::tile_width>;
-    using l_col_vec = col_vec<st_fl<K::qo_height, K::tile_width>>;
-    using o_tile    =         st_bf<K::qo_height, K::tile_width>;
-
-    using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
-    using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
-    using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
-    using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
-    using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
-
-    using globals      = fwd_globals<D>;
-
-    q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(seq_len)};
-    o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-
-    globals g{
-        qg_arg, kg_arg, vg_arg, lg_arg, og_arg, 
-        static_cast<int>(seq_len), static_cast<int>(hr), static_cast<int>(max_kv_blocks_per_q), 
-        q2k_block_sparse_index_ptr, q2k_block_sparse_num_ptr, block_size_ptr
-    };
-
-    // Shared memory size for the kernel
-    // 54000 bytes is calibrated for H100 shared memory constraints for these tile sizes
-    constexpr int mem_size = 54000;
-
-    dim3 grid(seq_len/(64), qo_heads, batch);
-
-    cudaFuncSetAttribute(
-        fwd_attend_ker<D>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize,
-        mem_size
-    );
-
-    fwd_attend_ker<D><<<grid, (128), mem_size, stream>>>(g);
-}
-
-template<int D>
-void block_sparse_attention_backward_impl(
-    bf16* d_q, bf16* d_k, bf16* d_v, bf16* d_o, bf16* d_og, float* d_l, float* d_d, float* d_qg, float* d_kg, float* d_vg,
-    int batch, int qo_heads, int kv_heads, int seq_len, int hr, int max_q_blocks_per_kv,
-    int32_t* k2q_block_sparse_index_ptr,
-    int32_t* k2q_block_sparse_num_ptr,
-    int32_t* block_size_ptr,
-    cudaStream_t stream
-) {
-    using G = bwd_attend_ker_tile_dims<D>;
-    using og_tile = st_bf<4*16, D>;
-    using o_tile  = st_bf<4*16, D>;
-    using d_tile  = col_vec<st_fl<4*16, D>>;
-
-    using og_global = gl<bf16,  -1, -1, -1, -1, og_tile>;
-    using o_global  = gl<bf16,  -1, -1, -1, -1, o_tile>;
-    using d_global  = gl<float, -1, -1, -1, -1, d_tile>;
-
-    using prep_globals = bwd_prep_globals<D>;
-
-    constexpr int mem_size_prep = kittens::MAX_SHARED_MEMORY;
-    int threads_prep = PREP_NUM_WARPS * kittens::WARP_THREADS;
-    dim3 grid_bwd_prep(seq_len/(PREP_NUM_WARPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
-
-    og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-
-    prep_globals bwd_g{prep_og_arg, prep_o_arg, prep_d_arg};
-
-    cudaFuncSetAttribute(
-        bwd_attend_prep_ker<D>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize,
-        mem_size_prep
-    );
-    bwd_attend_prep_ker<D><<<grid_bwd_prep, threads_prep, mem_size_prep, stream>>>(bwd_g);
-
-    using bwd_q_tile    =         st_bf<G::tile_h_qo, G::tile_width>;
-    using bwd_k_tile    =         st_bf<G::tile_h,    G::tile_width>;
-    using bwd_v_tile    =         st_bf<G::tile_h,    G::tile_width>;
-    using bwd_og_tile   =         st_bf<G::tile_h_qo, G::tile_width>;
-    using bwd_qg_tile   =         st_fl<G::tile_h_qo, G::tile_width>;
-    using bwd_kg_tile   =         st_fl<G::tile_h,    G::tile_width>;
-    using bwd_vg_tile   =         st_fl<G::tile_h,    G::tile_width>;
-    using bwd_l_tile    = row_vec<st_fl<G::tile_h_qo, G::tile_h>>;
-    using bwd_d_tile    = row_vec<st_fl<G::tile_h_qo, G::tile_h>>;
-
-    using bwd_q_global  = gl<bf16,  -1, -1, -1, -1, bwd_q_tile>;
-    using bwd_k_global  = gl<bf16,  -1, -1, -1, -1, bwd_k_tile>;
-    using bwd_v_global  = gl<bf16,  -1, -1, -1, -1, bwd_v_tile>;
-    using bwd_og_global = gl<bf16,  -1, -1, -1, -1, bwd_og_tile>;
-    using bwd_qg_global = gl<float, -1, -1, -1, -1, bwd_qg_tile>;
-    using bwd_kg_global = gl<float, -1, -1, -1, -1, bwd_kg_tile>;
-    using bwd_vg_global = gl<float, -1, -1, -1, -1, bwd_vg_tile>;
-    using bwd_l_global  = gl<float, -1, -1, -1, -1, bwd_l_tile>;
-    using bwd_d_global  = gl<float, -1, -1, -1, -1, bwd_d_tile>;
-
-    using bwd_global_args = bwd_globals<D>;
-
-    bwd_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), static_cast<uint32_t>(D)};
-    bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-    bwd_d_global  bwd_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-
-    bwd_global_args bwd_global{bwd_q_arg, bwd_k_arg, bwd_v_arg, bwd_og_arg, bwd_qg_arg, bwd_kg_arg, bwd_vg_arg, bwd_l_arg, bwd_d_arg, 
-                        static_cast<int>(seq_len), static_cast<int>(hr), static_cast<int>(max_q_blocks_per_kv),
-                        k2q_block_sparse_index_ptr, k2q_block_sparse_num_ptr, block_size_ptr};
-
-    dim3 grid_bwd_main(seq_len/64, qo_heads, batch);
-    int threads_main = 128;
-    // Calibrated shared memory sizes for different head dimensions
-    int bwd_mem_size = (D == 64) ? 72000 : 113000;
-
-    cudaFuncSetAttribute(
-        bwd_attend_ker<D>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize,
-        bwd_mem_size
-    );
-    bwd_attend_ker<D><<<grid_bwd_main, threads_main, bwd_mem_size, stream>>>(bwd_global);
-}
-
 #include "pyutils/torch_helpers.cuh"
 #include <ATen/cuda/CUDAContext.h>
 #include <iostream>
@@ -810,23 +672,32 @@ block_sparse_attention_forward(
     torch::Tensor v, 
     torch::Tensor q2k_block_sparse_index, 
     torch::Tensor q2k_block_sparse_num, 
-    torch::Tensor block_size
+    torch::Tensor kv_block_size
 )
 {
     CHECK_INPUT(q);
     CHECK_INPUT(k);
     CHECK_INPUT(v);
 
+    // q shape: (batch, qo_heads, q_seq_len, head_dim)
+    // k shape: (batch, kv_heads, kv_seq_len, head_dim)
+    // v shape: (batch, kv_heads, kv_seq_len, head_dim)
+    // q2k_block_sparse_index shape: (batch, qo_heads, num_q_blocks, max_kv_blocks_per_q)
+    // q2k_block_sparse_num shape: (batch, qo_heads, num_q_blocks)
+    // kv_block_size shape: (num_kv_blocks) This does not need other dimensions because across all batch/heads the padding is the same.
+    
     auto batch    = q.size(0);
-    auto seq_len  = q.size(2); 
+    auto q_seq_len  = q.size(2);
+    auto kv_seq_len = k.size(2);
     auto head_dim = q.size(3); 
     auto qo_heads = q.size(1);
     auto kv_heads = k.size(1);
     auto max_kv_blocks_per_q = q2k_block_sparse_index.size(3);
-    auto num_q_blocks = block_size.size(0);
+    auto num_q_blocks = q2k_block_sparse_index.size(2);
+    auto num_kv_blocks = kv_block_size.size(0);
     TORCH_CHECK(batch==1, "Batch size dim will be removed in the future, please set batch to 1");
-    TORCH_CHECK(num_q_blocks * 64 == seq_len, "This kernel supports variable block size, but it assumes the input sequence is properly padded.");
-    TORCH_CHECK(num_q_blocks == q2k_block_sparse_index.size(2), "Number of Q blocks does not match between q2k_block_sparse_index and block_size");
+    TORCH_CHECK(num_q_blocks * BLOCK_M == q_seq_len, "This kernel supports variable q block size, but it assumes the input sequence is properly padded.");
+    TORCH_CHECK(num_kv_blocks * BLOCK_M == kv_seq_len, "This kernel supports variable kv block size, but it assumes the input sequence is properly padded.");
     // check to see that these dimensions match for all inputs
     TORCH_CHECK(q.size(0) == batch, "Q batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(k.size(0) == batch, "K batch dimension - idx 0 - must match for all inputs");
@@ -834,11 +705,9 @@ block_sparse_attention_forward(
     TORCH_CHECK(q2k_block_sparse_index.size(0) == batch, "q2k_block_sparse_index batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(q2k_block_sparse_num.size(0) == batch, "q2k_block_sparse_num batch dimension - idx 0 - must match for all inputs");
 
-    TORCH_CHECK(q.size(2) == seq_len, "Q sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(k.size(2) == seq_len, "K sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(v.size(2) == seq_len, "V sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(q2k_block_sparse_index.size(2) == seq_len / BLOCK_M, "q2k_block_sparse_index idx 2 - must match seq_len / BLOCK_M");
-    TORCH_CHECK(q2k_block_sparse_num.size(2) == seq_len / BLOCK_M, "q2k_block_sparse_num idx 2 - must match seq_len / BLOCK_M");
+    TORCH_CHECK(v.size(2) == kv_seq_len, "V sequence length dimension - idx 2 - must match K inputs");
+    TORCH_CHECK(q2k_block_sparse_num.size(2) == num_q_blocks, "q2k_block_sparse_num idx 2 - must match num_q_blocks");
+
 
     TORCH_CHECK(q.size(3) == head_dim, "Q head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(k.size(3) == head_dim, "K head dimension - idx 3 - must match for all non-vector inputs");
@@ -864,12 +733,12 @@ block_sparse_attention_forward(
     // for the returned outputs
     torch::Tensor o     = torch::empty({static_cast<const uint>(batch), 
                                         static_cast<const uint>(qo_heads), 
-                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(q_seq_len), 
                                         static_cast<const uint>(head_dim)}, v.options());
     
     torch::Tensor l_vec = torch::empty({static_cast<const uint>(batch), 
                                         static_cast<const uint>(qo_heads), 
-                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(q_seq_len), 
                                         static_cast<const uint>(1)}, 
                                         torch::TensorOptions().dtype(torch::kFloat).device(q.device()).memory_format(at::MemoryFormat::Contiguous));
         
@@ -880,32 +749,110 @@ block_sparse_attention_forward(
     float* l_ptr = reinterpret_cast<float*>(l_vec.data_ptr<float>());
     float* d_l   = reinterpret_cast<float*>(l_ptr);
 
+    //cudadevicesynchronize();
     const c10::cuda::OptionalCUDAGuard device_guard(q.device());    
     const cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream(); 
 
-    // Temporated implementation to avoid code duplication between head_dim=64 and 128
     if (head_dim == 64) {
-        block_sparse_attention_forward_impl<64>(
-            d_q, d_k, d_v, d_l, d_o,
-            batch, qo_heads, kv_heads, seq_len, hr,
-            max_kv_blocks_per_q,
-            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()),
+        using q_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::qo_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+        using k_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::kv_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+        using v_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::kv_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+        using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<64>::qo_height, fwd_attend_ker_tile_dims<64>::tile_width>>;
+        using o_tile    =         st_bf<fwd_attend_ker_tile_dims<64>::qo_height, fwd_attend_ker_tile_dims<64>::tile_width>;
+
+        using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
+        using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
+        using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
+        using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
+        using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+        using globals      = fwd_globals<64>;
+
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(q_seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+
+        globals g{
+            qg_arg, 
+            kg_arg, 
+            vg_arg, 
+            lg_arg, 
+            og_arg, 
+            static_cast<int>(q_seq_len), 
+            static_cast<int>(hr), 
+            static_cast<int>(max_kv_blocks_per_q), 
+            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()), 
             reinterpret_cast<int32_t*>(q2k_block_sparse_num.data_ptr()),
-            reinterpret_cast<int32_t*>(block_size.data_ptr()),
-            stream
+            reinterpret_cast<int32_t*>(kv_block_size.data_ptr())
+        };
+
+        constexpr int mem_size = 54000;
+
+        dim3 grid(q_seq_len/(BLOCK_M), qo_heads, batch);
+
+        cudaFuncSetAttribute(
+            fwd_attend_ker<64>,
+            cudaFuncAttributeMaxDynamicSharedMemorySize,
+            mem_size
         );
-    } else if (head_dim == 128) {
-        block_sparse_attention_forward_impl<128>(
-            d_q, d_k, d_v, d_l, d_o,
-            batch, qo_heads, kv_heads, seq_len, hr,
-            max_kv_blocks_per_q,
-            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()),
+
+        fwd_attend_ker<64><<<grid, (128), mem_size, stream>>>(g);
+
+        CHECK_CUDA_ERROR(cudaGetLastError());
+        // cudaStreamSynchronize(stream);
+    }
+
+    if (head_dim == 128) {
+        using q_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using k_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::kv_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using v_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::kv_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+        using l_col_vec = col_vec<st_fl<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>>;
+        using o_tile    =         st_bf<fwd_attend_ker_tile_dims<128>::qo_height, fwd_attend_ker_tile_dims<128>::tile_width>;
+
+        using q_global = gl<bf16,  -1, -1, -1, -1, q_tile>;
+        using k_global = gl<bf16,  -1, -1, -1, -1, k_tile>;
+        using v_global = gl<bf16,  -1, -1, -1, -1, v_tile>;
+        using l_global = gl<float, -1, -1, -1, -1, l_col_vec>;
+        using o_global = gl<bf16,  -1, -1, -1, -1, o_tile>;
+
+        using globals      = fwd_globals<128>;
+
+        q_global qg_arg{d_q, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        k_global kg_arg{d_k, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        v_global vg_arg{d_v, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        l_global lg_arg{d_l, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,   static_cast<unsigned int>(q_seq_len)};
+        o_global og_arg{d_o, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+
+        globals g{
+            qg_arg, 
+            kg_arg, 
+            vg_arg, 
+            lg_arg, 
+            og_arg, 
+            static_cast<int>(q_seq_len), 
+            static_cast<int>(hr), 
+            static_cast<int>(max_kv_blocks_per_q), 
+            reinterpret_cast<int32_t*>(q2k_block_sparse_index.data_ptr()), 
             reinterpret_cast<int32_t*>(q2k_block_sparse_num.data_ptr()),
-            reinterpret_cast<int32_t*>(block_size.data_ptr()),
-            stream
+            reinterpret_cast<int32_t*>(kv_block_size.data_ptr())
+        };
+
+        constexpr int mem_size = 54000;
+
+        dim3 grid(q_seq_len/(BLOCK_M), qo_heads, batch);
+
+        cudaFuncSetAttribute(
+            fwd_attend_ker<128>,
+            cudaFuncAttributeMaxDynamicSharedMemorySize,
+            mem_size
         );
-    } else {
-        TORCH_CHECK(false, "Unsupported head_dim: ", head_dim, ". Only 64 and 128 are supported.");
+
+        fwd_attend_ker<128><<<grid, (128), mem_size, stream>>>(g);
+
+        CHECK_CUDA_ERROR(cudaGetLastError());
+        // cudaStreamSynchronize(stream);
     }
 
     return {o, l_vec};
@@ -921,7 +868,7 @@ block_sparse_attention_backward(torch::Tensor q,
                    torch::Tensor og,
                    torch::Tensor k2q_block_sparse_index,
                    torch::Tensor k2q_block_sparse_num,
-                   torch::Tensor block_size)
+                   torch::Tensor kv_block_size)
 {
     CHECK_INPUT(q);
     CHECK_INPUT(k);
@@ -930,11 +877,23 @@ block_sparse_attention_backward(torch::Tensor q,
     CHECK_INPUT(o);
     CHECK_INPUT(og);
 
+    // q: [batch, qo_heads, q_seq_len, head_dim]
+    // k: [batch, kv_heads, kv_seq_len, head_dim]
+    // v: [batch, kv_heads, kv_seq_len, head_dim]
+    // o: [batch, qo_heads, q_seq_len, head_dim]
+    // l_vec: [batch, qo_heads, q_seq_len, 1]
+    // og: [batch, qo_heads, q_seq_len, head_dim]
+    // k2q_block_sparse_index: [batch, kv_heads, num_kv_blocks, max_num_q_blocks]
+    // k2q_block_sparse_num: [batch, kv_heads, num_kv_blocks]
+    // kv_block_size: [num_kv_blocks]
+
     auto batch    = q.size(0);
-    auto seq_len  = q.size(2);
+    auto q_seq_len  = q.size(2);
+    auto kv_seq_len = k.size(2);
     auto head_dim = q.size(3);
     auto max_q_blocks_per_kv = k2q_block_sparse_index.size(3);
-    TORCH_CHECK(k2q_block_sparse_index.size(2) == block_size.size(0), "k2q_block_sparse_index.size(2) must match block_size.size(0)");
+    auto num_kv_blocks = kv_block_size.size(0);
+    TORCH_CHECK(k2q_block_sparse_index.size(2) == num_kv_blocks, "k2q_block_sparse_index.size(2) must match num_kv_blocks (kv_block_size.size(0))");
     // check to see that these dimensions match for all inputs
     TORCH_CHECK(q.size(0)     == batch, "Q  batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(k.size(0)     == batch, "K  batch dimension - idx 0 - must match for all inputs");
@@ -945,23 +904,18 @@ block_sparse_attention_backward(torch::Tensor q,
     TORCH_CHECK(k2q_block_sparse_index.size(0) == batch, "k2q_block_sparse_index batch dimension - idx 0 - must match for all inputs");
     TORCH_CHECK(k2q_block_sparse_num.size(0) == batch, "k2q_block_sparse_num batch dimension - idx 0 - must match for all inputs");
 
-
-    TORCH_CHECK(q.size(2)     == seq_len, "Q  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(k.size(2)     == seq_len, "K  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(v.size(2)     == seq_len, "V  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(l_vec.size(2) == seq_len, "L  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(o.size(2)     == seq_len, "O  sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(og.size(2)    == seq_len, "OG sequence length dimension - idx 2 - must match for all inputs");
-    TORCH_CHECK(k2q_block_sparse_index.size(2) == seq_len / BLOCK_N, "k2q_block_sparse_index idx 2 - must match seq_len / BLOCK_N");
-    TORCH_CHECK(k2q_block_sparse_num.size(2) == seq_len / BLOCK_N, "k2q_block_sparse_num idx 2 - must match seq_len / BLOCK_N");
+    TORCH_CHECK(v.size(2)     == kv_seq_len, "V  sequence length dimension - idx 2 - must match K sequence length");
+    TORCH_CHECK(l_vec.size(2) == q_seq_len, "L  sequence length dimension - idx 2 - must match Q sequence length");
+    TORCH_CHECK(o.size(2)     == q_seq_len, "O  sequence length dimension - idx 2 - must match Q sequence length");
+    TORCH_CHECK(og.size(2)    == q_seq_len, "OG sequence length dimension - idx 2 - must match Q sequence length");
+    TORCH_CHECK(k2q_block_sparse_index.size(2) == num_kv_blocks, "k2q_block_sparse_index idx 2 - must match num_kv_blocks (kv_block_size.size(0))");
+    TORCH_CHECK(k2q_block_sparse_num.size(2) == num_kv_blocks, "k2q_block_sparse_num idx 2 - must match num_kv_blocks (kv_block_size.size(0))");
 
     TORCH_CHECK(q.size(3)  == head_dim, "Q  head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(k.size(3)  == head_dim, "K  head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(v.size(3)  == head_dim, "V  head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(o.size(3)  == head_dim, "O  head dimension - idx 3 - must match for all non-vector inputs");
     TORCH_CHECK(og.size(3) == head_dim, "OG head dimension - idx 3 - must match for all non-vector inputs");
-    
-
 
     auto qo_heads = q.size(1);
     auto kv_heads = k.size(1);
@@ -988,20 +942,20 @@ block_sparse_attention_backward(torch::Tensor q,
 
     torch::Tensor qg = torch::zeros({static_cast<const uint>(batch), 
                                      static_cast<const uint>(qo_heads), 
-                                     static_cast<const uint>(seq_len), 
+                                     static_cast<const uint>(q_seq_len), 
                                      static_cast<const uint>(head_dim)},   l_vec.options());
     torch::Tensor kg = torch::zeros({static_cast<const uint>(batch), 
                                      static_cast<const uint>(kv_heads), 
-                                     static_cast<const uint>(seq_len), 
+                                     static_cast<const uint>(kv_seq_len), 
                                      static_cast<const uint>(head_dim)},   l_vec.options());
     torch::Tensor vg = torch::zeros({static_cast<const uint>(batch), 
                                      static_cast<const uint>(kv_heads), 
-                                     static_cast<const uint>(seq_len), 
+                                     static_cast<const uint>(kv_seq_len), 
                                      static_cast<const uint>(head_dim)},   l_vec.options());
     
     torch::Tensor d_vec = torch::empty({static_cast<const uint>(batch), 
                                         static_cast<const uint>(qo_heads), 
-                                        static_cast<const uint>(seq_len), 
+                                        static_cast<const uint>(q_seq_len), 
                                         static_cast<const uint>(1)},       l_vec.options());
 
     float*         qg_ptr = qg.data_ptr<float>();
@@ -1030,7 +984,7 @@ block_sparse_attention_backward(torch::Tensor q,
     //  cudaStreamSynchronize(stream);
 
     // TORCH_CHECK(seq_len % (4*kittens::TILE_DIM*4) == 0, "sequence length must be divisible by 256");
-    dim3 grid_bwd(seq_len/(PREP_NUM_WARPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
+    dim3 grid_bwd(q_seq_len/(PREP_NUM_WARPS*kittens::TILE_ROW_DIM<bf16>*4), qo_heads, batch);
 
     if (head_dim == 64)  {
         using og_tile = st_bf<4*16, 64>;
@@ -1043,9 +997,9 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_prep_globals = bwd_prep_globals<64>;
 
-        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
+        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
 
         bwd_prep_globals bwd_g{prep_og_arg, prep_o_arg, prep_d_arg};
 
@@ -1082,15 +1036,15 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_global_args = bwd_globals<64>;
 
-        bwd_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 64U};
-        bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-        bwd_d_global  bwd_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
+        bwd_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 64U};
+        bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 64U};
+        bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
+        bwd_d_global  bwd_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
 
         bwd_global_args bwd_global{bwd_q_arg, 
                         bwd_k_arg, 
@@ -1101,14 +1055,14 @@ block_sparse_attention_backward(torch::Tensor q,
                         bwd_vg_arg, 
                         bwd_l_arg, 
                         bwd_d_arg, 
-                        static_cast<int>(seq_len), 
+                        static_cast<int>(kv_seq_len), // N is not used in the kernel
                         static_cast<int>(hr),
                         static_cast<int>(max_q_blocks_per_kv),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_index.data_ptr()),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_num.data_ptr()),
-                        reinterpret_cast<int32_t*>(block_size.data_ptr())};
+                        reinterpret_cast<int32_t*>(kv_block_size.data_ptr())};
 
-        dim3 grid_bwd_2(seq_len/64, qo_heads, batch);
+        dim3 grid_bwd_2(kv_seq_len/BLOCK_N, qo_heads, batch);
         threads = 128;
 
         //cudadevicesynchronize();
@@ -1147,9 +1101,9 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_prep_globals = bwd_prep_globals<128>;
 
-        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
+        og_global prep_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        o_global  prep_o_arg {d_o,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        d_global  prep_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
 
         bwd_prep_globals bwd_g{prep_og_arg, prep_o_arg, prep_d_arg};
 
@@ -1186,15 +1140,15 @@ block_sparse_attention_backward(torch::Tensor q,
 
         using bwd_global_args = bwd_globals<128>;
 
-        bwd_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(seq_len), 128U};
-        bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
-        bwd_d_global  bwd_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(seq_len)};
+        bwd_q_global  bwd_q_arg {d_q,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        bwd_k_global  bwd_k_arg {d_k,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_v_global  bwd_v_arg {d_v,  static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_og_global bwd_og_arg{d_og, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        bwd_qg_global bwd_qg_arg{d_qg, static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), static_cast<unsigned int>(q_seq_len), 128U};
+        bwd_kg_global bwd_kg_arg{d_kg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_vg_global bwd_vg_arg{d_vg, static_cast<unsigned int>(batch), static_cast<unsigned int>(kv_heads), static_cast<unsigned int>(kv_seq_len), 128U};
+        bwd_l_global  bwd_l_arg {d_l,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
+        bwd_d_global  bwd_d_arg {d_d,  static_cast<unsigned int>(batch), static_cast<unsigned int>(qo_heads), 1U,  static_cast<unsigned int>(q_seq_len)};
 
         bwd_global_args bwd_global{bwd_q_arg, 
                         bwd_k_arg, 
@@ -1205,14 +1159,14 @@ block_sparse_attention_backward(torch::Tensor q,
                         bwd_vg_arg, 
                         bwd_l_arg, 
                         bwd_d_arg, 
-                        static_cast<int>(seq_len), 
+                        static_cast<int>(kv_seq_len), // N is not used in the kernel
                         static_cast<int>(hr),
                         static_cast<int>(max_q_blocks_per_kv),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_index.data_ptr()),
                         reinterpret_cast<int32_t*>(k2q_block_sparse_num.data_ptr()),
-                        reinterpret_cast<int32_t*>(block_size.data_ptr())};
+                        reinterpret_cast<int32_t*>(kv_block_size.data_ptr())};
 
-        dim3 grid_bwd_2(seq_len/64, qo_heads, batch);
+        dim3 grid_bwd_2(kv_seq_len/BLOCK_N, qo_heads, batch);
         threads = 128;
 
         //cudadevicesynchronize();
@@ -1233,4 +1187,4 @@ block_sparse_attention_backward(torch::Tensor q,
 
     return {qg, kg, vg};
     //cudadevicesynchronize();
-}
+}

fastvideo-kernel/csrc/turbodiffusion/norm/layernorm.cu

@@ -4,6 +4,7 @@
 #include <torch/all.h>
 #include <torch/python.h>
 #include <cutlass/cutlass.h>
+#include <cutlass/numeric_types.h>
 #include "common/common.hpp"
 #include "norm/layernorm.hpp"
 
@@ -14,10 +15,6 @@ auto layer_norm(
   std::optional<at::Tensor const> const B,
   std::optional<at::Tensor> Output
 ) {
-  using ElementIn = float;
-  using ElementOut = float;
-  using ElementWeight = float;
-
   int64_t const m = Input.size(0);
   int64_t const n = Input.size(1);
   torch::Device const input_device = Input.device();
@@ -26,31 +23,70 @@ auto layer_norm(
     Output.emplace(
       torch::empty(
         {m, n},
-        torch::TensorOptions().device(input_device).dtype(torch::kFloat32)
+        torch::TensorOptions().device(input_device).dtype(Input.scalar_type())
       )
     );
   }
 
+  TORCH_CHECK(Output.value().scalar_type() == Input.scalar_type(),
+              "Output dtype must match Input dtype. Got Output=",
+              Output.value().scalar_type(), ", Input=", Input.scalar_type());
+  if (W.has_value()) {
+    TORCH_CHECK(W.value().scalar_type() == Input.scalar_type(),
+                "W dtype must match Input dtype. Got W=",
+                W.value().scalar_type(), ", Input=", Input.scalar_type());
+  }
+  if (B.has_value()) {
+    TORCH_CHECK(B.value().scalar_type() == Input.scalar_type(),
+                "B dtype must match Input dtype. Got B=",
+                B.value().scalar_type(), ", Input=", Input.scalar_type());
+  }
 
   void *Iptr = Input.data_ptr();
   void *Wptr = W.has_value() ? W.value().data_ptr() : nullptr;
   void *Bptr = B.has_value() ? B.value().data_ptr() : nullptr;
   void *Optr = Output.value().data_ptr();
 
-  BOOL_SWITCH(B.has_value(), BIAS, [&]{
-    BOOL_SWITCH(W.has_value(), AFFINE, [&]{
-    CONFIG_SWITCH(n, [&]{
-    layernorm<
-    ElementIn, ElementOut, ElementWeight,
-    AFFINE, BIAS,
-    MAX_HIDDEN_SIZE, NUM_THR_PER_CTA> (
-      Iptr, Wptr, Bptr,
-      Optr, eps, m, n,
-      at::cuda::getCurrentCUDAStream().stream()
-        );
+  auto stream = at::cuda::getCurrentCUDAStream().stream();
+  if (Input.scalar_type() == at::kHalf) {
+    using ElementIn = cutlass::half_t;
+    using ElementOut = cutlass::half_t;
+    using ElementWeight = cutlass::half_t;
+    BOOL_SWITCH(B.has_value(), BIAS, [&]{
+      BOOL_SWITCH(W.has_value(), AFFINE, [&]{
+        CONFIG_SWITCH(n, [&]{
+          layernorm<ElementIn, ElementOut, ElementWeight, AFFINE, BIAS, MAX_HIDDEN_SIZE, NUM_THR_PER_CTA>(
+              Iptr, Wptr, Bptr, Optr, eps, m, n, stream);
+        });
       });
     });
-  });
+  } else if (Input.scalar_type() == at::kBFloat16) {
+    using ElementIn = cutlass::bfloat16_t;
+    using ElementOut = cutlass::bfloat16_t;
+    using ElementWeight = cutlass::bfloat16_t;
+    BOOL_SWITCH(B.has_value(), BIAS, [&]{
+      BOOL_SWITCH(W.has_value(), AFFINE, [&]{
+        CONFIG_SWITCH(n, [&]{
+          layernorm<ElementIn, ElementOut, ElementWeight, AFFINE, BIAS, MAX_HIDDEN_SIZE, NUM_THR_PER_CTA>(
+              Iptr, Wptr, Bptr, Optr, eps, m, n, stream);
+        });
+      });
+    });
+  } else if (Input.scalar_type() == at::kFloat) {
+    using ElementIn = float;
+    using ElementOut = float;
+    using ElementWeight = float;
+    BOOL_SWITCH(B.has_value(), BIAS, [&]{
+      BOOL_SWITCH(W.has_value(), AFFINE, [&]{
+        CONFIG_SWITCH(n, [&]{
+          layernorm<ElementIn, ElementOut, ElementWeight, AFFINE, BIAS, MAX_HIDDEN_SIZE, NUM_THR_PER_CTA>(
+              Iptr, Wptr, Bptr, Optr, eps, m, n, stream);
+        });
+      });
+    });
+  } else {
+    TORCH_CHECK(false, "Unsupported dtype for layer_norm_cuda: ", Input.scalar_type());
+  }
   
     
 

fastvideo-kernel/csrc/turbodiffusion/norm/layernorm.hpp

@@ -68,9 +68,22 @@ public:
     // mean reduction
     float u = _reduce_sum(x, shared_data) / params.n;
 
+    // IMPORTANT:
+    // Loader pads out-of-range lanes with 0. That is OK for the sum, but after
+    // subtracting mean, those padded lanes become -u and would incorrectly
+    // contribute to the variance. Mask them back to 0 before variance reduction.
+    // We launch exactly NumThrPerCta threads for a 1xMaxHiddenSize tile,
+    // so each thread is responsible for a contiguous chunk in N.
+    int thr_n_offset = tidx * NumElementPerThread;
     CUTLASS_PRAGMA_UNROLL
-    for (int i = 0; i < NumElementPerThread; ++i)
-      x[i] -= u;
+    for (int i = 0; i < NumElementPerThread; ++i) {
+      int idx = thr_n_offset + i;
+      if (idx < params.n) {
+        x[i] -= u;
+      } else {
+        x[i] = 0.f;
+      }
+    }
 
     __syncthreads();
     // var reduction

fastvideo-kernel/csrc/turbodiffusion/norm/rmsnorm.cu

@@ -4,6 +4,7 @@
 #include <torch/all.h>
 #include <torch/python.h>
 #include <cutlass/cutlass.h>
+#include <cutlass/numeric_types.h>
 #include <pybind11/pybind11.h>
 
 #include "common/common.hpp"
@@ -16,10 +17,6 @@ auto rms_norm(
   std::optional<at::Tensor>& Output
 ) {
   
-  using ElementIn = float;
-  using ElementOut = float;
-  using ElementWeight = float;
-  
   int64_t const m = Input.size(0);
   int64_t const n = Input.size(1);
   torch::Device const input_device = Input.device();
@@ -28,27 +25,51 @@ auto rms_norm(
     Output.emplace(
       torch::empty(
         {m, n},
-        torch::TensorOptions().device(input_device).dtype(torch::kFloat32)
+        torch::TensorOptions().device(input_device).dtype(Input.scalar_type())
       )
     );
   }
 
+  TORCH_CHECK(Output.value().scalar_type() == Input.scalar_type(),
+              "Output dtype must match Input dtype. Got Output=",
+              Output.value().scalar_type(), ", Input=", Input.scalar_type());
+  if (Weight.has_value()) {
+    TORCH_CHECK(Weight.value().scalar_type() == Input.scalar_type(),
+                "Weight dtype must match Input dtype. Got Weight=",
+                Weight.value().scalar_type(), ", Input=", Input.scalar_type());
+  }
+
   void *Iptr = Input.data_ptr();
   void *Wptr = Weight.has_value() ? Weight.value().data_ptr() : nullptr;
   void *Optr = Output.value().data_ptr();
 
-
-  CONFIG_SWITCH(n, [&]{
-    rmsnorm<
-    ElementIn, ElementOut, ElementWeight,
-    MAX_HIDDEN_SIZE, NUM_THR_PER_CTA
-    > (
-      Iptr, Wptr,
-      Optr, 
-      eps, m, n, 
-      at::cuda::getCurrentCUDAStream().stream()
-    );
-});
+  if (Input.scalar_type() == at::kHalf) {
+    using ElementIn = cutlass::half_t;
+    using ElementOut = cutlass::half_t;
+    using ElementWeight = cutlass::half_t;
+    CONFIG_SWITCH(n, [&]{
+      rmsnorm<ElementIn, ElementOut, ElementWeight, MAX_HIDDEN_SIZE, NUM_THR_PER_CTA>(
+          Iptr, Wptr, Optr, eps, m, n, at::cuda::getCurrentCUDAStream().stream());
+    });
+  } else if (Input.scalar_type() == at::kBFloat16) {
+    using ElementIn = cutlass::bfloat16_t;
+    using ElementOut = cutlass::bfloat16_t;
+    using ElementWeight = cutlass::bfloat16_t;
+    CONFIG_SWITCH(n, [&]{
+      rmsnorm<ElementIn, ElementOut, ElementWeight, MAX_HIDDEN_SIZE, NUM_THR_PER_CTA>(
+          Iptr, Wptr, Optr, eps, m, n, at::cuda::getCurrentCUDAStream().stream());
+    });
+  } else if (Input.scalar_type() == at::kFloat) {
+    using ElementIn = float;
+    using ElementOut = float;
+    using ElementWeight = float;
+    CONFIG_SWITCH(n, [&]{
+      rmsnorm<ElementIn, ElementOut, ElementWeight, MAX_HIDDEN_SIZE, NUM_THR_PER_CTA>(
+          Iptr, Wptr, Optr, eps, m, n, at::cuda::getCurrentCUDAStream().stream());
+    });
+  } else {
+    TORCH_CHECK(false, "Unsupported dtype for rms_norm_cuda: ", Input.scalar_type());
+  }
   
 
   return Output;

fastvideo-kernel/pyproject.toml

@@ -9,7 +9,7 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "fastvideo-kernel"
-version = "0.2.2"
+version = "0.2.4"
 description = "Unified CUDA kernels for FastVideo"
 readme = "README.md"
 requires-python = ">=3.10"

fastvideo-kernel/python/fastvideo_kernel/block_sparse_attn.py

@@ -0,0 +1,298 @@
+from __future__ import annotations
+
+import os
+from typing import Tuple
+
+import torch
+
+
+def _get_sm90_ops():
+    try:
+        from fastvideo_kernel._C import fastvideo_kernel_ops  # type: ignore
+    except Exception:
+        return None, None
+    return (
+        getattr(fastvideo_kernel_ops, "block_sparse_fwd", None),
+        getattr(fastvideo_kernel_ops, "block_sparse_bwd", None),
+    )
+
+
+def _is_sm90() -> bool:
+    if not torch.cuda.is_available():
+        return False
+    major, minor = torch.cuda.get_device_capability(0)
+    return major == 9 and minor == 0
+
+
+def _force_triton() -> bool:
+    # Force Triton even on SM90 and even if the compiled extension is available.
+    # Useful for CI / debugging / parity testing.
+    return os.environ.get("FASTVIDEO_KERNEL_VSA_FORCE_TRITON", "0") == "1"
+
+
+def _map_to_index_torch(block_map: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Pure-torch (no triton) conversion:
+      block_map: [B, H, Q, KV] bool  (or [H, Q, KV] which will be treated as B=1)
+      returns:
+        index: [B, H, Q, KV] int32 (packed KV indices, -1 padding)
+        num:   [B, H, Q] int32 (#kv blocks per q block)
+    """
+    if block_map.dim() == 3:
+        block_map = block_map.unsqueeze(0)
+    if block_map.dim() != 4:
+        raise ValueError(f"block_map must be [B,H,Q,KV] (or [H,Q,KV]), got shape={tuple(block_map.shape)}")
+    if block_map.dtype != torch.bool:
+        block_map = block_map.to(torch.bool)
+
+    B, H, Q, KV = block_map.shape
+    index = torch.full((B, H, Q, KV), -1, dtype=torch.int32, device=block_map.device)
+    num = torch.zeros((B, H, Q), dtype=torch.int32, device=block_map.device)
+
+    # Small sizes in practice (B=1, H<=16, Q/KV<=64), so a Python loop is fine.
+    for b in range(B):
+        for h in range(H):
+            for q in range(Q):
+                kv_idx = torch.nonzero(block_map[b, h, q], as_tuple=False).flatten().to(torch.int32)
+                n = int(kv_idx.numel())
+                if n:
+                    index[b, h, q, :n] = kv_idx
+                num[b, h, q] = n
+    return index, num
+
+
+@torch.library.custom_op(
+    "fastvideo_kernel::block_sparse_attn_triton",
+    mutates_args=(),
+    device_types="cuda",
+)
+def block_sparse_attn_triton(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    q = q.contiguous()
+    k = k.contiguous()
+    v = v.contiguous()
+    block_map = block_map.to(torch.bool)
+    q2k_idx, q2k_num = _map_to_index_torch(block_map)
+
+    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (  # local import
+        triton_block_sparse_attn_forward,
+    )
+
+    o, M = triton_block_sparse_attn_forward(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
+    return o, M
+
+
+@torch.library.register_fake("fastvideo_kernel::block_sparse_attn_triton")
+def _block_sparse_attn_triton_fake(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    o = torch.empty_like(q)
+    M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
+    return o, M
+
+
+@torch.library.custom_op(
+    "fastvideo_kernel::block_sparse_attn_backward_triton",
+    mutates_args=(),
+    device_types="cuda",
+)
+def block_sparse_attn_backward_triton(
+    grad_output: torch.Tensor,
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    o: torch.Tensor,
+    M: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    grad_output = grad_output.contiguous()
+    block_map = block_map.to(torch.bool)
+    q2k_idx, q2k_num = _map_to_index_torch(block_map)
+    k2q_idx, k2q_num = _map_to_index_torch(block_map.transpose(-1, -2).contiguous())
+
+    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (  # local import
+        triton_block_sparse_attn_backward,
+    )
+
+    dq, dk, dv = triton_block_sparse_attn_backward(
+        grad_output, q, k, v, o, M, q2k_idx, q2k_num, k2q_idx, k2q_num, variable_block_sizes
+    )
+    return dq, dk, dv
+
+
+@torch.library.register_fake("fastvideo_kernel::block_sparse_attn_backward_triton")
+def _block_sparse_attn_backward_triton_fake(
+    grad_output: torch.Tensor,
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    o: torch.Tensor,
+    M: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    dq = torch.empty_like(q)
+    dk = torch.empty_like(k)
+    dv = torch.empty_like(v)
+    return dq, dk, dv
+
+
+def _backward_triton(ctx, grad_o, grad_M):
+    q, k, v, o, M, block_map, variable_block_sizes = ctx.saved_tensors
+    dq, dk, dv = block_sparse_attn_backward_triton(grad_o, q, k, v, o, M, block_map, variable_block_sizes)
+    return dq, dk, dv, None, None
+
+
+def _setup_context_triton(ctx, inputs, output):
+    q, k, v, block_map, variable_block_sizes = inputs
+    o, M = output
+    ctx.save_for_backward(q, k, v, o, M, block_map, variable_block_sizes)
+
+
+block_sparse_attn_triton.register_autograd(_backward_triton, setup_context=_setup_context_triton)
+
+
+@torch.library.custom_op(
+    "fastvideo_kernel::block_sparse_attn_sm90",
+    mutates_args=(),
+    device_types="cuda",
+)
+def block_sparse_attn_sm90(
+    q_padded: torch.Tensor,
+    k_padded: torch.Tensor,
+    v_padded: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    block_sparse_fwd, _ = _get_sm90_ops()
+    if block_sparse_fwd is None:
+        raise ImportError("fastvideo_kernel_ops.block_sparse_fwd is not available")
+
+    q_padded = q_padded.contiguous()
+    k_padded = k_padded.contiguous()
+    v_padded = v_padded.contiguous()
+    block_map = block_map.to(torch.bool)
+    q2k_idx, q2k_num = _map_to_index_torch(block_map)
+
+    o_padded, lse_padded = block_sparse_fwd(
+        q_padded, k_padded, v_padded, q2k_idx, q2k_num, variable_block_sizes.int()
+    )
+    return o_padded, lse_padded
+
+
+@torch.library.register_fake("fastvideo_kernel::block_sparse_attn_sm90")
+def _block_sparse_attn_sm90_fake(
+    q_padded: torch.Tensor,
+    k_padded: torch.Tensor,
+    v_padded: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    o = torch.empty_like(q_padded)
+    lse = torch.empty((q_padded.shape[0], q_padded.shape[1], q_padded.shape[2], 1), device=q_padded.device, dtype=torch.float32)
+    return o, lse
+
+
+@torch.library.custom_op(
+    "fastvideo_kernel::block_sparse_attn_backward_sm90",
+    mutates_args=(),
+    device_types="cuda",
+)
+def block_sparse_attn_backward_sm90(
+    grad_output_padded: torch.Tensor,
+    q_padded: torch.Tensor,
+    k_padded: torch.Tensor,
+    v_padded: torch.Tensor,
+    o_padded: torch.Tensor,
+    lse_padded: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    _, block_sparse_bwd = _get_sm90_ops()
+    if block_sparse_bwd is None:
+        raise ImportError("fastvideo_kernel_ops.block_sparse_bwd is not available")
+
+    grad_output_padded = grad_output_padded.contiguous()
+    block_map = block_map.to(torch.bool)
+    k2q_idx, k2q_num = _map_to_index_torch(block_map.transpose(-1, -2).contiguous())
+
+    dq, dk, dv = block_sparse_bwd(
+        q_padded,
+        k_padded,
+        v_padded,
+        o_padded,
+        lse_padded,
+        grad_output_padded,
+        k2q_idx,
+        k2q_num,
+        variable_block_sizes.int(),
+    )
+    # C++ kernel returns fp32 grads; cast back to match PyTorch convention if needed
+    return dq.to(grad_output_padded.dtype), dk.to(grad_output_padded.dtype), dv.to(grad_output_padded.dtype)
+
+
+@torch.library.register_fake("fastvideo_kernel::block_sparse_attn_backward_sm90")
+def _block_sparse_attn_backward_sm90_fake(
+    grad_output_padded: torch.Tensor,
+    q_padded: torch.Tensor,
+    k_padded: torch.Tensor,
+    v_padded: torch.Tensor,
+    o_padded: torch.Tensor,
+    lse_padded: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    dq = torch.empty_like(q_padded)
+    dk = torch.empty_like(k_padded)
+    dv = torch.empty_like(v_padded)
+    return dq, dk, dv
+
+
+def _backward_sm90(ctx, grad_o, grad_lse):
+    q, k, v, o, lse, block_map, variable_block_sizes = ctx.saved_tensors
+    dq, dk, dv = block_sparse_attn_backward_sm90(
+        grad_o, q, k, v, o, lse, block_map, variable_block_sizes
+    )
+    return dq, dk, dv, None, None
+
+
+def _setup_context_sm90(ctx, inputs, output):
+    q, k, v, block_map, variable_block_sizes = inputs
+    o, lse = output
+    ctx.save_for_backward(q, k, v, o, lse, block_map, variable_block_sizes)
+
+
+block_sparse_attn_sm90.register_autograd(_backward_sm90, setup_context=_setup_context_sm90)
+
+
+def block_sparse_attn(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    block_map: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Unified block-sparse attention op with autograd support.
+    - On SM90 with compiled extension present: uses fastvideo_kernel_ops.block_sparse_fwd/bwd.
+    - Otherwise: uses Triton implementation (requires q/k/v to have same padded length today).
+    """
+    block_sparse_fwd, block_sparse_bwd = _get_sm90_ops()
+    if (not _force_triton()) and _is_sm90() and (block_sparse_fwd is not None) and (block_sparse_bwd is not None):
+        return block_sparse_attn_sm90(q, k, v, block_map, variable_block_sizes)
+    # Triton path: generally assumes q/k/v share the same padded length
+    if q.shape[2] != k.shape[2] or q.shape[2] != v.shape[2]:
+        raise RuntimeError("Triton fallback requires q/k/v to have the same padded length.")
+    return block_sparse_attn_triton(q, k, v, block_map, variable_block_sizes)
+
+

fastvideo-kernel/python/fastvideo_kernel/ops.py

@@ -1,5 +1,6 @@
 import math
 import torch
+from .block_sparse_attn import block_sparse_attn
 from .triton_kernels.block_sparse_attn_triton import triton_block_sparse_attn_forward
 from .triton_kernels.st_attn_triton import sliding_tile_attention_triton
 from .triton_kernels.index import map_to_index
@@ -45,14 +46,22 @@ def sliding_tile_attention(
     flag = shape_map[seq_shape]
 
     for head_idx, (t, h, w) in enumerate(window_size):
+        # Per-head slices are not contiguous in the batch dimension when batch>1
+        # (they keep the original head-stride). The TK kernel assumes contiguous
+        # [B, H, S, D] layout, so we materialize a contiguous [B,1,S,D] view.
+        q_h = q[:, head_idx:head_idx + 1].contiguous()
+        k_h = k[:, head_idx:head_idx + 1].contiguous()
+        v_h = v[:, head_idx:head_idx + 1].contiguous()
+        o_h = torch.empty_like(q_h)
         sta_fwd(
-            q[:, head_idx:head_idx + 1], k[:, head_idx:head_idx + 1],
-            v[:, head_idx:head_idx + 1], output[:, head_idx:head_idx + 1],
+            q_h, k_h,
+            v_h, o_h,
             t, h, w, text_length, False, has_text, flag
         )
+        output[:, head_idx:head_idx + 1] = o_h
 
     if has_text:
-        sta_fwd(q, k, v, output, 3, 3, 3, text_length, True, True, flag)
+        sta_fwd(q.contiguous(), k.contiguous(), v.contiguous(), output, 3, 3, 3, text_length, True, True, flag)
 
     return output[:, :, :seq_length]
 
@@ -62,6 +71,7 @@ def video_sparse_attn(
     k: torch.Tensor,
     v: torch.Tensor,
     variable_block_sizes: torch.Tensor,
+    q_variable_block_sizes: torch.Tensor,
     topk: int,
     block_size: int | tuple = 64,
     compress_attn_weight: torch.Tensor = None,
@@ -70,14 +80,42 @@ def video_sparse_attn(
         block_size = (block_size, block_size, block_size)
 
     block_elements = block_size[0] * block_size[1] * block_size[2]
-    batch, heads, seq_len, dim = q.shape
+    batch, heads, q_seq_len, dim = q.shape
+    kv_seq_len = k.shape[2]
+    if v.shape[2] != kv_seq_len:
+        raise ValueError(
+            f"Expected k and v to have the same sequence length, got "
+            f"k.shape[2]={kv_seq_len}, v.shape[2]={v.shape[2]}"
+        )
+    if k.shape[0] != batch or v.shape[0] != batch or k.shape[1] != heads or v.shape[1] != heads:
+        raise ValueError("Expected q/k/v to have the same batch and head dimensions.")
+
+    if q_seq_len % block_elements != 0 or kv_seq_len % block_elements != 0:
+        raise ValueError(
+            f"q_seq_len and kv_seq_len must be divisible by block_elements={block_elements}, "
+            f"got q_seq_len={q_seq_len}, kv_seq_len={kv_seq_len}"
+        )
+    q_num_blocks = q_seq_len // block_elements
+    kv_num_blocks = kv_seq_len // block_elements
+
+    if variable_block_sizes.numel() != kv_num_blocks:
+        raise ValueError(
+            f"variable_block_sizes must have length kv_num_blocks={kv_num_blocks}, "
+            f"got {variable_block_sizes.numel()}"
+        )
+
+    if q_variable_block_sizes.numel() != q_num_blocks:
+        raise ValueError(
+            f"q_variable_block_sizes must have length q_num_blocks={q_num_blocks}, "
+            f"got {q_variable_block_sizes.numel()}"
+        )
 
     # Compression branch
-    q_c = q.view(batch, heads, seq_len // block_elements, block_elements, dim)
-    k_c = k.view(batch, heads, seq_len // block_elements, block_elements, dim)
-    v_c = v.view(batch, heads, seq_len // block_elements, block_elements, dim)
+    q_c = q.view(batch, heads, q_num_blocks, block_elements, dim)
+    k_c = k.view(batch, heads, kv_num_blocks, block_elements, dim)
+    v_c = v.view(batch, heads, kv_num_blocks, block_elements, dim)
 
-    q_c = (q_c.float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(
+    q_c = (q_c.float().sum(dim=3) / q_variable_block_sizes.view(1, 1, -1, 1)).to(
         q.dtype)
     k_c = (k_c.float().sum(dim=3) / variable_block_sizes.view(1, 1, -1, 1)).to(
         k.dtype)
@@ -88,9 +126,9 @@ def video_sparse_attn(
     attn = torch.softmax(scores, dim=-1)
     out_c = torch.matmul(attn, v_c)
 
-    out_c = out_c.view(batch, heads, seq_len // block_elements, 1, dim)
+    out_c = out_c.view(batch, heads, q_num_blocks, 1, dim)
     out_c = out_c.repeat(1, 1, 1, block_elements,
-                         1).view(batch, heads, seq_len, dim)
+                         1).view(batch, heads, q_seq_len, dim)
 
     # Sparse branch
     topk_idx = torch.topk(scores, topk, dim=-1).indices
@@ -100,12 +138,17 @@ def video_sparse_attn(
     idx, num = map_to_index(mask)
     
     if block_sparse_fwd is not None:
-        out_s = block_sparse_fwd(
-            q, k, v, idx, num, variable_block_sizes.int()
-        )[0] # block_sparse_fwd returns vector<Tensor>
+        # Use autograd-enabled wrapper so backward works (and still uses SM90 kernel when available)
+        out_s = block_sparse_attn(q, k, v, mask, variable_block_sizes)[0]
     else:
-        out_s, _ = triton_block_sparse_attn_forward(q, k, v, idx, num,
-                                                    variable_block_sizes)
+        if q_seq_len != kv_seq_len:
+            raise RuntimeError(
+                "q/k have different lengths, but the compiled CUDA kernel (block_sparse_fwd) "
+                "is not available. The Triton fallback currently requires q and k/v to have "
+                "the same padded length."
+            )
+        # Triton-only forward (kept for environments without the wrapper deps)
+        out_s, _ = triton_block_sparse_attn_forward(q, k, v, idx, num, variable_block_sizes)
 
     if compress_attn_weight is not None:
         return out_c * compress_attn_weight + out_s

fastvideo-kernel/python/fastvideo_kernel/version.py

@@ -1 +1 @@
-__version__ = "0.2.2"
+__version__ = "0.2.4"

fastvideo-kernel/tests/test_vsa.py

@@ -42,37 +42,13 @@ def block_sparse_kernel_test(Q, K, V, block_sparse_mask, variable_block_sizes, q
     q_padded = vsa_pad(Q, q_non_pad_index, q_num_blocks, BLOCK_M)
     k_padded = vsa_pad(K, kv_non_pad_index, kv_num_blocks, BLOCK_M)
     v_padded = vsa_pad(V, kv_non_pad_index, kv_num_blocks, BLOCK_M)
-    # Use raw kernel or triton
-    try:
-        from fastvideo_kernel._C import fastvideo_kernel_ops
-        raw_kernel = getattr(fastvideo_kernel_ops, "block_sparse_fwd", None)
-    except ImportError:
-        raw_kernel = None
+    # Use autograd-enabled wrapper (internally dispatches to SM90 kernel or Triton)
+    from fastvideo_kernel.block_sparse_attn import block_sparse_attn
+    output_padded, _aux = block_sparse_attn(
+        q_padded, k_padded, v_padded, block_sparse_mask, variable_block_sizes
+    )
 
-    from fastvideo_kernel.triton_kernels.index import map_to_index
-
-    
-    # Convert mask to indices
-    # block_sparse_mask is [H, M, N] bool
-    # We need to map it to index.
-    # block_sparse_mask needs to be expanded/reshaped? 
-    # generate_block_sparse_mask_for_function returns [H, NumBlocksQ, NumBlocksKV]
-    
-    # Ops.py logic:
-    # mask = torch.zeros_like(scores, dtype=torch.bool).scatter_(-1, topk_idx, True)
-    # idx, num = map_to_index(mask)
-    
-    idx, num = map_to_index(block_sparse_mask.unsqueeze(0)) # Add batch dim [1, H, M, N]
-    
-    if raw_kernel:
-        out_s = raw_kernel(q_padded, k_padded, v_padded, idx, num, variable_block_sizes.int())
-        output = out_s[0]
-    else:
-        # Fallback to triton testing if C++ not available
-        from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import triton_block_sparse_attn_forward
-        output, _ = triton_block_sparse_attn_forward(q_padded, k_padded, v_padded, idx, num, variable_block_sizes)
-
-    output = output[:, :, q_non_pad_index, :]
+    output = output_padded[:, :, q_non_pad_index, :]
     output.backward(dO)
     return output, Q.grad, K.grad, V.grad
 
@@ -264,7 +240,6 @@ def generate_error_graphs_qkdiff(h, d, error_mode='all'):
 
     print("-" * 150)
 
-@pytest.mark.skip()
 def test_video_sparse_attention_backward():
     if not torch.cuda.is_available():
         return

fastvideo-kernel/tests/test_vsa_forward.py

@@ -3,6 +3,7 @@ import sys
 from typing import Tuple
 
 import torch
+import pytest
 
 from .utils import (
     generate_block_sparse_mask_for_function,
@@ -57,23 +58,14 @@ def block_sparse_forward_test(
     k_padded = ref.vsa_pad(K, kv_non_pad_index, kv_num_blocks, BLOCK_M)
     v_padded = ref.vsa_pad(V, kv_non_pad_index, kv_num_blocks, BLOCK_M)
 
-    # Use raw kernel or triton
+    # Use autograd-enabled wrapper (internally dispatches SM90 C++ vs Triton)
+    from fastvideo_kernel.block_sparse_attn import block_sparse_attn
     try:
-        from fastvideo_kernel._C import fastvideo_kernel_ops
-        raw_kernel = getattr(fastvideo_kernel_ops, "block_sparse_fwd", None)
-    except ImportError:
-        raw_kernel = None
-        
-    from fastvideo_kernel.triton_kernels.index import map_to_index
-    idx, num = map_to_index(block_sparse_mask)
-    
-    if raw_kernel:
-        out_padded = raw_kernel(q_padded, k_padded, v_padded, idx, num, variable_block_sizes.int())[0]
-    else:
-        from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import triton_block_sparse_attn_forward
-        out_padded, _ = triton_block_sparse_attn_forward(
-            q_padded, k_padded, v_padded, idx, num, variable_block_sizes
+        out_padded, _aux = block_sparse_attn(
+            q_padded, k_padded, v_padded, block_sparse_mask, variable_block_sizes
         )
+    except RuntimeError as e:
+        pytest.skip(str(e))
 
     # Remove padding on the query side
     out = out_padded[:, :, q_non_pad_index, :]
@@ -156,11 +148,6 @@ def run_forward_qk_diff(
 ) -> Tuple[float, float]:
     """
     Forward-only correctness test for the case S_q != S_kv.
-
-    NOTE:
-      - The Triton backend supports different Q/KV logical lengths via padding.
-      - The SM90 (H100) CUDA backend currently assumes the same number of blocks
-        for Q and KV, so we skip this test there.
     """
     assert torch.cuda.is_available(), "VSA kernels require CUDA"
 

fastvideo/attention/backends/video_sparse_attn.py

@@ -276,8 +276,9 @@ class VideoSparseAttentionImpl(AttentionImpl):
             query,
             key,
             value,
-            variable_block_sizes=attn_metadata.variable_block_sizes,
-            topk=cur_topk,
+            attn_metadata.variable_block_sizes,
+            attn_metadata.variable_block_sizes,
+            cur_topk,
             block_size=VSA_TILE_SIZE,
             compress_attn_weight=gate_compress).transpose(1, 2)
 

fastvideo/configs/models/encoders/__init__.py

@@ -7,10 +7,11 @@ from fastvideo.configs.models.encoders.clip import (
 from fastvideo.configs.models.encoders.llama import LlamaConfig
 from fastvideo.configs.models.encoders.t5 import T5Config, T5LargeConfig
 from fastvideo.configs.models.encoders.qwen2_5 import Qwen2_5_VLConfig
+from fastvideo.configs.models.encoders.reason1 import Reason1ArchConfig, Reason1Config
 
 __all__ = [
     "EncoderConfig", "TextEncoderConfig", "ImageEncoderConfig",
     "BaseEncoderOutput", "CLIPTextConfig", "CLIPVisionConfig",
     "WAN2_1ControlCLIPVisionConfig", "LlamaConfig", "T5Config", "T5LargeConfig",
-    "Qwen2_5_VLConfig"
+    "Qwen2_5_VLConfig", "Reason1ArchConfig", "Reason1Config"
 ]

fastvideo/configs/models/encoders/reason1.py

@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Config for Reason1 (Qwen2.5-VL) text encoder."""
+
+from dataclasses import dataclass, field
+from typing import Any
+
+from fastvideo.configs.models.encoders.base import TextEncoderArchConfig, TextEncoderConfig
+
+
+@dataclass
+class Reason1ArchConfig(TextEncoderArchConfig):
+    """Architecture settings (defaults match Qwen2.5-VL-7B-Instruct)."""
+
+    architectures: list[str] = field(
+        default_factory=lambda: ["Qwen2_5_VLForConditionalGeneration"])
+    model_type: str = "qwen2_5_vl"
+
+    vocab_size: int = 152064
+    hidden_size: int = 3584
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 28
+    num_key_value_heads: int = 4
+    intermediate_size: int = 18944
+
+    text_len: int = 512
+    hidden_state_skip_layer: int = 0
+    bos_token_id: int = 151643
+    pad_token_id: int = 151643
+    eos_token_id: int = 151645
+
+    image_token_id: int = 151655
+    video_token_id: int = 151656
+    vision_token_id: int = 151654
+    vision_start_token_id: int = 151652
+    vision_end_token_id: int = 151653
+
+    vision_config: dict[str, Any] | None = None
+
+    rope_theta: float = 1000000.0
+    rope_scaling: dict[str, Any] | None = field(default_factory=lambda: {
+        "type": "mrope",
+        "mrope_section": [16, 24, 24]
+    })
+    max_position_embeddings: int = 128000
+    max_window_layers: int = 28
+
+    embedding_concat_strategy: str = "mean_pooling"
+    n_layers_per_group: int = 5
+    num_embedding_padding_tokens: int = 512
+
+    attention_dropout: float = 0.0
+    hidden_act: str = "silu"
+    initializer_range: float = 0.02
+    rms_norm_eps: float = 1e-6
+
+    use_sliding_window: bool = False
+    sliding_window: int = 32768
+
+    tie_word_embeddings: bool = False
+    use_cache: bool = False
+    output_hidden_states: bool = True
+
+    torch_dtype: str = "bfloat16"
+    _attn_implementation: str = "flash_attention_2"
+
+
+@dataclass
+class Reason1Config(TextEncoderConfig):
+    """Reason1 text encoder config."""
+
+    arch_config: Reason1ArchConfig = field(default_factory=Reason1ArchConfig)
+    tokenizer_type: str = "Qwen/Qwen2.5-VL-7B-Instruct"

fastvideo/configs/models/vaes/__init__.py

@@ -1,4 +1,5 @@
 from fastvideo.configs.models.vaes.cosmosvae import CosmosVAEConfig
+from fastvideo.configs.models.vaes.cosmos2_5vae import Cosmos25VAEConfig
 from fastvideo.configs.models.vaes.hunyuanvae import HunyuanVAEConfig
 from fastvideo.configs.models.vaes.hunyuan15vae import Hunyuan15VAEConfig
 from fastvideo.configs.models.vaes.stepvideovae import StepVideoVAEConfig
@@ -9,5 +10,6 @@ __all__ = [
     "WanVAEConfig",
     "StepVideoVAEConfig",
     "CosmosVAEConfig",
+    "Cosmos25VAEConfig",
     "Hunyuan15VAEConfig",
 ]

fastvideo/configs/models/vaes/cosmos2_5vae.py

@@ -0,0 +1,223 @@
+"""Cosmos 2.5 (Wan2.1-style) VAE config and checkpoint-key mapping."""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models.vaes.base import VAEArchConfig, VAEConfig
+
+
+@dataclass
+class Cosmos25VAEArchConfig(VAEArchConfig):
+    _name_or_path: str = ""
+    base_dim: int = 96
+    decoder_base_dim: int | None = None
+    z_dim: int = 16
+    dim_mult: tuple[int, ...] = (1, 2, 4, 4)
+    num_res_blocks: int = 2
+    attn_scales: tuple[float, ...] = ()
+    temperal_downsample: tuple[bool, ...] = (False, True, True)
+    dropout: float = 0.0
+    is_residual: bool = False
+    in_channels: int = 3
+    out_channels: int = 3
+    patch_size: int | None = None
+    scale_factor_temporal: int = 4
+    scale_factor_spatial: int = 8
+    clip_output: bool = True
+
+    latents_mean: tuple[float, ...] = (
+        -0.7571,
+        -0.7089,
+        -0.9113,
+        0.1075,
+        -0.1745,
+        0.9653,
+        -0.1517,
+        1.5508,
+        0.4134,
+        -0.0715,
+        0.5517,
+        -0.3632,
+        -0.1922,
+        -0.9497,
+        0.2503,
+        -0.2921,
+    )
+    latents_std: tuple[float, ...] = (
+        2.8184,
+        1.4541,
+        2.3275,
+        2.6558,
+        1.2196,
+        1.7708,
+        2.6052,
+        2.0743,
+        3.2687,
+        2.1526,
+        2.8652,
+        1.5579,
+        1.6382,
+        1.1253,
+        2.8251,
+        1.9160,
+    )
+
+    # Simple 1:1 renames. More complex decoder remapping is handled by
+    # `map_official_key()`.
+    param_names_mapping: dict[str, str] = field(
+        default_factory=lambda: {
+            r"^conv1\.(.*)$": r"quant_conv.\1",
+            r"^conv2\.(.*)$": r"post_quant_conv.\1",
+            r"^encoder\.conv1\.(.*)$": r"encoder.conv_in.\1",
+            r"^decoder\.conv1\.(.*)$": r"decoder.conv_in.\1",
+            r"^encoder\.head\.0\.gamma$": r"encoder.norm_out.gamma",
+            r"^encoder\.head\.2\.(.*)$": r"encoder.conv_out.\1",
+            r"^decoder\.head\.0\.gamma$": r"decoder.norm_out.gamma",
+            r"^decoder\.head\.2\.(.*)$": r"decoder.conv_out.\1",
+        })
+
+    @staticmethod
+    def map_official_key(key: str) -> str | None:
+        """Map a single official checkpoint key into FastVideo key space."""
+
+        def map_residual_subkey(prefix: str, sub: str) -> str | None:
+            if re.match(r"^residual\.0\.gamma$", sub):
+                return f"{prefix}.norm1.gamma"
+            m = re.match(r"^residual\.2\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.conv1.{m.group(1)}"
+            if re.match(r"^residual\.3\.gamma$", sub):
+                return f"{prefix}.norm2.gamma"
+            m = re.match(r"^residual\.6\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.conv2.{m.group(1)}"
+            m = re.match(r"^shortcut\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.conv_shortcut.{m.group(1)}"
+            return None
+
+        def map_attn_subkey(prefix: str, sub: str) -> str | None:
+            if re.match(r"^norm\.gamma$", sub):
+                return f"{prefix}.norm.gamma"
+            m = re.match(r"^to_qkv\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.to_qkv.{m.group(1)}"
+            m = re.match(r"^proj\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.proj.{m.group(1)}"
+            return None
+
+        def map_resample_subkey(prefix: str, sub: str) -> str | None:
+            m = re.match(r"^resample\.1\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.resample.1.{m.group(1)}"
+            m = re.match(r"^time_conv\.(weight|bias)$", sub)
+            if m:
+                return f"{prefix}.time_conv.{m.group(1)}"
+            return None
+
+        m = re.match(r"^conv1\.(weight|bias)$", key)
+        if m:
+            return f"quant_conv.{m.group(1)}"
+        m = re.match(r"^conv2\.(weight|bias)$", key)
+        if m:
+            return f"post_quant_conv.{m.group(1)}"
+        m = re.match(r"^(encoder|decoder)\.conv1\.(weight|bias)$", key)
+        if m:
+            return f"{m.group(1)}.conv_in.{m.group(2)}"
+        m = re.match(r"^(encoder|decoder)\.head\.0\.gamma$", key)
+        if m:
+            return f"{m.group(1)}.norm_out.gamma"
+        m = re.match(r"^(encoder|decoder)\.head\.2\.(weight|bias)$", key)
+        if m:
+            return f"{m.group(1)}.conv_out.{m.group(2)}"
+
+        m = re.match(r"^(encoder|decoder)\.middle\.0\.(.*)$", key)
+        if m:
+            return map_residual_subkey(f"{m.group(1)}.mid_block.resnets.0",
+                                       m.group(2))
+        m = re.match(r"^(encoder|decoder)\.middle\.1\.(.*)$", key)
+        if m:
+            return map_attn_subkey(f"{m.group(1)}.mid_block.attentions.0",
+                                   m.group(2))
+        m = re.match(r"^(encoder|decoder)\.middle\.2\.(.*)$", key)
+        if m:
+            return map_residual_subkey(f"{m.group(1)}.mid_block.resnets.1",
+                                       m.group(2))
+
+        m = re.match(r"^encoder\.downsamples\.(\d+)\.(.*)$", key)
+        if m:
+            idx = int(m.group(1))
+            sub = m.group(2)
+            if sub.startswith("residual.") or sub.startswith("shortcut."):
+                return map_residual_subkey(f"encoder.down_blocks.{idx}", sub)
+            if sub.startswith("resample.") or sub.startswith("time_conv."):
+                return map_resample_subkey(f"encoder.down_blocks.{idx}", sub)
+            return None
+
+        m = re.match(r"^decoder\.upsamples\.(\d+)\.(.*)$", key)
+        if m:
+            uidx = int(m.group(1))
+            sub = m.group(2)
+
+            if uidx in (0, 1, 2):
+                block_i, res_i = 0, uidx
+            elif uidx == 3:
+                block_i, res_i = 0, None
+            elif uidx in (4, 5, 6):
+                block_i, res_i = 1, uidx - 4
+            elif uidx == 7:
+                block_i, res_i = 1, None
+            elif uidx in (8, 9, 10):
+                block_i, res_i = 2, uidx - 8
+            elif uidx == 11:
+                block_i, res_i = 2, None
+            elif uidx in (12, 13, 14):
+                block_i, res_i = 3, uidx - 12
+            else:
+                return None
+
+            if res_i is None:
+                return map_resample_subkey(
+                    f"decoder.up_blocks.{block_i}.upsamplers.0",
+                    sub,
+                )
+
+            return map_residual_subkey(
+                f"decoder.up_blocks.{block_i}.resnets.{res_i}",
+                sub,
+            )
+
+        return None
+
+    temporal_compression_ratio: int = 4
+    spatial_compression_ratio: int = 8
+
+    def __post_init__(self):
+        self.scaling_factor: torch.Tensor = 1.0 / torch.tensor(
+            self.latents_std).view(1, self.z_dim, 1, 1, 1)
+        self.shift_factor: torch.Tensor = torch.tensor(self.latents_mean).view(
+            1, self.z_dim, 1, 1, 1)
+        self.temporal_compression_ratio = self.scale_factor_temporal
+        self.spatial_compression_ratio = self.scale_factor_spatial
+
+
+@dataclass
+class Cosmos25VAEConfig(VAEConfig):
+    """Cosmos2.5 VAE config."""
+
+    arch_config: Cosmos25VAEArchConfig = field(
+        default_factory=Cosmos25VAEArchConfig)
+
+    use_feature_cache: bool = True
+    use_tiling: bool = False
+    use_temporal_tiling: bool = False
+    use_parallel_tiling: bool = False
+
+    def __post_init__(self):
+        self.blend_num_frames = (self.tile_sample_min_num_frames -
+                                 self.tile_sample_stride_num_frames) * 2

fastvideo/configs/pipelines/__init__.py

@@ -1,6 +1,7 @@
 from fastvideo.configs.pipelines.base import (PipelineConfig,
                                               SlidingTileAttnConfig)
 from fastvideo.configs.pipelines.cosmos import CosmosConfig
+from fastvideo.configs.pipelines.cosmos2_5 import Cosmos25Config
 from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.registry import (
@@ -15,5 +16,5 @@ __all__ = [
     "Hunyuan15T2V480PConfig", "Hunyuan15T2V720PConfig", "SlidingTileAttnConfig",
     "WanT2V480PConfig", "WanI2V480PConfig", "WanT2V720PConfig",
     "WanI2V720PConfig", "StepVideoT2VConfig", "SelfForcingWanT2V480PConfig",
-    "CosmosConfig", "get_pipeline_config_cls_from_name"
+    "CosmosConfig", "Cosmos25Config", "get_pipeline_config_cls_from_name"
 ]

fastvideo/configs/pipelines/cosmos2_5.py

@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: Apache-2.0
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits import Cosmos25VideoConfig
+from fastvideo.configs.models.dits.cosmos2_5 import Cosmos25ArchConfig
+from fastvideo.configs.models.encoders import BaseEncoderOutput
+from fastvideo.configs.models.encoders.reason1 import Reason1Config, Reason1ArchConfig
+from fastvideo.configs.models.vaes import Cosmos25VAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig, STA_Mode
+
+
+def _identity_preprocess_text(prompt: str) -> str:
+    return prompt
+
+
+def reason1_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
+    hidden_states = getattr(outputs, "hidden_states", None)
+    if hidden_states is None:
+        raise ValueError("Reason1 postprocess requires outputs.hidden_states")
+
+    hs = list(hidden_states)[1:]
+    normed = []
+    for h in hs:
+        h = h.float()
+        h = (h - h.mean(dim=-1, keepdim=True)) / (h.std(dim=-1, keepdim=True) +
+                                                  1e-8)
+        normed.append(h)
+    return torch.cat(normed, dim=-1).to(hidden_states[0].dtype)
+
+
+@dataclass
+class Cosmos25Config(PipelineConfig):
+    """Configuration for Cosmos 2.5 (Predict2.5) video generation pipeline."""
+
+    dit_config: DiTConfig = field(default_factory=lambda: Cosmos25VideoConfig(
+        arch_config=Cosmos25ArchConfig(
+            num_attention_heads=16,
+            attention_head_dim=128,
+            in_channels=16,
+            out_channels=16,
+            num_layers=28,
+            patch_size=[1, 2, 2],
+            max_size=[128, 240, 240],
+            rope_scale=[1.0, 3.0, 3.0],
+            text_embed_dim=1024,
+            mlp_ratio=4.0,
+            adaln_lora_dim=256,
+            use_adaln_lora=True,
+            concat_padding_mask=True,
+            extra_pos_embed_type=None,
+            use_crossattn_projection=True,
+            rope_enable_fps_modulation=False,
+            qk_norm="rms_norm",
+        )))
+
+    vae_config: VAEConfig = field(default_factory=Cosmos25VAEConfig)
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(
+        default_factory=lambda: (Reason1Config(arch_config=Reason1ArchConfig(
+            embedding_concat_strategy="full_concat")), ))
+
+    preprocess_text_funcs: tuple[Callable[[str], str], ...] = field(
+        default_factory=lambda: (_identity_preprocess_text, ))
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
+                                  ...] = field(default_factory=lambda:
+                                               (reason1_postprocess_text, ))
+
+    dit_precision: str = "bf16"
+    vae_precision: str = "bf16"
+    text_encoder_precisions: tuple[str, ...] = field(
+        default_factory=lambda: ("bf16", ))
+
+    embedded_cfg_scale: float = 0.0
+    flow_shift: float = 5.0
+
+    vae_tiling: bool = False
+    vae_sp: bool = False
+
+    STA_mode: STA_Mode = STA_Mode.NONE
+    skip_time_steps: int = 0
+
+    def __post_init__(self):
+        self.vae_config.load_encoder = True
+        self.vae_config.load_decoder = True
+        self._vae_latent_dim = 16

fastvideo/configs/pipelines/registry.py

@@ -6,6 +6,7 @@ from collections.abc import Callable
 
 from fastvideo.configs.pipelines.base import PipelineConfig
 from fastvideo.configs.pipelines.cosmos import CosmosConfig
+from fastvideo.configs.pipelines.cosmos2_5 import Cosmos25Config
 from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
@@ -55,6 +56,7 @@ PIPE_NAME_TO_CONFIG: dict[str, type[PipelineConfig]] = {
     "Wan-AI/Wan2.2-T2V-A14B-Diffusers": Wan2_2_T2V_A14B_Config,
     "Wan-AI/Wan2.2-I2V-A14B-Diffusers": Wan2_2_I2V_A14B_Config,
     "nvidia/Cosmos-Predict2-2B-Video2World": CosmosConfig,
+    "KyleShao/Cosmos-Predict2.5-2B-Diffusers": Cosmos25Config,
     "FastVideo/Matrix-Game-2.0-Base-Diffusers": MatrixGameI2V480PConfig,
     "FastVideo/Matrix-Game-2.0-GTA-Diffusers": MatrixGameI2V480PConfig,
     "FastVideo/Matrix-Game-2.0-TempleRun-Diffusers": MatrixGameI2V480PConfig,
@@ -94,7 +96,10 @@ PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
     "stepvideo":
     lambda id: "stepvideo" in id.lower(),
     "cosmos":
-    lambda id: "cosmos" in id.lower(),
+    lambda id: "cosmos" in id.lower() and ("2.5" not in id.lower(
+    ) and "2_5" not in id.lower() and "25" not in id.lower()),
+    "cosmos25":
+    lambda id: "cosmos25" in id.lower(),
     "turbodiffusion":
     lambda id: "turbodiffusion" in id.lower() or "turbowan" in id.lower(),
     # Add other pipeline architecture detectors
@@ -105,6 +110,7 @@ PIPELINE_FALLBACK_CONFIG: dict[str, type[PipelineConfig]] = {
     "longcatimagetovideo": LongCatT2V480PConfig,
     "longcatvideocontinuation": LongCatT2V480PConfig,
     "longcat": LongCatT2V480PConfig,
+    "cosmos25": Cosmos25Config,
     "hunyuan":
     HunyuanConfig,  # Base Hunyuan config as fallback for any Hunyuan variant
     "matrixgame": MatrixGameI2V480PConfig,

fastvideo/configs/sample/cosmos2_5.py

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass
+
+from fastvideo.configs.sample.base import SamplingParam
+
+
+@dataclass
+class Cosmos_Predict2_5_2B_Diffusers_SamplingParam(SamplingParam):
+    """Defaults for Cosmos 2.5 (Predict2.5) text-to-video diffusers-format model."""
+
+    height: int = 480
+    width: int = 832
+    num_frames: int = 121
+    fps: int = 24
+
+    guidance_scale: float = 7.0
+    # Official Cosmos2.5 sampling uses empty string as unconditional.
+    negative_prompt: str = ""
+    num_inference_steps: int = 35

fastvideo/configs/sample/registry.py

@@ -9,6 +9,7 @@ from fastvideo.configs.sample.hunyuan15 import Hunyuan15_480P_SamplingParam, Hun
 from fastvideo.configs.sample.stepvideo import StepVideoT2VSamplingParam
 
 from fastvideo.configs.sample.cosmos import Cosmos_Predict2_2B_Video2World_SamplingParam
+from fastvideo.configs.sample.cosmos2_5 import Cosmos_Predict2_5_2B_Diffusers_SamplingParam
 
 # isort: off
 from fastvideo.configs.sample.wan import (
@@ -96,6 +97,10 @@ SAMPLING_PARAM_REGISTRY: dict[str, Any] = {
     "nvidia/Cosmos-Predict2-2B-Video2World":
     Cosmos_Predict2_2B_Video2World_SamplingParam,
 
+    # Cosmos2.5
+    "KyleShao/Cosmos-Predict2.5-2B-Diffusers":
+    Cosmos_Predict2_5_2B_Diffusers_SamplingParam,
+
     # MatrixGame2.0 models
     "FastVideo/Matrix-Game-2.0-Base-Diffusers":
     MatrixGame2_SamplingParam,
@@ -135,6 +140,10 @@ SAMPLING_PARAM_DETECTOR: dict[str, Callable[[str], bool]] = {
     lambda id: "matrixgame" in id.lower() or "matrix-game" in id.lower(),
     "turbodiffusion":
     lambda id: "turbodiffusion" in id.lower() or "turbowan" in id.lower(),
+    "cosmos25":
+    lambda id: "cosmos2_5" in id.lower(),
+    "cosmos":
+    lambda id: "cosmos" in id.lower() and "2_5" not in id.lower(),
     # Add other pipeline architecture detectors
 }
 
@@ -153,6 +162,8 @@ SAMPLING_FALLBACK_PARAM: dict[str, Any] = {
     "matrixgame": MatrixGame2_SamplingParam,
     "turbodiffusion":
     TurboDiffusionT2V_1_3B_SamplingParam,  # Default to T2V for fallback
+    "cosmos25": Cosmos_Predict2_5_2B_Diffusers_SamplingParam,
+    "cosmos": Cosmos_Predict2_2B_Video2World_SamplingParam,
     # Other fallbacks by architecture
 }
 
@@ -176,9 +187,6 @@ def get_sampling_param_cls_for_name(pipeline_name_or_path: str) -> Any | None:
 
     if os.path.exists(pipeline_name_or_path):
         config = verify_model_config_and_directory(pipeline_name_or_path)
-        logger.warning(
-            "FastVideo may not correctly identify the optimal sampling param for this model, as the local directory may have been renamed."
-        )
     else:
         config = maybe_download_model_index(pipeline_name_or_path)
 

fastvideo/dataset/__init__.py

@@ -8,6 +8,7 @@ from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset,
 from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255,
                                          TemporalRandomCrop)
 from fastvideo.dataset.validation_dataset import ValidationDataset
+from fastvideo.dataset.rl_prompt_dataset import build_rl_prompt_dataloader
 
 
 def getdataset(args) -> VideoCaptionMergedDataset:
@@ -47,5 +48,6 @@ def gettextdataset(args) -> TextDataset:
 
 __all__ = [
     "build_parquet_map_style_dataloader", "ValidationDataset",
-    "VideoCaptionMergedDataset", "TextDataset"
+    "VideoCaptionMergedDataset", "TextDataset",
+    "build_rl_prompt_dataloader"
 ]

fastvideo/dataset/rl_prompt_dataset.py

@@ -0,0 +1,174 @@
+# SPDX-License-Identifier: Apache-2.0
+import torch
+from torch.utils.data import Dataset, DataLoader, Sampler
+import json
+import os
+
+
+class TextPromptDataset(Dataset):
+    """Dataset for loading text prompts from a simple text file (one prompt per line)."""
+    
+    def __init__(self, dataset, split='train'):
+        self.file_path = os.path.join(dataset, f'{split}.txt')
+        with open(self.file_path, 'r') as f:
+            self.prompts = [line.strip() for line in f.readlines()]
+        
+    def __len__(self):
+        return len(self.prompts)
+    
+    def __getitem__(self, idx):
+        return {"prompt": self.prompts[idx], "metadata": {}}
+
+    @staticmethod
+    def collate_fn(examples):
+        prompts = [example["prompt"] for example in examples]
+        metadatas = [example["metadata"] for example in examples]
+        return prompts, metadatas
+
+
+class GenevalPromptDataset(Dataset):
+    """Dataset for loading prompts with metadata from JSONL files (e.g., GenEval format)."""
+    
+    def __init__(self, dataset, split='train'):
+        self.file_path = os.path.join(dataset, f'{split}_metadata.jsonl')
+        with open(self.file_path, 'r', encoding='utf-8') as f:
+            self.metadatas = [json.loads(line) for line in f]
+            self.prompts = [item['prompt'] for item in self.metadatas]
+        
+    def __len__(self):
+        return len(self.prompts)
+    
+    def __getitem__(self, idx):
+        return {"prompt": self.prompts[idx], "metadata": self.metadatas[idx]}
+
+    @staticmethod
+    def collate_fn(examples):
+        prompts = [example["prompt"] for example in examples]
+        metadatas = [example["metadata"] for example in examples]
+        return prompts, metadatas
+
+
+class KRepeatSampler(Sampler):
+    """Sampler that repeats each sample k times, ensuring synchronized random selection. For single-node training, set num_replicas=1 and rank=0."""
+    
+    def __init__(self, dataset, batch_size, k, num_replicas, rank, seed=0):
+        self.dataset = dataset
+        self.batch_size = batch_size        # Batch size per GPU/card
+        self.k = k                          # Number of repetitions per sample
+        self.num_replicas = num_replicas    # Total number of GPUs/cards
+        self.rank = rank                    # Current GPU/card rank
+        self.seed = seed                    # Random seed for synchronization
+        
+        # Calculate the number of unique samples needed for each iteration
+        self.total_samples = self.num_replicas * self.batch_size
+        assert self.total_samples % self.k == 0, f"k can not div n*b, k{k}-num_replicas{num_replicas}-batch_size{batch_size}"
+        self.m = self.total_samples // self.k  # different number of samples
+        self.step = 0
+
+    def __iter__(self):
+        while True:
+            # Generate a deterministic random sequence to ensure all cards are synchronized
+            g = torch.Generator()
+            g.manual_seed(self.seed + self.step)
+            
+            # Randomly select m unique samples
+            indices = torch.randperm(len(self.dataset), generator=g)[:self.m].tolist()
+            
+            # Repeat each sample k times to generate a total of n*b samples
+            repeated_indices = [idx for idx in indices for _ in range(self.k)]
+            
+            # Shuffle the order to ensure even distribution
+            shuffled_indices = torch.randperm(len(repeated_indices), generator=g).tolist()
+            shuffled_samples = [repeated_indices[i] for i in shuffled_indices]
+            
+            # Split samples among all cards
+            per_card_samples = []
+            for i in range(self.num_replicas):
+                start = i * self.batch_size
+                end = start + self.batch_size
+                per_card_samples.append(shuffled_samples[start:end])
+            
+            # Return the sample indices for the current card
+            yield per_card_samples[self.rank]
+        
+    def __len__(self):
+        return len(self.dataset) // self.batch_size
+    
+    def set_step(self, step):
+        """Used to synchronize the random state for different epochs."""
+        self.step = step
+
+
+def build_rl_prompt_dataloader(
+    dataset_path: str,
+    dataset_type: str = "text",
+    split: str = "train",
+    train_batch_size: int = 8,
+    test_batch_size: int = 8,
+    k: int = 1,
+    seed: int = 42,
+    train_num_workers: int = 1,
+    test_num_workers: int = 8,
+    num_replicas: int = 1,
+    rank: int = 0,
+) -> tuple[DataLoader, DataLoader]:
+    """
+    Factory function to create train and test dataloaders for RL prompt datasets.
+    
+    Args:
+        dataset_path: Path to dataset directory
+        dataset_type: "text" for TextPromptDataset or "geneval" for GenevalPromptDataset
+        split: Dataset split ("train" or "test")
+        train_batch_size: Batch size per GPU for training
+        test_batch_size: Batch size for testing
+        k: Number of times to repeat each sample (num_image_per_prompt)
+        seed: Random seed for sampler synchronization
+        train_num_workers: Number of workers for training dataloader
+        test_num_workers: Number of workers for test dataloader
+        num_replicas: Number of replicas (default 1 for single-node)
+        rank: Rank of current process (default 0 for single-node)
+    
+    Returns:
+        Tuple of (train_dataloader, test_dataloader)
+    """
+    # Create datasets based on type
+    if dataset_type == "text":
+        train_dataset = TextPromptDataset(dataset_path, 'train')
+        test_dataset = TextPromptDataset(dataset_path, 'test')
+        collate_fn = TextPromptDataset.collate_fn
+    elif dataset_type == "geneval":
+        train_dataset = GenevalPromptDataset(dataset_path, 'train')
+        test_dataset = GenevalPromptDataset(dataset_path, 'test')
+        collate_fn = GenevalPromptDataset.collate_fn
+    else:
+        raise ValueError(f"Unknown dataset_type: {dataset_type}. Must be 'text' or 'geneval'")
+    
+    # Create infinite-loop training sampler
+    train_sampler = KRepeatSampler(
+        dataset=train_dataset,
+        batch_size=train_batch_size,
+        k=k,
+        num_replicas=num_replicas,
+        rank=rank,
+        seed=seed
+    )
+    
+    # Create training dataloader with batch_sampler (infinite loop)
+    train_dataloader = DataLoader(
+        train_dataset,
+        batch_sampler=train_sampler,
+        num_workers=train_num_workers,
+        collate_fn=collate_fn,
+    )
+    
+    # Create standard test dataloader
+    test_dataloader = DataLoader(
+        test_dataset,
+        batch_size=test_batch_size,
+        collate_fn=collate_fn,
+        shuffle=False,
+        num_workers=test_num_workers,
+    )
+    
+    return train_dataloader, test_dataloader, train_dataset, test_dataset
+

fastvideo/fastvideo_args.py

@@ -133,7 +133,7 @@ class FastVideoArgs:
     # CPU offload parameters
     dit_cpu_offload: bool = True
     use_fsdp_inference: bool = False
-    dit_layerwise_offload: bool = False
+    dit_layerwise_offload: bool = True
     text_encoder_cpu_offload: bool = True
     image_encoder_cpu_offload: bool = True
     vae_cpu_offload: bool = True
@@ -740,6 +740,271 @@ def get_current_fastvideo_args() -> FastVideoArgs:
     return _current_fastvideo_args
 
 
+@dataclasses.dataclass
+class RLArgs:
+    """
+    Reinforcement Learning (RL) specific arguments
+    """
+    # ============================================================================
+    # SHARED RL CONFIGURATION
+    rl_mode: bool = False # Enable RL training mode
+    rl_algorithm: str = "grpo"  # RL algorithm to use: "grpo", "ppo", "dpo"
+
+    # Trajectory collection
+    num_rollouts: int = 4  # Number of rollouts to collect per training step
+    rollout_steps: str = "20,30"  # Random intermediate steps for sampling (comma-separated)
+    noise_injection_min: int = 10  # Minimum timestep for noise injection
+    noise_injection_max: int = 40  # Maximum timestep for noise injection
+    use_sde_sampling: bool = True  # Use SDE sampling (Flow-GRPO-Fast)
+    num_denoising_steps: int = 2  # Number of denoising steps per trajectory (1-2 for fast)
+
+    # Advantage estimation
+    gamma: float = 0.99  # Discount factor for returns
+    lambda_param: float = 0.95  # GAE lambda parameter
+    use_gae: bool = True  # Use Generalized Advantage Estimation
+    normalize_advantages: bool = True  # Normalize advantages before policy update
+
+    # Reward models
+    reward_models: dict[str, float] = field(default_factory=lambda: {"dummy": 1.0})  # reward models (names, weight)
+    value_model_path: str = ""  # Path to value model (can be empty to train from scratch)
+    value_model_share_backbone: bool = False  # Share transformer backbone between policy and value
+
+    # Training schedule
+    warmup_steps: int = 1000  # Collect SFT-style data before starting RL
+    collect_on_policy: bool = True  # Collect fresh rollouts each step (on-policy)
+    timestep_fraction: float = 0.99  # Fraction of timesteps to train on
+    num_inner_epochs: int = 1  # Number of inner epochs per outer epoch
+
+    # KL regularization
+    kl_beta: float = 0.004  # KL loss coefficient (GRPO uses KL loss, DPO uses larger beta)
+    kl_reward: float = 0.0  # KL reward coefficient (alternative to KL loss, typically 0)
+
+    # SFT integration
+    sft_weight: float = 0.0  # SFT loss weight for supervised learning in RL training
+    sft_batch_size: int = 3  # Batch size for SFT data
+
+    # CFG
+    guidance_scale = 1.0 # use guidance_scale > 1.0 to enable CFG
+
+    # Statistics tracking
+    global_std: bool = False  # Use global std across all samples vs per-group std
+    per_prompt_stat_tracking: bool = True  # Track statistics per prompt
+
+    # Training options
+    use_diffusion_loss: bool = True  # Use diffusion loss in training
+
+    # ============================================================================
+    # GRPO-SPECIFIC CONFIGURATION
+
+    # Policy optimization
+    grpo_policy_clip_range: float = 0.001  # PPO-style clipping range for policy ratio
+    grpo_value_clip_range: float = 0.2  # Value function clipping range
+    grpo_num_policy_epochs: int = 1  # Number of policy update epochs (GRPO typically uses 1)
+    grpo_num_value_epochs: int = 1  # Number of value function update epochs
+    grpo_target_kl: float = 0.01  # Target KL divergence for early stopping
+    grpo_entropy_coef: float = 0.0  # Entropy coefficient for exploration
+    grpo_value_loss_coef: float = 0.5  # Value loss coefficient
+
+    # GRPO-Guard safety mechanisms
+    grpo_use_grpo_guard: bool = True  # Enable GRPO-Guard safety mechanisms
+    grpo_ratio_norm_correction: bool = True  # RatioNorm: correct importance ratio bias
+    grpo_gradient_reweighting: bool = True  # Reweight gradients across denoising steps
+    grpo_max_importance_ratio: float = 10.0  # Clip importance ratios above this value
+
+    # ============================================================================
+    # DPO-SPECIFIC CONFIGURATION
+
+    dpo_beta: float = 100.0  # DPO regularization parameter (typically much larger than GRPO beta)
+    dpo_ref_update_step: int = 10000000  # Reference model update frequency for OnlineDPO
+    dpo_label_smoothing: float = 0.0  # Label smoothing for DPO loss
+
+    @staticmethod
+    def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
+        """Add RL-specific CLI arguments to the parser."""
+        # RL (Reinforcement Learning) arguments
+        parser.add_argument("--rl-mode",
+                            action=StoreBoolean,
+                            help="Enable RL training mode")
+        parser.add_argument("--rl-algorithm",
+                            type=str,
+                            default=RLArgs.rl_algorithm,
+                            choices=["grpo", "ppo", "dpo"],
+                            help="RL algorithm to use (grpo, ppo, dpo)")
+
+        # Trajectory collection (Flow-GRPO-Fast)
+        parser.add_argument("--rl-num-rollouts",
+                            type=int,
+                            default=RLArgs.num_rollouts,
+                            help="Number of rollouts to collect per training step")
+        parser.add_argument("--rl-rollout-steps",
+                            type=str,
+                            default=RLArgs.rollout_steps,
+                            help="Random intermediate steps for sampling (comma-separated)")
+        parser.add_argument("--rl-noise-injection-min",
+                            type=int,
+                            default=RLArgs.noise_injection_min,
+                            help="Minimum timestep for noise injection")
+        parser.add_argument("--rl-noise-injection-max",
+                            type=int,
+                            default=RLArgs.noise_injection_max,
+                            help="Maximum timestep for noise injection")
+        parser.add_argument("--rl-use-sde-sampling",
+                            action=StoreBoolean,
+                            help="Use SDE sampling (Flow-GRPO-Fast)")
+        parser.add_argument("--rl-num-denoising-steps",
+                            type=int,
+                            default=RLArgs.num_denoising_steps,
+                            help="Number of denoising steps per trajectory (1-2 for fast)")
+
+        # Advantage estimation
+        parser.add_argument("--rl-gamma",
+                            type=float,
+                            default=RLArgs.gamma,
+                            help="Discount factor for returns")
+        parser.add_argument("--rl-lambda",
+                            type=float,
+                            default=RLArgs.lambda_param,
+                            help="GAE lambda parameter")
+        parser.add_argument("--rl-use-gae",
+                            action=StoreBoolean,
+                            help="Use Generalized Advantage Estimation")
+        parser.add_argument("--rl-normalize-advantages",
+                            action=StoreBoolean,
+                            help="Normalize advantages before policy update")
+
+        # Policy optimization (GRPO/PPO)
+        parser.add_argument("--rl-policy-clip-range",
+                            type=float,
+                            default=RLArgs.grpo_policy_clip_range,
+                            dest="grpo_policy_clip_range",  # Map to RLArgs field name
+                            help="PPO-style clipping range for policy ratio")
+        parser.add_argument("--rl-value-clip-range",
+                            type=float,
+                            default=RLArgs.grpo_value_clip_range,
+                            help="Value function clipping range")
+        parser.add_argument("--rl-num-policy-epochs",
+                            type=int,
+                            default=RLArgs.grpo_num_policy_epochs,
+                            help="Number of policy update epochs (GRPO typically uses 1)")
+        parser.add_argument("--rl-num-value-epochs",
+                            type=int,
+                            default=RLArgs.grpo_num_value_epochs,
+                            help="Number of value function update epochs")
+        parser.add_argument("--rl-target-kl",
+                            type=float,
+                            default=RLArgs.grpo_target_kl,
+                            help="Target KL divergence for early stopping")
+        parser.add_argument("--rl-entropy-coef",
+                            type=float,
+                            default=RLArgs.grpo_entropy_coef,
+                            help="Entropy coefficient for exploration")
+        parser.add_argument("--rl-value-loss-coef",
+                            type=float,
+                            default=RLArgs.grpo_value_loss_coef,
+                            help="Value loss coefficient")
+
+        # GRPO-Guard (safety mechanisms)
+        parser.add_argument("--rl-use-grpo-guard",
+                            action=StoreBoolean,
+                            help="Enable GRPO-Guard safety mechanisms")
+        parser.add_argument("--rl-ratio-norm-correction",
+                            action=StoreBoolean,
+                            help="RatioNorm: correct importance ratio bias")
+        parser.add_argument("--rl-gradient-reweighting",
+                            action=StoreBoolean,
+                            help="Reweight gradients across denoising steps")
+        parser.add_argument("--rl-max-importance-ratio",
+                            type=float,
+                            default=RLArgs.grpo_max_importance_ratio,
+                            help="Clip importance ratios above this value")
+
+        # Reward models
+        parser.add_argument("--reward-models",
+                            type=str,
+                            default='{"dummy": 1.0}',
+                            help="Reward models as JSON dict (e.g., '{\"video_ocr\": 1.0, \"pickscore\": 0.5}')")
+        parser.add_argument("--value-model-path",
+                            type=str,
+                            default=RLArgs.value_model_path,
+                            help="Path to value model (can be empty to train from scratch)")
+        parser.add_argument("--value-model-share-backbone",
+                            action=StoreBoolean,
+                            help="Share transformer backbone between policy and value")
+
+        # Training schedule
+        parser.add_argument("--rl-warmup-steps",
+                            type=int,
+                            default=RLArgs.warmup_steps,
+                            help="Collect SFT-style data before starting RL")
+        parser.add_argument("--rl-collect-on-policy",
+                            action=StoreBoolean,
+                            help="Collect fresh rollouts each step (on-policy)")
+        parser.add_argument("--rl-timestep-fraction",
+                            type=float,
+                            default=RLArgs.timestep_fraction,
+                            help="Fraction of timesteps to train on")
+        parser.add_argument("--rl-num-inner-epochs",
+                            type=int,
+                            default=RLArgs.num_inner_epochs,
+                            help="Number of inner epochs per outer epoch")
+
+        # KL regularization
+        parser.add_argument("--rl-kl-beta",
+                            type=float,
+                            default=RLArgs.kl_beta,
+                            dest="kl_beta",  # Map CLI arg to RLArgs field name
+                            help="KL loss coefficient (GRPO uses KL loss, DPO uses larger beta)")
+        parser.add_argument("--rl-kl-reward",
+                            type=float,
+                            default=RLArgs.kl_reward,
+                            help="KL reward coefficient (alternative to KL loss, typically 0)")
+
+        # SFT integration
+        parser.add_argument("--rl-sft-weight",
+                            type=float,
+                            default=RLArgs.sft_weight,
+                            help="SFT loss weight for supervised learning in RL training")
+        parser.add_argument("--rl-sft-batch-size",
+                            type=int,
+                            default=RLArgs.sft_batch_size,
+                            help="Batch size for SFT data")
+
+        # CFG settings
+        parser.add_argument("--guidance-scale",
+                            type=float,
+                            default=1.0,
+                            help="Guidance scale for CFG")
+
+        # Statistics tracking
+        parser.add_argument("--rl-global-std",
+                            action=StoreBoolean,
+                            help="Use global std across all samples vs per-group std")
+        parser.add_argument("--rl-per-prompt-stat-tracking",
+                            action=StoreBoolean,
+                            help="Track statistics per prompt")
+
+        # Training options
+        parser.add_argument("--rl-use-diffusion-loss",
+                            action=StoreBoolean,
+                            help="Use diffusion loss in training")
+
+        # DPO-specific
+        parser.add_argument("--dpo-beta",
+                            type=float,
+                            default=RLArgs.dpo_beta,
+                            help="DPO regularization parameter (typically much larger than GRPO beta)")
+        parser.add_argument("--dpo-ref-update-step",
+                            type=int,
+                            default=RLArgs.dpo_ref_update_step,
+                            help="Reference model update frequency for OnlineDPO")
+        parser.add_argument("--dpo-label-smoothing",
+                            type=float,
+                            default=RLArgs.dpo_label_smoothing,
+                            help="Label smoothing for DPO loss")
+
+        return parser
+
+
 @dataclasses.dataclass
 class TrainingArgs(FastVideoArgs):
     """
@@ -752,6 +1017,11 @@ class TrainingArgs(FastVideoArgs):
     num_height: int = 0
     num_width: int = 0
     num_frames: int = 0
+    
+    # RL dataset configuration (for RL prompt datasets)
+    rl_dataset_path: str = ""  # Path to RL prompt dataset directory (defaults to data_path if not set)
+    rl_dataset_type: str = "text"  # "text" or "geneval"
+    rl_num_image_per_prompt: int = 4  # k parameter for KRepeatSampler (num_image_per_prompt)
 
     train_batch_size: int = 0
     num_latent_t: int = 0
@@ -862,6 +1132,9 @@ class TrainingArgs(FastVideoArgs):
     last_step_only: bool = False  # Only use the last timestep for training
     context_noise: int = 0  # Context noise level for cache updates
 
+    # Nested RL configuration
+    rl_args: RLArgs = dataclasses.field(default_factory=RLArgs)
+
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace) -> "TrainingArgs":
         provided_args = clean_cli_args(args)
@@ -886,6 +1159,25 @@ class TrainingArgs(FastVideoArgs):
                 kwargs[attr] = WorkloadType.from_string(
                     workload_type_value) if isinstance(
                         workload_type_value, str) else workload_type_value
+            elif attr == 'rl_args':
+                # Construct nested RLArgs from CLI arguments
+                rl_kwargs = {}
+                for rl_field in dataclasses.fields(RLArgs):
+                    rl_attr = rl_field.name
+                    if hasattr(args, rl_attr):
+                        value = getattr(args, rl_attr)
+                        # Special handling for reward_models: parse JSON string to dict
+                        if rl_attr == 'reward_models' and isinstance(value, str):
+                            rl_kwargs[rl_attr] = json.loads(value) if value else {}
+                        else:
+                            rl_kwargs[rl_attr] = value
+                    else:
+                        # Use default value from RLArgs
+                        if rl_field.default_factory is not dataclasses.MISSING:
+                            rl_kwargs[rl_attr] = rl_field.default_factory()
+                        elif rl_field.default is not dataclasses.MISSING:
+                            rl_kwargs[rl_attr] = rl_field.default
+                kwargs[attr] = RLArgs(**rl_kwargs)
             # Use getattr with default value from the dataclass for potentially missing attributes
             else:
                 # Get the field to check its default value
@@ -915,11 +1207,26 @@ class TrainingArgs(FastVideoArgs):
         parser.add_argument("--data-path",
                             type=str,
                             required=True,
-                            help="Path to parquet files")
+                            help="Path to parquet files (or RL prompt dataset directory for RL training)")
         parser.add_argument("--dataloader-num-workers",
                             type=int,
                             required=True,
                             help="Number of workers for dataloader")
+        
+        # RL dataset arguments (optional, defaults to data_path)
+        parser.add_argument("--rl-dataset-path",
+                            type=str,
+                            default="",
+                            help="Path to RL prompt dataset directory (defaults to --data-path if not set)")
+        parser.add_argument("--rl-dataset-type",
+                            type=str,
+                            default="text",
+                            choices=["text", "geneval"],
+                            help="RL dataset type: 'text' for TextPromptDataset or 'geneval' for GenevalPromptDataset")
+        parser.add_argument("--rl-num-image-per-prompt",
+                            type=int,
+                            default=4,
+                            help="Number of times to repeat each prompt (k parameter for KRepeatSampler)")
         parser.add_argument("--num-height",
                             type=int,
                             required=True,
@@ -1284,6 +1591,9 @@ class TrainingArgs(FastVideoArgs):
                             default=TrainingArgs.context_noise,
                             help="Context noise level for cache updates")
 
+        # RL (Reinforcement Learning) arguments
+        RLArgs.add_cli_args(parser)
+
         return parser
 
 

fastvideo/hooks/hooks.py

@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import functools
+from typing import Any
+from torch import nn
+
+
+class ForwardHook:
+    """
+    Base class for forward hooks.
+    Hooks are used in the way:
+        modified_args, modified_kwargs = hook.pre_forward(module, *args, **kwargs)
+        output = module.forward(*modified_args, **modified_kwargs)
+        modified_output = hook.post_forward(module, output)
+    """
+
+    @classmethod
+    def name(cls) -> str:
+        raise NotImplementedError
+
+    def on_attach(self, module: nn.Module):  # noqa: B027
+        """Called once when the hook is attached to the module."""
+        pass
+
+    def on_detach(self, module: nn.Module):  # noqa: B027
+        """
+        Called once when the hook is detached from the module.
+        Note: this function is not guaranteed to be called if the module is
+              deleted before the hook is detached.
+        """
+        pass
+
+    def pre_forward(self, module: nn.Module, *args,
+                    **kwargs) -> tuple[tuple[Any, ...], dict[str, Any]]:
+        """Called before the module's forward method is executed."""
+        return args, kwargs
+
+    def post_forward(self, module: nn.Module, output: Any) -> Any:
+        """Called after the module's forward method is executed."""
+        return output
+
+
+class ModuleHookManager:
+    module_hook_attribute = "_hook_manager"
+
+    def __init__(self, module: nn.Module):
+        self.module = module
+        self.forward_hooks: dict[str, ForwardHook] = {}
+        self.original_forward = module.forward
+
+    @classmethod
+    def get_from(cls, module: nn.Module) -> "ModuleHookManager | None":
+        if hasattr(module, cls.module_hook_attribute):
+            return getattr(module, cls.module_hook_attribute)
+        return None
+
+    @classmethod
+    def get_from_or_default(cls, module: nn.Module) -> "ModuleHookManager":
+        if not hasattr(module, cls.module_hook_attribute):
+            setattr(module, cls.module_hook_attribute, cls(module))
+
+            def forward_hook_wrapper(mod: nn.Module, *args, **kwargs):
+                manager: ModuleHookManager = getattr(mod,
+                                                     cls.module_hook_attribute)
+                for hook in manager.forward_hooks.values():
+                    args, kwargs = hook.pre_forward(mod, *args, **kwargs)
+                output = manager.original_forward(*args, **kwargs)
+                for hook in reversed(manager.forward_hooks.values()):
+                    output = hook.post_forward(mod, output)
+                return output
+
+            module.forward = functools.partial(forward_hook_wrapper, module)
+
+        return getattr(module, cls.module_hook_attribute)
+
+    @staticmethod
+    def remove_from_manager(module: nn.Module) -> None:
+        if hasattr(module, ModuleHookManager.module_hook_attribute):
+            manager: ModuleHookManager = getattr(
+                module, ModuleHookManager.module_hook_attribute)
+            module.forward = manager.original_forward
+            delattr(module, ModuleHookManager.module_hook_attribute)
+
+    def _check_manager_attached(self) -> None:
+        if not hasattr(self.module, self.module_hook_attribute):
+            raise ValueError("ModuleHookManager is not attached to the module.")
+        if getattr(self.module, self.module_hook_attribute) is not self:
+            raise ValueError(
+                "ModuleHookManager attached to the module is different.")
+
+    def append_forward_hook(self, hook: ForwardHook):
+        self._check_manager_attached()
+        if hook.name() in self.forward_hooks:
+            raise ValueError(
+                f"Hook with name {hook.name()} is already registered.")
+        # after python 3.7, dicts maintain insertion order
+        self.forward_hooks[hook.name()] = hook
+        hook.on_attach(self.module)
+
+    def replace_forward_hook(self,
+                             hook_name: str,
+                             new_hook: ForwardHook,
+                             run_on_attach: bool = True):
+        self._check_manager_attached()
+        if hook_name not in self.forward_hooks:
+            raise ValueError(f"No hook with name {hook_name} found.")
+        old_hook = self.forward_hooks[hook_name]
+        if run_on_attach:
+            old_hook.on_detach(self.module)
+        self.forward_hooks[hook_name] = new_hook
+        new_hook.on_attach(self.module)
+
+    def remove_forward_hook(self, hook_name: str, run_detach: bool = True):
+        self._check_manager_attached()
+        if hook_name not in self.forward_hooks:
+            raise ValueError(f"No hook with name {hook_name} found.")
+        if run_detach:
+            self.forward_hooks[hook_name].on_detach(self.module)
+        del self.forward_hooks[hook_name]
+
+    def get_forward_hook(self, hook_name: str) -> ForwardHook | None:
+        return self.forward_hooks.get(hook_name, None)

fastvideo/hooks/layerwise_offload.py

@@ -0,0 +1,164 @@
+from contextlib import contextmanager
+from typing import Any
+import torch
+from torch import nn
+from fastvideo.hooks.hooks import ForwardHook, ModuleHookManager
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def _tensor_placeholder(tensor: torch.Tensor,
+                        device: torch.device) -> torch.Tensor:
+    """Create a rank-preserving empty placeholder on the specified device."""
+    shape = (0, ) if tensor.ndim <= 0 else (0, ) * tensor.ndim
+    return torch.empty(shape, device=device, dtype=tensor.dtype)
+
+
+class LayerwiseOffloadState:
+
+    def __init__(
+        self,
+        async_copy_stream: torch.cuda.Stream,
+        device: torch.device,
+        next_state: "LayerwiseOffloadState | None" = None,
+    ) -> None:
+        self.async_copy_stream = async_copy_stream
+        self.next_state = next_state
+        self.gpu_named_parameters: dict[str, torch.Tensor] = {}
+        self.cpu_named_parameters: dict[str, torch.Tensor] = {}
+        self.module_ref: nn.Module = None  # type: ignore
+        self.device: torch.device = device
+
+    def _will_offload(self, name: str) -> bool:
+        return True
+
+    @torch.compiler.disable
+    def on_init(self, module: nn.Module):
+        self.module_ref = module
+        for name, param in self.module_ref.named_parameters():
+            if self._will_offload(name):
+                self.cpu_named_parameters[name] = (
+                    param.data.detach().to("cpu").pin_memory())
+                param.data = _tensor_placeholder(param.data, self.device)
+
+    @torch.compiler.disable
+    def wait_and_replace_params(self):
+        torch.cuda.current_stream().wait_stream(self.async_copy_stream)
+        # now gpu_named_parameters are ready
+        for name, param in self.module_ref.named_parameters():
+            if not self._will_offload(name):
+                continue
+            if name not in self.gpu_named_parameters:
+                # first load with blocking load
+                self.gpu_named_parameters[name] = self.cpu_named_parameters[
+                    name].to(self.device)
+            param.data = self.gpu_named_parameters[name]
+
+    @torch.compiler.disable
+    def prefetch_params(self):
+        compute_stream = torch.cuda.current_stream()
+        with torch.cuda.stream(self.async_copy_stream):
+            for name, param in self.module_ref.named_parameters():
+                if not self._will_offload(name):
+                    continue
+                assert name not in self.gpu_named_parameters
+                gpu_param = self.cpu_named_parameters[name].to(
+                    self.device, non_blocking=True)
+                gpu_param.record_stream(
+                    compute_stream
+                )  # ensure tensor will not be freed until forward is completed
+                self.gpu_named_parameters[name] = gpu_param
+
+    @torch.compiler.disable
+    def release_gpu_params(self):
+        for name, param in self.module_ref.named_parameters():
+            if self._will_offload(name):
+                param.data = _tensor_placeholder(param.data, self.device)
+                del self.gpu_named_parameters[name]
+        assert len(self.gpu_named_parameters) == 0
+
+
+class LayerwiseOffloadHook(ForwardHook):
+    """A hook that enables layerwise CPU offloading during forward pass."""
+
+    def __init__(self, state: LayerwiseOffloadState) -> None:
+        self.state = state
+
+    def on_attach(self, module: nn.Module):
+        self.state.on_init(module)  # pyright: ignore
+
+    def on_detach(self, module: nn.Module):
+        named_parameters = dict(module.named_parameters())
+        for name, cpu_tensor in self.state.cpu_named_parameters.items():
+            if name not in self.state.gpu_named_parameters:
+                if name in named_parameters:
+                    named_parameters[name].data = cpu_tensor.to(
+                        device=self.state.device)
+                else:
+                    logger.warning(
+                        "Parameter {} not found in module during detachment.",
+                        name,
+                    )
+
+    @classmethod
+    def name(cls) -> str:
+        return "LayerwiseOffloadHook"
+
+    def pre_forward(self, module: nn.Module, *args, **kwargs):
+        self.state.wait_and_replace_params()  # pyright: ignore
+        if self.state.next_state is not None:
+            self.state.next_state.prefetch_params()  # pyright: ignore
+        return args, kwargs
+
+    def post_forward(self, module: torch.nn.Module, output: Any):
+        self.state.release_gpu_params()  # pyright: ignore
+        return output
+
+    @contextmanager
+    def mutate_params_scope(self):
+        try:
+            # load params to GPU and keep them there
+            self.state.wait_and_replace_params()  # pyright: ignore
+            yield
+        finally:
+            # instead of releasing, we should overwrite the original params since they have been modified
+            self.state.cpu_named_parameters.clear()
+            self.state.gpu_named_parameters.clear()
+            self.state.on_init(self.state.module_ref)  # pyright: ignore
+
+
+def enable_layerwise_offload(model: nn.Module, is_replace: bool = False):
+    if torch.cuda.is_available():
+        device = torch.device("cuda", torch.cuda.current_device())
+    else:
+        logger.warning(
+            "CUDA is not available. Layerwise offloading is disabled.")
+        return
+    state_list = []
+    async_stream = torch.cuda.Stream()
+    for name, submodule in model.named_children():
+        if isinstance(submodule, nn.ModuleList):
+            for idx, module_entry in enumerate(submodule):
+                state = LayerwiseOffloadState(async_copy_stream=async_stream,
+                                              device=device)
+                state_list.append(state)
+                hook_mgr = ModuleHookManager.get_from_or_default(module_entry)
+                hook = LayerwiseOffloadHook(state)
+                if is_replace:
+                    existing_hook = hook_mgr.forward_hooks.get(hook.name())
+                    if existing_hook is not None:
+                        hook_mgr.replace_forward_hook(hook.name(), hook)
+                    else:
+                        raise AssertionError(
+                            f"Expect hook exists in {name} for replacement.")
+                else:
+                    hook_mgr.append_forward_hook(hook)
+            break
+    if len(state_list) == 0:
+        raise ValueError(
+            "No nn.ModuleList found in the model for layerwise offloading.")
+
+    # circular linking of states
+    for i in range(len(state_list)):
+        state_list[i].next_state = state_list[(i + 1) % len(state_list)]

fastvideo/models/dits/wanvideo.py

@@ -734,18 +734,8 @@ class WanTransformer3DModel(CachableDiT):
                         block, hidden_states, encoder_hidden_states,
                         timestep_proj, freqs_cis, attention_mask)
             else:
-                offload_mgr = getattr(self, "_layerwise_offload_manager", None)
-                use_offload = offload_mgr is not None and getattr(offload_mgr, "enabled", False)
-
-                for i, block in enumerate(self.blocks):
-                    scope = offload_mgr.layer_scope(
-                        prefetch_layer_idx=i + 1 if i + 1 < len(self.blocks) else None,
-                        release_layer_idx=i,
-                        non_blocking=True,
-                    ) if use_offload else nullcontext()
-                    
-                    with scope:
-                        hidden_states = block(hidden_states, encoder_hidden_states,
+                for block in self.blocks:
+                    hidden_states = block(hidden_states, encoder_hidden_states,
                                               timestep_proj, freqs_cis, attention_mask)
             # if teacache is enabled, we need to cache the original hidden states
 

fastvideo/models/encoders/reason1.py

@@ -0,0 +1,353 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Reason1 (Qwen2.5-VL) text encoder."""
+
+import os
+from dataclasses import dataclass
+from collections.abc import Iterable
+
+import torch
+from transformers import AutoProcessor
+
+from fastvideo.configs.models.encoders import BaseEncoderOutput, Reason1Config
+from fastvideo.logger import init_logger
+from fastvideo.models.encoders.base import TextEncoder
+from fastvideo.models.loader.weight_utils import default_weight_loader
+from fastvideo.platforms import AttentionBackendEnum
+
+from fastvideo.models.encoders.qwen2_5_vl_custom import (
+    Qwen2_5_VLForConditionalGenerationSimple,
+    Qwen2_5_VLConfig,
+    get_rope_index,
+)
+
+logger = init_logger(__name__)
+
+
+@dataclass(frozen=True)
+class _WeightsSource:
+    """Mimic `TextEncoderLoader.Source` (avoid import cycles)."""
+
+    model_or_path: str
+    prefix: str = ""
+    fall_back_to_pt: bool = True
+    allow_patterns_overrides: list[str] | None = None
+
+
+
+
+class Reason1TextEncoder(TextEncoder):
+    """Reason1 (Qwen2.5-VL) text encoder."""
+
+    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+    )
+    
+    def __init__(self, config: Reason1Config, prefix: str = "", checkpoint_path: str | None = None):
+        super().__init__(config)
+        
+        self.prefix = prefix
+        self.quant_config = None  # For future quantization support
+        
+        self.embedding_concat_strategy = config.arch_config.embedding_concat_strategy
+        self.n_layers_per_group = config.arch_config.n_layers_per_group
+        self.num_embedding_padding_tokens = config.arch_config.num_embedding_padding_tokens
+        
+        config_path = checkpoint_path if checkpoint_path else config.tokenizer_type
+        
+        logger.info("Initializing Reason1TextEncoder (Qwen2.5-VL) from %s", config_path)
+        try:
+            from transformers import AutoConfig as HFAutoConfig
+            hf_config = HFAutoConfig.from_pretrained(
+                config_path,
+                trust_remote_code=True,
+            )
+        except Exception as e:
+            logger.warning("Failed to load HF config from %s (%s). Using default Qwen2.5-VL-7B config.",
+                           config_path, e)
+            hf_config = Qwen2_5_VLConfig(
+                    hidden_size=3584,
+                    intermediate_size=18944,
+                    max_window_layers=28,
+                    num_attention_heads=28,
+                    num_hidden_layers=28,
+                    num_key_value_heads=4,
+                    tie_word_embeddings=False,
+                    vocab_size=152064,
+            )
+        
+        hf_config.output_hidden_states = True
+        
+        if hasattr(config.arch_config, '_attn_implementation') and config.arch_config._attn_implementation:
+            hf_config._attn_implementation = config.arch_config._attn_implementation
+        else:
+            hf_config._attn_implementation = "flash_attention_2"
+        logger.info("Reason1 attention implementation: %s", getattr(hf_config, "_attn_implementation", None))
+        
+        with torch.device("meta"):
+            self.model = Qwen2_5_VLForConditionalGenerationSimple(hf_config)
+        
+        self.processor = AutoProcessor.from_pretrained(
+            config_path,
+            trust_remote_code=True,
+        )
+
+        weights_override = os.getenv("FASTVIDEO_REASON1_WEIGHTS_PATH")
+        if weights_override:
+            self.secondary_weights = (
+                _WeightsSource(
+                    model_or_path=weights_override,
+                    prefix="",
+                    fall_back_to_pt=True,
+                    allow_patterns_overrides=None,
+                ),
+            )
+            logger.info("Reason1TextEncoder: overlaying weights from %s", weights_override)
+        
+        self._weights_loaded = False
+    
+    def forward(
+        self,
+        input_ids: torch.Tensor | None,
+        position_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        output_hidden_states: bool | None = None,
+        **kwargs,
+    ) -> BaseEncoderOutput:
+        # Cosmos2.5 alignment: keep attention_mask=None.
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=None,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=True,
+            return_dict=True,
+            pixel_values=kwargs.get('pixel_values', None),
+            pixel_values_videos=kwargs.get('pixel_values_videos', None),
+            image_grid_thw=kwargs.get('image_grid_thw', None),
+            video_grid_thw=kwargs.get('video_grid_thw', None),
+        )
+        
+        hidden_states = outputs.hidden_states
+        last_hidden_state = hidden_states[-1]
+        
+        return BaseEncoderOutput(
+            last_hidden_state=last_hidden_state,
+            hidden_states=hidden_states if output_hidden_states else None,
+            attention_mask=None,
+        )
+    
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        first_weight = None
+        weights_list = []
+        for name, weight in weights:
+            if first_weight is None:
+                first_weight = weight
+                self.model = self.model.to_empty(device=weight.device)
+                self.model.init_weights(buffer_device=weight.device)
+            weights_list.append((name, weight))
+        
+        params_dict = dict(self.model.named_parameters())
+        loaded_params: set[str] = set()
+        skipped_weights = {"lm_head": 0, "visual": 0, "decoder": 0}
+        
+        for name, loaded_weight in weights_list:
+            if "lm_head" in name:
+                skipped_weights["lm_head"] += 1
+                continue
+            if "visual" in name:
+                skipped_weights["visual"] += 1
+                continue
+            if "decoder" in name:
+                skipped_weights["decoder"] += 1
+                continue
+            
+            # Handle stacked params mapping (for quantized models)
+            for param_name, weight_name, shard_id in self.config.arch_config.stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                
+                if name not in params_dict:
+                    continue
+                
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                param_name_with_prefix = f"model.{name}" if self.prefix == "" else f"{self.prefix}.{name}"
+                loaded_params.add(param_name_with_prefix)
+                break
+            else:
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                
+                if name not in params_dict:
+                    continue
+                
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+                param_name_with_prefix = f"model.{name}" if self.prefix == "" else f"{self.prefix}.{name}"
+                loaded_params.add(param_name_with_prefix)
+        if first_weight is not None:
+            self.model = self.model.to(first_weight.device)
+        
+        all_params = set(f"model.{name}" if self.prefix == "" else f"{self.prefix}.{name}" 
+                        for name in params_dict.keys())
+        loaded_params.update(all_params)
+        
+        # Mark weights as loaded
+        self._weights_loaded = True
+        return loaded_params
+    
+    def compute_text_embeddings_online(
+        self,
+        data_batch: dict[str, list[str]],
+        input_caption_key: str,
+    ) -> torch.Tensor:
+        prompts = data_batch[input_caption_key]
+        return self.compute_text_embeddings(prompts)
+    
+    def compute_text_embeddings(
+        self,
+        prompts: list[str],
+        device: str | torch.device = "cuda",
+    ) -> torch.Tensor:
+        """Compute embeddings for a list of prompts."""
+        input_ids_batch = []
+        
+        tok = getattr(self.processor, "tokenizer", None)
+        if tok is None:
+            raise RuntimeError("Reason1TextEncoder requires processor.tokenizer")
+        pad_id = getattr(tok, "pad_id", None)
+        if pad_id is None:
+            pad_id = getattr(tok, "pad_token_id", None)
+        if pad_id is None:
+            pad_id = getattr(self.model.config, "pad_token_id", None)
+        if pad_id is None:
+            pad_id = 0
+        
+        for prompt in prompts:
+            conversations = [
+                {
+                    "role": "system",
+                    "content": [
+                        {
+                            "type": "text",
+                            "text": "You are a helpful assistant who will provide prompts to an image generator.",
+                        }
+                    ],
+                },
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "text",
+                            "text": prompt,
+                        }
+                    ],
+                },
+            ]
+            
+            try:
+                tokenizer_output = tok.apply_chat_template(
+                    conversations,
+                    tokenize=True,
+                    add_generation_prompt=False,
+                    add_vision_id=False,
+                )
+            except TypeError:
+                tokenizer_output = tok.apply_chat_template(
+                conversations,
+                tokenize=True,
+                add_generation_prompt=False,
+            )
+
+            if isinstance(tokenizer_output, dict) and "input_ids" in tokenizer_output:
+                input_ids = tokenizer_output["input_ids"]
+                if hasattr(input_ids, "tolist"):
+                    input_ids = input_ids.tolist()
+            else:
+                input_ids = tokenizer_output
+                if hasattr(input_ids, "tolist"):
+                    input_ids = input_ids.tolist()
+                if isinstance(input_ids, list) and len(input_ids) == 1 and isinstance(
+                        input_ids[0], list):
+                    input_ids = input_ids[0]
+                if not isinstance(input_ids, list):
+                    raise RuntimeError(
+                        f"Unexpected chat_template output type: {type(tokenizer_output)}"
+                    )
+            
+            if self.num_embedding_padding_tokens > len(input_ids):
+                pad_len = self.num_embedding_padding_tokens - len(input_ids)
+                input_ids = input_ids + [pad_id] * pad_len
+            else:
+                input_ids = input_ids[:self.num_embedding_padding_tokens]
+            
+            input_ids = torch.LongTensor(input_ids).to(device=device)
+            input_ids_batch.append(input_ids)
+        
+        input_ids_batch = torch.stack(input_ids_batch, dim=0)
+
+        # Cosmos2.5 alignment: keep attention_mask=None.
+        target_device = input_ids_batch.device
+        try:
+            embed_device = self.model.model.embed_tokens.weight.device  # type: ignore[attr-defined]
+        except Exception:
+            embed_device = None
+        if embed_device is not None and embed_device != target_device:
+            self.model = self.model.to(target_device)
+        
+        with torch.no_grad():
+            position_ids, _ = get_rope_index(
+                self.model.config,
+                input_ids_batch,
+                image_grid_thw=None,
+                video_grid_thw=None,
+                second_per_grid_ts=None,
+                attention_mask=None,
+            )
+            position_ids = position_ids.to(target_device)
+                
+            outputs = self.model.model(
+                input_ids=input_ids_batch,
+                position_ids=position_ids,
+                attention_mask=None,
+                output_hidden_states=True,
+                return_dict=True,
+                use_cache=False,
+            )
+            hidden_states = outputs.hidden_states
+        
+        normalized_hidden_states = []
+        for layer_idx in range(1, len(hidden_states)):
+            normalized_state = self._mean_normalize(hidden_states[layer_idx])
+            normalized_hidden_states.append(normalized_state)
+        
+        if self.embedding_concat_strategy == "full_concat":
+            text_embeddings = torch.cat(normalized_hidden_states, dim=-1)
+        elif self.embedding_concat_strategy == "mean_pooling":
+            text_embeddings = torch.stack(normalized_hidden_states).mean(dim=0)
+        elif self.embedding_concat_strategy == "pool_every_n_layers_and_concat":
+            pooled_embeddings = []
+            for i in range(0, len(normalized_hidden_states), self.n_layers_per_group):
+                group = normalized_hidden_states[i : i + self.n_layers_per_group]
+                pooled = torch.stack(group).mean(dim=0)
+                pooled_embeddings.append(pooled)
+            text_embeddings = torch.cat(pooled_embeddings, dim=-1)
+        else:
+            raise ValueError(
+                f"Unknown embedding_concat_strategy: {self.embedding_concat_strategy}"
+            )
+        
+        return text_embeddings
+    @staticmethod
+    def _mean_normalize(tensor: torch.Tensor) -> torch.Tensor:
+        return (tensor - tensor.mean(dim=-1, keepdim=True)) / (
+            tensor.std(dim=-1, keepdim=True) + 1e-8
+        )
+

fastvideo/models/loader/component_loader.py

@@ -15,8 +15,7 @@ import torch.distributed as dist
 import torch.nn as nn
 from safetensors.torch import load_file as safetensors_load_file
 from torch.distributed import init_device_mesh
-from transformers import AutoImageProcessor, AutoTokenizer
-from transformers import UMT5EncoderModel
+from transformers import AutoImageProcessor, AutoProcessor, AutoTokenizer
 from transformers.utils import SAFE_WEIGHTS_INDEX_NAME
 
 from fastvideo.configs.models import EncoderConfig
@@ -35,8 +34,8 @@ from fastvideo.models.loader.weight_utils import (
     safetensors_weights_iterator,
 )
 from fastvideo.models.registry import ModelRegistry
-from fastvideo.utils import PRECISION_TO_TYPE
-from fastvideo.models.layerwise_offload import LayerwiseOffloadManager
+from fastvideo.utils import PRECISION_TO_TYPE, is_pin_memory_available
+from fastvideo.hooks.layerwise_offload import enable_layerwise_offload
 
 logger = init_logger(__name__)
 
@@ -91,10 +90,12 @@ class ComponentLoader(ABC):
 
         if module_type in module_loaders:
             loader_cls, expected_library = module_loaders[module_type]
-            # Assert that the library matches what's expected for this module type
-            assert transformers_or_diffusers == expected_library, (
-                f"{module_type} must be loaded from {expected_library}, got {transformers_or_diffusers}"
-            )
+            # Allow fastvideo.* libraries for custom implementations (e.g. Cosmos2_5Pipeline)
+            # that aren't available in diffusers/transformers yet
+            is_fastvideo_module = transformers_or_diffusers.startswith("fastvideo.")
+            if not is_fastvideo_module:
+                # Assert that the library matches what's expected for this module type
+                assert transformers_or_diffusers == expected_library, f"{module_type} must be loaded from {expected_library}, got {transformers_or_diffusers}"
             return loader_cls()
 
         # For unknown module types, use a generic loader
@@ -279,7 +280,7 @@ class TextEncoderLoader(ComponentLoader):
         target_device: torch.device,
         fastvideo_args: FastVideoArgs,
         dtype: str = "fp16",
-        use_text_encoder_override: bool = False, # prevent subclasses from misusing
+        use_text_encoder_override: bool = False,  # prevent subclasses from misusing
     ):
         use_cpu_offload = (
             fastvideo_args.text_encoder_cpu_offload
@@ -296,7 +297,10 @@ class TextEncoderLoader(ComponentLoader):
             )
 
         # Set quantization config if specified
-        if use_text_encoder_override and fastvideo_args.override_text_encoder_quant is not None:
+        if (
+            use_text_encoder_override
+            and fastvideo_args.override_text_encoder_quant is not None
+        ):
             if fastvideo_args.override_text_encoder_safetensors is None:
                 raise ValueError(
                     "override_text_encoder_quant is set but override_text_encoder_safetensors is None"
@@ -313,7 +317,10 @@ class TextEncoderLoader(ComponentLoader):
                 model: TextEncoder = model_cls(model_config)  # type: ignore
 
             weights_to_load = {name for name, _ in model.named_parameters()}
-            if use_text_encoder_override and fastvideo_args.override_text_encoder_safetensors is not None:
+            if (
+                use_text_encoder_override
+                and fastvideo_args.override_text_encoder_safetensors is not None
+            ):
                 loaded_weights: set[str] = model.load_weights(
                     safetensors_weights_iterator(
                         [fastvideo_args.override_text_encoder_safetensors],
@@ -340,6 +347,7 @@ class TextEncoderLoader(ComponentLoader):
             from fastvideo.platforms import current_platform
 
             if use_cpu_offload:
+                pin_cpu_memory = fastvideo_args.pin_cpu_memory and is_pin_memory_available()
                 # Disable FSDP for MPS as it's not compatible
                 if current_platform.is_mps():
                     logger.info(
@@ -357,7 +365,7 @@ class TextEncoderLoader(ComponentLoader):
                         reshard_after_forward=True,
                         mesh=mesh["offload"],
                         fsdp_shard_conditions=model._fsdp_shard_conditions,
-                        pin_cpu_memory=fastvideo_args.pin_cpu_memory,
+                        pin_cpu_memory=pin_cpu_memory,
                     )
                 else:
                     mesh = init_device_mesh(
@@ -371,7 +379,7 @@ class TextEncoderLoader(ComponentLoader):
                         reshard_after_forward=True,
                         mesh=mesh["offload"],
                         fsdp_shard_conditions=model._fsdp_shard_conditions,
-                        pin_cpu_memory=fastvideo_args.pin_cpu_memory,
+                        pin_cpu_memory=pin_cpu_memory,
                     )
             # We only enable strict check for non-quantized models
             # that have loaded weights tracking currently.
@@ -449,6 +457,33 @@ class TokenizerLoader(ComponentLoader):
         """Load the tokenizer based on the model path, and inference args."""
         logger.info("Loading tokenizer from %s", model_path)
 
+        # Cosmos2.5 stores an AutoProcessor config in `tokenizer/config.json` (not a tokenizer
+        # config). Use its `_name_or_path` (e.g. Qwen/Qwen2.5-VL-7B-Instruct) as the source.
+        tokenizer_cfg_path = os.path.join(model_path, "config.json")
+        if os.path.exists(tokenizer_cfg_path):
+            try:
+                with open(tokenizer_cfg_path, "r") as f:
+                    tokenizer_cfg = json.load(f)
+                if isinstance(tokenizer_cfg, dict) and (
+                    tokenizer_cfg.get("_class_name") == "AutoProcessor"
+                    or "processor_type" in tokenizer_cfg
+                ):
+                    src = tokenizer_cfg.get("_name_or_path", "")
+                    if isinstance(src, str) and src.strip():
+                        processor = AutoProcessor.from_pretrained(
+                            src.strip(),
+                            trust_remote_code=True,
+                        )
+                        logger.info(
+                            "Loaded tokenizer/processor from %s: %s",
+                            src,
+                            processor.__class__.__name__,
+                        )
+                        return processor
+            except Exception:
+                # If parsing fails, fall through to AutoTokenizer below.
+                pass
+
         tokenizer = AutoTokenizer.from_pretrained(
             model_path,  # "<path to model>/tokenizer"
             # in v0, this was same string as encoder_name "ClipTextModel"
@@ -491,19 +526,40 @@ class VAELoader(ComponentLoader):
             if fastvideo_args.pipeline_config.vae_precision
             else torch.bfloat16
         ):
+            # Cosmos2.5 uses a Wan2.1 VAE stored as `tokenizer.safetensors` under the VAE folder.
+            is_cosmos25 = fastvideo_args.pipeline_config.__class__.__name__ == "Cosmos25Config"
+            if class_name == "AutoencoderKLWan" and is_cosmos25:
+                from fastvideo.models.vaes.cosmos25wanvae import Cosmos25WanVAE
+
+                dtype = PRECISION_TO_TYPE[fastvideo_args.pipeline_config.vae_precision]
+                vae = Cosmos25WanVAE(device=target_device, dtype=dtype)
+
+                weight_path = os.path.join(model_path, "tokenizer.safetensors")
+                if not os.path.exists(weight_path):
+                    raise FileNotFoundError(
+                        f"Missing Cosmos2.5 VAE weights: {weight_path}"
+                    )
+                sd = safetensors_load_file(weight_path)
+                vae.load_state_dict(sd, strict=False)
+                return vae.eval()
+
             vae_cls, _ = ModelRegistry.resolve_model_cls(class_name)
             vae = vae_cls(vae_config).to(target_device)
 
         # Find all safetensors files
         safetensors_list = glob.glob(
-            os.path.join(str(model_path), "*.safetensors")
-        )
-        loaded = {}
-        for sf_file in safetensors_list:
-            loaded.update(safetensors_load_file(sf_file))
-        vae.load_state_dict(
-            loaded, strict=False
-        )  # We might only load encoder or decoder
+            os.path.join(str(model_path), "*.safetensors"))
+        if not safetensors_list:
+            raise ValueError(f"No safetensors files found in {model_path}")
+        # Common case: a single `.safetensors` checkpoint file.
+        # Some models may be sharded into multiple files; in that case we merge.
+        if len(safetensors_list) == 1:
+            loaded = safetensors_load_file(safetensors_list[0])
+        else:
+            loaded = {}
+            for sf_file in safetensors_list:
+                loaded.update(safetensors_load_file(sf_file))
+        vae.load_state_dict(loaded, strict=False)
 
         return vae.eval()
 
@@ -581,7 +637,17 @@ class TransformerLoader(ComponentLoader):
         ]
 
         # Load the model using FSDP loader
+        logger.info("Loading model from %s, default_dtype: %s", cls_name,
+                    default_dtype)
         assert fastvideo_args.hsdp_shard_dim is not None
+        # Cosmos2.5 checkpoints can include extra entries not present in the
+        # instantiated model (e.g. pos_embedder ranges / *_extra_state). Load
+        # non-strictly for Cosmos2.5 only; keep upstream strict behavior for others.
+        strict_load = not (
+            cls_name.startswith("Cosmos25")
+            or cls_name == "Cosmos25Transformer3DModel"
+            or getattr(fastvideo_args.pipeline_config, "prefix", "") == "Cosmos25"
+        )
         model = maybe_load_fsdp_model(
             model_cls=model_cls,
             init_params={"config": dit_config, "hf_config": hf_config},
@@ -589,6 +655,7 @@ class TransformerLoader(ComponentLoader):
             device=get_local_torch_device(),
             hsdp_replicate_dim=fastvideo_args.hsdp_replicate_dim,
             hsdp_shard_dim=fastvideo_args.hsdp_shard_dim,
+            strict=strict_load,
             cpu_offload=fastvideo_args.dit_cpu_offload,
             pin_cpu_memory=fastvideo_args.pin_cpu_memory,
             fsdp_inference=fastvideo_args.use_fsdp_inference,
@@ -611,35 +678,8 @@ class TransformerLoader(ComponentLoader):
 
         model = model.eval()
 
-        if fastvideo_args.dit_layerwise_offload and hasattr(model, "blocks"):
-            # Check if this is a Wan model (only Wan models support layerwise offload)
-            is_wan_model = "Wan" in cls_name
-            if not is_wan_model:
-                logger.warning(
-                    "Layerwise offload is currently only supported for Wan models. "
-                    "Model class '%s' does not support layerwise offload. "
-                    "Disabling layerwise offload for this model.",
-                    cls_name
-                )
-            else:
-                try:
-                    num_layers = len(getattr(model, "blocks"))
-                except TypeError:
-                    num_layers = None
-                if isinstance(num_layers, int) and num_layers > 0:
-                    # Ensure model is on the correct device (CUDA) before initializing manager
-                    # This ensures non-managed parameters (embeddings, final norms) are on GPU
-                    model = model.to(get_local_torch_device())
-                    mgr = LayerwiseOffloadManager(
-                        model,
-                        module_list_attr="blocks",
-                        num_layers=num_layers,
-                        enabled=True,
-                        pin_cpu_memory=fastvideo_args.pin_cpu_memory,
-                        auto_initialize=True,
-                    )
-                    setattr(model, "_layerwise_offload_manager", mgr)
-
+        if fastvideo_args.inference_mode and fastvideo_args.dit_layerwise_offload:
+            enable_layerwise_offload(model)
         return model
 
 

fastvideo/models/loader/fsdp_load.py

@@ -23,7 +23,7 @@ from fastvideo.logger import init_logger
 from fastvideo.models.loader.utils import (get_param_names_mapping,
                                            hf_to_custom_state_dict)
 from fastvideo.models.loader.weight_utils import safetensors_weights_iterator
-from fastvideo.utils import set_mixed_precision_policy
+from fastvideo.utils import set_mixed_precision_policy, is_pin_memory_available
 
 logger = init_logger(__name__)
 
@@ -67,6 +67,7 @@ def maybe_load_fsdp_model(
     default_dtype: torch.dtype,
     param_dtype: torch.dtype,
     reduce_dtype: torch.dtype,
+    strict: bool = True,
     cpu_offload: bool = False,
     fsdp_inference: bool = False,
     output_dtype: torch.dtype | None = None,
@@ -106,6 +107,7 @@ def maybe_load_fsdp_model(
         logger.info("Disabling FSDP for MPS platform as it's not compatible")
 
     if use_fsdp:
+        pin_cpu_memory = pin_cpu_memory and is_pin_memory_available()
         world_size = hsdp_replicate_dim * hsdp_shard_dim
         if not training_mode and not fsdp_inference:
             hsdp_replicate_dim = world_size
@@ -141,7 +143,7 @@ def maybe_load_fsdp_model(
         weight_iterator,
         device,
         default_dtype,
-        strict=True,
+        strict=strict,
         cpu_offload=cpu_offload,
         param_names_mapping=param_names_mapping_fn,
     )
@@ -151,6 +153,7 @@ def maybe_load_fsdp_model(
                 f"Unexpected param or buffer {n} on meta device.")
         # Avoid unintended computation graph accumulation during inference
         if isinstance(p, torch.nn.Parameter):
+            
             p.requires_grad = False
 
     compile_in_loader = enable_torch_compile and training_mode
@@ -293,6 +296,26 @@ def load_model_from_full_model_state_dict(
     for target_param_name, full_tensor in custom_param_sd.items():
         meta_sharded_param = meta_sd.get(target_param_name)
         if meta_sharded_param is None:
+            # Some checkpoints include extra entries that are not part of the
+            # instantiated model's state_dict (e.g. `_extra_state` keys from
+            # some FSDP checkpoint formats). These can be safely skipped.
+            if (target_param_name.endswith("._extra_state")
+                    or target_param_name.endswith("_extra_state")):
+                logger.warning(
+                    "Skipping non-parameter checkpoint key: %s",
+                    target_param_name,
+                )
+                continue
+
+            # For non-strict loads, treat this as an "unexpected key" and skip it
+            # (mirrors torch.nn.Module.load_state_dict(strict=False)).
+            if not strict:
+                logger.warning(
+                    "Skipping unexpected checkpoint key (not present in model): %s",
+                    target_param_name,
+                )
+                continue
+
             raise ValueError(
                 f"Parameter {target_param_name} not found in custom model state dict. The hf to custom mapping may be incorrect."
             )

fastvideo/models/registry.py

@@ -30,8 +30,9 @@ _TEXT_TO_VIDEO_DIT_MODELS = {
     "CausalWanTransformer3DModel": ("dits", "causal_wanvideo", "CausalWanTransformer3DModel"),
     "StepVideoModel": ("dits", "stepvideo", "StepVideoModel"),
     "CosmosTransformer3DModel": ("dits", "cosmos", "CosmosTransformer3DModel"),
-    "LongCatVideoTransformer3DModel": ("dits", "longcat_video_dit", "LongCatVideoTransformer3DModel"),
-    "LongCatTransformer3DModel": ("dits", "longcat", "LongCatTransformer3DModel"),
+    "Cosmos25Transformer3DModel": ("dits", "cosmos2_5", "Cosmos25Transformer3DModel"),
+    "LongCatVideoTransformer3DModel": ("dits", "longcat_video_dit", "LongCatVideoTransformer3DModel"),  # Wrapper (Phase 1)
+    "LongCatTransformer3DModel": ("dits", "longcat", "LongCatTransformer3DModel"),  # Native (Phase 2)
 }
 
 _IMAGE_TO_VIDEO_DIT_MODELS = {
@@ -50,6 +51,9 @@ _TEXT_ENCODER_MODELS = {
     "STEP1TextEncoder": ("encoders", "stepllm", "STEP1TextEncoder"),
     "BertModel": ("encoders", "clip", "CLIPTextModel"),
     "Qwen2_5_VLTextModel": ("encoders", "qwen2_5", "Qwen2_5_VLTextModel"),
+    "Reason1TextEncoder": ("encoders", "reason1", "Reason1TextEncoder"),
+    "Qwen2_5_VLForConditionalGeneration":
+    ("encoders", "reason1", "Reason1TextEncoder"),
 }
 
 _IMAGE_ENCODER_MODELS: dict[str, tuple] = {
@@ -72,6 +76,8 @@ _SCHEDULERS = {
      "FlowMatchEulerDiscreteScheduler"),
     "UniPCMultistepScheduler":
     ("schedulers", "scheduling_unipc_multistep", "UniPCMultistepScheduler"),
+    "FlowUniPCMultistepScheduler":
+    ("schedulers", "scheduling_flow_unipc_multistep", "FlowUniPCMultistepScheduler"),
     "SelfForcingFlowMatchScheduler":
     ("schedulers", "scheduling_self_forcing_flow_match",
      "SelfForcingFlowMatchScheduler"),

fastvideo/models/schedulers/scheduling_flow_unipc_multistep.py

@@ -109,6 +109,9 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
         sigmas = 1.0 - alphas
         sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32)
 
+        # Needed when final_sigmas_type == "sigma_min" (kept for compatibility).
+        self.alphas_cumprod = torch.from_numpy(alphas).to(dtype=torch.float32)
+
         if not use_dynamic_shifting:
             # when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution
             assert shift is not None
@@ -171,6 +174,8 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
         sigmas: list[float] | None = None,
         mu: float | None | None = None,
         shift: float | None | None = None,
+        use_karras_sigmas: bool | None = None,
+        use_kerras_sigma: bool | None = None,
     ):
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
@@ -186,21 +191,44 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
                 " you have to pass a value for `mu` when `use_dynamic_shifting` is set to be `True`"
             )
 
-        if sigmas is None:
-            assert num_inference_steps is not None
-            sigmas = np.linspace(self.sigma_max, self.sigma_min,
-                                 num_inference_steps +
-                                 1).copy()[:-1]  # pyright: ignore
+
+        # Cosmos official uses `use_kerras_sigma=True` and a specific EDM sigma schedule.
+        # Some external code uses the misspelling `use_kerras_sigma`; support both.
+        if use_karras_sigmas is None and use_kerras_sigma is not None:
+            use_karras_sigmas = use_kerras_sigma
+
+        if use_karras_sigmas:
+            # Force to use the exact sigma used in official EDM sampler:
+            # sigma_max=200, sigma_min=0.01, rho=7
+            sigma_max = 200.0
+            sigma_min = 0.01
+            rho = 7.0
+            # Match the official Cosmos implementation: Karras/EDM schedule with
+            # `num_inference_steps + 1` points, then `final_sigmas_type="zero"`
+            # appends the terminal sigma (0.0).
+            ramp = np.arange(num_inference_steps + 1,
+                             dtype=np.float32) / float(num_inference_steps)
+            min_inv_rho = sigma_min**(1 / rho)
+            max_inv_rho = sigma_max**(1 / rho)
+            sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho))**rho
+            # Convert EDM sigma to flow-matching sigma in [0, 1).
+            sigmas = sigmas / (1.0 + sigmas)
+        else:
+            if sigmas is None:
+                assert num_inference_steps is not None
+                sigmas = np.linspace(self.sigma_max, self.sigma_min,
+                                     num_inference_steps +
+                                     1).copy()[:-1]  # pyright: ignore
 
         if self.config.use_dynamic_shifting:
             assert mu is not None
             sigmas = self.time_shift(mu, 1.0, sigmas)  # pyright: ignore
         else:
-            if shift is None:
-                shift = self.config.shift
-            assert isinstance(sigmas, np.ndarray)
-            sigmas = shift * sigmas / (1 +
-                                       (shift - 1) * sigmas)  # pyright: ignore
+            if not use_karras_sigmas:
+                if shift is None:
+                    shift = self.config.shift
+                assert isinstance(sigmas, np.ndarray)
+                sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)  # pyright: ignore
 
         if self.config.final_sigmas_type == "sigma_min":
             sigma_last = ((1 - self.alphas_cumprod[0]) /
@@ -418,8 +446,12 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
         alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
         alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
 
-        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
-        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+        # Numerical safety
+        eps = 1e-12
+        lambda_t = torch.log(torch.clamp(alpha_t, min=eps)) - torch.log(
+            torch.clamp(sigma_t, min=eps))
+        lambda_s0 = torch.log(torch.clamp(alpha_s0, min=eps)) - torch.log(
+            torch.clamp(sigma_s0, min=eps))
 
         h = lambda_t - lambda_s0
         device = sample.device
@@ -430,7 +462,8 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
             si = self.step_index - i  # pyright: ignore
             mi = model_output_list[-(i + 1)]
             alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si])
-            lambda_si = torch.log(alpha_si) - torch.log(sigma_si)
+            lambda_si = torch.log(torch.clamp(alpha_si, min=eps)) - torch.log(
+                torch.clamp(sigma_si, min=eps))
             rk = (lambda_si - lambda_s0) / h
             rks.append(rk)
             assert mi is not None
@@ -563,8 +596,11 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
         alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
         alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
 
-        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
-        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+        eps = 1e-12
+        lambda_t = torch.log(torch.clamp(alpha_t, min=eps)) - torch.log(
+            torch.clamp(sigma_t, min=eps))
+        lambda_s0 = torch.log(torch.clamp(alpha_s0, min=eps)) - torch.log(
+            torch.clamp(sigma_s0, min=eps))
 
         h = lambda_t - lambda_s0
         device = this_sample.device
@@ -575,7 +611,8 @@ class FlowUniPCMultistepScheduler(SchedulerMixin, ConfigMixin, BaseScheduler):
             si = self.step_index - (i + 1)  # pyright: ignore
             mi = model_output_list[-(i + 1)]
             alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si])
-            lambda_si = torch.log(alpha_si) - torch.log(sigma_si)
+            lambda_si = torch.log(torch.clamp(alpha_si, min=eps)) - torch.log(
+                torch.clamp(sigma_si, min=eps))
             rk = (lambda_si - lambda_s0) / h
             rks.append(rk)
             assert mi is not None

fastvideo/models/vaes/cosmos25wanvae.py

@@ -0,0 +1,735 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+"""
+Cosmos 2.5 / Wan2.1 VAE adapter.
+
+Why this exists:
+- Cosmos2.5 uses a Wan2.1-style VAE, but the *diffusion model* operates in a
+  **normalized latent space**:
+    z_norm = (z - mean) / std
+
+  Meanwhile, FastVideo's `AutoencoderKLWan` operates in the VAE's native latent
+  space (denormalized):
+    z = z_norm * std + mean
+
+This adapter provides a single, stable interface for FastVideo pipelines:
+- `encode(x)` returns an object with `.mean` / `.sample()` / `.mode()`
+- `decode(z)` returns a tensor in pixel space
+
+It also exposes flags used by pipeline stages to avoid double (de)normalization:
+- `handles_latent_norm = True`   -> stages should NOT normalize encoder latents
+- `handles_latent_denorm = True` -> stages should NOT denormalize before decode
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+@dataclass
+class _TensorLatentDist:
+    """Minimal distribution-like wrapper used by pipeline stages."""
+
+    mean: torch.Tensor
+
+    def mode(self) -> torch.Tensor:
+        return self.mean
+
+    def sample(self, generator: Any | None = None) -> torch.Tensor:  # generator for API compatibility
+        # The official interface encodes deterministically; for compatibility we
+        # return the mean. (Stochastic posterior sampling isn't required for
+        # Cosmos2.5 inference.)
+        _ = generator
+        return self.mean
+
+
+class Cosmos25WanVAEAdapter(nn.Module):
+    """
+    Adapter that makes a Wan2.1-style VAE follow Cosmos2.5's latent contract:
+    - `encode()` returns **normalized** latents
+    - `decode()` expects **normalized** latents
+    """
+
+    # Pipeline stage hints (see latent_preparation.py / decoding.py / image_encoding.py)
+    handles_latent_norm: bool = True
+    handles_latent_denorm: bool = True
+    latent_norm_mode: str = "internal"  # informational
+
+    def __init__(
+        self,
+        inner: Any,
+        *,
+        latents_mean: Optional[torch.Tensor] = None,
+        latents_std: Optional[torch.Tensor] = None,
+    ) -> None:
+        super().__init__()
+        self.inner = inner
+
+        # Preserve `config` when available; some pipeline utilities expect it.
+        self.config = getattr(inner, "config", None)
+
+        # If not provided, try to derive from `config.latents_mean/std`.
+        cfg = self.config
+        if latents_mean is None and cfg is not None and hasattr(cfg, "latents_mean"):
+            latents_mean = torch.tensor(cfg.latents_mean, dtype=torch.float32).view(1, -1, 1, 1, 1)
+        if latents_std is None and cfg is not None and hasattr(cfg, "latents_std"):
+            latents_std = torch.tensor(cfg.latents_std, dtype=torch.float32).view(1, -1, 1, 1, 1)
+
+        if latents_mean is None or latents_std is None:
+            raise RuntimeError(
+                "Cosmos25WanVAEAdapter requires latents_mean/latents_std (either passed explicitly or available on inner.config)."
+            )
+
+        # Register as buffers so `.to(...)` moves them with the module.
+        self.register_buffer("_latents_mean", latents_mean, persistent=False)
+        self.register_buffer("_latents_std", latents_std, persistent=False)
+
+    def _to_latent_stats(self, like: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        mean = self._latents_mean.to(device=like.device, dtype=like.dtype)
+        std = self._latents_std.to(device=like.device, dtype=like.dtype)
+        return mean, std
+
+    def get_latent_num_frames(self, num_pixel_frames: int) -> int:
+        # Keep parity with official interface.
+        if hasattr(self.inner, "get_latent_num_frames"):
+            return int(self.inner.get_latent_num_frames(num_pixel_frames))
+        return 1 + (num_pixel_frames - 1) // 4
+
+    def encode(self, x: torch.Tensor) -> _TensorLatentDist:
+        """
+        Returns *normalized* latents (Cosmos contract).
+        """
+        enc_out = self.inner.encode(x)
+
+        # Support common encoder output shapes:
+        # - DiagonalGaussianDistribution (FastVideo VAE): has `.mean` / `.sample()` / `.mode()`
+        # - diffusers EncoderOutput: has `.latent_dist`
+        # - raw tensor
+        if hasattr(enc_out, "latent_dist"):
+            dist = enc_out.latent_dist
+            z_mean = dist.mode() if hasattr(dist, "mode") else dist.mean
+        elif hasattr(enc_out, "mode") and hasattr(enc_out, "mean"):
+            z_mean = enc_out.mode()
+        elif isinstance(enc_out, torch.Tensor):
+            z_mean = enc_out
+        else:
+            attrs = [a for a in dir(enc_out) if not a.startswith("_")]
+            raise RuntimeError(
+                f"Unsupported VAE encoder output type: {type(enc_out)}. attrs={attrs}"
+            )
+
+        mean, std = self._to_latent_stats(z_mean)
+        z_norm = (z_mean - mean) / std
+        return _TensorLatentDist(z_norm)
+
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        """
+        Expects *normalized* latents (Cosmos contract).
+        """
+        mean, std = self._to_latent_stats(z)
+        z_denorm = z * std + mean
+        out = self.inner.decode(z_denorm)
+        return out.sample if hasattr(out, "sample") else out
+
+
+#
+# Official-like Wan2.1 VAE implementation (ported from cosmos_predict2 wan2pt1.py)
+# -------------------------------------------------------------------------------
+# Motivation:
+# - We already solved checkpoint *key mapping* and can load official weights.
+# - Remaining output drift vs the official tokenizer is largely decoder-side.
+# - FastVideo's `AutoencoderKLWan` uses a different temporal upsample path
+#   (`DupUp3D` + `first_chunk` slicing), while the official tokenizer uses
+#   `Resample(mode="upsample3d")` with a time-conv + interleave reshape.
+#
+# This section ports the core modules (CausalConv3d/Resample/etc.) so we can run
+# a VAE that is behaviorally closer to the official implementation WITHOUT
+# importing any official repo classes at runtime.
+#
+
+CACHE_T = 2
+
+
+class Cosmos25CausalConv3d(nn.Conv3d):
+    """
+    Official-like causal 3D convolution.
+
+    Matches `CausalConv3d` in the official tokenizer: uses explicit F.pad and
+    supports a `cache_x` prefix for causal chunking.
+    """
+
+    def __init__(self, *args: Any, **kwargs: Any) -> None:
+        super().__init__(*args, **kwargs)
+        # padding order for F.pad: (W_left, W_right, H_left, H_right, T_left, T_right)
+        self._padding: tuple[int, ...] = (
+            self.padding[2],
+            self.padding[2],
+            self.padding[1],
+            self.padding[1],
+            2 * self.padding[0],
+            0,
+        )
+        self.padding = (0, 0, 0)
+
+    def forward(self, x: torch.Tensor, cache_x: torch.Tensor | None = None) -> torch.Tensor:
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+        return super().forward(x)
+
+
+class Cosmos25RMSNorm(nn.Module):
+    """Official-like RMS_norm (uses learnable gamma and optional bias)."""
+
+    def __init__(self, dim: int, channel_first: bool = True, images: bool = True, bias: bool = False) -> None:
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.0
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        dim = 1 if self.channel_first else -1
+        return F.normalize(x, dim=dim) * self.scale * self.gamma + self.bias
+
+
+class Cosmos25Upsample(nn.Upsample):
+    """Official-like Upsample that is safe for bf16 (casts to fp32 internally)."""
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:  # type: ignore[override]
+        return super().forward(x.float()).type_as(x)
+
+
+class Cosmos25Resample(nn.Module):
+    """
+    Official-like Resample used for both spatial and temporal up/downsampling.
+    """
+
+    def __init__(self, dim: int, mode: str) -> None:
+        assert mode in ("none", "upsample2d", "upsample3d", "downsample2d", "downsample3d")
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                Cosmos25Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim // 2, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                Cosmos25Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim // 2, 3, padding=1),
+            )
+            self.time_conv = Cosmos25CausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = Cosmos25CausalConv3d(dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x: torch.Tensor, feat_cache: list[Any] | None = None, feat_idx: list[int] = [0]) -> torch.Tensor:
+        b, c, t, h, w = x.size()
+
+        # Temporal upsample uses a time-conv and then interleaves frames.
+        if self.mode == "upsample3d" and feat_cache is not None:
+            idx = feat_idx[0]
+            if feat_cache[idx] is None:
+                feat_cache[idx] = "Rep"
+                feat_idx[0] += 1
+            else:
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] != "Rep":
+                    cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] == "Rep":
+                    cache_x = torch.cat([torch.zeros_like(cache_x).to(cache_x.device), cache_x], dim=2)
+
+                if feat_cache[idx] == "Rep":
+                    x = self.time_conv(x)
+                else:
+                    x = self.time_conv(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+
+                x = x.reshape(b, 2, c, t, h, w)
+                x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]), 3)
+                x = x.reshape(b, c, t * 2, h, w)
+
+        t = x.shape[2]
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        x = self.resample(x)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=t)
+
+        # Temporal downsample: time_conv consumes last-frame cache.
+        if self.mode == "downsample3d" and feat_cache is not None:
+            idx = feat_idx[0]
+            if feat_cache[idx] is None:
+                feat_cache[idx] = x.clone()
+                feat_idx[0] += 1
+            else:
+                cache_x = x[:, :, -1:, :, :].clone()
+                x = self.time_conv(torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+
+        return x
+
+
+class Cosmos25ResidualBlock(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, dropout: float = 0.0) -> None:
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.residual = nn.Sequential(
+            Cosmos25RMSNorm(in_dim, images=False),
+            nn.SiLU(),
+            Cosmos25CausalConv3d(in_dim, out_dim, 3, padding=1),
+            Cosmos25RMSNorm(out_dim, images=False),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            Cosmos25CausalConv3d(out_dim, out_dim, 3, padding=1),
+        )
+        self.shortcut = Cosmos25CausalConv3d(in_dim, out_dim, 1) if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x: torch.Tensor, feat_cache: list[Any] | None = None, feat_idx: list[int] = [0]) -> torch.Tensor:
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if isinstance(layer, Cosmos25CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class Cosmos25AttentionBlock(nn.Module):
+    """Official-like causal self-attention with a single head."""
+
+    def __init__(self, dim: int) -> None:
+        super().__init__()
+        self.dim = dim
+        self.norm = Cosmos25RMSNorm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+        nn.init.zeros_(self.proj.weight)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        identity = x
+        b, c, t, h, w = x.size()
+        x2 = rearrange(x, "b c t h w -> (b t) c h w")
+        x2 = self.norm(x2)
+        q, k, v = (
+            self.to_qkv(x2)
+            .reshape(b * t, 1, c * 3, -1)
+            .permute(0, 1, 3, 2)
+            .contiguous()
+            .chunk(3, dim=-1)
+        )
+        x2 = F.scaled_dot_product_attention(q, k, v)
+        x2 = x2.squeeze(1).permute(0, 2, 1).reshape(b * t, c, h, w)
+        x2 = self.proj(x2)
+        x2 = rearrange(x2, "(b t) c h w-> b c t h w", t=t)
+        return x2 + identity
+
+
+class Cosmos25Encoder3d(nn.Module):
+    def __init__(
+        self,
+        dim: int = 96,
+        z_dim: int = 32,
+        dim_mult: list[int] = [1, 2, 4, 4],
+        num_res_blocks: int = 2,
+        attn_scales: list[float] = [],
+        temperal_downsample: list[bool] = [False, True, True],
+        dropout: float = 0.0,
+    ) -> None:
+        super().__init__()
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        self.conv1 = Cosmos25CausalConv3d(3, dims[0], 3, padding=1)
+
+        downsamples: list[nn.Module] = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            for _ in range(num_res_blocks):
+                downsamples.append(Cosmos25ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(Cosmos25AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            if i != len(dim_mult) - 1:
+                mode = "downsample3d" if temperal_downsample[i] else "downsample2d"
+                downsamples.append(Cosmos25Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        self.middle = nn.Sequential(
+            Cosmos25ResidualBlock(out_dim, out_dim, dropout),
+            Cosmos25AttentionBlock(out_dim),
+            Cosmos25ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        self.head = nn.Sequential(
+            Cosmos25RMSNorm(out_dim, images=False),
+            nn.SiLU(),
+            Cosmos25CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+    def forward(self, x: torch.Tensor, feat_cache: list[Any] | None = None, feat_idx: list[int] = [0]) -> torch.Tensor:
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)  # type: ignore[misc]
+            else:
+                x = layer(x)  # type: ignore[misc]
+
+        for layer in self.middle:
+            if isinstance(layer, Cosmos25ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)  # type: ignore[misc]
+
+        for layer in self.head:
+            if isinstance(layer, Cosmos25CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)  # type: ignore[misc]
+        return x
+
+
+class Cosmos25Decoder3d(nn.Module):
+    def __init__(
+        self,
+        dim: int = 96,
+        z_dim: int = 16,
+        dim_mult: list[int] = [1, 2, 4, 4],
+        num_res_blocks: int = 2,
+        attn_scales: list[float] = [],
+        temperal_upsample: list[bool] = [False, True, True],
+        dropout: float = 0.0,
+    ) -> None:
+        super().__init__()
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2 ** (len(dim_mult) - 2)
+
+        self.conv1 = Cosmos25CausalConv3d(z_dim, dims[0], 3, padding=1)
+        self.middle = nn.Sequential(
+            Cosmos25ResidualBlock(dims[0], dims[0], dropout),
+            Cosmos25AttentionBlock(dims[0]),
+            Cosmos25ResidualBlock(dims[0], dims[0], dropout),
+        )
+
+        upsamples: list[nn.Module] = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            if i in (1, 2, 3):
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(Cosmos25ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(Cosmos25AttentionBlock(out_dim))
+                in_dim = out_dim
+            if i != len(dim_mult) - 1:
+                mode = "upsample3d" if temperal_upsample[i] else "upsample2d"
+                upsamples.append(Cosmos25Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        self.head = nn.Sequential(
+            Cosmos25RMSNorm(out_dim, images=False),
+            nn.SiLU(),
+            Cosmos25CausalConv3d(out_dim, 3, 3, padding=1),
+        )
+
+    def forward(self, x: torch.Tensor, feat_cache: list[Any] | None = None, feat_idx: list[int] = [0]) -> torch.Tensor:
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if isinstance(layer, Cosmos25ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)  # type: ignore[misc]
+
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)  # type: ignore[misc]
+            else:
+                x = layer(x)  # type: ignore[misc]
+
+        for layer in self.head:
+            if isinstance(layer, Cosmos25CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)  # type: ignore[misc]
+        return x
+
+
+def _count_cosmos25_conv3d(model: nn.Module) -> int:
+    return sum(1 for m in model.modules() if isinstance(m, Cosmos25CausalConv3d))
+
+
+class Cosmos25WanVAE(nn.Module):
+    """
+    A FastVideo-native copy of the *official-like* Wan2.1 VAE core.
+
+    Key properties:
+    - Module naming matches official tokenizer (`encoder`, `decoder`, `conv1`, `conv2`)
+      so it can consume `tokenizer.pth` keys directly.
+    - `encode()` returns **normalized** latents and `decode()` expects **normalized**
+      latents (Cosmos2.5 contract), matching `Wan2pt1VAEInterface`.
+    """
+
+    handles_latent_norm: bool = True
+    handles_latent_denorm: bool = True
+
+    def __init__(
+        self,
+        *,
+        device: torch.device | str = "cpu",
+        dtype: torch.dtype = torch.float32,
+        temporal_window: int = 4,
+        latents_mean: Optional[torch.Tensor] = None,
+        latents_std: Optional[torch.Tensor] = None,
+    ) -> None:
+        super().__init__()
+
+        # Official hyperparams for Cosmos2.5 tokenizer (Wan2.1 VAE).
+        cfg = dict(
+            dim=96,
+            z_dim=16,
+            dim_mult=[1, 2, 4, 4],
+            num_res_blocks=2,
+            attn_scales=[],
+            temperal_downsample=[False, True, True],
+            dropout=0.0,
+            temporal_window=temporal_window,
+        )
+        self.z_dim = 16
+        self.temporal_window = temporal_window
+
+        self.encoder = Cosmos25Encoder3d(
+            dim=cfg["dim"],
+            z_dim=cfg["z_dim"] * 2,
+            dim_mult=cfg["dim_mult"],
+            num_res_blocks=cfg["num_res_blocks"],
+            attn_scales=cfg["attn_scales"],
+            temperal_downsample=cfg["temperal_downsample"],
+            dropout=cfg["dropout"],
+        )
+        self.conv1 = Cosmos25CausalConv3d(self.z_dim * 2, self.z_dim * 2, 1)
+        self.conv2 = Cosmos25CausalConv3d(self.z_dim, self.z_dim, 1)
+        self.decoder = Cosmos25Decoder3d(
+            dim=cfg["dim"],
+            z_dim=cfg["z_dim"],
+            dim_mult=cfg["dim_mult"],
+            num_res_blocks=cfg["num_res_blocks"],
+            attn_scales=cfg["attn_scales"],
+            temperal_upsample=list(cfg["temperal_downsample"])[::-1],
+            dropout=cfg["dropout"],
+        )
+
+        # Default Cosmos2.5 latent stats (shared with configs).
+        if latents_mean is None:
+            latents_mean = torch.tensor(
+                [
+                    -0.7571,
+                    -0.7089,
+                    -0.9113,
+                    0.1075,
+                    -0.1745,
+                    0.9653,
+                    -0.1517,
+                    1.5508,
+                    0.4134,
+                    -0.0715,
+                    0.5517,
+                    -0.3632,
+                    -0.1922,
+                    -0.9497,
+                    0.2503,
+                    -0.2921,
+                ],
+                dtype=torch.float32,
+            ).view(1, 16, 1, 1, 1)
+        if latents_std is None:
+            latents_std = torch.tensor(
+                [
+                    2.8184,
+                    1.4541,
+                    2.3275,
+                    2.6558,
+                    1.2196,
+                    1.7708,
+                    2.6052,
+                    2.0743,
+                    3.2687,
+                    2.1526,
+                    2.8652,
+                    1.5579,
+                    1.6382,
+                    1.1253,
+                    2.8251,
+                    1.9160,
+                ],
+                dtype=torch.float32,
+            ).view(1, 16, 1, 1, 1)
+
+        self.register_buffer("_latents_mean", latents_mean, persistent=False)
+        self.register_buffer("_latents_std", latents_std, persistent=False)
+
+        self.to(device=device, dtype=dtype)
+        self.clear_cache()
+
+    def clear_cache(self) -> None:
+        # Decoder cache
+        self._conv_num = _count_cosmos25_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map: list[Any] = [None] * self._conv_num
+        # Encoder cache
+        self._enc_conv_num = _count_cosmos25_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map: list[Any] = [None] * self._enc_conv_num
+
+    def _scale(self, like: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        mean = self._latents_mean.to(device=like.device, dtype=like.dtype)
+        std = self._latents_std.to(device=like.device, dtype=like.dtype)
+        return mean, 1.0 / std
+
+    def _i0_encode(self, x: torch.Tensor) -> torch.Tensor:
+        return self.encoder(x[:, :, :1, :, :], feat_cache=self._enc_feat_map, feat_idx=self._enc_conv_idx)
+
+    def _i0_decode(self, x: torch.Tensor) -> torch.Tensor:
+        return self.decoder(x[:, :, 0:1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+
+    def encode(self, x: torch.Tensor) -> _TensorLatentDist:
+        """
+        Encode to *normalized* latents (Cosmos contract).
+        """
+        self.clear_cache()
+        t = x.shape[2]
+        iters = 1 + (t - 1) // self.temporal_window
+
+        for i in range(iters):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self._i0_encode(x)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + self.temporal_window * (i - 1) : 1 + self.temporal_window * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+
+        if (t - 1) % self.temporal_window:
+            self._enc_conv_idx = [0]
+            out_ = self.encoder(
+                x[:, :, 1 + self.temporal_window * (iters - 1) :, :, :],
+                feat_cache=self._enc_feat_map,
+                feat_idx=self._enc_conv_idx,
+            )
+            out = torch.cat([out, out_], 2)
+
+        mu, _log_var = self.conv1(out).chunk(2, dim=1)
+        mean, inv_std = self._scale(mu)
+        z_norm = (mu - mean) * inv_std
+        self.clear_cache()
+        return _TensorLatentDist(z_norm)
+
+    def decode(self, latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode from *normalized* latents (Cosmos contract).
+        """
+        self.clear_cache()
+        mean, inv_std = self._scale(latent)
+        z = latent / inv_std + mean  # z = z_norm * std + mean
+
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self._i0_decode(x)
+            else:
+                out_ = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+
+        self.clear_cache()
+        return out
+
+    # --- Interface helpers (match official Wan2pt1VAEInterface) ---
+    def get_latent_num_frames(self, num_pixel_frames: int) -> int:
+        return 1 + (int(num_pixel_frames) - 1) // 4
+
+    def get_pixel_num_frames(self, num_latent_frames: int) -> int:
+        return (int(num_latent_frames) - 1) * 4 + 1
+
+    @property
+    def spatial_compression_factor(self) -> int:
+        return 8
+
+    @property
+    def temporal_compression_factor(self) -> int:
+        return 4
+
+    @property
+    def latent_ch(self) -> int:
+        return 16
+
+
+
+
+
+
+
+
+

fastvideo/pipelines/basic/cosmos/cosmos2_5_pipeline.py

@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Cosmos 2.5 pipeline entry (staged pipeline)."""
+
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
+from fastvideo.pipelines.stages import (ConditioningStage,
+                                        Cosmos25DenoisingStage,
+                                        Cosmos25LatentPreparationStage,
+                                        DecodingStage, InputValidationStage,
+                                        Cosmos25TextEncodingStage,
+                                        Cosmos25TimestepPreparationStage)
+
+logger = init_logger(__name__)
+
+
+class Cosmos2_5Pipeline(ComposedPipelineBase):
+    """Cosmos 2.5 video generation pipeline."""
+
+    _required_config_modules = [
+        "text_encoder", "tokenizer", "vae", "transformer", "scheduler",
+        "safety_checker"
+    ]
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
+        logger.info("Creating Cosmos 2.5 pipeline stages...")
+
+        self.add_stage(stage_name="input_validation_stage",
+                       stage=InputValidationStage())
+
+        self.add_stage(
+            stage_name="prompt_encoding_stage",
+            stage=Cosmos25TextEncodingStage(
+                text_encoder=self.get_module("text_encoder"), ),
+        )
+
+        self.add_stage(stage_name="conditioning_stage",
+                       stage=ConditioningStage())
+
+        self.add_stage(stage_name="timestep_preparation_stage",
+                       stage=Cosmos25TimestepPreparationStage(
+                           scheduler=self.get_module("scheduler")))
+
+        self.add_stage(stage_name="latent_preparation_stage",
+                       stage=Cosmos25LatentPreparationStage(
+                           scheduler=self.get_module("scheduler"),
+                           transformer=self.get_module("transformer"),
+                           vae=self.get_module("vae")))
+
+        self.add_stage(stage_name="denoising_stage",
+                       stage=Cosmos25DenoisingStage(
+                           transformer=self.get_module("transformer"),
+                           scheduler=self.get_module("scheduler")))
+
+        self.add_stage(stage_name="decoding_stage",
+                       stage=DecodingStage(vae=self.get_module("vae")))
+        logger.info("Cosmos 2.5 pipeline stages created")
+
+
+# Entry point for pipeline registry
+EntryClass = Cosmos2_5Pipeline

fastvideo/pipelines/composed_pipeline_base.py

@@ -201,8 +201,6 @@ class ComposedPipelineBase(ABC):
             # fwd, bwd, and other operations' precision.
             assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
 
-        logger.info("fastvideo_args in from_pretrained: %s", fastvideo_args)
-
         pipe = cls(model_path,
                    fastvideo_args,
                    required_config_modules=required_config_modules,

fastvideo/pipelines/lora_pipeline.py

@@ -1,7 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 from collections import defaultdict
 from collections.abc import Hashable
+from contextlib import nullcontext
 from typing import Any
+from collections.abc import Generator
 
 import torch
 import torch.distributed as dist
@@ -12,8 +14,13 @@ from torch.distributed.tensor import DTensor
 
 from fastvideo.distributed import get_local_torch_device
 from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
-from fastvideo.layers.lora.linear import (BaseLayerWithLoRA, get_lora_layer,
-                                          replace_submodule)
+from fastvideo.hooks.hooks import ModuleHookManager
+from fastvideo.hooks.layerwise_offload import LayerwiseOffloadHook
+from fastvideo.layers.lora.linear import (
+    BaseLayerWithLoRA,
+    get_lora_layer,
+    replace_submodule,
+)
 from fastvideo.logger import init_logger
 from fastvideo.models.loader.utils import get_param_names_mapping
 from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
@@ -22,17 +29,89 @@ from fastvideo.utils import maybe_download_lora
 logger = init_logger(__name__)
 
 
+def _get_hook_ctx(module: nn.Module | None):
+    if module is None:
+        return nullcontext()
+    hook_mgr = ModuleHookManager.get_from(module)
+    if hook_mgr is not None:
+        offload_hook = hook_mgr.forward_hooks.get(LayerwiseOffloadHook.name())
+        if offload_hook is not None:
+            return offload_hook.mutate_params_scope()  # type: ignore
+    return nullcontext()
+
+
+def _named_module_by_prefix(
+    module: nn.Module, prefixes: list[str]
+) -> list[tuple[str | None, list[tuple[str, nn.Module]]]]:
+    none_list: list[tuple[str, nn.Module]] = []
+    prefix_list: list[tuple[str, list[tuple[str, nn.Module]]]] = [
+        (prefix, []) for prefix in prefixes
+    ]
+    for name, submodule in module.named_modules():
+        for cur_prefix, cur_list in prefix_list:
+            # we should exclude e.g. block.1 and block.12.attn
+            if name.startswith(cur_prefix + "."):
+                cur_list.append((name, submodule))
+                break
+        else:
+            none_list.append((name, submodule))
+    return prefix_list + [(None, none_list)]  # type: ignore
+
+
+class LoRAModelLayers:
+
+    def __init__(self, block_list: list[tuple[str, nn.Module]]) -> None:
+        # block_name -> {layer_name -> layer}
+        self.block_to_lora_layers: dict[str, dict[str, BaseLayerWithLoRA]] = {}
+        # layer_name -> block_name
+        self.lora_layers_to_block: dict[str, str | None] = {}
+        self.other_lora_layers: dict[str, BaseLayerWithLoRA] = {}
+        self.block_mapping = dict(block_list)
+
+    def add_lora_layer(self, block_name: str | None, layer_name: str,
+                       layer: BaseLayerWithLoRA):
+        if block_name is None:
+            self.other_lora_layers[layer_name] = layer
+            self.lora_layers_to_block[layer_name] = None
+        else:
+            if block_name not in self.block_to_lora_layers:
+                self.block_to_lora_layers[block_name] = {}
+            self.block_to_lora_layers[block_name][layer_name] = layer
+            self.lora_layers_to_block[layer_name] = block_name
+
+    def all_lora_layers(
+        self, ) -> Generator[tuple[str, BaseLayerWithLoRA], Any, None]:
+        for block_layers in self.block_to_lora_layers.values():
+            for name, layer in block_layers.items():
+                yield name, layer
+        for name, layer in self.other_lora_layers.items():
+            yield name, layer
+
+    def lora_layers_by_block(
+        self,
+    ) -> Generator[
+            tuple[nn.Module | None, dict[str, BaseLayerWithLoRA]],
+            Any,
+            None,
+    ]:
+        for block_name, layers in self.block_to_lora_layers.items():
+            yield self.block_mapping[block_name], layers
+        yield None, self.other_lora_layers
+
+
 class LoRAPipeline(ComposedPipelineBase):
     """
     Pipeline that supports injecting LoRA adapters into the diffusion transformer.
     TODO: support training.
     """
+
     lora_adapters: dict[str, dict[str, torch.Tensor]] = defaultdict(
         dict
     )  # state dicts of loaded lora adapters (includes lora_A, lora_B, and lora_alpha)
     cur_adapter_name: str = ""
     cur_adapter_path: str = ""
-    lora_layers: dict[str, dict[str, BaseLayerWithLoRA]] = {}
+    # model_name -> layers
+    lora_layers: dict[str, LoRAModelLayers] = {}
     fastvideo_args: FastVideoArgs | TrainingArgs
     exclude_lora_layers: dict[str, list[str]] = {}
     device: torch.device = get_local_torch_device()
@@ -48,10 +127,10 @@ class LoRAPipeline(ComposedPipelineBase):
         self.device = get_local_torch_device()
         # build list of trainable transformers
         for transformer_name in self.trainable_transformer_names:
-            if transformer_name in self.modules and self.modules[
-                    transformer_name] is not None:
-                self.trainable_transformer_modules[
-                    transformer_name] = self.modules[transformer_name]
+            if (transformer_name in self.modules
+                    and self.modules[transformer_name] is not None):
+                self.trainable_transformer_modules[transformer_name] = (
+                    self.modules[transformer_name])
             # check for transformer_2 in case of Wan2.2 MoE or fake_score_transformer_2
             if transformer_name.endswith("_2"):
                 raise ValueError(
@@ -59,19 +138,23 @@ class LoRAPipeline(ComposedPipelineBase):
                 )
 
             secondary_transformer_name = transformer_name + "_2"
-            if secondary_transformer_name in self.modules and self.modules[
-                    secondary_transformer_name] is not None:
+            if (secondary_transformer_name in self.modules
+                    and self.modules[secondary_transformer_name] is not None):
                 self.trainable_transformer_modules[
                     secondary_transformer_name] = self.modules[
                         secondary_transformer_name]
 
-        logger.info("trainable_transformer_modules: %s",
-                    self.trainable_transformer_modules.keys())
+        logger.info(
+            "trainable_transformer_modules: %s",
+            self.trainable_transformer_modules.keys(),
+        )
 
-        for transformer_name, transformer_module in self.trainable_transformer_modules.items(
-        ):
-            self.exclude_lora_layers[
-                transformer_name] = transformer_module.config.arch_config.exclude_lora_layers
+        for (
+                transformer_name,
+                transformer_module,
+        ) in self.trainable_transformer_modules.items():
+            self.exclude_lora_layers[transformer_name] = (
+                transformer_module.config.arch_config.exclude_lora_layers)
         self.lora_target_modules = self.fastvideo_args.lora_target_modules
         self.lora_path = self.fastvideo_args.lora_path
         self.lora_nickname = self.fastvideo_args.lora_nickname
@@ -83,20 +166,33 @@ class LoRAPipeline(ComposedPipelineBase):
                 self.fastvideo_args.lora_alpha = self.fastvideo_args.lora_rank
             self.lora_rank = self.fastvideo_args.lora_rank  # type: ignore
             self.lora_alpha = self.fastvideo_args.lora_alpha  # type: ignore
-            logger.info("Using LoRA training with rank %d and alpha %d",
-                        self.lora_rank, self.lora_alpha)
+            logger.info(
+                "Using LoRA training with rank %d and alpha %d",
+                self.lora_rank,
+                self.lora_alpha,
+            )
             if self.lora_target_modules is None:
                 self.lora_target_modules = [
-                    "q_proj", "k_proj", "v_proj", "o_proj", "to_q", "to_k",
-                    "to_v", "to_out", "to_qkv", "to_gate_compress"
+                    "q_proj",
+                    "k_proj",
+                    "v_proj",
+                    "o_proj",
+                    "to_q",
+                    "to_k",
+                    "to_v",
+                    "to_out",
+                    "to_qkv",
+                    "to_gate_compress",
                 ]
                 logger.info(
                     "Using default lora_target_modules for all transformers: %s",
-                    self.lora_target_modules)
+                    self.lora_target_modules,
+                )
             else:
                 logger.warning(
                     "Using custom lora_target_modules for all transformers, which may not be intended: %s",
-                    self.lora_target_modules)
+                    self.lora_target_modules,
+                )
 
             self.convert_to_lora_layers()
         # Inference
@@ -104,7 +200,8 @@ class LoRAPipeline(ComposedPipelineBase):
             self.convert_to_lora_layers()
             self.set_lora_adapter(
                 self.lora_nickname,  # type: ignore
-                self.lora_path)  # type: ignore
+                self.lora_path,
+            )  # type: ignore
 
     def is_target_layer(self, module_name: str) -> bool:
         if self.lora_target_modules is None:
@@ -114,9 +211,9 @@ class LoRAPipeline(ComposedPipelineBase):
 
     def set_trainable(self) -> None:
 
-        def set_lora_grads(lora_layers: dict[str, BaseLayerWithLoRA],
+        def set_lora_grads(lora_layers: LoRAModelLayers,
                            device_mesh: DeviceMesh):
-            for name, layer in lora_layers.items():
+            for name, layer in lora_layers.all_lora_layers():
                 layer.lora_A.requires_grad_(True)
                 layer.lora_B.requires_grad_(True)
                 layer.base_layer.requires_grad_(False)
@@ -131,10 +228,15 @@ class LoRAPipeline(ComposedPipelineBase):
             super().set_trainable()
             return
 
-        device_mesh = init_device_mesh("cuda", (dist.get_world_size(), 1),
-                                       mesh_dim_names=["fake", "replicate"])
-        for transformer_name, transformer_module in self.trainable_transformer_modules.items(
-        ):
+        device_mesh = init_device_mesh(
+            "cuda",
+            (dist.get_world_size(), 1),
+            mesh_dim_names=["fake", "replicate"],
+        )
+        for (
+                transformer_name,
+                transformer_module,
+        ) in self.trainable_transformer_modules.items():
             transformer_module.train()
             transformer_module.requires_grad_(False)
             if transformer_name in self.lora_layers:
@@ -151,32 +253,71 @@ class LoRAPipeline(ComposedPipelineBase):
         if self.lora_initialized:
             return
         self.lora_initialized = True
-        for transformer_name, transformer_module in self.trainable_transformer_modules.items(
-        ):
+        for (
+                transformer_name,
+                transformer_module,
+        ) in self.trainable_transformer_modules.items():
             converted_count = 0
+            # init bookkeeping structures
             if transformer_name not in self.lora_layers:
-                self.lora_layers[transformer_name] = {}
-            logger.info("Converting %s to LoRA Transformer", transformer_name)
-            for name, layer in transformer_module.named_modules():
-                if not self.is_target_layer(name):
-                    continue
-
-                excluded = False
-                for exclude_layer in self.exclude_lora_layers[transformer_name]:
-                    if exclude_layer in name:
-                        excluded = True
+                # get block list
+                block_list = []
+                for name, submodule in transformer_module.named_children():
+                    if isinstance(submodule, nn.ModuleList):
+                        block_list = [(f"{name}.{i}", m)
+                                      for i, m in enumerate(submodule)]
                         break
-                if excluded:
-                    continue
+                self.lora_layers[transformer_name] = LoRAModelLayers(block_list)
+            logger.info("Converting %s to LoRA Transformer", transformer_name)
+            # scan every module and convert to LoRA layer if applicable
 
-                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[transformer_name][name] = layer
-                    replace_submodule(transformer_module, name, layer)
-                    converted_count += 1
+            for block_name, block_modules in _named_module_by_prefix(
+                    transformer_module,
+                    list(self.lora_layers[transformer_name].block_mapping),
+            ):
+                if block_name is not None and (
+                        not self.fastvideo_args.training_mode
+                        and self.fastvideo_args.dit_layerwise_offload):
+                    scope_ctx = _get_hook_ctx(
+                        self.lora_layers[transformer_name].
+                        block_mapping[block_name])
+                else:
+                    scope_ctx = nullcontext()
+                with scope_ctx:
+                    for name, layer in block_modules:
+                        if not self.is_target_layer(name):
+                            continue
+
+                        excluded = False
+                        for exclude_layer in self.exclude_lora_layers[
+                                transformer_name]:
+                            if exclude_layer in name:
+                                excluded = True
+                                break
+                        if excluded:
+                            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:
+                            block_name_split = name.split(".", 2)
+                            if len(block_name_split) > 2:
+                                block_name = (block_name_split[0] + "." +
+                                              block_name_split[1])
+                            else:
+                                block_name = None
+                            if (block_name
+                                    not in self.lora_layers[transformer_name].
+                                    block_mapping):
+                                block_name = None
+                            self.lora_layers[transformer_name].add_lora_layer(
+                                block_name, name, layer)
+                            replace_submodule(transformer_module, name, layer)
+                            converted_count += 1
             logger.info("Converted %d layers to LoRA layers", converted_count)
 
     def set_lora_adapter(self,
@@ -209,7 +350,7 @@ class LoRAPipeline(ComposedPipelineBase):
 
             # Extract alpha values and weights in a single pass
             to_merge_params: defaultdict[Hashable,
-                                         dict[Any, Any]] = defaultdict(dict)
+                                         dict[Any, Any]] = (defaultdict(dict))
             for name, weight in lora_state_dict.items():
                 # Extract weights (lora_A, lora_B, and lora_alpha)
                 name = name.replace("diffusion_model.", "")
@@ -223,13 +364,14 @@ class LoRAPipeline(ComposedPipelineBase):
                     target_name, _, _ = param_names_mapping_fn(layer_name)
                     # Store alpha alongside weights with same target_name base
                     alpha_key = target_name + ".lora_alpha"
-                    self.lora_adapters[lora_nickname][alpha_key] = weight.item(
-                    ) if weight.numel() == 1 else float(weight.mean())
+                    self.lora_adapters[lora_nickname][alpha_key] = (
+                        weight.item()
+                        if weight.numel() == 1 else float(weight.mean()))
                     continue
 
                 name, _, _ = lora_param_names_mapping_fn(name)
-                target_name, merge_index, num_params_to_merge = param_names_mapping_fn(
-                    name)
+                target_name, merge_index, num_params_to_merge = (
+                    param_names_mapping_fn(name))
                 # for (in_dim, r) @ (r, out_dim), we only merge (r, out_dim * n) where n is the number of linear layers to fuse
                 # see param mapping in HunyuanVideoArchConfig
                 if merge_index is not None and "lora_B" in name:
@@ -261,45 +403,84 @@ class LoRAPipeline(ComposedPipelineBase):
 
         # Merge the new adapter
         adapted_count = 0
-        for transformer_name, transformer_lora_layers in self.lora_layers.items(
-        ):
-            for name, layer in transformer_lora_layers.items():
-                lora_A_name = name + ".lora_A"
-                lora_B_name = name + ".lora_B"
-                lora_alpha_name = name + ".lora_alpha"
-                if lora_A_name in self.lora_adapters[lora_nickname]\
-                    and lora_B_name in self.lora_adapters[lora_nickname]:
-                    # Get alpha value for this layer (defaults to None if not present)
-                    lora_A = self.lora_adapters[lora_nickname][lora_A_name]
-                    lora_B = self.lora_adapters[lora_nickname][lora_B_name]
-                    # Simple lookup - alpha stored with same naming scheme as lora_A/lora_B
-                    alpha = self.lora_adapters[lora_nickname].get(
-                        lora_alpha_name) if adapter_updated else None
-
-                    layer.set_lora_weights(
-                        lora_A,
-                        lora_B,
-                        lora_alpha=alpha,
-                        training_mode=self.fastvideo_args.training_mode,
-                        lora_path=lora_path)
-                    adapted_count += 1
-                else:
-                    if rank == 0:
-                        logger.warning(
-                            "LoRA adapter %s does not contain the weights for layer %s. LoRA will not be applied to it.",
-                            lora_path, name)
-                    layer.disable_lora = True
-        logger.info("Rank %d: LoRA adapter %s applied to %d layers", rank,
-                    lora_path, adapted_count)
+        for (
+                transformer_name,
+                transformer_lora_layers,
+        ) in self.lora_layers.items():
+            for (
+                    module,
+                    layers,
+            ) in transformer_lora_layers.lora_layers_by_block():
+                with _get_hook_ctx(module):
+                    for name, layer in layers.items():
+                        lora_A_name = name + ".lora_A"
+                        lora_B_name = name + ".lora_B"
+                        lora_alpha_name = name + ".lora_alpha"
+                        if (lora_A_name in self.lora_adapters[lora_nickname]
+                                and lora_B_name
+                                in self.lora_adapters[lora_nickname]):
+                            # Get alpha value for this layer (defaults to None if not present)
+                            lora_A = self.lora_adapters[lora_nickname][
+                                lora_A_name]
+                            lora_B = self.lora_adapters[lora_nickname][
+                                lora_B_name]
+                            # Simple lookup - alpha stored with same naming scheme as lora_A/lora_B
+                            alpha = (self.lora_adapters[lora_nickname].get(
+                                lora_alpha_name) if adapter_updated else None)
+                            try:
+                                layer.set_lora_weights(
+                                    lora_A,
+                                    lora_B,
+                                    lora_alpha=alpha,
+                                    training_mode=self.fastvideo_args.
+                                    training_mode,
+                                    lora_path=lora_path,
+                                )
+                            except Exception as e:
+                                logger.error(
+                                    "Error setting LoRA weights for layer %s: %s",
+                                    name,
+                                    str(e),
+                                )
+                                raise e
+                            adapted_count += 1
+                        else:
+                            if rank == 0:
+                                logger.warning(
+                                    "LoRA adapter %s does not contain the weights for layer %s. LoRA will not be applied to it.",
+                                    lora_path,
+                                    name,
+                                )
+                            layer.disable_lora = True
+        logger.info(
+            "Rank %d: LoRA adapter %s applied to %d layers",
+            rank,
+            lora_path,
+            adapted_count,
+        )
 
     def merge_lora_weights(self) -> None:
-        for transformer_name, transformer_lora_layers in self.lora_layers.items(
-        ):
-            for name, layer in transformer_lora_layers.items():
-                layer.merge_lora_weights()
+        for (
+                transformer_name,
+                transformer_lora_layers,
+        ) in self.lora_layers.items():
+            for (
+                    module,
+                    layers,
+            ) in transformer_lora_layers.lora_layers_by_block():
+                with _get_hook_ctx(module):
+                    for name, layer in layers.items():
+                        layer.merge_lora_weights()
 
     def unmerge_lora_weights(self) -> None:
-        for transformer_name, transformer_lora_layers in self.lora_layers.items(
-        ):
-            for name, layer in transformer_lora_layers.items():
-                layer.unmerge_lora_weights()
+        for (
+                transformer_name,
+                transformer_lora_layers,
+        ) in self.lora_layers.items():
+            for (
+                    module,
+                    layers,
+            ) in transformer_lora_layers.lora_layers_by_block():
+                with _get_hook_ctx(module):
+                    for name, layer in layers.items():
+                        layer.unmerge_lora_weights()

fastvideo/pipelines/pipeline_batch_info.py

@@ -67,6 +67,21 @@ class ForwardBatch:
     execution, allowing methods to update specific components without needing
     to manage numerous individual parameters.
     """
+
+    @dataclass
+    class RLData:
+        """RL-specific data collection options and outputs."""
+        enabled: bool = False
+        collect_log_probs: bool = True
+        collect_kl: bool = False
+        kl_reward: float = 0.0
+        store_trajectory: bool = True
+        keep_trajectory_on_cpu: bool = False
+        log_probs: torch.Tensor | None = None
+        kl: torch.Tensor | None = None
+        trajectory_latents: torch.Tensor | None = None
+        trajectory_timesteps: torch.Tensor | None = None
+
     # TODO(will): double check that args are separate from fastvideo_args
     # properly. Also maybe think about providing an abstraction for pipeline
     # specific arguments.
@@ -197,6 +212,9 @@ class ForwardBatch:
     logging_info: PipelineLoggingInfo = field(
         default_factory=PipelineLoggingInfo)
 
+    # RL data collection
+    rl_data: "ForwardBatch.RLData" = field(default_factory=RLData)
+
     def __post_init__(self):
         """Initialize dependent fields after dataclass initialization."""
 
@@ -267,6 +285,36 @@ class TrainingBatch:
     latent_vis_dict: dict[str, Any] = field(default_factory=dict)
     fake_score_latent_vis_dict: dict[str, Any] = field(default_factory=dict)
 
+    # RL/GRPO-specific attributes
+    reward_scores: torch.Tensor | None = None  # Computed rewards from reward models
+    log_probs: torch.Tensor | None = None  # Current policy log probabilities [B, num_steps] or [B]
+    old_log_probs: torch.Tensor | None = None  # Old policy log probs (for importance ratio) [B, num_steps] or [B]
+    advantages: torch.Tensor | None = None  # GAE advantages [B, num_steps] or [B]
+    returns: torch.Tensor | None = None  # TD returns (advantages + values) [B, num_steps] or [B]
+    values: torch.Tensor | None = None  # Value function predictions [B]
+    old_values: torch.Tensor | None = None  # Old value predictions (for clipping) [B]
+
+    # GRPO sampling-specific attributes
+    kl: torch.Tensor | None = None  # KL divergences from sampling [B, num_steps] (if kl_reward > 0)
+    prompt_ids: torch.Tensor | None = None  # Prompt token IDs for stat tracking [B, seq_len]
+    prompt_embeds: torch.Tensor | None = None  # Prompt embeddings used in sampling [B, seq_len, hidden_dim]
+    negative_prompt_embeds: torch.Tensor | None = None  # Negative prompt embeddings for CFG [B, seq_len, hidden_dim]
+
+    # RL loss components
+    policy_loss: float = 0.0  # GRPO/PPO policy loss
+    value_loss: float = 0.0  # Value function loss
+    kl_divergence: float = 0.0  # KL(new_policy || old_policy)
+    importance_ratio: float = 1.0  # exp(log_prob - old_log_prob)
+    clip_fraction: float = 0.0  # Fraction of ratios that were clipped
+
+    # RL metrics
+    advantage_mean: float = 0.0  # Mean advantage (should be ~0 after normalization)
+    advantage_std: float = 1.0  # Std of advantages
+    reward_mean: float = 0.0  # Mean reward across batch
+    reward_std: float = 0.0  # Std of rewards
+    value_mean: float = 0.0  # Mean value prediction
+    entropy: float = 0.0  # Policy entropy (for exploration)
+
 
 @dataclass
 class PreprocessBatch(ForwardBatch):

fastvideo/pipelines/pipeline_registry.py

@@ -29,6 +29,7 @@ _PIPELINE_NAME_TO_ARCHITECTURE_NAME: dict[str, str] = {
     "HunyuanVideoPipeline": "hunyuan",
     "HunyuanVideo15Pipeline": "hunyuan15",
     "Cosmos2VideoToWorldPipeline": "cosmos",
+    "Cosmos2_5Pipeline": "cosmos",
     "MatrixGamePipeline": "matrixgame",
     "MatrixGameCausalDMDPipeline": "matrixgame",
     "LongCatPipeline": "longcat",

fastvideo/pipelines/stages/__init__.py

@@ -10,7 +10,8 @@ from fastvideo.pipelines.stages.base import PipelineStage
 from fastvideo.pipelines.stages.causal_denoising import CausalDMDDenosingStage
 from fastvideo.pipelines.stages.conditioning import ConditioningStage
 from fastvideo.pipelines.stages.decoding import DecodingStage
-from fastvideo.pipelines.stages.denoising import (CosmosDenoisingStage,
+from fastvideo.pipelines.stages.denoising import (Cosmos25DenoisingStage,
+                                                  CosmosDenoisingStage,
                                                   DenoisingStage,
                                                   DmdDenoisingStage)
 from fastvideo.pipelines.stages.encoding import EncodingStage
@@ -19,14 +20,16 @@ from fastvideo.pipelines.stages.image_encoding import (
     ImageVAEEncodingStage, VideoVAEEncodingStage, Hy15ImageEncodingStage)
 from fastvideo.pipelines.stages.input_validation import InputValidationStage
 from fastvideo.pipelines.stages.latent_preparation import (
-    CosmosLatentPreparationStage, LatentPreparationStage)
+    Cosmos25LatentPreparationStage, CosmosLatentPreparationStage,
+    LatentPreparationStage)
 from fastvideo.pipelines.stages.matrixgame_denoising import (
     MatrixGameCausalDenoisingStage)
 from fastvideo.pipelines.stages.stepvideo_encoding import (
     StepvideoPromptEncodingStage)
-from fastvideo.pipelines.stages.text_encoding import TextEncodingStage
+from fastvideo.pipelines.stages.text_encoding import (Cosmos25TextEncodingStage,
+                                                      TextEncodingStage)
 from fastvideo.pipelines.stages.timestep_preparation import (
-    TimestepPreparationStage)
+    Cosmos25TimestepPreparationStage, TimestepPreparationStage)
 
 # LongCat stages
 from fastvideo.pipelines.stages.longcat_video_vae_encoding import LongCatVideoVAEEncodingStage
@@ -37,14 +40,17 @@ __all__ = [
     "PipelineStage",
     "InputValidationStage",
     "TimestepPreparationStage",
+    "Cosmos25TimestepPreparationStage",
     "LatentPreparationStage",
     "CosmosLatentPreparationStage",
+    "Cosmos25LatentPreparationStage",
     "ConditioningStage",
     "DenoisingStage",
     "DmdDenoisingStage",
     "CausalDMDDenosingStage",
     "MatrixGameCausalDenoisingStage",
     "CosmosDenoisingStage",
+    "Cosmos25DenoisingStage",
     "EncodingStage",
     "DecodingStage",
     "ImageEncodingStage",
@@ -54,6 +60,7 @@ __all__ = [
     "ImageVAEEncodingStage",
     "VideoVAEEncodingStage",
     "TextEncodingStage",
+    "Cosmos25TextEncodingStage",
     "StepvideoPromptEncodingStage",
     # LongCat stages
     "LongCatVideoVAEEncodingStage",

fastvideo/pipelines/stages/decoding.py

@@ -51,39 +51,42 @@ class DecodingStage(PipelineStage):
         return result
 
     def _denormalize_latents(self, latents: torch.Tensor) -> torch.Tensor:
-        # denormalization for MatrixGame VAE
-        # z = z * std + mean during decode
-        if (hasattr(self.vae.config, 'latents_mean')
-                and hasattr(self.vae.config, 'latents_std')):
-            # Convert config values to tensors
-            latents_mean = torch.tensor(self.vae.config.latents_mean,
+        """Convert normalized latents into the VAE's expected latent space."""
+        # Some VAEs handle latent (de)normalization internally.
+        if bool(getattr(self.vae, "handles_latent_denorm", False)):
+            return latents
+
+        cfg = getattr(self.vae, "config", None)
+
+        # MatrixGame-style: z = z * std + mean
+        if (cfg is not None and hasattr(cfg, "latents_mean")
+                and hasattr(cfg, "latents_std")):
+            latents_mean = torch.tensor(cfg.latents_mean,
                                         device=latents.device,
                                         dtype=latents.dtype).view(
                                             1, -1, 1, 1, 1)
-
-            latents_std = torch.tensor(self.vae.config.latents_std,
+            latents_std = torch.tensor(cfg.latents_std,
                                        device=latents.device,
                                        dtype=latents.dtype).view(
                                            1, -1, 1, 1, 1)
+            return latents * latents_std + latents_mean
 
-            # Apply denormalization: z = z * std + mean
-            latents = latents * latents_std + latents_mean
-        elif hasattr(self.vae, 'scaling_factor'):
-            # Standard VAE scaling
+        # Diffusers-style: scaling_factor (+ optional shift_factor)
+        if hasattr(self.vae, "scaling_factor"):
             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
 
-            # Apply shifting if needed
-            if (hasattr(self.vae, "shift_factor")
-                    and self.vae.shift_factor is not None):
+            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)
+                    latents = latents + self.vae.shift_factor.to(
+                        latents.device, latents.dtype)
                 else:
-                    latents += self.vae.shift_factor
+                    latents = latents + self.vae.shift_factor
+
         return latents
 
     @torch.no_grad()
@@ -273,4 +276,4 @@ class DecodingStage(PipelineStage):
                 del pipeline.modules["vae"]
             fastvideo_args.model_loaded["vae"] = False
 
-        return batch
+        return batch

fastvideo/pipelines/stages/denoising.py

@@ -4,11 +4,14 @@ Denoising stage for diffusion pipelines.
 """
 
 import inspect
+import math
 import weakref
 from collections.abc import Iterable
+from contextlib import nullcontext
 from typing import Any
 
 import torch
+from diffusers.utils.torch_utils import randn_tensor
 from tqdm.auto import tqdm
 
 from fastvideo.attention import get_attn_backend
@@ -52,6 +55,84 @@ except ImportError:
 logger = init_logger(__name__)
 
 
+def sde_step_with_logprob(
+    scheduler,
+    model_output: torch.FloatTensor,
+    timestep: float | torch.FloatTensor,
+    sample: torch.FloatTensor,
+    prev_sample: torch.FloatTensor | None = None,
+    generator: torch.Generator | None = None,
+    deterministic: bool = False,
+    return_pixel_log_prob: bool = False,
+    return_dt_and_std_dev_t: bool = False
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, ...]:
+    """
+    Predict the sample from the previous timestep by reversing the SDE and
+    compute log probabilities for the transition.
+    """
+    if isinstance(timestep, torch.Tensor):
+        if timestep.ndim == 0:
+            timestep = timestep.unsqueeze(0)
+        step_indices = [
+            scheduler.index_for_timestep(t.item()) for t in timestep
+        ]
+    else:
+        step_indices = [scheduler.index_for_timestep(timestep)]
+
+    prev_step_indices = [step + 1 for step in step_indices]
+
+    sigmas = scheduler.sigmas.to(sample.device, sample.dtype)
+    sigma = sigmas[step_indices].view(-1, 1, 1, 1, 1)
+    sigma_prev = sigmas[prev_step_indices].view(-1, 1, 1, 1, 1)
+    sigma_max = sigmas[0].item()
+    sigma_min = sigmas[-1].item()
+
+    dt = sigma_prev - sigma
+
+    std_dev_t = sigma_min + (sigma_max - sigma_min) * sigma
+    prev_sample_mean = (sample * (1 + std_dev_t**2 / (2 * sigma) * dt) +
+                        model_output * (1 + std_dev_t**2 * (1 - sigma) /
+                                        (2 * sigma)) * dt)
+
+    if prev_sample is not None and generator is not None:
+        raise ValueError(
+            "Cannot pass both generator and prev_sample. Please make sure that either `generator` or"
+            " `prev_sample` stays `None`.")
+
+    if prev_sample is None:
+        variance_noise = randn_tensor(
+            model_output.shape,
+            generator=generator,
+            device=model_output.device,
+            dtype=model_output.dtype,
+        )
+        sqrt_dt = torch.sqrt(-1 * dt)
+        prev_sample = prev_sample_mean + std_dev_t * sqrt_dt * variance_noise
+    else:
+        sqrt_dt = torch.sqrt(-1 * dt)
+
+    if deterministic:
+        prev_sample = sample + dt * model_output
+        sqrt_dt = torch.sqrt(-1 * dt)
+
+    if return_pixel_log_prob:
+        raise NotImplementedError(
+            "Pixel-level log prob is not supported in this helper.")
+
+    std_dev_sqrt_dt = std_dev_t * sqrt_dt
+    log_prob = (
+        -((prev_sample.detach() - prev_sample_mean)**2) /
+        (2 *
+         (std_dev_sqrt_dt**2)) - torch.log(std_dev_sqrt_dt + 1e-8) - torch.log(
+             torch.sqrt(2 * torch.as_tensor(math.pi, device=sample.device))))
+
+    log_prob = log_prob.mean(dim=tuple(range(1, log_prob.ndim)))
+
+    if return_dt_and_std_dev_t:
+        return prev_sample, log_prob, prev_sample_mean, std_dev_t, sqrt_dt
+    return prev_sample, log_prob, prev_sample_mean, std_dev_t * sqrt_dt
+
+
 class DenoisingStage(PipelineStage):
     """
     Stage for running the denoising loop in diffusion pipelines.
@@ -203,9 +284,11 @@ class DenoisingStage(PipelineStage):
         else:
             boundary_timestep = None
         latent_model_input = latents.to(target_dtype)
-        assert latent_model_input.shape[0] == 1, "only support batch size 1"
+        rl_data = batch.rl_data if batch.rl_data and batch.rl_data.enabled else None
 
         if fastvideo_args.pipeline_config.ti2v_task and batch.pil_image is not None:
+            assert latent_model_input.shape[
+                0] == 1, "TI2V task only supports batch size 1"
             # TI2V directly replaces the first frame of the latent with
             # the image latent instead of appending along the channel dim
             assert batch.image_latent is None, "TI2V task should not have image latents"
@@ -243,6 +326,12 @@ class DenoisingStage(PipelineStage):
         # Initialize lists for ODE trajectory
         trajectory_timesteps: list[torch.Tensor] = []
         trajectory_latents: list[torch.Tensor] = []
+        rl_timesteps: list[torch.Tensor] = []
+        rl_latents: list[torch.Tensor] = []
+        rl_log_probs: list[torch.Tensor] = []
+        rl_kl: list[torch.Tensor] = []
+        if rl_data is not None and rl_data.store_trajectory:
+            rl_latents.append(latents)
 
         # Run denoising loop
         with self.progress_bar(total=num_inference_steps) as progress_bar:
@@ -329,6 +418,24 @@ class DenoisingStage(PipelineStage):
                     1000.0 if fastvideo_args.pipeline_config.embedded_cfg_scale
                     is not None else None)
 
+                def run_transformer(model, encoder_hidden_states, cond_kwargs,
+                                    is_cfg_negative: bool):
+                    batch.is_cfg_negative = is_cfg_negative
+                    with set_forward_context(
+                            current_timestep=i,
+                            attn_metadata=attn_metadata,
+                            forward_batch=batch,
+                    ):
+                        return model(
+                            latent_model_input,
+                            encoder_hidden_states,
+                            t_expand,
+                            guidance=guidance_expand,
+                            **image_kwargs,
+                            **cond_kwargs,
+                            **action_kwargs,
+                        )
+
                 # Predict noise residual
                 with torch.autocast(device_type="cuda",
                                     dtype=target_dtype,
@@ -390,40 +497,13 @@ class DenoisingStage(PipelineStage):
                     # support torch dynamo compilation. They pass in
                     # attn_metadata, vllm_config, and num_tokens. We can pass in
                     # fastvideo_args or training_args, and attn_metadata.
-                    batch.is_cfg_negative = False
-                    with set_forward_context(
-                            current_timestep=i,
-                            attn_metadata=attn_metadata,
-                            forward_batch=batch,
-                            # fastvideo_args=fastvideo_args
-                    ):
-                        # Run transformer
-                        noise_pred = current_model(
-                            latent_model_input,
-                            prompt_embeds,
-                            t_expand,
-                            guidance=guidance_expand,
-                            **image_kwargs,
-                            **pos_cond_kwargs,
-                            **action_kwargs,
-                        )
+                    noise_pred = run_transformer(current_model, prompt_embeds,
+                                                 pos_cond_kwargs, False)
 
                     if batch.do_classifier_free_guidance:
-                        batch.is_cfg_negative = True
-                        with set_forward_context(
-                                current_timestep=i,
-                                attn_metadata=attn_metadata,
-                                forward_batch=batch,
-                        ):
-                            noise_pred_uncond = current_model(
-                                latent_model_input,
-                                neg_prompt_embeds,
-                                t_expand,
-                                guidance=guidance_expand,
-                                **image_kwargs,
-                                **neg_cond_kwargs,
-                                **action_kwargs,
-                            )
+                        noise_pred_uncond = run_transformer(
+                            current_model, neg_prompt_embeds, neg_cond_kwargs,
+                            True)
 
                         noise_pred_text = noise_pred
                         noise_pred = noise_pred_uncond + current_guidance_scale * (
@@ -438,11 +518,58 @@ class DenoisingStage(PipelineStage):
                                 guidance_rescale=batch.guidance_rescale,
                             )
                     # Compute the previous noisy sample
+                    prev_latents = latents
                     latents = self.scheduler.step(noise_pred,
                                                   t,
                                                   latents,
                                                   **extra_step_kwargs,
                                                   return_dict=False)[0]
+                    if rl_data is not None:
+                        if rl_data.collect_log_probs:
+                            _, log_prob, prev_latents_mean, std_dev_t, _ = sde_step_with_logprob(
+                                self.scheduler,
+                                noise_pred.float(),
+                                t,
+                                prev_latents.float(),
+                                prev_sample=latents.float(),
+                                deterministic=False,
+                                return_dt_and_std_dev_t=True,
+                            )
+                            rl_log_probs.append(log_prob)
+
+                        if rl_data.collect_kl and rl_data.kl_reward > 0:
+                            adapter_ctx = nullcontext()
+                            if hasattr(current_model, "disable_adapter"):
+                                adapter_ctx = current_model.disable_adapter()
+                            with adapter_ctx:
+                                noise_pred_ref = run_transformer(
+                                    current_model, prompt_embeds,
+                                    pos_cond_kwargs, False)
+                                if batch.do_classifier_free_guidance:
+                                    noise_pred_uncond_ref = run_transformer(
+                                        current_model, neg_prompt_embeds,
+                                        neg_cond_kwargs, True)
+                                    noise_pred_text_ref = noise_pred_ref
+                                    noise_pred_ref = noise_pred_uncond_ref + current_guidance_scale * (
+                                        noise_pred_text_ref -
+                                        noise_pred_uncond_ref)
+                            _, _, prev_latents_mean_ref, std_dev_t_ref, _ = sde_step_with_logprob(
+                                self.scheduler,
+                                noise_pred_ref.float(),
+                                t,
+                                prev_latents.float(),
+                                prev_sample=latents.float(),
+                                deterministic=False,
+                                return_dt_and_std_dev_t=True,
+                            )
+                            if not torch.allclose(std_dev_t, std_dev_t_ref):
+                                logger.warning(
+                                    "std_dev_t mismatch in RL KL computation at step %s",
+                                    i)
+                            kl = (prev_latents_mean -
+                                  prev_latents_mean_ref)**2 / (2 * std_dev_t**2)
+                            kl = kl.mean(dim=tuple(range(1, kl.ndim)))
+                            rl_kl.append(kl)
                     if fastvideo_args.pipeline_config.ti2v_task and batch.pil_image is not None:
                         latents = latents.squeeze(0)
                         latents = (1. - mask2[0]) * z + mask2[0] * latents
@@ -452,6 +579,15 @@ class DenoisingStage(PipelineStage):
                 if batch.return_trajectory_latents:
                     trajectory_timesteps.append(t)
                     trajectory_latents.append(latents)
+                if rl_data is not None:
+                    rl_timesteps.append(t)
+                    if rl_data.store_trajectory:
+                        rl_latents.append(latents)
+                    if rl_data.collect_kl and rl_data.kl_reward <= 0:
+                        rl_kl.append(
+                            torch.zeros(latents.shape[0],
+                                        device=latents.device,
+                                        dtype=latents.dtype))
 
                 # Update progress bar
                 if i == len(timesteps) - 1 or (
@@ -472,6 +608,25 @@ class DenoisingStage(PipelineStage):
         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 rl_data is not None:
+            if rl_timesteps:
+                rl_data.trajectory_timesteps = torch.stack(rl_timesteps, dim=0)
+                if rl_data.keep_trajectory_on_cpu:
+                    rl_data.trajectory_timesteps = rl_data.trajectory_timesteps.cpu(
+                    )
+            if rl_data.store_trajectory and rl_latents:
+                rl_data.trajectory_latents = torch.stack(rl_latents, dim=1)
+                if rl_data.keep_trajectory_on_cpu:
+                    rl_data.trajectory_latents = rl_data.trajectory_latents.cpu(
+                    )
+            if rl_log_probs:
+                rl_data.log_probs = torch.stack(rl_log_probs, dim=1)
+                if rl_data.keep_trajectory_on_cpu:
+                    rl_data.log_probs = rl_data.log_probs.cpu()
+            if rl_kl:
+                rl_data.kl = torch.stack(rl_kl, dim=1)
+                if rl_data.keep_trajectory_on_cpu:
+                    rl_data.kl = rl_data.kl.cpu()
 
         # Update batch with final latents
         batch.latents = latents
@@ -1018,6 +1173,145 @@ class CosmosDenoisingStage(DenoisingStage):
         return result
 
 
+class Cosmos25DenoisingStage(CosmosDenoisingStage):
+    """Denoising stage for Cosmos 2.5 DiT (expects 1D/2D timestep, not 5D)."""
+
+    def forward(
+        self,
+        batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> ForwardBatch:
+        pipeline = self.pipeline() if self.pipeline else None
+        if not fastvideo_args.model_loaded["transformer"]:
+            loader = TransformerLoader()
+            self.transformer = loader.load(
+                fastvideo_args.model_paths["transformer"], fastvideo_args)
+            if pipeline:
+                pipeline.add_module("transformer", self.transformer)
+            fastvideo_args.model_loaded["transformer"] = True
+
+        extra_step_kwargs = self.prepare_extra_func_kwargs(
+            self.scheduler.step,
+            {
+                "generator": batch.generator,
+                "eta": batch.eta
+            },
+        )
+
+        if hasattr(self.transformer, 'module'):
+            transformer_dtype = next(self.transformer.module.parameters()).dtype
+        else:
+            transformer_dtype = next(self.transformer.parameters()).dtype
+        target_dtype = transformer_dtype
+        autocast_enabled = (target_dtype != torch.float32
+                            ) and not fastvideo_args.disable_autocast
+
+        latents = batch.latents
+        if latents is None:
+            raise ValueError(
+                "latents must be provided for Cosmos25DenoisingStage")
+        guidance_scale = batch.guidance_scale
+
+        # Use timesteps prepared by Cosmos25TimestepPreparationStage when available.
+        if batch.timesteps is None:
+            self.scheduler.set_timesteps(batch.num_inference_steps,
+                                         device=latents.device)
+            timesteps = self.scheduler.timesteps
+        else:
+            timesteps = batch.timesteps.to(latents.device)
+
+        # Match official behavior: pass fps as a tensor.
+        fps_val = batch.fps if isinstance(batch.fps, int | float) else 24
+        fps_tensor = torch.tensor([fps_val],
+                                  device=latents.device,
+                                  dtype=target_dtype)
+
+        # Cosmos2.5 denoises a 4D latent (C,T,H,W) and the scheduler.step expects (B,C,T,H,W).
+        latents_4d = latents[0]
+
+        # Masks from latent prep stage
+        condition_mask = batch.cond_mask.to(target_dtype) if hasattr(
+            batch, 'cond_mask') else None
+        padding_mask = batch.padding_mask.to(target_dtype) if hasattr(
+            batch, 'padding_mask') else None
+        if condition_mask is None:
+            _, t, h, w = latents_4d.shape
+            condition_mask = torch.zeros(1,
+                                         1,
+                                         t,
+                                         h,
+                                         w,
+                                         device=latents.device,
+                                         dtype=target_dtype)
+        if padding_mask is None:
+            _, _, h, w = latents_4d.shape
+            padding_mask = torch.ones(1,
+                                      1,
+                                      h,
+                                      w,
+                                      device=latents.device,
+                                      dtype=target_dtype)
+
+        # Cosmos2.5 timestep scaling (see compare_pipelines.py): t * 0.001
+        timestep_scale = 0.001
+
+        with self.progress_bar(total=len(timesteps)) as progress_bar:
+            for i, t in enumerate(timesteps):
+                t_val = float(t)
+                timestep_val = t_val * timestep_scale
+                timestep = torch.tensor([[timestep_val]],
+                                        device=latents.device,
+                                        dtype=target_dtype)
+
+                model_hidden_states = latents_4d.unsqueeze(0)
+
+                with (
+                        set_forward_context(current_timestep=int(t_val),
+                                            attn_metadata=None,
+                                            forward_batch=batch),
+                        torch.autocast(device_type="cuda",
+                                       dtype=target_dtype,
+                                       enabled=autocast_enabled),
+                ):
+                    cond_v = self.transformer(
+                        hidden_states=model_hidden_states.to(target_dtype),
+                        encoder_hidden_states=batch.prompt_embeds[0].to(
+                            target_dtype),
+                        timestep=timestep,
+                        fps=fps_tensor,
+                        condition_mask=condition_mask,
+                        padding_mask=padding_mask,
+                        return_dict=False,
+                    )[0]
+
+                    if batch.do_classifier_free_guidance and batch.negative_prompt_embeds:
+                        uncond_v = self.transformer(
+                            hidden_states=model_hidden_states.to(target_dtype),
+                            encoder_hidden_states=batch.
+                            negative_prompt_embeds[0].to(target_dtype),
+                            timestep=timestep,
+                            fps=fps_tensor,
+                            condition_mask=condition_mask,
+                            padding_mask=padding_mask,
+                            return_dict=False,
+                        )[0]
+                        v = uncond_v + guidance_scale * (cond_v - uncond_v)
+                    else:
+                        v = cond_v
+
+                prev = self.scheduler.step(v.unsqueeze(0),
+                                           t,
+                                           latents_4d.unsqueeze(0),
+                                           **extra_step_kwargs,
+                                           return_dict=False)[0]
+                latents_4d = prev.squeeze(0)
+
+                progress_bar.update()
+
+        batch.latents = latents_4d.unsqueeze(0)
+        return batch
+
+
 class DmdDenoisingStage(DenoisingStage):
     """
     Denoising stage for DMD.

fastvideo/pipelines/stages/latent_preparation.py

@@ -5,6 +5,7 @@ Latent preparation stage for diffusion pipelines.
 
 from typing import Any
 
+import numpy as np
 import torch
 from diffusers.utils.torch_utils import randn_tensor
 
@@ -427,6 +428,210 @@ class CosmosLatentPreparationStage(PipelineStage):
 
         return batch
 
+
+class Cosmos25LatentPreparationStage(CosmosLatentPreparationStage):
+    """Latent preparation for Cosmos 2.5 DiT input conventions."""
+
+    @staticmethod
+    def _arch_invariant_randn(
+        shape: tuple[int, ...],
+        *,
+        seed: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> torch.Tensor:
+        """Architecture-invariant RNG (matches cosmos_predict2.misc.arch_invariant_rand)."""
+        rng = np.random.RandomState(seed)
+        arr = rng.standard_normal(shape).astype(np.float32)
+        return torch.from_numpy(arr).to(device=device, dtype=dtype)
+
+    def forward(
+        self,
+        batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> ForwardBatch:
+        # Differences vs `CosmosLatentPreparationStage`: channel convention, seed usage,
+        # and `padding_mask` for concat_padding_mask=True.
+
+        # Determine batch size
+        if isinstance(batch.prompt, list):
+            batch_size = len(batch.prompt)
+        elif batch.prompt is not None:
+            batch_size = 1
+        else:
+            batch_size = batch.prompt_embeds[0].shape[0]
+
+        batch_size *= batch.num_videos_per_prompt
+
+        # Match `compare_pipelines.py`: initialize noise in fp32, then run the
+        # denoising computation in bf16.
+        dtype = torch.float32
+        device = get_local_torch_device()
+        generator = batch.generator
+        latents = batch.latents
+        num_frames = batch.num_frames
+        height = batch.height
+        width = batch.width
+
+        if height is None or width is None:
+            raise ValueError("Height and width must be provided")
+
+        vae_scale_factor_spatial = 8
+        vae_scale_factor_temporal = 4
+
+        latent_height = height // 8
+        latent_width = width // vae_scale_factor_spatial
+        num_latent_frames = (num_frames - 1) // vae_scale_factor_temporal + 1
+
+        # Cosmos 2.5 convention: transformer in_channels == latent channels
+        num_channels_latents = self.transformer.config.in_channels
+
+        shape = (batch_size, num_channels_latents, num_latent_frames,
+                 latent_height, latent_width)
+
+        init_latents = None
+        conditioning_latents = None
+        video = None
+
+        if hasattr(batch, 'video') and batch.video is not None:
+            video = batch.video
+        elif hasattr(batch, 'pil_image') and batch.pil_image is not None:
+            vae_scale_factor_spatial = 8
+            image_processor = ImageProcessor(
+                vae_scale_factor=vae_scale_factor_spatial)
+            processed_image = image_processor.preprocess(
+                batch.pil_image, height, width)
+            video = processed_image.unsqueeze(2)
+            video = video.to(device=device, dtype=torch.bfloat16)
+        elif hasattr(
+                batch,
+                'preprocessed_image') and batch.preprocessed_image is not None:
+            if isinstance(batch.preprocessed_image, torch.Tensor):
+                if batch.preprocessed_image.dim() == 4:
+                    video = batch.preprocessed_image.unsqueeze(2)
+                elif batch.preprocessed_image.dim() == 5:
+                    video = batch.preprocessed_image
+        else:
+            logger.info(
+                "CosmosLatentPreparationStage - No video input sources found")
+
+        if video is not None:
+            num_cond_frames = video.size(2)
+            if num_cond_frames >= num_frames:
+                num_cond_latent_frames = (num_frames -
+                                          1) // vae_scale_factor_temporal + 1
+                video = video[:, :, -num_frames:]
+            else:
+                num_cond_latent_frames = (num_cond_frames -
+                                          1) // vae_scale_factor_temporal + 1
+                num_padding_frames = num_frames - num_cond_frames
+                last_frame = video[:, :, -1:]
+                padding = last_frame.repeat(1, 1, num_padding_frames, 1, 1)
+                video = torch.cat([video, padding], dim=2)
+
+            if self.vae is not None:
+                self.vae = self.vae.to(device)
+                self.vae = self.vae.to(dtype=video.dtype)
+
+                def retrieve_latents(
+                        encoder_output: Any,
+                        generator: Any | None = None) -> torch.Tensor:
+                    if hasattr(encoder_output, "latent_dist"):
+                        return encoder_output.latent_dist.sample(generator)
+                    elif hasattr(encoder_output, "latents"):
+                        return encoder_output.latents
+                    elif hasattr(encoder_output, "sample"):
+                        return encoder_output.sample(generator)
+                    elif isinstance(encoder_output, torch.Tensor):
+                        return encoder_output
+                    else:
+                        attrs = [
+                            attr for attr in dir(encoder_output)
+                            if not attr.startswith('_')
+                        ]
+                        raise AttributeError(
+                            f"Could not access latents of provided encoder_output. Available attributes: {attrs}"
+                        )
+
+                if isinstance(generator, list):
+                    init_latents = [
+                        retrieve_latents(self.vae.encode(video[i].unsqueeze(0)),
+                                         generator=torch.Generator(
+                                             device="cpu").manual_seed(100))
+                        for i in range(batch_size)
+                    ]
+                else:
+                    init_latents = [
+                        retrieve_latents(
+                            self.vae.encode(vid.unsqueeze(0)),
+                            torch.Generator(device="cpu").manual_seed(100))
+                        for vid in video
+                    ]
+
+                init_latents = torch.cat(init_latents, dim=0).to(dtype)
+
+                cfg = getattr(self.vae, "config", None)
+                if (not bool(getattr(self.vae, "handles_latent_norm", False))
+                        and cfg is not None and hasattr(cfg, 'latents_mean')
+                        and hasattr(cfg, 'latents_std')):
+                    latents_mean = torch.tensor(cfg.latents_mean).view(
+                        1, cfg.z_dim, 1, 1, 1).to(device, dtype)
+                    latents_std = torch.tensor(cfg.latents_std).view(
+                        1, cfg.z_dim, 1, 1, 1).to(device, dtype)
+                    init_latents = (init_latents - latents_mean
+                                    ) / latents_std * self.scheduler.sigma_data
+
+                conditioning_latents = init_latents
+                self.vae.to("cpu")
+        else:
+            num_cond_latent_frames = 0
+
+        if latents is None:
+            seed = int(batch.seed if batch.seed is not None else 0)
+            # Use arch-invariant RNG to match Cosmos2.5 reference sampling.
+            latents_fp32 = self._arch_invariant_randn(shape,
+                                                      seed=seed,
+                                                      device=device,
+                                                      dtype=torch.float32)
+            latents = latents_fp32.to(torch.bfloat16)
+        else:
+            # If latents are supplied, keep compute dtype consistent with Cosmos sampling.
+            latents = latents.to(device=device, dtype=torch.bfloat16)
+
+        # Cosmos2.5 starts from unit Gaussian noise (no extra sigma_max scaling).
+
+        padding_shape = (batch_size, 1, num_latent_frames, latent_height,
+                         latent_width)
+        ones_padding = latents.new_ones(padding_shape)
+        zeros_padding = latents.new_zeros(padding_shape)
+
+        cond_indicator = latents.new_zeros(1, 1, latents.size(2), 1, 1)
+        cond_indicator[:, :, :num_cond_latent_frames] = 1.0
+        cond_mask = cond_indicator * ones_padding + (
+            1 - cond_indicator) * zeros_padding
+
+        uncond_indicator = None
+        uncond_mask = None
+        if batch.do_classifier_free_guidance:
+            uncond_indicator = latents.new_zeros(1, 1, latents.size(2), 1, 1)
+            uncond_indicator[:, :, :num_cond_latent_frames] = 1.0
+            uncond_mask = uncond_indicator * ones_padding + (
+                1 - uncond_indicator) * zeros_padding
+
+        # Cosmos 2.5 requires a spatial padding mask when concat_padding_mask=True
+        padding_mask = latents.new_ones(batch_size, 1, latent_height,
+                                        latent_width)
+
+        batch.latents = latents
+        batch.raw_latent_shape = latents.shape
+        batch.conditioning_latents = conditioning_latents
+        batch.cond_indicator = cond_indicator
+        batch.uncond_indicator = uncond_indicator
+        batch.cond_mask = cond_mask
+        batch.uncond_mask = uncond_mask
+        batch.padding_mask = padding_mask
+        return batch
+
     def adjust_video_length(self, batch: ForwardBatch,
                             fastvideo_args: FastVideoArgs) -> int:
         """

fastvideo/pipelines/stages/text_encoding.py

@@ -328,3 +328,69 @@ class TextEncodingStage(PipelineStage):
             lambda x: not batch.do_classifier_free_guidance or V.
             list_of_tensors_with_min_dims(x, 2))
         return result
+
+
+class Cosmos25TextEncodingStage(PipelineStage):
+    """Cosmos 2.5 text encoding stage.
+
+    Cosmos 2.5 uses Reason1 (Qwen2.5-VL) and relies on the encoder's
+    `compute_text_embeddings_online()`.
+    """
+
+    def __init__(self, text_encoder) -> None:
+        super().__init__()
+        self.text_encoder = text_encoder
+
+    @torch.no_grad()
+    def forward(self, batch: ForwardBatch,
+                fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        assert batch.prompt is not None
+        prompts = [batch.prompt] if isinstance(batch.prompt,
+                                               str) else batch.prompt
+
+        encoder = self.text_encoder
+        if not hasattr(encoder, "compute_text_embeddings_online"):
+            raise RuntimeError(
+                "Cosmos25TextEncodingStage requires text_encoder.compute_text_embeddings_online()"
+            )
+
+        with set_forward_context(current_timestep=0, attn_metadata=None):
+            prompt_embeds = encoder.compute_text_embeddings_online(
+                {"text": prompts}, "text")
+
+        batch.prompt_embeds = [prompt_embeds]
+
+        if batch.do_classifier_free_guidance:
+            neg = batch.negative_prompt
+            neg_prompts = ([neg] *
+                           len(prompts)) if isinstance(neg, str) else neg
+            with set_forward_context(current_timestep=0, attn_metadata=None):
+                neg_embeds = encoder.compute_text_embeddings_online(
+                    {"text": neg_prompts}, "text")
+            batch.negative_prompt_embeds = [neg_embeds]
+        else:
+            batch.negative_prompt_embeds = []
+
+        return batch
+
+    def verify_input(self, batch: ForwardBatch,
+                     fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("prompt", batch.prompt, V.string_or_list_strings)
+        result.add_check(
+            "negative_prompt",
+            batch.negative_prompt,
+            lambda x:
+            (not batch.do_classifier_free_guidance) or isinstance(x, str),
+        )
+        return result
+
+    def verify_output(self, batch: ForwardBatch,
+                      fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("prompt_embeds", batch.prompt_embeds,
+                         V.list_of_tensors_min_dims(2))
+        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/timestep_preparation.py

@@ -123,3 +123,32 @@ class TimestepPreparationStage(PipelineStage):
         result.add_check("timesteps", batch.timesteps,
                          [V.is_tensor, V.with_dims(1)])
         return result
+
+
+class Cosmos25TimestepPreparationStage(TimestepPreparationStage):
+    """Cosmos 2.5 timestep preparation with scheduler-specific kwargs."""
+
+    def forward(
+        self,
+        batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> ForwardBatch:
+        scheduler = self.scheduler
+        device = get_local_torch_device()
+        num_inference_steps = batch.num_inference_steps
+
+        extra_kwargs: dict = {}
+        sig = inspect.signature(scheduler.set_timesteps)
+        if "shift" in sig.parameters:
+            extra_kwargs["shift"] = fastvideo_args.pipeline_config.flow_shift
+        # Prefer the canonical diffusers kwarg name if available.
+        if "use_karras_sigmas" in sig.parameters:
+            extra_kwargs["use_karras_sigmas"] = True
+        elif "use_kerras_sigma" in sig.parameters:
+            extra_kwargs["use_kerras_sigma"] = True
+
+        scheduler.set_timesteps(num_inference_steps,
+                                device=device,
+                                **extra_kwargs)
+        batch.timesteps = scheduler.timesteps
+        return batch

fastvideo/tests/hooks/test_hooks_api.py

@@ -0,0 +1,60 @@
+from fastvideo.hooks.hooks import ForwardHook, ModuleHookManager
+from torch import nn
+from typing import Any
+import torch
+
+
+class EventHook(ForwardHook):
+    def __init__(self, content: str, event_list: list[str]):
+        self.content = content
+        self.event_list = event_list
+
+    def name(self) -> str:
+        return f"EventHook_{self.content}"
+
+    def pre_forward(self, module: nn.Module, *args, **kwargs):
+        print(
+            f"[{self.content}] Pre-forward called with args[0].shape: {args[0].shape}"
+        )
+        self.event_list.append(f"[pre]{self.content}")
+        return args, kwargs
+
+    def post_forward(self, module: nn.Module, output: Any):
+        print(
+            f"[{self.content}] Post-forward called with outputs.shape: {output.shape}"
+        )
+        self.event_list.append(f"[post]{self.content}")
+        return output
+
+
+def test_hook_execution_order():
+    """Test that hooks are executed in the correct order: LIFO for pre-hooks, FIFO for post-hooks."""
+    # Create a simple model
+    model = nn.Linear(10, 20)
+
+    # Create event list to track hook execution order
+    events = []
+
+    # Create and push hooks in order: A then B
+
+    manager = ModuleHookManager.get_from_or_default(model)
+
+    hook_a = EventHook("A", events)
+    hook_b = EventHook("B", events)
+
+    manager.append_forward_hook(hook_a)
+    manager.append_forward_hook(hook_b)
+
+    # Perform a forward pass
+    input_tensor = torch.randn(2, 10)
+    model(input_tensor)
+
+    # Verify the execution order is [pre_a, pre_b, post_b, post_a]
+    # Pre-hooks should be FILO (First In Last Out): A then B
+    # Post-hooks should be LIFO (Last In First Out): B then A
+    expected_events = ["[pre]A", "[pre]B", "[post]B", "[post]A"]
+
+    assert events == expected_events, (
+        f"Expected {expected_events}, but got {events}"
+    )
+    print(f"✓ Hook execution order test passed: {events}")

fastvideo/tests/hooks/test_layerwise_offload.py

@@ -0,0 +1,395 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+import torch
+import torch.nn as nn
+
+from fastvideo.hooks.layerwise_offload import (
+    LayerwiseOffloadHook,
+    enable_layerwise_offload,
+)
+from fastvideo.hooks.hooks import ModuleHookManager
+
+
+class SimpleBlock(nn.Module):
+    """A simple block with linear layers for testing."""
+
+    def __init__(self, hidden_size: int, dtype: torch.dtype = torch.float32):
+        super().__init__()
+        self.linear1 = nn.Linear(hidden_size, hidden_size, dtype=dtype)
+        self.linear2 = nn.Linear(hidden_size, hidden_size, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.linear1(x)
+        x = torch.relu(x)
+        x = self.linear2(x)
+        return x
+
+
+class SimpleModelWithModuleList(nn.Module):
+    """A simple model with ModuleList for testing layerwise offloading."""
+
+    def __init__(
+        self,
+        num_blocks: int = 4,
+        hidden_size: int = 128,
+        dtype: torch.dtype = torch.float32,
+    ):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [SimpleBlock(hidden_size, dtype=dtype) for _ in range(num_blocks)]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        for block in self.blocks:
+            x = block(x)
+        return x
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_basic():
+    """Test basic functionality of layerwise offloading."""
+    device = torch.device("cuda")
+    hidden_size = 128
+    batch_size = 2
+    seq_len = 16
+    num_blocks = 4
+
+    # Create model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+
+    # Get reference output without offloading
+    input_tensor = torch.randn(batch_size, seq_len, hidden_size, device=device)
+    with torch.no_grad():
+        reference_output = model(input_tensor.clone())
+
+    # Enable layerwise offloading
+    enable_layerwise_offload(model)
+
+    # Verify parameters are offloaded to CPU
+    for block in model.blocks:
+        for param in block.parameters():
+            # Parameters should be placeholder tensors (empty)
+            assert param.numel() == 0, (
+                "Parameters should be offloaded (empty tensors)"
+            )
+
+    # Run forward pass with offloading
+    with torch.no_grad():
+        offloaded_output = model(input_tensor.clone())
+
+    # Check output correctness
+    assert torch.allclose(
+        reference_output, offloaded_output, rtol=1e-4, atol=1e-5
+    ), "Output with offloading should match reference output"
+
+    print(
+        f" Layerwise offload basic test passed: max diff = {(reference_output - offloaded_output).abs().max().item()}"
+    )
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_bf16():
+    """Test layerwise offloading with bfloat16 precision."""
+    device = torch.device("cuda")
+    hidden_size = 256
+    batch_size = 1
+    seq_len = 32
+    num_blocks = 3
+
+    # Create model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.bfloat16
+    ).to(device)
+
+    # Get reference output without offloading
+    input_tensor = torch.randn(
+        batch_size, seq_len, hidden_size, device=device, dtype=torch.bfloat16
+    )
+    with torch.no_grad():
+        reference_output = model(input_tensor.clone())
+
+    # Enable layerwise offloading
+    enable_layerwise_offload(model)
+
+    # Run forward pass with offloading
+    with torch.no_grad():
+        offloaded_output = model(input_tensor.clone())
+
+    # Check output correctness (looser tolerance for bf16)
+    assert torch.allclose(
+        reference_output, offloaded_output, rtol=1e-2, atol=1e-3
+    ), "Output with offloading should match reference output for bf16"
+
+    print(
+        f" Layerwise offload bf16 test passed: max diff = {(reference_output - offloaded_output).abs().max().item()}"
+    )
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_multiple_forward_passes():
+    """Test that layerwise offloading works correctly across multiple forward passes."""
+    device = torch.device("cuda")
+    hidden_size = 64
+    batch_size = 2
+    seq_len = 8
+    num_blocks = 3
+    num_iterations = 5
+
+    # Create model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+
+    # Get reference outputs without offloading
+    torch.manual_seed(42)
+    reference_outputs = []
+    for i in range(num_iterations):
+        input_tensor = torch.randn(
+            batch_size, seq_len, hidden_size, device=device
+        )
+        with torch.no_grad():
+            reference_outputs.append(model(input_tensor.clone()))
+
+    # Enable layerwise offloading
+    enable_layerwise_offload(model)
+
+    # Run multiple forward passes with offloading
+    torch.manual_seed(42)
+    for i in range(num_iterations):
+        input_tensor = torch.randn(
+            batch_size, seq_len, hidden_size, device=device
+        )
+        with torch.no_grad():
+            offloaded_output = model(input_tensor.clone())
+
+        # Check output correctness for each iteration
+        assert torch.allclose(
+            reference_outputs[i], offloaded_output, rtol=1e-4, atol=1e-5
+        ), f"Output mismatch in iteration {i}"
+
+    print(
+        f" Multiple forward passes test passed for {num_iterations} iterations"
+    )
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_parameter_integrity():
+    """Test that parameters are correctly restored during forward pass."""
+    device = torch.device("cuda")
+    hidden_size = 128
+    num_blocks = 2
+
+    # Create model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+
+    # Store original parameter values per block
+    original_params_per_block = []
+    for block in model.blocks:
+        block_params = {}
+        for name, param in block.named_parameters():
+            block_params[name] = param.data.clone()
+        original_params_per_block.append(block_params)
+
+    # Enable layerwise offloading
+    enable_layerwise_offload(model)
+
+    # Create hook managers and verify they exist
+    for block_idx, block in enumerate(model.blocks):
+        manager = ModuleHookManager.get_from(block)
+        assert manager is not None, "Hook manager should be attached to blocks"
+        hook: LayerwiseOffloadHook | None = manager.get_forward_hook(
+            "LayerwiseOffloadHook"
+        )
+        assert hook is not None, "LayerwiseOffloadHook should be registered"
+
+        # Verify parameters are stored in CPU
+        state = hook.state
+        assert len(state.cpu_named_parameters) > 0, (
+            "CPU parameters should be stored"
+        )
+
+        # Verify CPU parameters match original values for this block
+        original_params = original_params_per_block[block_idx]
+        for name, cpu_param in state.cpu_named_parameters.items():
+            assert name in original_params, (
+                f"CPU parameter {name} not found in original params for block {block_idx}"
+            )
+            assert torch.allclose(
+                cpu_param.cpu(), original_params[name].cpu(), rtol=1e-5, atol=1e-6
+            ), f"CPU parameter {name} should match original for block {block_idx}"
+
+    print(" Parameter integrity test passed")
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_no_modulelist_error():
+    """Test that enabling offload on a model without ModuleList raises an error."""
+
+    class ModelWithoutModuleList(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.linear = nn.Linear(128, 128)
+
+        def forward(self, x):
+            return self.linear(x)
+
+    model = ModelWithoutModuleList().to("cuda")
+
+    with pytest.raises(
+        ValueError,
+        match="No nn.ModuleList found in the model for layerwise offloading",
+    ):
+        enable_layerwise_offload(model)
+
+    print(" No ModuleList error test passed")
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_memory_reduction():
+    """Test that layerwise offloading reduces GPU memory usage."""
+    device = torch.device("cuda")
+    hidden_size = 512
+    num_blocks = 8
+    batch_size = 1
+    seq_len = 64
+
+    # Create model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+
+    # Measure initial GPU memory
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+
+    input_tensor = torch.randn(batch_size, seq_len, hidden_size, device=device)
+    with torch.no_grad():
+        _ = model(input_tensor)
+
+    memory_without_offload = torch.cuda.max_memory_allocated()
+
+    # Reset model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+
+    # Enable offloading
+    enable_layerwise_offload(model)
+
+    # Measure GPU memory with offloading
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+
+    input_tensor = torch.randn(batch_size, seq_len, hidden_size, device=device)
+    with torch.no_grad():
+        _ = model(input_tensor)
+
+    memory_with_offload = torch.cuda.max_memory_allocated()
+
+    # Memory with offload should be less (parameters are offloaded)
+    # Note: This is a weak check as memory usage depends on many factors
+    print(
+        f"Memory without offload: {memory_without_offload / 1024**2:.2f} MB, "
+        f"with offload: {memory_with_offload / 1024**2:.2f} MB"
+    )
+
+    # We expect some reduction but this test is more informational
+    assert memory_with_offload < memory_without_offload * 1.5, (
+        "Memory usage should not increase significantly with offloading"
+    )
+
+    print(" Memory reduction test passed")
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_gradient_disabled():
+    """Test that layerwise offloading works correctly with gradients disabled."""
+    device = torch.device("cuda")
+    hidden_size = 128
+    batch_size = 2
+    seq_len = 16
+    num_blocks = 3
+
+    # Create model
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+    model.eval()
+
+    # Enable layerwise offloading
+    enable_layerwise_offload(model)
+
+    input_tensor = torch.randn(batch_size, seq_len, hidden_size, device=device)
+
+    # Forward pass should work without gradients
+    with torch.no_grad():
+        output = model(input_tensor)
+
+    assert output.requires_grad is False, "Output should not require gradients"
+    assert output.shape == (
+        batch_size,
+        seq_len,
+        hidden_size,
+    ), "Output shape should match input"
+
+    print(" Gradient disabled test passed")
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
+def test_layerwise_offload_different_batch_sizes():
+    """Test layerwise offloading with different batch sizes."""
+    device = torch.device("cuda")
+    hidden_size = 128
+    seq_len = 16
+    num_blocks = 3
+
+    model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+
+    # Get reference model without offloading
+    reference_model = SimpleModelWithModuleList(
+        num_blocks=num_blocks, hidden_size=hidden_size, dtype=torch.float32
+    ).to(device)
+    reference_model.load_state_dict(model.state_dict())
+
+    # Enable layerwise offloading
+    enable_layerwise_offload(model)
+
+    # Test different batch sizes
+    for batch_size in [1, 2, 4, 8]:
+        input_tensor = torch.randn(
+            batch_size, seq_len, hidden_size, device=device
+        )
+
+        with torch.no_grad():
+            reference_output = reference_model(input_tensor.clone())
+            offloaded_output = model(input_tensor.clone())
+
+        assert torch.allclose(
+            reference_output, offloaded_output, rtol=1e-4, atol=1e-5
+        ), f"Output mismatch for batch_size={batch_size}"
+
+    print(" Different batch sizes test passed")
+
+
+if __name__ == "__main__":
+    # Run tests manually for debugging
+    if torch.cuda.is_available():
+        print("Running layerwise offloading tests...")
+        test_layerwise_offload_basic()
+        test_layerwise_offload_bf16()
+        test_layerwise_offload_multiple_forward_passes()
+        test_layerwise_offload_parameter_integrity()
+        test_layerwise_offload_no_modulelist_error()
+        test_layerwise_offload_memory_reduction()
+        test_layerwise_offload_gradient_disabled()
+        test_layerwise_offload_different_batch_sizes()
+        print("\n All tests passed!")
+    else:
+        print("CUDA not available, skipping tests")

fastvideo/tests/inference/lora/test_lora_inference_similarity.py

@@ -93,8 +93,11 @@ def test_merge_lora_weights(model_id):
 
     lora_nickname = lora_config["lora_nickname"]
     lora_path = lora_config["lora_path"]
+    # When layerwise offload is enabled, placeholder tensors cannot be compared directly.
     args = FastVideoArgs.from_kwargs(
         model_path=model_id,
+        dit_layerwise_offload=False,
+        use_fsdp_inference=True,
         dit_cpu_offload=True,
         dit_precision="bf16",
     )

fastvideo/tests/modal/pr_test.py

@@ -82,12 +82,12 @@ def run_transformer_tests():
 @app.function(
     gpu="L40S:4", 
     image=image, 
-    timeout=3600,
+    timeout=6000, 
     secrets=[modal.Secret.from_dict({"HF_API_KEY": os.environ.get("HF_API_KEY", "")})],
     volumes={"/root/data": model_vol} 
 )
 def run_ssim_tests():
-    run_test("export MODEL_PATH='/root/data/weights' && hf auth login --token $HF_API_KEY && pytest ./fastvideo/tests/ssim -vs")
+    run_test("export HF_HOME='/root/data/.cache' && export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True && hf auth login --token $HF_API_KEY && pytest ./fastvideo/tests/ssim -vs")
 
 @app.function(gpu="L40S:4", image=image, timeout=900, secrets=[modal.Secret.from_dict({"WANDB_API_KEY": os.environ.get("WANDB_API_KEY", "")})])
 def run_training_tests():

fastvideo/tests/ssim/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-License-Identifier: Apache-2.0
+
+

fastvideo/tests/ssim/generated_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/steps4_Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of _ssim.json

@@ -1,11 +0,0 @@
-{
-  "mean_ssim": 0.7606927804004999,
-  "min_ssim": 0.7035917639732361,
-  "max_ssim": 0.7920367121696472,
-  "reference_video": "/mnt/fast-disks/hao_lab/loay/FastVideo/fastvideo/tests/ssim/L40S_reference_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4",
-  "generated_video": "/mnt/fast-disks/hao_lab/loay/FastVideo/fastvideo/tests/ssim/generated_videos/TurboWan2.1-T2V-1.3B-Diffusers/SLA_ATTN/Will Smith casually eats noodles, his relaxed demeanor contrasting with the energetic background of.mp4",
-  "parameters": {
-    "num_inference_steps": 4,
-    "prompt": "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."
-  }
-}

fastvideo/tests/ssim/test_inference_similarity.py

@@ -26,17 +26,17 @@ else:
 
 # Base parameters from the shell script
 HUNYUAN_PARAMS = {
-    "num_gpus": 2,
+    "num_gpus": 4,
     "model_path": "FastVideo/FastHunyuan-diffusers",
     "height": 720,
     "width": 1280,
     "num_frames": 45,
-    "num_inference_steps": 6,
+    "num_inference_steps": 2,
     "guidance_scale": 1,
     "embedded_cfg_scale": 6,
     "flow_shift": 17,
     "seed": 1024,
-    "sp_size": 2,
+    "sp_size": 4,
     "tp_size": 1,
     "vae_sp": True,
     "fps": 24,
@@ -48,7 +48,7 @@ WAN_T2V_PARAMS = {
     "height": 480,
     "width": 832,
     "num_frames": 45,
-    "num_inference_steps": 20,
+    "num_inference_steps": 4,
     "guidance_scale": 3,
     "embedded_cfg_scale": 6,
     "flow_shift": 7.0,
@@ -67,7 +67,7 @@ WAN_I2V_PARAMS = {
     "height": 480,
     "width": 832,
     "num_frames": 45,
-    "num_inference_steps": 6,
+    "num_inference_steps": 2,
     "guidance_scale": 5.0,
     "embedded_cfg_scale": 6,
     "flow_shift": 7.0,
@@ -232,7 +232,9 @@ def test_inference_similarity(prompt, ATTENTION_BACKEND, model_id):
         "flow_shift": BASE_PARAMS["flow_shift"],
         "sp_size": BASE_PARAMS["sp_size"],
         "tp_size": BASE_PARAMS["tp_size"],
-        "dit_cpu_offload": True,
+        "use_fsdp_inference": True,
+        "dit_cpu_offload": False,
+        "dit_layerwise_offload": False,
     }
     if BASE_PARAMS.get("vae_sp"):
         init_kwargs["vae_sp"] = True

fastvideo/tests/ssim/test_longcat_similarity.py

@@ -0,0 +1,429 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+SSIM-based similarity tests for LongCat video generation.
+
+Tests three LongCat modes:
+- T2V (Text-to-Video): 480p video from text prompt
+- I2V (Image-to-Video): 480p video from image + text prompt  
+- VC (Video Continuation): 480p video continuation from input video + text prompt
+
+Sampling parameters are derived from:
+- examples/inference/basic/basic_longcat_t2v.py
+- examples/inference/basic/basic_longcat_i2v.py
+- examples/inference/basic/basic_longcat_vc.py
+
+Note: num_inference_steps is reduced for CI speed (4 steps vs 50 in examples).
+"""
+import os
+
+import pytest
+import torch
+
+from fastvideo import VideoGenerator
+from fastvideo.logger import init_logger
+from fastvideo.tests.utils import compute_video_ssim_torchvision, write_ssim_results
+
+logger = init_logger(__name__)
+
+# Device-specific reference folder
+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
+elif "H100" in device_name:
+    device_reference_folder = "H100" + device_reference_folder_suffix
+else:
+    logger.warning(f"Unsupported device for ssim tests: {device_name}")
+
+# Common negative prompt from example scripts
+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"
+)
+
+# =============================================================================
+# LongCat T2V Parameters (from basic_longcat_t2v.py)
+# =============================================================================
+LONGCAT_T2V_PARAMS = {
+    "num_gpus": 1,
+    "model_path": "FastVideo/LongCat-Video-T2V-Diffusers",
+    "height": 480,
+    "width": 480,
+    "num_frames": 43,
+    "num_inference_steps": 4,  # Reduced from 50 for CI speed
+    "guidance_scale": 4.0,
+    "fps": 15,
+    "seed": 42,
+    "negative_prompt": NEGATIVE_PROMPT,
+}
+
+# =============================================================================
+# LongCat I2V Parameters (from basic_longcat_i2v.py)
+# =============================================================================
+LONGCAT_I2V_PARAMS = {
+    "num_gpus": 1,
+    "model_path": "FastVideo/LongCat-Video-I2V-Diffusers",
+    "height": 480,
+    "width": 480,  # Square for I2V
+    "num_frames": 43,
+    "num_inference_steps": 4,  # Reduced from 50 for CI speed
+    "guidance_scale": 4.0,
+    "fps": 15,
+    "seed": 42,
+    "negative_prompt": NEGATIVE_PROMPT,
+}
+
+# =============================================================================
+# LongCat VC Parameters (from basic_longcat_vc.py)
+# =============================================================================
+LONGCAT_VC_PARAMS = {
+    "num_gpus": 1,
+    "model_path": "FastVideo/LongCat-Video-VC-Diffusers",
+    "height": 480,
+    "width": 480,
+    "num_frames": 43,
+    "num_inference_steps": 4,  # Reduced from 50 for CI speed
+    "guidance_scale": 4.0,
+    "fps": 15,
+    "seed": 42,
+    "num_cond_frames": 13,
+    "negative_prompt": NEGATIVE_PROMPT,
+}
+
+# Test prompts
+T2V_TEST_PROMPTS = [
+    "In a realistic photography style, a white boy around seven or eight years old "
+    "sits on a park bench, wearing a light blue T-shirt, denim shorts, and white sneakers. "
+    "He holds an ice cream cone with vanilla and chocolate flavors, and beside him is a "
+    "medium-sized golden Labrador. Smiling, the boy offers the ice cream to the dog, "
+    "who eagerly licks it with its tongue. The sun is shining brightly, and the background "
+    "features a green lawn and several tall trees, creating a warm and loving scene.",
+]
+
+I2V_TEST_PROMPTS = [
+    "A woman sits at a wooden table by the window in a cozy café. She reaches out "
+    "with her right hand, picks up the white coffee cup from the saucer, and gently "
+    "brings it to her lips to take a sip. After drinking, she places the cup back on "
+    "the table and looks out the window, enjoying the peaceful atmosphere.",
+]
+
+I2V_IMAGE_PATHS = [
+    "assets/girl.png",
+]
+
+VC_TEST_PROMPTS = [
+    "A person rides a motorcycle along a long, straight road that stretches between "
+    "a body of water and a forested hillside. The rider steadily accelerates, keeping "
+    "the motorcycle centered between the guardrails, while the scenery passes by on "
+    "both sides. The video captures the journey from the rider's perspective, emphasizing "
+    "the sense of motion and adventure.",
+]
+
+VC_VIDEO_PATHS = [
+    "assets/motorcycle.mp4",
+]
+
+
+def _resolve_asset_path(asset_path: str) -> str:
+    """Resolve asset path relative to FastVideo root."""
+    # Check if absolute or already exists
+    if os.path.isabs(asset_path) or os.path.exists(asset_path):
+        return asset_path
+    # Try relative to workspace root
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    repo_root = os.path.abspath(os.path.join(script_dir, "..", "..", ".."))
+    return os.path.join(repo_root, asset_path)
+
+
+@pytest.mark.parametrize("prompt", T2V_TEST_PROMPTS)
+@pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN"])
+def test_longcat_t2v_similarity(prompt: str, ATTENTION_BACKEND: str):
+    """
+    Test LongCat T2V inference and compare output to reference videos using SSIM.
+    
+    Parameters derived from examples/inference/basic/basic_longcat_t2v.py
+    """
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = ATTENTION_BACKEND
+
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    model_id = "LongCat-Video-T2V"
+    
+    output_dir = os.path.join(script_dir, "generated_videos", model_id, ATTENTION_BACKEND)
+    output_video_name = f"{prompt[:100].strip()}.mp4"
+    os.makedirs(output_dir, exist_ok=True)
+
+    init_kwargs = {
+        "num_gpus": LONGCAT_T2V_PARAMS["num_gpus"],
+        "use_fsdp_inference": True,
+        "dit_cpu_offload": True,
+        "vae_cpu_offload": True,
+        "text_encoder_cpu_offload": True,
+        "enable_bsa": False,
+    }
+
+    generation_kwargs = {
+        "output_path": output_dir,
+        "height": LONGCAT_T2V_PARAMS["height"],
+        "width": LONGCAT_T2V_PARAMS["width"],
+        "num_frames": LONGCAT_T2V_PARAMS["num_frames"],
+        "num_inference_steps": LONGCAT_T2V_PARAMS["num_inference_steps"],
+        "guidance_scale": LONGCAT_T2V_PARAMS["guidance_scale"],
+        "fps": LONGCAT_T2V_PARAMS["fps"],
+        "seed": LONGCAT_T2V_PARAMS["seed"],
+        "negative_prompt": LONGCAT_T2V_PARAMS["negative_prompt"],
+    }
+
+    generator = VideoGenerator.from_pretrained(
+        model_path=LONGCAT_T2V_PARAMS["model_path"], **init_kwargs
+    )
+    generator.generate_video(prompt, **generation_kwargs)
+    generator.shutdown()
+
+    generated_video_path = os.path.join(output_dir, output_video_name)
+    assert os.path.exists(generated_video_path), (
+        f"Output video was not generated at {generated_video_path}"
+    )
+
+    # Find reference video
+    reference_folder = os.path.join(
+        script_dir, device_reference_folder, model_id, ATTENTION_BACKEND
+    )
+    if not os.path.exists(reference_folder):
+        raise FileNotFoundError(
+            f"Reference video folder does not exist: {reference_folder}"
+        )
+
+    reference_video_name = None
+    for filename in os.listdir(reference_folder):
+        if filename.endswith(".mp4") and prompt[:100].strip() in filename:
+            reference_video_name = filename
+            break
+
+    if not reference_video_name:
+        raise FileNotFoundError(
+            f"Reference video not found for prompt: {prompt[:50]}... with backend: {ATTENTION_BACKEND}"
+        )
+
+    reference_video_path = os.path.join(reference_folder, reference_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}")
+
+    write_ssim_results(
+        output_dir, ssim_values, reference_video_path, generated_video_path,
+        LONGCAT_T2V_PARAMS["num_inference_steps"], prompt
+    )
+
+    min_acceptable_ssim = 0.90
+    assert mean_ssim >= min_acceptable_ssim, (
+        f"SSIM value {mean_ssim} is below threshold {min_acceptable_ssim} "
+        f"for {model_id} with backend {ATTENTION_BACKEND}"
+    )
+
+
+@pytest.mark.parametrize("prompt", I2V_TEST_PROMPTS)
+@pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN"])
+def test_longcat_i2v_similarity(prompt: str, ATTENTION_BACKEND: str):
+    """
+    Test LongCat I2V inference and compare output to reference videos using SSIM.
+    
+    Parameters derived from examples/inference/basic/basic_longcat_i2v.py
+    """
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = ATTENTION_BACKEND
+
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    model_id = "LongCat-Video-I2V"
+    
+    output_dir = os.path.join(script_dir, "generated_videos", model_id, ATTENTION_BACKEND)
+    output_video_name = f"{prompt[:100].strip()}.mp4"
+    os.makedirs(output_dir, exist_ok=True)
+
+    # Get image path for this prompt
+    prompt_idx = I2V_TEST_PROMPTS.index(prompt)
+    image_path = _resolve_asset_path(I2V_IMAGE_PATHS[prompt_idx])
+
+    init_kwargs = {
+        "num_gpus": LONGCAT_I2V_PARAMS["num_gpus"],
+        "use_fsdp_inference": True,
+        "dit_cpu_offload": True,
+        "vae_cpu_offload": True,
+        "text_encoder_cpu_offload": True,
+        "enable_bsa": False,
+    }
+
+    generation_kwargs = {
+        "output_path": output_dir,
+        "image_path": image_path,
+        "height": LONGCAT_I2V_PARAMS["height"],
+        "width": LONGCAT_I2V_PARAMS["width"],
+        "num_frames": LONGCAT_I2V_PARAMS["num_frames"],
+        "num_inference_steps": LONGCAT_I2V_PARAMS["num_inference_steps"],
+        "guidance_scale": LONGCAT_I2V_PARAMS["guidance_scale"],
+        "fps": LONGCAT_I2V_PARAMS["fps"],
+        "seed": LONGCAT_I2V_PARAMS["seed"],
+        "negative_prompt": LONGCAT_I2V_PARAMS["negative_prompt"],
+    }
+
+    generator = VideoGenerator.from_pretrained(
+        model_path=LONGCAT_I2V_PARAMS["model_path"], **init_kwargs
+    )
+    generator.generate_video(prompt, **generation_kwargs)
+    generator.shutdown()
+
+    generated_video_path = os.path.join(output_dir, output_video_name)
+    assert os.path.exists(generated_video_path), (
+        f"Output video was not generated at {generated_video_path}"
+    )
+
+    # Find reference video
+    reference_folder = os.path.join(
+        script_dir, device_reference_folder, model_id, ATTENTION_BACKEND
+    )
+    if not os.path.exists(reference_folder):
+        raise FileNotFoundError(
+            f"Reference video folder does not exist: {reference_folder}"
+        )
+
+    reference_video_name = None
+    for filename in os.listdir(reference_folder):
+        if filename.endswith(".mp4") and prompt[:100].strip() in filename:
+            reference_video_name = filename
+            break
+
+    if not reference_video_name:
+        raise FileNotFoundError(
+            f"Reference video not found for prompt: {prompt[:50]}... with backend: {ATTENTION_BACKEND}"
+        )
+
+    reference_video_path = os.path.join(reference_folder, reference_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}")
+
+    write_ssim_results(
+        output_dir, ssim_values, reference_video_path, generated_video_path,
+        LONGCAT_I2V_PARAMS["num_inference_steps"], prompt
+    )
+
+    min_acceptable_ssim = 0.90
+    assert mean_ssim >= min_acceptable_ssim, (
+        f"SSIM value {mean_ssim} is below threshold {min_acceptable_ssim} "
+        f"for {model_id} with backend {ATTENTION_BACKEND}"
+    )
+
+
+@pytest.mark.parametrize("prompt", VC_TEST_PROMPTS)
+@pytest.mark.parametrize("ATTENTION_BACKEND", ["FLASH_ATTN"])
+def test_longcat_vc_similarity(prompt: str, ATTENTION_BACKEND: str):
+    """
+    Test LongCat VC (Video Continuation) inference and compare output to reference videos using SSIM.
+    
+    Parameters derived from examples/inference/basic/basic_longcat_vc.py
+    """
+    os.environ["FASTVIDEO_ATTENTION_BACKEND"] = ATTENTION_BACKEND
+
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    model_id = "LongCat-Video-VC"
+    
+    output_dir = os.path.join(script_dir, "generated_videos", model_id, ATTENTION_BACKEND)
+    output_video_name = f"{prompt[:100].strip()}.mp4"
+    os.makedirs(output_dir, exist_ok=True)
+
+    # Get video path for this prompt
+    prompt_idx = VC_TEST_PROMPTS.index(prompt)
+    video_path = _resolve_asset_path(VC_VIDEO_PATHS[prompt_idx])
+    
+    if not os.path.exists(video_path):
+        pytest.skip(f"Input video not found at {video_path}")
+
+    init_kwargs = {
+        "num_gpus": LONGCAT_VC_PARAMS["num_gpus"],
+        "use_fsdp_inference": False,
+        "dit_cpu_offload": False,
+        "vae_cpu_offload": True,
+        "text_encoder_cpu_offload": True,
+        "pin_cpu_memory": False,
+        "enable_bsa": False,
+    }
+
+    generation_kwargs = {
+        "output_path": output_dir,
+        "video_path": video_path,
+        "num_cond_frames": LONGCAT_VC_PARAMS["num_cond_frames"],
+        "height": LONGCAT_VC_PARAMS["height"],
+        "width": LONGCAT_VC_PARAMS["width"],
+        "num_frames": LONGCAT_VC_PARAMS["num_frames"],
+        "num_inference_steps": LONGCAT_VC_PARAMS["num_inference_steps"],
+        "guidance_scale": LONGCAT_VC_PARAMS["guidance_scale"],
+        "fps": LONGCAT_VC_PARAMS["fps"],
+        "seed": LONGCAT_VC_PARAMS["seed"],
+        "negative_prompt": LONGCAT_VC_PARAMS["negative_prompt"],
+    }
+
+    generator = VideoGenerator.from_pretrained(
+        model_path=LONGCAT_VC_PARAMS["model_path"], **init_kwargs
+    )
+    generator.generate_video(prompt, **generation_kwargs)
+    generator.shutdown()
+
+    generated_video_path = os.path.join(output_dir, output_video_name)
+    assert os.path.exists(generated_video_path), (
+        f"Output video was not generated at {generated_video_path}"
+    )
+
+    # Find reference video
+    reference_folder = os.path.join(
+        script_dir, device_reference_folder, model_id, ATTENTION_BACKEND
+    )
+    if not os.path.exists(reference_folder):
+        raise FileNotFoundError(
+            f"Reference video folder does not exist: {reference_folder}"
+        )
+
+    reference_video_name = None
+    for filename in os.listdir(reference_folder):
+        if filename.endswith(".mp4") and prompt[:100].strip() in filename:
+            reference_video_name = filename
+            break
+
+    if not reference_video_name:
+        raise FileNotFoundError(
+            f"Reference video not found for prompt: {prompt[:50]}... with backend: {ATTENTION_BACKEND}"
+        )
+
+    reference_video_path = os.path.join(reference_folder, reference_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}")
+
+    write_ssim_results(
+        output_dir, ssim_values, reference_video_path, generated_video_path,
+        LONGCAT_VC_PARAMS["num_inference_steps"], prompt
+    )
+
+    min_acceptable_ssim = 0.90
+    assert mean_ssim >= min_acceptable_ssim, (
+        f"SSIM value {mean_ssim} is below threshold {min_acceptable_ssim} "
+        f"for {model_id} with backend {ATTENTION_BACKEND}"
+    )

fastvideo/tests/ssim/test_matrixgame_similarity.py

@@ -88,6 +88,7 @@ def test_matrixgame_similarity(prompt, ATTENTION_BACKEND, model_id):
     init_kwargs = {
         "num_gpus": BASE_PARAMS["num_gpus"],
         "use_fsdp_inference": True,
+        "dit_layerwise_offload": False,
         "dit_cpu_offload": False,
         "vae_cpu_offload": False,
         "text_encoder_cpu_offload": True,
@@ -108,9 +109,7 @@ def test_matrixgame_similarity(prompt, ATTENTION_BACKEND, model_id):
         "save_video": True,
     }
 
-    generator = VideoGenerator.from_pretrained(
-        model_path=BASE_PARAMS["model_path"], **init_kwargs
-    )
+    generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs)
     generator.generate_video(prompt, **generation_kwargs)
 
     if isinstance(generator.executor, MultiprocExecutor):

fastvideo/tests/ssim/test_turbodiffusion_similarity.py

@@ -32,7 +32,7 @@ else:
 
 # TurboDiffusion parameters (1-4 step generation with RCM scheduler + SLA attention)
 TURBODIFFUSION_PARAMS = {
-    "num_gpus": 2,
+    "num_gpus": 4,
     "model_path": "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers",
     "height": 480,
     "width": 832,
@@ -40,7 +40,7 @@ TURBODIFFUSION_PARAMS = {
     "num_inference_steps": 4,  # TurboDiffusion uses 1-4 steps
     "guidance_scale": 1.0,  # No CFG for TurboDiffusion
     "seed": 42,
-    "sp_size": 2,
+    "sp_size": 4,
     "tp_size": 1,
     "fps": 24,
 }
@@ -94,10 +94,7 @@ def test_turbodiffusion_inference_similarity(prompt, model_id):
         "fps": BASE_PARAMS["fps"],
     }
 
-    generator = VideoGenerator.from_pretrained(
-        model_path=BASE_PARAMS["model_path"],
-        **init_kwargs
-    )
+    generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs)
     generator.generate_video(prompt, **generation_kwargs)
 
     if isinstance(generator.executor, MultiprocExecutor):
@@ -220,6 +217,8 @@ def test_turbodiffusion_i2v_inference_similarity(prompt, model_id):
         "override_pipeline_cls_name": "TurboDiffusionI2VPipeline",
         # Keep both transformers in VRAM - avoids CPU RAM bottleneck
         "dit_cpu_offload": False,
+        "use_fsdp_inference": True,
+        "dit_layerwise_offload": False,
     }
 
     generation_kwargs = {

fastvideo/tests/training/VSA/test_training_loss_VSA.py

@@ -6,6 +6,11 @@ import json
 from huggingface_hub import snapshot_download
 import torch
 
+# Ensure backend selection happens during import-time initialization
+os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
+# Force VSA to use Triton implementation even on H100 / when CUDA extension is available
+# os.environ["FASTVIDEO_KERNEL_VSA_FORCE_TRITON"] = "1"
+
 # Import the training pipeline
 sys.path.append(str(Path(__file__).parent.parent.parent.parent.parent))
 from fastvideo.training.wan_training_pipeline import main
@@ -19,8 +24,6 @@ h200_reference_wandb_summary_file = "fastvideo/tests/training/VSA/h200_reference
 NUM_NODES = "1"
 NUM_GPUS_PER_NODE = "2"
 
-os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
-
 def run_worker():
     """Worker function that will be run on each GPU"""
     # Create and populate args
@@ -118,10 +121,10 @@ def test_distributed_training():
     wandb_summary = json.load(open(summary_file))
 
     fields_and_thresholds = {
-        'avg_step_time': 1.0,
-        'grad_norm': 0.1,
-        'step_time': 1.0,
-        'train_loss': 0.02
+        'avg_step_time': 3,
+        'grad_norm': 0.2,
+        'step_time': 2.5,
+        'train_loss': 0.04
     }
 
     failures = []

fastvideo/tests/transformers/test_wanvideo.py

@@ -17,10 +17,11 @@ from fastvideo.configs.models.dits import WanVideoConfig
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 
 
+
 logger = init_logger(__name__)
 
 os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
+os.environ["MASTER_PORT"] = "29701"
 
 BASE_MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
@@ -121,4 +122,24 @@ def test_wan_transformer():
     assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
 
     # Check if outputs are similar (allowing for small numerical differences)
-    assert_close(output1, output2, atol=1e-1, rtol=1e-2)
+    max_diff = torch.max(torch.abs(output1 - output2))
+    mean_diff = torch.mean(torch.abs(output1 - output2))
+    logger.info("Max Diff: %s", max_diff.item())
+    logger.info("Mean Diff: %s", mean_diff.item())
+    assert max_diff < 1e-1, f"Maximum difference between outputs: {max_diff.item()}"
+    # mean diff
+    assert mean_diff < 1e-2, f"Mean difference between outputs: {mean_diff.item()}"
+
+
+if __name__ == "__main__":
+    from fastvideo.distributed import (
+        cleanup_dist_env_and_memory,
+        maybe_init_distributed_environment_and_model_parallel,
+    )
+    # Allow running this test file directly without pytest.
+    maybe_init_distributed_environment_and_model_parallel(1, 1)
+    try:
+        test_wan_transformer()
+        logger.info("test_wan_transformer finished successfully.")
+    finally:
+        cleanup_dist_env_and_memory()

fastvideo/training/__init__.py

@@ -1,5 +1,12 @@
 from .distillation_pipeline import DistillationPipeline
 from .training_pipeline import TrainingPipeline
 from .wan_training_pipeline import WanTrainingPipeline
+from fastvideo.training.rl import RLPipeline, create_rl_pipeline
 
-__all__ = ["TrainingPipeline", "WanTrainingPipeline", "DistillationPipeline"]
+__all__ = [
+    "TrainingPipeline",
+    "WanTrainingPipeline",
+    "DistillationPipeline",
+    "RLPipeline",
+    "create_rl_pipeline",
+]

fastvideo/training/rl/__init__.py

@@ -0,0 +1,6 @@
+from .rl_pipeline import RLPipeline, create_rl_pipeline
+
+__all__ = [
+    "RLPipeline",
+    "create_rl_pipeline",
+]

fastvideo/training/rl/rewards/__init__.py

@@ -0,0 +1,11 @@
+from .rewards import (
+    create_reward_models,
+    MultiRewardAggregator,
+    ValueModel
+)
+
+__all__ = [
+    "create_reward_models",
+    "MultiRewardAggregator",
+    "ValueModel",
+]

fastvideo/training/rl/rewards/base.py

@@ -0,0 +1,63 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+    Abstract base class for VIDEO reward models.
+
+    All VIDEO reward models should inherit from this class and implement
+    the compute_reward() method.
+
+    IMPORTANT: Reward models must process FULL VIDEO SEQUENCES, not individual frames.
+    Input shape is [B, T, C, H, W] where T is the temporal (frame) dimension.
+
+    For video-specific rewards, consider:
+    - Temporal coherence across frames
+    - Motion quality and smoothness
+    - Video-text alignment (not just frame-text)
+    - Multi-frame aesthetic quality
+    """
+
+from typing import Any
+
+from abc import ABC, abstractmethod
+import torch
+import torch.nn as nn
+
+class BaseRewardModel(ABC, nn.Module):
+    def __init__(self, model_path: str | None = None, device: str = "cuda"):
+        super().__init__()
+        self.model_path = model_path
+        self.device = device
+
+    @abstractmethod
+    def compute_reward(
+        self,
+        videos: torch.Tensor,  # [B, T, C, H, W] decoded video sequences
+        prompts: list[str] | None,  # Text prompts
+        **kwargs: Any
+    ) -> torch.Tensor:
+        """
+        Compute rewards for generated VIDEO sequences.
+
+        IMPORTANT: This method must process the FULL temporal sequence [B, T, C, H, W].
+        Do NOT evaluate individual frames independently and average.
+
+        Args:
+            videos: Decoded video tensors [B, T, C, H, W] in range [0, 1]
+                   B = batch size
+                   T = number of frames (temporal dimension)
+                   C = channels (typically 3 for RGB)
+                   H, W = height, width
+            prompts: List of text prompts (length B) describing each video
+            **kwargs: Additional model-specific arguments
+
+        Returns:
+            rewards: Tensor of shape [B] with reward scores for each video sequence
+
+        Example:
+            >>> videos = torch.rand(4, 17, 3, 256, 256)  # 4 videos, 17 frames each
+            >>> prompts = ["A cat jumping", "A dog running", ...]
+            >>> rewards = model.compute_reward(videos, prompts)
+        """
+        raise NotImplementedError("Subclasses must implement compute_reward()")
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(model_path={self.model_path})"

fastvideo/training/rl/rewards/ocr.py

@@ -0,0 +1,206 @@
+from paddleocr import PaddleOCR
+import torch
+import numpy as np
+from Levenshtein import distance
+from typing import Any
+from PIL import Image
+
+from fastvideo.training.rl.rewards.base import BaseRewardModel
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class OcrScorerVideo(BaseRewardModel):
+    """
+    OCR reward model for multi-frame video OCR evaluation.
+
+    This model evaluates multiple frames across the video sequence,
+    sampling frames at a specified interval and averaging the OCR scores.
+    """
+
+    def __init__(self,
+                 model_path: str | None = None,
+                 device: str = "cpu",
+                 frame_interval: int = 4):
+        """
+        OCR reward calculator for videos
+
+        Args:
+            model_path: Not used for PaddleOCR (kept for BaseRewardModel compatibility)
+            device: Device string (used to determine use_gpu if not explicitly set)
+            frame_interval: Sample every Nth frame (default: 4)
+        """
+        super().__init__(model_path=model_path, device=device)
+
+        self.frame_interval = frame_interval
+        self.ocr = PaddleOCR(
+            use_angle_cls=False,
+            lang="en",
+            use_gpu=False,
+            show_log=False  # Disable unnecessary log output
+        )
+
+        logger.info("Initialized OcrScorerVideo (device=%s, frame_interval=%d)",
+                    device, frame_interval)
+
+    def _process_single_video(self, video_tensor: torch.Tensor,
+                              prompt: str) -> float:
+        """
+        Process a single video tensor and return its OCR reward.
+
+        Args:
+            video_tensor: Video tensor of shape [C, T, H, W]
+            prompt: Text prompt containing target OCR text in quotes
+
+        Returns:
+            Average reward across positive-scoring frames
+        """
+        # Extract target text from prompt
+        try:
+            target_text = prompt.split('"')[1].replace(' ', '').lower()
+        except IndexError:
+            logger.warning("Failed to extract quoted text from prompt: %s",
+                           prompt)
+            target_text = prompt.replace(' ', '').lower()
+
+        if not target_text:
+            return 0.0
+
+        # video_tensor is [C, T, H, W]
+        C, T, H, W = video_tensor.shape
+
+        # Convert to numpy and move to CPU if needed
+        video_np = video_tensor.detach().cpu().numpy()
+
+        # Convert from [C, T, H, W] to [T, H, W, C] for easier frame extraction
+        video_np = np.transpose(video_np, (1, 2, 3, 0))  # [T, H, W, C]
+        logger.info(f"in ocr 1.5, video_np[0][0]: {video_np[0][0]}")
+
+        # Normalize to [0, 255] uint8 if needed
+        if video_np.max() <= 1.0:
+            video_np = (video_np * 255).astype(np.uint8)
+        else:
+            video_np = video_np.astype(np.uint8)
+
+        frame_rewards = []
+
+        # Sample frames at specified interval
+        for frame_idx in range(0, T, self.frame_interval):
+            frame = video_np[frame_idx]  # [H, W, C]
+            logger.info(f"in ocr 2, frame.shape: {frame.shape}")
+            # Run OCR
+            try:
+                result = self.ocr.ocr(frame, cls=False)
+                logger.info(f"in ocr 3, result: {result}")
+                if result and result[0]:
+                    recognized_text = "".join(
+                        [line[1][0] for line in result[0] if line[1][1] > 0])
+                else:
+                    recognized_text = ""
+            except Exception as e:
+                logger.info("OCR failed on frame %d: %s", frame_idx, str(e))
+                recognized_text = ''
+
+            logger.info(f"in ocr 4, recognized_text: {recognized_text}")
+
+            recognized_text = recognized_text.replace(' ', '').lower()
+            if target_text in recognized_text:
+                dist = 0
+            else:
+                dist = distance(recognized_text, target_text)
+            dist = min(dist, len(target_text))
+            reward = 1.0 - dist / len(target_text)
+
+            logger.info(f"in ocr 5, reward: {reward}")
+            if reward > 0:
+                frame_rewards.append(reward)
+
+        logger.info(f"in ocr 6, frame_rewards: {frame_rewards}")
+
+        return sum([reward / len(frame_rewards)
+                    for reward in frame_rewards]) if frame_rewards else 0.0
+
+    @torch.no_grad()
+    def compute_reward(self, videos: torch.Tensor, prompts: list[str],
+                       **kwargs: Any) -> torch.Tensor:
+        """
+        Calculate OCR reward by evaluating sampled frames across the video.
+
+        Args:
+            videos: Video tensor of shape [B, C, T, H, W]
+                   B = batch size
+                   C = channels (typically 3 for RGB)
+                   T = number of frames (temporal dimension)
+                   H, W = height, width
+            prompts: List of text prompts containing target OCR text in quotes (length B)
+            **kwargs: Additional arguments
+
+        Returns:
+            Reward tensor [B] with averaged OCR similarity scores across frames
+        """
+        # Ensure videos is a torch tensor with correct shape
+        assert isinstance(
+            videos,
+            torch.Tensor), f"videos must be torch.Tensor, got {type(videos)}"
+        assert videos.ndim == 5, f"videos must have 5 dimensions [B, C, T, H, W], got shape {videos.shape}"
+
+        logger.info(f"in ocr 1, videos.shape: {videos.shape}")
+
+        B, C, T, H, W = videos.shape
+        assert len(
+            prompts
+        ) == B, f"Number of prompts ({len(prompts)}) must match batch size ({B})"
+
+        rewards = []
+        for b in range(B):
+            # Extract single video: [C, T, H, W]
+            video = videos[b]
+            reward = self._process_single_video(video, prompts[b])
+            rewards.append(reward)
+
+        logger.info(f"in ocr 7, rewards: {rewards}")
+
+        rewards = torch.tensor(rewards, dtype=torch.float32, device=self.device)
+
+        logger.info(f"in ocr 8, rewards: {rewards}")
+
+        # Check for NaN or Inf values
+        if torch.isnan(rewards).any() or torch.isinf(rewards).any():
+            logger.warning(
+                "NaN or Inf detected in OCR rewards, returning zero tensor")
+            return torch.zeros_like(rewards)
+
+        return rewards
+
+
+if __name__ == "__main__":
+    example_image_path = "flowgrpo_cmd.png"
+    example_image = Image.open(example_image_path)
+    example_prompt = '/f1ow_grpo$'
+
+    # Convert image to RGB if needed
+    if example_image.mode != 'RGB':
+        example_image = example_image.convert('RGB')
+
+    # Convert PIL Image to numpy array [H, W, C]
+    image_np = np.array(example_image)
+
+    # Normalize to [0, 1] range and convert to float32
+    image_np = image_np.astype(np.float32) / 255.0
+
+    # Convert to torch tensor and reshape: [H, W, C] -> [C, H, W]
+    image_tensor = torch.from_numpy(image_np).permute(2, 0, 1)
+
+    # Add temporal dimension: [C, H, W] -> [C, T, H, W] where T=1
+    video_tensor = image_tensor.unsqueeze(1)  # [C, 1, H, W]
+
+    # Add batch dimension: [C, T, H, W] -> [B, C, T, H, W] where B=1
+    video_tensor = video_tensor.unsqueeze(0)  # [1, C, 1, H, W]
+
+    # Instantiate scorer
+    scorer = OcrScorerVideo(device="cpu")
+
+    # Call compute_reward method with video tensor
+    reward = scorer.compute_reward(video_tensor, [example_prompt])
+    print(f"OCR Reward: {reward.item()}")

fastvideo/training/rl/rewards/rewards.py

@@ -0,0 +1,338 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Base infrastructure for VIDEO reward models in RL/GRPO training.
+
+IMPORTANT: This module is designed exclusively for VIDEO generation models.
+All reward models must operate on video sequences [B, T, C, H, W], not single frames.
+
+This module provides:
+1. Multi-reward aggregation for video
+2. Value model wrapper
+3. Integration with FastVideo video generation infrastructure
+"""
+
+from typing import Any
+
+import torch
+import torch.nn as nn
+
+from fastvideo.logger import init_logger
+from fastvideo.training.rl.rewards.ocr import OcrScorerVideo
+from fastvideo.training.rl.rewards.base import BaseRewardModel
+logger = init_logger(__name__)
+
+
+class MultiRewardAggregator(nn.Module):
+    """
+    Aggregates multiple reward models with configurable weights.
+
+    This implements the multi-reward aggregation strategy from flow_grpo,
+    allowing combination of different reward signals (aesthetic quality,
+    text-video alignment, compositional understanding, etc.)
+    """
+
+    def __init__(
+        self,
+        reward_models: list[BaseRewardModel],
+        reward_weights: list[float] | None = None,
+        normalize_rewards: bool = True
+    ):
+        """
+        Initialize multi-reward aggregator.
+
+        Args:
+            reward_models: List of reward model instances
+            reward_weights: Weights for each reward model (default: uniform)
+            normalize_rewards: Whether to normalize rewards before aggregation
+        """
+        super().__init__()
+        self.reward_models = nn.ModuleList(reward_models)
+
+        if reward_weights is None:
+            reward_weights = [1.0 / len(reward_models)] * len(reward_models)
+
+        assert len(reward_weights) == len(reward_models), \
+            f"Number of weights ({len(reward_weights)}) must match number of models ({len(reward_models)})"
+
+        assert abs(sum(reward_weights) - 1.0) < 1e-6, \
+            f"Reward weights must sum to 1.0, got {sum(reward_weights)}"
+
+        self.reward_weights = reward_weights
+        self.normalize_rewards = normalize_rewards
+
+        logger.info(
+            "Initialized MultiRewardAggregator with %d models: %s",
+            len(reward_models),
+            [(type(m).__name__, w) for m, w in zip(reward_models, reward_weights, strict=False)]
+        )
+
+    def compute_reward(
+        self,
+        videos: torch.Tensor,
+        prompts: list[str],
+        return_individual: bool = False,
+        **kwargs: Any
+    ) -> torch.Tensor | dict[str, torch.Tensor]:
+        """
+        Compute aggregated reward from multiple models.
+
+        Args:
+            videos: Decoded video tensors [B, C, T, H, W]
+            prompts: List of text prompts
+            return_individual: If True, return dict with individual rewards
+            **kwargs: Additional arguments passed to reward models
+
+        Returns:
+            If return_individual=False: aggregated_rewards [B]
+            If return_individual=True: dict with "aggregated" and individual model rewards
+        """
+        batch_size = videos.shape[0]
+        individual_rewards: dict[str, torch.Tensor] = {}
+
+        # Collect rewards from all models
+        all_rewards = []
+        for i, (model, weight) in enumerate(zip(self.reward_models, self.reward_weights, strict=False)):
+            reward = model.compute_reward(videos, prompts, **kwargs)
+            assert reward.shape == (batch_size,), \
+                f"Reward model {i} returned shape {reward.shape}, expected ({batch_size},)"
+
+            # Optionally normalize individual rewards
+            if self.normalize_rewards:
+                reward = (reward - reward.mean()) / (reward.std() + 1e-8)
+
+            individual_rewards[f"reward_{type(model).__name__}"] = reward
+            all_rewards.append(weight * reward)
+
+        # Aggregate with weights
+        aggregated = sum(all_rewards)
+
+        if return_individual:
+            individual_rewards["aggregated"] = aggregated
+            return individual_rewards
+
+        return aggregated
+
+    def __repr__(self) -> str:
+        models_str = ", ".join([
+            f"{type(m).__name__}(w={w:.3f})"
+            for m, w in zip(self.reward_models, self.reward_weights, strict=False)
+        ])
+        return f"MultiRewardAggregator({models_str})"
+
+
+class ValueModel(nn.Module):
+    """
+    Value function model wrapper for RL training.
+
+    The value model can either:
+    1. Share the transformer backbone with the policy (memory efficient)
+    2. Use a separate transformer (more flexible)
+
+    For now, this is a placeholder that will be expanded based on
+    the chosen architecture strategy.
+    """
+
+    def __init__(
+        self,
+        transformer: nn.Module,
+        share_backbone: bool = False,
+        hidden_size: int | None = None
+    ):
+        """
+        Initialize value model.
+
+        Args:
+            transformer: Transformer model (policy or separate)
+            share_backbone: Whether to share backbone with policy
+            hidden_size: Hidden size for value head (inferred if None)
+        """
+        super().__init__()
+        self.transformer = transformer
+        self.share_backbone = share_backbone
+
+        # Value head will be added later based on transformer architecture
+        # For now, just store the transformer reference
+        logger.info(
+            "Initialized ValueModel (share_backbone=%s)",
+            share_backbone
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.Tensor,
+        **kwargs: Any
+    ) -> torch.Tensor:
+        """
+        Forward pass to compute value predictions.
+
+        Args:
+            hidden_states: Latent states [B, C, T, H, W]
+            encoder_hidden_states: Text embeddings [B, L, D]
+            timestep: Timesteps [B]
+            **kwargs: Additional transformer arguments
+
+        Returns:
+            values: Value predictions [B]
+        """
+        # TODO: Implement value prediction
+        # For now, return dummy values
+        batch_size = hidden_states.shape[0]
+        return torch.zeros(batch_size, device=hidden_states.device)
+
+
+class DummyRewardModel(BaseRewardModel):
+    """
+    Dummy VIDEO reward model for testing and development.
+
+    Returns random rewards in the range [0, 1] for VIDEO inputs.
+    This is a placeholder for testing the RL pipeline before real video reward models
+    are implemented.
+
+    NOTE: This does NOT actually evaluate video quality - it's just for testing!
+    """
+
+    def __init__(self, mean: float = 0.5, std: float = 0.1):
+        super().__init__(model_path=None)
+        self.mean = mean
+        self.std = std
+        logger.info("Initialized DummyRewardModel (VIDEO) - mean=%.2f, std=%.2f", mean, std)
+        logger.warning(
+            "DummyRewardModel is for TESTING ONLY - does not evaluate actual video quality!"
+        )
+
+    def compute_reward(
+        self,
+        videos: torch.Tensor,  # [B, T, C, H, W]
+        prompts: list[str],
+        **kwargs: Any
+    ) -> torch.Tensor:
+        """
+        Return random rewards for testing.
+
+        Args:
+            videos: Video sequences [B, T, C, H, W]
+            prompts: Text prompts
+
+        Returns:
+            Random rewards [B] in range [0, 1]
+        """
+        batch_size = videos.shape[0]
+        num_frames = videos.shape[1]
+
+        logger.debug(
+            "DummyRewardModel processing %d videos with %d frames each",
+            batch_size,
+            num_frames
+        )
+
+        # Generate random rewards (not based on actual video content!)
+        rewards = torch.randn(batch_size, device=videos.device) * self.std + self.mean
+        return rewards.clamp(0.0, 1.0)
+
+    def load_model(self) -> None:
+        """No model to load for dummy."""
+        pass
+
+
+def create_reward_models(
+    reward_models: dict,
+    device: str = "cuda"
+) -> MultiRewardAggregator:
+    """
+    Factory function to create VIDEO reward models from configuration strings.
+
+    IMPORTANT: Only creates VIDEO reward models. Image-only reward models
+    (PickScore, ImageReward, GenEval, etc.) are NOT supported.
+
+    Args:
+        reward_models: dictionary of reward model names to weights
+                      Example: {"paddle_ocr": 0.5, "video_score": 0.5}
+        device: Device to load models on
+
+    Returns:
+        MultiRewardAggregator with loaded VIDEO reward models
+
+    Supported VIDEO Reward Types:
+        - "paddle_ocr": PaddleOCR multi-frame video text recognition
+        - "video_score": Video aesthetic quality (multi-frame) - TODO
+        - "video_text_alignment": CLIP-based video-text similarity - TODO
+        - "temporal_coherence": Frame-to-frame consistency - TODO
+        - "motion_quality": Motion smoothness and realism - TODO
+        - "dummy": Random rewards for testing (VIDEO-aware)
+
+    NOT Supported (Image-Only):
+        - "pickscore": Image aesthetic (use "video_score" instead)
+        - "imagereward": Image quality (use "video_score" instead)
+        - "geneval": Image compositional (no video equivalent yet)
+        - Any single-frame reward models
+
+    Example:
+        >>> models = create_reward_models(
+        ...     reward_models={
+        ...         "paddle_ocr": 0.5,
+        ...         "video_text_alignment": 0.5
+        ...     },
+        ...     device="cuda"
+        ... )
+    """
+
+
+    assert reward_models, "No reward models specified. Please select at least 1 reward model"
+
+    types = [t.strip() for t in reward_models.keys()]
+    weights = list(reward_models.values())
+
+    assert len(types) == len(weights), \
+        f"Number of models ({len(types)}) must match number of weights ({len(weights)})"
+
+    # Create reward models based on types
+    models_list: list[BaseRewardModel] = []
+    for reward_type in types:
+        if reward_type == "dummy":
+            model = DummyRewardModel()
+
+        elif reward_type == "paddle_ocr":
+            logger.info("Creating PaddleOCR reward model")
+            model = OcrScorerVideo(device=device)
+
+        elif reward_type == "video_score":
+            # TODO: Implement VideoScore reward model (Phase 2)
+            logger.warning(
+                "VideoScore reward not implemented yet, using DummyRewardModel"
+            )
+            model = DummyRewardModel()
+        elif reward_type == "video_text_alignment":
+            # TODO: Implement VideoTextAlignment reward model (Phase 2)
+            logger.warning(
+                "VideoTextAlignment reward not implemented yet, using DummyRewardModel"
+            )
+            model = DummyRewardModel()
+        elif reward_type == "temporal_coherence":
+            # TODO: Implement TemporalCoherence reward model (Phase 2)
+            logger.warning(
+                "TemporalCoherence reward not implemented yet, using DummyRewardModel"
+            )
+            model = DummyRewardModel()
+        elif reward_type == "motion_quality":
+            # TODO: Implement MotionQuality reward model (Phase 2)
+            logger.warning(
+                "MotionQuality reward not implemented yet, using DummyRewardModel"
+            )
+            model = DummyRewardModel()
+        else:
+            logger.warning(
+                "Unknown VIDEO reward type '%s', using DummyRewardModel",
+                reward_type
+            )
+            model = DummyRewardModel()
+
+        models_list.append(model)
+
+    logger.info(
+        "Created MultiRewardAggregator with %d VIDEO reward models",
+        len(models_list)
+    )
+
+    return MultiRewardAggregator(models_list, weights, normalize_rewards=True)

fastvideo/training/rl/rl_utils.py

@@ -0,0 +1,385 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Utility functions for RL/GRPO training.
+"""
+
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def compute_gae(
+    rewards: torch.Tensor,
+    values: torch.Tensor,
+    next_values: torch.Tensor,
+    dones: torch.Tensor | None = None,
+    gamma: float = 0.99,
+    lambda_: float = 0.95
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Compute Generalized Advantage Estimation (GAE-lambda).
+
+    GAE reduces variance in advantage estimation while allowing some bias.
+    This is a key component of modern policy gradient methods like PPO and GRPO.
+
+    Args:
+        rewards: Rewards at each step [B, T] or [B]
+        values: Value predictions at each step [B, T] or [B]
+        next_values: Value predictions at next step [B, T] or [B]
+        dones: Episode termination flags [B, T] or [B] (1 if done, 0 otherwise)
+        gamma: Discount factor
+        lambda_: GAE lambda parameter (0=TD(0), 1=Monte Carlo)
+
+    Returns:
+        advantages: GAE advantages [B, T] or [B]
+        returns: TD(lambda) returns [B, T] or [B]
+
+    Reference:
+        Schulman et al. "High-Dimensional Continuous Control Using Generalized Advantage Estimation"
+        https://arxiv.org/abs/1506.02438
+    """
+    if dones is None:
+        dones = torch.zeros_like(rewards)
+
+    # Compute TD residuals: delta_t = r_t + gamma * V(s_{t+1}) - V(s_t)
+    deltas = rewards + gamma * next_values * (1.0 - dones) - values
+
+    # If single step (no time dimension), return directly
+    if deltas.dim() == 1:
+        advantages = deltas
+        returns = advantages + values
+        return advantages, returns
+
+    # Multi-step: compute GAE recursively
+    batch_size, num_steps = deltas.shape
+    advantages = torch.zeros_like(deltas)
+    gae = torch.zeros(batch_size, device=deltas.device)
+
+    # Backward pass to compute GAE
+    for t in reversed(range(num_steps)):
+        gae = deltas[:, t] + gamma * lambda_ * (1.0 - dones[:, t]) * gae
+        advantages[:, t] = gae
+
+    # Returns are advantages + values
+    returns = advantages + values
+
+    return advantages, returns
+
+
+def normalize_advantages(
+    advantages: torch.Tensor,
+    epsilon: float = 1e-8
+) -> torch.Tensor:
+    """
+    Normalize advantages to have zero mean and unit variance.
+
+    This is a common practice in PPO and GRPO to stabilize training.
+
+    Args:
+        advantages: Raw advantages [B, ...]
+        epsilon: Small constant for numerical stability
+
+    Returns:
+        normalized_advantages: Normalized advantages [B, ...]
+    """
+    mean = advantages.mean()
+    std = advantages.std()
+    return (advantages - mean) / (std + epsilon)
+
+#TODO(jiali): refactor into algorithm
+def compute_grpo_policy_loss(
+    log_probs: torch.Tensor,
+    old_log_probs: torch.Tensor,
+    advantages: torch.Tensor,
+    clip_range: float = 0.2,
+    use_ratio_norm: bool = True,
+    max_importance_ratio: float = 10.0
+) -> tuple[torch.Tensor, dict[str, Any]]:
+    """
+    Compute GRPO policy loss with importance sampling and clipping.
+
+    This implements the core GRPO objective with safety mechanisms from GRPO-Guard:
+    - Importance ratio clipping (PPO-style)
+    - RatioNorm correction (GRPO-Guard)
+    - Ratio clamping for extreme values
+
+    Args:
+        log_probs: Log probabilities from current policy [B]
+        old_log_probs: Log probabilities from old policy [B]
+        advantages: Advantages [B]
+        clip_range: Clipping range for importance ratios
+        use_ratio_norm: Apply RatioNorm correction (GRPO-Guard)
+        max_importance_ratio: Maximum importance ratio before clamping
+
+    Returns:
+        loss: Policy loss (scalar)
+        info: Dictionary with diagnostic information
+
+    Reference:
+        - PPO: Schulman et al. "Proximal Policy Optimization Algorithms"
+        - GRPO-Guard: RatioNorm and gradient reweighting
+    """
+    # Compute importance ratio: r_t = pi_new(a|s) / pi_old(a|s)
+    log_ratio = log_probs - old_log_probs
+    ratio = torch.exp(log_ratio)
+
+    # Clamp extreme ratios for numerical stability
+    ratio = torch.clamp(ratio, 1.0 / max_importance_ratio, max_importance_ratio)
+
+    # RatioNorm correction (GRPO-Guard)
+    # Corrects bias in importance sampling when ratio >> 1
+    if use_ratio_norm:
+        ratio_mean = ratio.mean()
+        ratio = ratio / (ratio_mean + 1e-8)
+
+    # Clipped surrogate objective
+    ratio_clipped = torch.clamp(ratio, 1.0 - clip_range, 1.0 + clip_range)
+    surrogate1 = ratio * advantages
+    surrogate2 = ratio_clipped * advantages
+    policy_loss = -torch.min(surrogate1, surrogate2).mean()
+
+    # Compute diagnostics
+    with torch.no_grad():
+        # Clip fraction: how often ratios were clipped
+        clip_fraction = ((ratio < 1.0 - clip_range) | (ratio > 1.0 + clip_range)).float().mean()
+
+        # KL divergence (approximate)
+        kl_div = log_ratio.mean()
+
+        # Importance ratio stats
+        importance_ratio_mean = ratio.mean()
+        importance_ratio_std = ratio.std()
+
+    info = {
+        "policy_loss": policy_loss.item(),
+        "clip_fraction": clip_fraction.item(),
+        "kl_divergence": kl_div.item(),
+        "importance_ratio_mean": importance_ratio_mean.item(),
+        "importance_ratio_std": importance_ratio_std.item(),
+    }
+
+    return policy_loss, info
+
+
+def compute_value_loss(
+    values: torch.Tensor,
+    returns: torch.Tensor,
+    old_values: torch.Tensor | None = None,
+    clip_range: float = 0.2,
+    use_clipping: bool = True
+) -> tuple[torch.Tensor, dict[str, Any]]:
+    """
+    Compute value function loss with optional clipping.
+
+    Args:
+        values: Value predictions from current model [B]
+        returns: Target returns (from GAE) [B]
+        old_values: Value predictions from old model [B] (for clipping)
+        clip_range: Clipping range for value updates
+        use_clipping: Whether to use clipped value loss (PPO-style)
+
+    Returns:
+        loss: Value loss (scalar)
+        info: Dictionary with diagnostic information
+    """
+    # Standard MSE loss
+    value_loss_unclipped = F.mse_loss(values, returns, reduction="none")
+
+    # Clipped value loss (PPO-style)
+    if use_clipping and old_values is not None:
+        values_clipped = old_values + torch.clamp(
+            values - old_values,
+            -clip_range,
+            clip_range
+        )
+        value_loss_clipped = F.mse_loss(values_clipped, returns, reduction="none")
+        value_loss = torch.max(value_loss_unclipped, value_loss_clipped).mean()
+    else:
+        value_loss = value_loss_unclipped.mean()
+
+    # Compute diagnostics
+    with torch.no_grad():
+        explained_variance = 1.0 - (returns - values).var() / (returns.var() + 1e-8)
+
+    info = {
+        "value_loss": value_loss.item(),
+        "explained_variance": explained_variance.item(),
+        "value_mean": values.mean().item(),
+        "value_std": values.std().item(),
+    }
+
+    return value_loss, info
+
+
+def compute_policy_entropy(log_probs: torch.Tensor) -> torch.Tensor:
+    """
+    Compute policy entropy for exploration bonus.
+
+    Args:
+        log_probs: Log probabilities [B]
+
+    Returns:
+        entropy: Mean entropy across batch (scalar)
+    """
+    # For continuous actions: H = -log_prob (assuming Gaussian)
+    # For discrete: H = -sum(p * log(p))
+    # Here we use a simple approximation
+    entropy = -log_probs.mean()
+    return entropy
+
+
+def apply_gradient_reweighting(
+    gradients: torch.Tensor,
+    timesteps: torch.Tensor,
+    num_train_timesteps: int = 1000
+) -> torch.Tensor:
+    """
+    Apply GRPO-Guard gradient reweighting across denoising steps.
+
+    This reweights gradients based on the timestep to balance learning
+    across different noise levels.
+
+    Args:
+        gradients: Gradients to reweight [B, ...]
+        timesteps: Timesteps at which gradients were computed [B]
+        num_train_timesteps: Total number of training timesteps
+
+    Returns:
+        reweighted_gradients: Reweighted gradients [B, ...]
+    """
+    # Compute timestep weights (higher weight for later timesteps)
+    # This is a simple linear weighting, can be made more sophisticated
+    timestep_weights = 1.0 + (timesteps.float() / num_train_timesteps)
+    timestep_weights = timestep_weights.view(-1, *([1] * (gradients.dim() - 1)))
+
+    return gradients * timestep_weights
+
+
+def sample_random_timesteps(
+    batch_size: int,
+    min_timestep: int,
+    max_timestep: int,
+    device: torch.device,
+    generator: torch.Generator | None = None
+) -> torch.Tensor:
+    """
+    Sample random timesteps for noise injection (Flow-GRPO-Fast).
+
+    Args:
+        batch_size: Number of samples
+        min_timestep: Minimum timestep
+        max_timestep: Maximum timestep
+        device: Device for tensor
+        generator: Random generator for reproducibility
+
+    Returns:
+        timesteps: Random timesteps [B]
+    """
+    if generator is not None:
+        timesteps = torch.randint(
+            min_timestep,
+            max_timestep + 1,
+            (batch_size,),
+            device=device,
+            generator=generator
+        )
+    else:
+        timesteps = torch.randint(
+            min_timestep,
+            max_timestep + 1,
+            (batch_size,),
+            device=device
+        )
+
+    return timesteps
+
+
+def compute_reward_statistics(
+    rewards: torch.Tensor
+) -> dict[str, float]:
+    """
+    Compute statistics for reward distribution.
+
+    Args:
+        rewards: Reward values [B]
+
+    Returns:
+        stats: Dictionary with mean, std, min, max
+    """
+    return {
+        "reward_mean": rewards.mean().item(),
+        "reward_std": rewards.std().item(),
+        "reward_min": rewards.min().item(),
+        "reward_max": rewards.max().item(),
+    }
+
+
+def check_early_stopping(
+    kl_divergence: float,
+    target_kl: float
+) -> bool:
+    """
+    Check if training should stop early based on KL divergence.
+
+    Args:
+        kl_divergence: Current KL divergence
+        target_kl: Target KL threshold
+
+    Returns:
+        should_stop: True if KL exceeds target
+    """
+    if kl_divergence > target_kl:
+        logger.warning(
+            "Early stopping triggered: KL divergence %.4f > target %.4f",
+            kl_divergence,
+            target_kl
+        )
+        return True
+    return False
+
+
+def compute_log_probs_from_model_output(
+    model_output: torch.Tensor,
+    target: torch.Tensor,
+    noise_level: float = 0.1
+) -> torch.Tensor:
+    """
+    Compute log probabilities from model predictions.
+
+    For diffusion models, we approximate log probabilities using the
+    negative squared error (assuming Gaussian likelihood).
+
+    Args:
+        model_output: Model predictions [B, C, T, H, W]
+        target: Target values [B, C, T, H, W]
+        noise_level: Assumed noise level (std) for Gaussian likelihood
+
+    Returns:
+        log_probs: Log probabilities [B]
+    """
+    # Compute mean squared error per sample
+    mse = ((model_output - target) ** 2).flatten(1).mean(dim=1)
+
+    # Log probability under Gaussian: log p(x) = -0.5 * (x - mu)^2 / sigma^2 + const
+    log_probs = -0.5 * mse / (noise_level ** 2)
+
+    return log_probs
+
+
+def check_for_nan_inf(tensor: torch.Tensor, name: str) -> None:
+    """
+    Check tensor for NaN or Inf values and raise error if found.
+
+    Args:
+        tensor: Tensor to check
+        name: Name for error message
+    """
+    if torch.isnan(tensor).any():
+        raise ValueError(f"{name} contains NaN values")
+    if torch.isinf(tensor).any():
+        raise ValueError(f"{name} contains Inf values")

fastvideo/training/rl/stat_tracking.py

@@ -0,0 +1,189 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Per-prompt statistics tracking for GRPO training.
+
+This module ports the PerPromptStatTracker from FlowGRPO to FastVideo.
+It tracks reward statistics per unique prompt and computes normalized advantages.
+
+Ported from:
+- flow_grpo/flow_grpo/stat_tracking.py
+
+Key adaptations:
+1. Uses FastVideo's logging instead of print statements
+2. Works with single GPU (no distributed logic)
+3. Supports numpy arrays and torch tensors
+"""
+
+import numpy as np
+from typing import Union
+
+import torch
+
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class PerPromptStatTracker:
+    """
+    Tracks reward statistics per unique prompt for advantage normalization.
+    
+    This class maintains running statistics (mean, std) for each unique prompt
+    and computes normalized advantages using either per-prompt or global statistics.
+    
+    Used in GRPO training to normalize advantages within groups of samples
+    generated from the same prompt, which helps stabilize training when different
+    prompts have different reward scales.
+    """
+
+    def __init__(self, global_std: bool = False):
+        """
+        Initialize the per-prompt stat tracker.
+        
+        Args:
+            global_std: If True, use global std across all rewards for normalization.
+                       If False, use per-prompt std (default, recommended for GRPO).
+        """
+        self.global_std = global_std
+        self.stats: dict[str, list] = {}  # Maps prompt -> list of rewards
+        self.history_prompts: set[int] = set()  # Set of hashed prompts seen
+
+    def update(
+        self,
+        prompts: Union[list[str], np.ndarray],
+        rewards: Union[list[float], np.ndarray, torch.Tensor],
+        type: str = 'grpo'
+    ) -> np.ndarray:
+        """
+        Update statistics and compute normalized advantages.
+        
+        Args:
+            prompts: List or array of prompt strings (one per sample)
+            rewards: Array or tensor of reward values (one per sample)
+            type: Advantage computation type:
+                - 'grpo': Normalize by (reward - mean) / std (default)
+                - 'rwr': Return rewards as-is (reward-weighted regression)
+                - 'sft': Binary advantages (1 for max, 0 otherwise)
+                - 'dpo': DPO-style advantages (1 for max, -1 for min)
+        
+        Returns:
+            advantages: Normalized advantages array [num_samples] or [num_samples, ...]
+                       Shape matches rewards shape
+        """
+        # Convert to numpy arrays
+        prompts = np.array(prompts)
+        if isinstance(rewards, torch.Tensor):
+            rewards = rewards.detach().cpu().numpy()
+        rewards = np.array(rewards, dtype=np.float64)
+        
+        # Ensure rewards are 1D (one reward per sample)
+        # FlowGRPO expects rewards to be aggregated per sample
+        if rewards.ndim > 1:
+            # If multi-dimensional, flatten or take mean
+            # For [B, num_steps] shape, we typically want one reward per sample
+            # So we take the mean across timesteps
+            if rewards.ndim == 2:
+                # Assume shape is [B, num_steps] - take mean across timesteps
+                rewards = rewards.mean(axis=1)
+            else:
+                # Flatten and take mean for higher dimensions
+                rewards = rewards.reshape(len(prompts), -1).mean(axis=1)
+        
+        # Ensure prompts and rewards have matching lengths
+        assert len(prompts) == len(rewards), \
+            f"Prompts ({len(prompts)}) and rewards ({len(rewards)}) must have same length"
+        
+        unique_prompts = np.unique(prompts)
+        advantages = np.zeros_like(rewards, dtype=np.float64)
+        
+        # First pass: collect rewards for each prompt
+        for prompt in unique_prompts:
+            prompt_mask = prompts == prompt
+            prompt_rewards = rewards[prompt_mask]
+            
+            # Store rewards in stats
+            if prompt not in self.stats:
+                self.stats[prompt] = []
+            self.stats[prompt].extend(prompt_rewards.tolist())
+            self.history_prompts.add(hash(prompt))
+        
+        # Second pass: compute statistics and advantages
+        for prompt in unique_prompts:
+            prompt_mask = prompts == prompt
+            prompt_rewards = rewards[prompt_mask]
+            
+            # Stack all historical rewards for this prompt
+            if len(self.stats[prompt]) > 0:
+                all_prompt_rewards = np.array(self.stats[prompt])
+            else:
+                all_prompt_rewards = prompt_rewards
+            
+            # Compute mean and std
+            mean = np.mean(all_prompt_rewards, axis=0, keepdims=True)
+            
+            if self.global_std:
+                # Use global std across all rewards
+                std = np.std(rewards, axis=0, keepdims=True) + 1e-4
+            else:
+                # Use per-prompt std
+                std = np.std(all_prompt_rewards, axis=0, keepdims=True) + 1e-4
+            
+            # Compute advantages based on type
+            if type == 'grpo':
+                # GRPO: normalize by (reward - mean) / std
+                advantages[prompt_mask] = (prompt_rewards - mean) / std
+            elif type == 'rwr':
+                # Reward-weighted regression: use rewards as-is
+                advantages[prompt_mask] = prompt_rewards
+            elif type == 'sft':
+                # Supervised fine-tuning: binary (1 for max, 0 otherwise)
+                max_reward = np.max(prompt_rewards)
+                advantages[prompt_mask] = (prompt_rewards == max_reward).astype(np.float64)
+            elif type == 'dpo':
+                # DPO-style: 1 for max, -1 for min
+                prompt_rewards_tensor = torch.tensor(prompt_rewards)
+                max_idx = torch.argmax(prompt_rewards_tensor)
+                min_idx = torch.argmin(prompt_rewards_tensor)
+                
+                # If all rewards are the same, use first two indices
+                if max_idx == min_idx:
+                    min_idx = torch.tensor(0)
+                    max_idx = torch.tensor(1) if len(prompt_rewards_tensor) > 1 else torch.tensor(0)
+                
+                result = torch.zeros_like(prompt_rewards_tensor, dtype=torch.float64)
+                result[max_idx] = 1.0
+                result[min_idx] = -1.0
+                advantages[prompt_mask] = result.numpy()
+            else:
+                raise ValueError(f"Unknown advantage type: {type}. Must be one of: 'grpo', 'rwr', 'sft', 'dpo'")
+        
+        return advantages
+
+    def get_stats(self) -> tuple[float, int]:
+        """
+        Get statistics about tracked prompts.
+        
+        Returns:
+            avg_group_size: Average number of samples per unique prompt
+            history_prompts: Number of unique prompts seen (across all updates)
+        """
+        if not self.stats:
+            avg_group_size = 0.0
+        else:
+            total_samples = sum(len(v) for v in self.stats.values())
+            avg_group_size = total_samples / len(self.stats)
+        
+        history_prompts = len(self.history_prompts)
+        
+        return avg_group_size, history_prompts
+
+    def clear(self) -> None:
+        """
+        Clear all statistics (but keep history_prompts for tracking).
+        
+        This is typically called after each epoch to reset per-epoch statistics
+        while maintaining a record of all prompts seen during training.
+        """
+        self.stats = {}
+        logger.debug("Cleared per-prompt statistics (kept %d unique prompts in history)", 
+                    len(self.history_prompts))

fastvideo/training/rl/wan_grpo_utils.py

@@ -0,0 +1,877 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+GRPO utilities for Wan model in FastVideo.
+
+This module ports the SDE step and pipeline functions from FlowGRPO to work with
+FastVideo's scheduler and pipeline interfaces.
+
+Ported from:
+- flow_grpo/flow_grpo/diffusers_patch/wan_pipeline_with_logprob.py
+
+Key adaptations:
+1. Uses FastVideo's FlowUniPCMultistepScheduler instead of diffusers' UniPCMultistepScheduler
+2. Works with FastVideo's WanPipeline (ComposedPipelineBase) instead of diffusers' WanPipeline
+3. Direct module access via pipeline.get_module() instead of pipeline attributes
+4. Simplified prompt encoding (direct text encoder usage instead of pipeline stages)
+"""
+
+import math
+import time
+from typing import Any
+
+import torch
+from tqdm import tqdm
+from diffusers.utils.torch_utils import randn_tensor
+
+from fastvideo.forward_context import set_forward_context
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_flow_unipc_multistep import (
+    FlowUniPCMultistepScheduler)
+from fastvideo.utils import get_compute_dtype
+
+# for test_wan_transformer2
+import os
+from diffusers import WanTransformer3DModel
+from fastvideo.utils import maybe_download_model
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.configs.pipelines import PipelineConfig
+from fastvideo.configs.models.dits import WanVideoConfig
+from fastvideo.models.loader.component_loader import TransformerLoader
+
+logger = init_logger(__name__)
+
+
+def test_wan_transformer():
+    BASE_MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+    MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
+                                      local_dir=os.path.join(
+                                          'data', BASE_MODEL_PATH))
+    TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
+
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    precision = torch.bfloat16
+    precision_str = "bf16"
+    args = FastVideoArgs(model_path=TRANSFORMER_PATH,
+                         dit_cpu_offload=True,
+                         pipeline_config=PipelineConfig(
+                             dit_config=WanVideoConfig(),
+                             dit_precision=precision_str))
+    args.device = device
+
+    loader = TransformerLoader()
+    model2 = loader.load(TRANSFORMER_PATH, args).to(dtype=precision)
+
+    model1 = WanTransformer3DModel.from_pretrained(
+        TRANSFORMER_PATH, device=device,
+        torch_dtype=precision).to(device, dtype=precision).requires_grad_(False)
+
+    total_params = sum(p.numel() for p in model1.parameters())
+    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
+    weight_sum_model1 = sum(
+        p.to(torch.float64).sum().item() for p in model1.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model1 = weight_sum_model1 / total_params
+    logger.info("Model 1 weight sum: %s", weight_sum_model1)
+    logger.info("Model 1 weight mean: %s", weight_mean_model1)
+
+    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
+    total_params_model2 = sum(p.numel() for p in model2.parameters())
+    weight_sum_model2 = sum(
+        p.to(torch.float64).sum().item() for p in model2.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model2 = weight_sum_model2 / total_params_model2
+    logger.info("Model 2 weight sum: %s", weight_sum_model2)
+    logger.info("Model 2 weight mean: %s", weight_mean_model2)
+
+    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
+    logger.info("Weight sum difference: %s", weight_sum_diff)
+    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
+    logger.info("Weight mean difference: %s", weight_mean_diff)
+
+    # Set both models to eval mode
+    model1 = model1.eval()
+    model2 = model2.eval()
+
+    # Create identical inputs for both models
+    batch_size = 1
+    seq_len = 30
+
+    # Video latents [B, C, T, H, W]
+    hidden_states = torch.randn(batch_size,
+                                16,
+                                21,
+                                160,
+                                90,
+                                device=device,
+                                dtype=precision)
+
+    # Text embeddings [B, L, D] (including global token)
+    encoder_hidden_states = torch.randn(batch_size,
+                                        seq_len + 1,
+                                        4096,
+                                        device=device,
+                                        dtype=precision)
+
+    # Timestep
+    timestep = torch.tensor([500], device=device, dtype=precision)
+
+    forward_batch = ForwardBatch(data_type="dummy", )
+
+    with torch.amp.autocast('cuda', dtype=precision):
+        output1 = model1(
+            hidden_states=hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            timestep=timestep,
+            return_dict=False,
+        )[0]
+        with set_forward_context(
+                current_timestep=0,
+                attn_metadata=None,
+                forward_batch=forward_batch,
+        ):
+            output2 = model2(hidden_states=hidden_states,
+                             encoder_hidden_states=encoder_hidden_states,
+                             timestep=timestep)
+
+    # Check if outputs have the same shape
+    assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
+    assert output1.dtype == output2.dtype, f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
+
+    # Check if outputs are similar (allowing for small numerical differences)
+    max_diff = torch.max(torch.abs(output1 - output2))
+    mean_diff = torch.mean(torch.abs(output1 - output2))
+    logger.info("Max Diff: %s", max_diff.item())
+    logger.info("Mean Diff: %s", mean_diff.item())
+    assert max_diff < 1e-1, f"Maximum difference between outputs: {max_diff.item()}"
+    # mean diff
+    assert mean_diff < 1e-2, f"Mean difference between outputs: {mean_diff.item()}"
+    '''
+    INFO 01-19 22:53:46 [wan_grpo_utils.py:74] Model 1 weight sum: 395834.3506456231████                                                    | 1/2 [00:00<00:00,  7.84it/s]
+    INFO 01-19 22:53:46 [wan_grpo_utils.py:75] Model 1 weight mean: 0.0002789536598289884
+    INFO 01-19 22:53:47 [wan_grpo_utils.py:83] Model 2 weight sum: 395834.3506456231
+    INFO 01-19 22:53:47 [wan_grpo_utils.py:84] Model 2 weight mean: 0.0002789536598289884
+    INFO 01-19 22:53:47 [wan_grpo_utils.py:87] Weight sum difference: 0.0
+    INFO 01-19 22:53:47 [wan_grpo_utils.py:89] Weight mean difference: 0.0
+    INFO 01-19 22:53:54 [wan_grpo_utils.py:145] Max Diff: 0.08203125
+    INFO 01-19 22:53:54 [wan_grpo_utils.py:146] Mean Diff: 0.01129150390625
+    '''
+
+
+def test_wan_transformer2(model2):
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    precision = torch.bfloat16
+
+    logger.info("loading model1 transformer weight")
+    BASE_MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+    MODEL_PATH = maybe_download_model(BASE_MODEL_PATH,
+                                      local_dir=os.path.join(
+                                          'data', BASE_MODEL_PATH))
+    TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
+    model1 = WanTransformer3DModel.from_pretrained(
+        TRANSFORMER_PATH,
+        device=device,
+        torch_dtype=precision,
+    ).to(device, dtype=precision).requires_grad_(False)
+
+    total_params = sum(p.numel() for p in model1.parameters())
+    # Calculate weight sum for model1 (converting to float64 to avoid overflow)
+    weight_sum_model1 = sum(
+        p.to(torch.float64).sum().item() for p in model1.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model1 = weight_sum_model1 / total_params
+    logger.info("Model 1 weight sum: %s", weight_sum_model1)
+    logger.info("Model 1 weight mean: %s", weight_mean_model1)
+
+    # Calculate weight sum for model2 (converting to float64 to avoid overflow)
+    total_params_model2 = sum(p.numel() for p in model2.parameters())
+    weight_sum_model2 = sum(
+        p.to(torch.float64).sum().item() for p in model2.parameters())
+    # Also calculate mean for more stable comparison
+    weight_mean_model2 = weight_sum_model2 / total_params_model2
+    logger.info("Model 2 weight sum: %s", weight_sum_model2)
+    logger.info("Model 2 weight mean: %s", weight_mean_model2)
+
+    weight_sum_diff = abs(weight_sum_model1 - weight_sum_model2)
+    logger.info("Weight sum difference: %s", weight_sum_diff)
+    weight_mean_diff = abs(weight_mean_model1 - weight_mean_model2)
+    logger.info("Weight mean difference: %s", weight_mean_diff)
+
+    # Set both models to eval mode
+    model1 = model1.eval()
+    model2 = model2.eval()
+
+    # Create identical inputs for both models
+    batch_size = 1
+    seq_len = 30
+
+    # Video latents [B, C, T, H, W]
+    hidden_states = torch.randn(
+        batch_size,
+        16,
+        21,
+        160,
+        90,
+        device=device,
+        dtype=precision,
+    )
+
+    # Text embeddings [B, L, D] (including global token)
+    encoder_hidden_states = torch.randn(
+        batch_size,
+        seq_len + 1,
+        4096,
+        device=device,
+        dtype=precision,
+    )
+
+    # Timestep
+    timestep = torch.tensor([500], device=device, dtype=precision)
+
+    forward_batch = ForwardBatch(data_type="dummy", )
+
+    with torch.amp.autocast("cuda", dtype=precision):
+        output1 = model1(
+            hidden_states=hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            timestep=timestep,
+            return_dict=False,
+        )[0]
+        with set_forward_context(
+                current_timestep=0,
+                attn_metadata=None,
+                forward_batch=forward_batch,
+        ):
+            output2 = model2(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                timestep=timestep,
+            )
+
+    # Print basic stats for debugging (cast to float32 for stability)
+    out1 = output1.detach().float()
+    out2 = output2.detach().float()
+    logger.info(
+        "output1 stats: min=%s max=%s mean=%s std=%s",
+        out1.min().item(),
+        out1.max().item(),
+        out1.mean().item(),
+        out1.std(unbiased=False).item(),
+    )
+    logger.info(
+        "output2 stats: min=%s max=%s mean=%s std=%s",
+        out2.min().item(),
+        out2.max().item(),
+        out2.mean().item(),
+        out2.std(unbiased=False).item(),
+    )
+
+    # Check if outputs have the same shape
+    assert (output1.shape == output2.shape
+            ), f"Output shapes don't match: {output1.shape} vs {output2.shape}"
+    assert (output1.dtype == output2.dtype
+            ), f"Output dtype don't match: {output1.dtype} vs {output2.dtype}"
+
+    # Check if outputs are similar (allowing for small numerical differences)
+    max_diff = torch.max(torch.abs(output1 - output2))
+    mean_diff = torch.mean(torch.abs(output1 - output2))
+    logger.info("Max Diff: %s", max_diff.item())
+    logger.info("Mean Diff: %s", mean_diff.item())
+    assert max_diff < 1e-1, f"Maximum difference between outputs: {max_diff.item()}"
+    # mean diff
+    assert mean_diff < 1e-2, f"Mean difference between outputs: {mean_diff.item()}"
+    '''
+    when --dit_precision "bf16", use_fsdp hardcoded to False:
+    INFO 01-19 22:01:24 [wan_grpo_utils.py:65] Model 1 weight sum: 395834.3506456231████████████████████████████████████████████████████████| 2/2 [00:00<00:00,  3.25it/s]
+    INFO 01-19 22:01:24 [wan_grpo_utils.py:66] Model 1 weight mean: 0.0002789536598289884
+    INFO 01-19 22:01:24 [wan_grpo_utils.py:75] Model 2 weight sum: 395125.463677882
+    INFO 01-19 22:01:24 [wan_grpo_utils.py:76] Model 2 weight mean: 0.0002739000890162162
+    INFO 01-19 22:01:24 [wan_grpo_utils.py:79] Weight sum difference: 708.8869677411276
+    INFO 01-19 22:01:24 [wan_grpo_utils.py:81] Weight mean difference: 5.053570812772192e-06
+    INFO 01-19 22:01:32 [wan_grpo_utils.py:139] output1 stats: min=-2.28125 max=1.921875 mean=-0.16638492047786713 std=0.458170622587204
+    INFO 01-19 22:01:32 [wan_grpo_utils.py:146] output2 stats: min=-2.296875 max=1.90625 mean=-0.166452556848526 std=0.4579130709171295
+    INFO 01-19 22:01:32 [wan_grpo_utils.py:165] Max Diff: 0.08984375
+    INFO 01-19 22:01:32 [wan_grpo_utils.py:166] Mean Diff: 0.0120849609375
+
+    when --dit_precision "fp32", use_fsdp not changed:
+
+    '''
+
+
+def sde_step_with_logprob(
+    scheduler: FlowUniPCMultistepScheduler,
+    model_output: torch.FloatTensor,
+    timestep: float | torch.FloatTensor,
+    sample: torch.FloatTensor,
+    prev_sample: torch.FloatTensor | None = None,
+    generator: torch.Generator | None = None,
+    deterministic: bool = False,
+    return_pixel_log_prob: bool = False,
+    return_dt_and_std_dev_t: bool = False
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, ...]:
+    """
+    Predict the sample from the previous timestep by reversing the SDE. 
+    This function propagates the flow process from the learned model outputs 
+    (most often the predicted velocity) and computes log probabilities.
+
+    Ported from FlowGRPO's sde_step_with_logprob to work with FastVideo's
+    FlowUniPCMultistepScheduler.
+
+    Args:
+        scheduler: FastVideo FlowUniPCMultistepScheduler instance
+        model_output: The direct output from learned flow model
+        timestep: The current discrete timestep in the diffusion chain
+        sample: A current instance of a sample created by the diffusion process
+        prev_sample: Optional previous sample (if provided, used instead of sampling)
+        generator: Optional random number generator
+        deterministic: If True, no noise is added (deterministic sampling)
+        return_pixel_log_prob: If True, return pixel-level log probabilities (not used)
+        return_dt_and_std_dev_t: If True, return dt and std_dev_t separately
+
+    Returns:
+        If return_dt_and_std_dev_t=True:
+            (prev_sample, log_prob, prev_sample_mean, std_dev_t, sqrt_dt)
+        Otherwise:
+            (prev_sample, log_prob, prev_sample_mean, std_dev_t * sqrt_dt)
+    """
+
+    # # Convert all variables to fp32 for numerical stability
+    # model_output = model_output.float()
+    # sample = sample.float()
+    # if prev_sample is not None:
+    #     prev_sample = prev_sample.float()
+
+    # Get step indices for current and previous timesteps
+    # Handle both single timestep and batch of timesteps
+    if isinstance(timestep, torch.Tensor):
+        if timestep.ndim == 0:
+            timestep = timestep.unsqueeze(0)
+        step_indices = [
+            scheduler.index_for_timestep(t.item()) for t in timestep
+        ]
+    else:
+        step_indices = [scheduler.index_for_timestep(timestep)]
+
+    prev_step_indices = [step + 1 for step in step_indices]
+
+    # Move sigmas to sample device
+    sigmas = scheduler.sigmas.to(sample.device)
+    # myregion debug: hardcode sigmas to flow_grpo's
+    sigmas = torch.Tensor([
+        0.9997, 0.9824, 0.9639, 0.9441, 0.9227, 0.8996, 0.8746, 0.8475, 0.8178,
+        0.7853, 0.7496, 0.7102, 0.6663, 0.6173, 0.5621, 0.4997, 0.4283, 0.3459,
+        0.2498, 0.1362, 0.0000
+    ]).to(sample.device, sample.dtype)
+    # end region
+
+    # Get sigma values for current and previous steps
+    sigma = sigmas[step_indices].view(-1, 1, 1, 1, 1)
+    sigma_prev = sigmas[prev_step_indices].view(-1, 1, 1, 1, 1)
+    sigma_max = sigmas[0].item()  # First sigma (highest)
+    sigma_min = sigmas[-1].item()  # Last sigma (lowest)
+
+    dt = sigma_prev - sigma
+
+    # myregion debug
+    print(f"[DEBUG]: sigma_max: {sigma_max}, sigma_min: {sigma_min}, dt: {dt}")
+    print(f"[DEBUG]: in sde_step_with_logprob(), timestep: {timestep}")
+    print(f"[DEBUG]: in sde_step_with_logprob(), sigmas: {sigmas}")
+    print(f"[DEBUG]: in sde_step_with_logprob(), step_indices: {step_indices}")
+    print(
+        f"[DEBUG]: in sde_step_with_logprob(), prev_step_indices: {prev_step_indices}"
+    )
+    '''
+    [DEBUG]: in sde_step_with_logprob(), timestep: tensor([428, 428, 428, 428], device='cuda:0')
+    [DEBUG]: in sde_step_with_logprob(), sigmas: tensor([0.9999, 0.9826, 0.9642, 0.9443, 0.9230, 0.8999, 0.8749, 0.8477, 0.8181,
+            0.7856, 0.7499, 0.7104, 0.6665, 0.6175, 0.5624, 0.4999, 0.4285, 0.3461,
+            0.2499, 0.1363, 0.0000], device='cuda:0')
+    [DEBUG]: in sde_step_with_logprob(), step_indices: [16, 16, 16, 16]
+    [DEBUG]: in sde_step_with_logprob(), prev_step_indices: [17, 17, 17, 17]
+
+    DEBUG]: in sde_step_with_logprob(), timestep: tensor([249], device='cuda:0')███████████████▎       | 18/20 [00:04<00:00,  3.87step/s, step_time=0.26s, timestep=346.0]
+    [DEBUG]: in sde_step_with_logprob(), sigmas: tensor([0.9999, 0.9826, 0.9642, 0.9443, 0.9230, 0.8999, 0.8749, 0.8477, 0.8181,
+            0.7856, 0.7499, 0.7104, 0.6665, 0.6175, 0.5624, 0.4999, 0.4285, 0.3461,
+            0.2499, 0.1363, 0.0000], device='cuda:0')
+    [DEBUG]: in sde_step_with_logprob(), step_indices: [18]
+    [DEBUG]: in sde_step_with_logprob(), prev_step_indices: [19]
+
+    [DEBUG]: in sde_step_with_logprob(), timestep: tensor([617, 617, 617, 617], device='cuda:0')
+    [DEBUG]: in sde_step_with_logprob(), sigmas: tensor([0.9999, 0.9826, 0.9642, 0.9443, 0.9230, 0.8999, 0.8749, 0.8477, 0.8181,
+            0.7856, 0.7499, 0.7104, 0.6665, 0.6175, 0.5624, 0.4999, 0.4285, 0.3461,
+            0.2499, 0.1363, 0.0000], device='cuda:0')
+    [DEBUG]: in sde_step_with_logprob(), step_indices: [13, 13, 13, 13]
+    [DEBUG]: in sde_step_with_logprob(), prev_step_indices: [14, 14, 14, 14]
+    '''
+    # endregion
+
+    # Compute std_dev_t and prev_sample_mean using SDE formulation
+    std_dev_t = sigma_min + (sigma_max - sigma_min) * sigma
+    prev_sample_mean = (sample * (1 + std_dev_t**2 / (2 * sigma) * dt) +
+                        model_output * (1 + std_dev_t**2 * (1 - sigma) /
+                                        (2 * sigma)) * dt)
+
+    if prev_sample is not None and generator is not None:
+        raise ValueError(
+            "Cannot pass both generator and prev_sample. Please make sure that either `generator` or"
+            " `prev_sample` stays `None`.")
+
+    # Sample prev_sample if not provided
+    if prev_sample is None:
+        variance_noise = randn_tensor(
+            model_output.shape,
+            generator=generator,
+            device=model_output.device,
+            dtype=model_output.dtype,
+        )
+        sqrt_dt = torch.sqrt(-1 * dt)  # dt is negative (going backwards)
+        prev_sample = prev_sample_mean + std_dev_t * sqrt_dt * variance_noise
+    else:
+        sqrt_dt = torch.sqrt(-1 * dt)
+
+    # No noise is added during evaluation (deterministic)
+    if deterministic:
+        prev_sample = sample + dt * model_output
+        sqrt_dt = torch.sqrt(-1 * dt)
+
+    # Compute log probability: log p(prev_sample | sample, model_output)
+    # Assuming Gaussian distribution: N(prev_sample_mean, (std_dev_t * sqrt_dt)^2)
+    std_dev_sqrt_dt = std_dev_t * sqrt_dt
+    log_prob = (
+        -((prev_sample.detach() - prev_sample_mean)**2) /
+        (2 * (std_dev_sqrt_dt**2)) - torch.log(
+            std_dev_sqrt_dt + 1e-8)  # Add small epsilon for numerical stability
+        - torch.log(
+            torch.sqrt(2 * torch.as_tensor(math.pi, device=sample.device))))
+
+    # Mean along all but batch dimension
+    log_prob = log_prob.mean(dim=tuple(range(1, log_prob.ndim)))
+
+    if return_dt_and_std_dev_t:
+        return prev_sample, log_prob, prev_sample_mean, std_dev_t, sqrt_dt
+    return prev_sample, log_prob, prev_sample_mean, std_dev_t * sqrt_dt
+
+
+def wan_pipeline_with_logprob(
+    pipeline,
+    prompt: str | list[str] = None,
+    negative_prompt: str | list[str] = None,
+    height: int = 480,
+    width: int = 832,
+    num_frames: int = 81,
+    num_inference_steps: int = 50,
+    guidance_scale: float = 5.0,
+    num_videos_per_prompt: int | None = 1,
+    generator: torch.Generator | list[torch.Generator] | None = None,
+    latents: torch.Tensor | None = None,
+    prompt_embeds: torch.Tensor | None = None,
+    negative_prompt_embeds: torch.Tensor | None = None,
+    output_type: str | None = "pt",
+    return_dict: bool = False,
+    attention_kwargs: dict[str, Any] | None = None,
+    max_sequence_length: int = 512,
+    deterministic: bool = False,
+    kl_reward: float = 0.0,
+    return_pixel_log_prob: bool = False,
+) -> tuple[torch.Tensor, list[torch.Tensor], list[torch.Tensor],
+           list[torch.Tensor], torch.Tensor | None]:
+    """
+    Wan pipeline with log probability computation for GRPO training.
+
+    Ported from FlowGRPO's wan_pipeline_with_logprob to work with FastVideo's WanPipeline.
+    This function generates videos and computes log probabilities at each denoising step.
+
+    Args:
+        pipeline: FastVideo WanPipeline instance
+        prompt: Text prompt(s) for generation
+        negative_prompt: Negative prompt(s) for classifier-free guidance
+        height: Height of generated video
+        width: Width of generated video
+        num_frames: Number of frames in generated video
+        num_inference_steps: Number of denoising steps
+        guidance_scale: Classifier-free guidance scale
+        num_videos_per_prompt: Number of videos to generate per prompt
+        generator: Random generator for reproducibility
+        latents: Optional initial latents
+        prompt_embeds: Optional pre-computed prompt embeddings
+        negative_prompt_embeds: Optional pre-computed negative prompt embeddings
+        output_type: Output type ("pt" for PyTorch tensor, "np" for numpy, "latent" for latents only)
+        return_dict: Whether to return dict (not used, always returns tuple)
+        attention_kwargs: Optional attention kwargs
+        max_sequence_length: Maximum sequence length for text encoding
+        deterministic: If True, use deterministic sampling (no noise)
+        kl_reward: KL reward coefficient (if > 0, computes KL divergence)
+        return_pixel_log_prob: If True, return pixel-level log probabilities (not used)
+
+    Returns:
+        Tuple of:
+            - video: Generated video tensor [B, C, T, H, W] or latents if output_type="latent"
+            - all_latents: List of latents at each step [num_steps+1] of shape [B, C, T, H, W]
+            - all_log_probs: List of log probabilities at each step [num_steps] of shape [B]
+            - all_kl: List of KL divergences at each step [num_steps] of shape [B] (if kl_reward > 0)
+            - prompt_ids: Tokenized prompt IDs [B, seq_len] (None if prompt_embeds were provided)
+    """
+    # Get device from transformer
+    transformer = pipeline.get_module("transformer")
+
+    # myregion debug: test transformer output
+    logger.info("testing transformer, running test_wan_transformer2")
+    test_wan_transformer()
+    # test_wan_transformer2(transformer)
+    # endregion
+
+    # hardcode dtype for debug
+    # transformer_dtype = torch.float32
+    # use get_compute_dtype() to get dtype based on mixed precision
+    transformer_dtype = get_compute_dtype()
+    logger.info(f"[DEBUG]: transformer_dtype: {transformer_dtype}")
+
+    # Get scheduler and other modules
+    scheduler = pipeline.get_module("scheduler")
+    vae = pipeline.get_module("vae")
+
+    # Determine batch size
+    if prompt is not None and isinstance(prompt, str):
+        batch_size = 1
+    elif prompt is not None and isinstance(prompt, list):
+        batch_size = len(prompt)
+    elif prompt_embeds is not None:
+        batch_size = prompt_embeds.shape[0]
+    else:
+        raise ValueError("Either prompt or prompt_embeds must be provided")
+
+    # Encode prompts if not provided
+    prompt_ids = None
+    if prompt_embeds is None:
+        # Encode prompts directly using text encoder and tokenizer
+        # This is a simplified encoding - for full pipeline encoding, use TextEncodingStage
+        text_encoder = pipeline.get_module("text_encoder")
+        tokenizer = pipeline.get_module("tokenizer")
+
+        # Normalize to list
+        if isinstance(prompt, str):
+            prompts_list = [prompt]
+        else:
+            prompts_list = prompt
+
+        # Tokenize prompts
+        text_inputs = tokenizer(prompts_list,
+                                padding="max_length",
+                                max_length=max_sequence_length,
+                                truncation=True,
+                                return_tensors="pt").to(pipeline.device)
+
+        # Store prompt_ids for return
+        prompt_ids = text_inputs["input_ids"]
+
+        # Encode with text encoder
+        with torch.no_grad():
+            outputs = text_encoder(
+                text_inputs["input_ids"],
+                attention_mask=text_inputs["attention_mask"],
+                output_hidden_states=True,
+            )
+            # Get last hidden state (Wan typically uses last hidden state)
+            prompt_embeds = outputs.last_hidden_state
+
+        # Encode negative prompts if CFG is enabled
+        if guidance_scale > 1.0:
+            if negative_prompt is None:
+                negative_prompt = [""] * len(prompts_list)
+            elif isinstance(negative_prompt, str):
+                negative_prompt = [negative_prompt]
+
+            neg_text_inputs = tokenizer(negative_prompt,
+                                        padding="max_length",
+                                        max_length=max_sequence_length,
+                                        truncation=True,
+                                        return_tensors="pt").to(pipeline.device)
+
+            with torch.no_grad():
+                neg_outputs = text_encoder(
+                    neg_text_inputs["input_ids"],
+                    attention_mask=neg_text_inputs["attention_mask"],
+                    output_hidden_states=True,
+                )
+                negative_prompt_embeds = neg_outputs.last_hidden_state
+        else:
+            negative_prompt_embeds = None
+
+# myregion Debug: Print shapes of prompt embeddings
+        logger.info(
+            f"After encoding - prompt_embeds shape: {prompt_embeds.shape if prompt_embeds is not None else None}"
+        )
+        logger.info(
+            f"After encoding - negative_prompt_embeds shape: {negative_prompt_embeds.shape if negative_prompt_embeds is not None else None}"
+        )
+        logger.info(
+            f"After encoding - prompt_embeds dtype: {prompt_embeds.dtype if prompt_embeds is not None else None}"
+        )
+        logger.info(
+            f"After encoding - negative_prompt_embeds dtype: {negative_prompt_embeds.dtype if negative_prompt_embeds is not None else None}"
+        )
+        '''
+        INFO 01-17 05:31:13 [wan_grpo_utils.py:290] After encoding - prompt_embeds shape: torch.Size([4, 512, 4096])
+        INFO 01-17 05:31:13 [wan_grpo_utils.py:291] After encoding - negative_prompt_embeds shape: None
+        INFO 01-17 05:31:13 [wan_grpo_utils.py:292] After encoding - prompt_embeds dtype: torch.float32
+        INFO 01-17 05:31:13 [wan_grpo_utils.py:293] After encoding - negative_prompt_embeds dtype: None
+        '''
+# endregion
+# logger.info("wan_pipeline_with_logprob's transformer class type: %s", type(transformer))
+# logger.info("Variables in transformer: %s", str(dir(transformer)))
+    prompt_embeds = prompt_embeds.to(transformer_dtype)
+    if negative_prompt_embeds is not None:
+        negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+    # Prepare timesteps
+    scheduler.set_timesteps(num_inference_steps, device=pipeline.device)
+    timesteps = scheduler.timesteps
+
+    # Prepare latent variables
+    num_channels_latents = transformer.config.in_channels
+    vae = pipeline.get_module("vae")
+    # Get VAE scale factors
+    vae_scale_factor_spatial = vae.spatial_compression_ratio
+    vae_scale_factor_temporal = vae.temporal_compression_ratio
+
+    if latents is None:
+        # Generate random latents
+        # Note: num_frames in latents accounts for temporal compression
+        num_latent_frames = (num_frames - 1) // vae_scale_factor_temporal + 1
+        latents_shape = (
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            num_latent_frames,
+            height // vae_scale_factor_spatial,
+            width // vae_scale_factor_spatial,
+        )
+        if generator is not None:
+            if isinstance(generator, list):
+                latents = [
+                    torch.randn(
+                        latents_shape[1:],
+                        generator=gen,
+                        device=pipeline.device,
+                        dtype=transformer_dtype,
+                    ) for gen in generator
+                ]
+                latents = torch.stack(latents, dim=0)
+            else:
+                latents = torch.randn(
+                    latents_shape,
+                    generator=generator,
+                    device=pipeline.device,
+                    dtype=transformer_dtype,
+                )
+        else:
+            latents = torch.randn(latents_shape,
+                                  device=pipeline.device,
+                                  dtype=transformer_dtype)
+    else:
+        latents = latents.to(device=pipeline.device, dtype=transformer_dtype)
+
+
+# myregion Debug: Print latents shape, dtype, and value range
+    logger.info("=" * 80)
+    logger.info("Latents Debug Information:")
+    logger.info(f"  Shape: {latents.shape}")
+    logger.info(f"  Dtype: {latents.dtype}")
+    logger.info(f"  Min value: {latents.min().item():.6f}")
+    logger.info(f"  Max value: {latents.max().item():.6f}")
+    logger.info(f"  Mean value: {latents.mean().item():.6f}")
+    logger.info(f"  Std value: {latents.std().item():.6f}")
+    logger.info(f"  Device: {latents.device}")
+    logger.info("=" * 80)
+    '''
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:355] ================================================================================
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:356] Latents Debug Information:
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:357]   Shape: torch.Size([4, 16, 9, 30, 52])
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:358]   Dtype: torch.bfloat16
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:359]   Min value: -4.500000
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:360]   Max value: 4.656250
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:361]   Mean value: 0.000111
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:362]   Std value: 1.000000
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:363]   Device: cuda:0
+    INFO 01-17 07:41:33 [wan_grpo_utils.py:364] ================================================================================
+    '''
+    # endregion
+
+    all_latents = [latents]
+    all_log_probs = []
+    all_kl = []
+
+    # myregion Debug
+    logger.info("Tensor type issue debugging:")
+    logger.info(f"latents: {type(latents)}")
+    logger.info(f"prompt_embeds: {type(prompt_embeds)}")
+    logger.info(
+        f"[DEBUG]: before denoising loop: type(timesteps): {type(timesteps)}")
+    logger.info(
+        f"[DEBUG]: before denoising loop: timesteps.shape: {timesteps.shape}")
+    # endregion
+
+    # Progress bar for denoising loop
+    progress_bar = tqdm(enumerate(timesteps),
+                        total=len(timesteps),
+                        desc="Denoising steps",
+                        unit="step")
+
+    for i, t in progress_bar:
+        step_start_time = time.time()
+        latents_ori = latents.clone()
+        timestep = t.expand(latents.shape[0]) if isinstance(
+            t, torch.Tensor) else torch.tensor([t] * latents.shape[0],
+                                               device=pipeline.device)
+
+        logger.info(
+            f"[DEBUG]: before set_forward_context: current_timestep=i:{i}")
+        # Predict noise with transformer
+        with set_forward_context(
+                current_timestep=t.item(),
+                attn_metadata=None,
+                forward_batch=None,
+        ):
+            noise_pred = transformer(
+                hidden_states=latents,
+                timestep=timestep,
+                encoder_hidden_states=prompt_embeds,
+                attention_kwargs=attention_kwargs,
+                return_dict=False,
+            )[0]
+        noise_pred = noise_pred.to(prompt_embeds.dtype)
+
+        # Classifier-free guidance
+        if guidance_scale > 1.0:
+            with set_forward_context(
+                    current_timestep=i,
+                    attn_metadata=None,
+                    forward_batch=None,
+            ):
+                noise_uncond = transformer(
+                    hidden_states=latents,
+                    timestep=timestep,
+                    encoder_hidden_states=negative_prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+            noise_pred = noise_uncond + guidance_scale * (noise_pred -
+                                                          noise_uncond)
+
+        # SDE step with log probability
+        latents, log_prob, prev_latents_mean, std_dev_t = sde_step_with_logprob(
+            scheduler,
+            noise_pred,  #.float(),
+            t.unsqueeze(0) if isinstance(t, torch.Tensor) else t,
+            latents,  #.float(),
+            deterministic=deterministic,
+            return_pixel_log_prob=return_pixel_log_prob)
+        # sde_step_with_logprob returns fp32
+        # latents = latents.to(transformer_dtype)
+        prev_latents = latents.clone()
+
+        all_latents.append(latents)
+        all_log_probs.append(log_prob)
+
+        # Compute KL divergence if kl_reward > 0 (for KL reward in sampling)
+        if kl_reward > 0 and not deterministic:
+            # Use reference model (disable adapter if using LoRA)
+            latent_model_input_ref = torch.cat(
+                [latents_ori] * 2) if guidance_scale > 1.0 else latents_ori
+            with set_forward_context(
+                    current_timestep=i,
+                    attn_metadata=None,
+                    forward_batch=None,
+            ):
+                with transformer.disable_adapter() if hasattr(
+                        transformer, 'disable_adapter') else torch.no_grad():
+                    noise_pred_ref = transformer(
+                        hidden_states=latent_model_input_ref,
+                        timestep=timestep,
+                        encoder_hidden_states=prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+            noise_pred_ref = noise_pred_ref.to(prompt_embeds.dtype)
+
+            # Perform guidance for reference model
+            if guidance_scale > 1.0:
+                noise_pred_uncond_ref, noise_pred_text_ref = noise_pred_ref.chunk(
+                    2)
+                noise_pred_ref = noise_pred_uncond_ref + guidance_scale * (
+                    noise_pred_text_ref - noise_pred_uncond_ref)
+
+            # Compute reference log prob
+            _, ref_log_prob, ref_prev_latents_mean, ref_std_dev_t = sde_step_with_logprob(
+                scheduler,
+                noise_pred_ref.float(),
+                t.unsqueeze(0) if isinstance(t, torch.Tensor) else t,
+                latents_ori.float(),
+                prev_sample=prev_latents.float(),
+                deterministic=deterministic,
+            )
+
+            # Compute KL divergence: KL = (mean_diff)^2 / (2 * std^2)
+            assert torch.allclose(
+                std_dev_t, ref_std_dev_t
+            ), "std_dev_t should match between current and reference"
+            kl = (prev_latents_mean - ref_prev_latents_mean)**2 / (2 *
+                                                                   std_dev_t**2)
+            kl = kl.mean(dim=tuple(range(1, kl.ndim)))
+            all_kl.append(kl)
+        else:
+            # No KL reward, set to zero
+            all_kl.append(torch.zeros(len(latents), device=latents.device))
+
+        # Update progress bar with timing information
+        step_time = time.time() - step_start_time
+        progress_bar.set_postfix({
+            "step_time":
+            f"{step_time:.2f}s",
+            "timestep":
+            f"{t.item() if isinstance(t, torch.Tensor) else t:.1f}"
+        })
+
+    # Decode latents to video if needed
+    if output_type != "latent":
+        latents = latents.to(vae.dtype)
+
+        # Apply VAE normalization (Wan VAE specific)
+        # Wan VAE requires denormalization before decoding
+        if hasattr(vae, 'config') and hasattr(vae.config,
+                                              'latents_mean') and hasattr(
+                                                  vae.config, 'latents_std'):
+            # Get z_dim from config or VAE
+            z_dim = getattr(vae.config, 'z_dim', latents.shape[1])
+            latents_mean = (torch.tensor(vae.config.latents_mean,
+                                         device=latents.device,
+                                         dtype=latents.dtype).view(
+                                             1, z_dim, 1, 1, 1))
+            latents_std = (
+                1.0 / torch.tensor(vae.config.latents_std,
+                                   device=latents.device,
+                                   dtype=latents.dtype).view(1, z_dim, 1, 1, 1))
+            latents = latents / latents_std + latents_mean
+        elif hasattr(vae, 'latents_mean') and hasattr(vae, 'latents_std'):
+            # Alternative: check if latents_mean/std are direct attributes
+            z_dim = latents.shape[1]
+            latents_mean = (torch.tensor(vae.latents_mean,
+                                         device=latents.device,
+                                         dtype=latents.dtype).view(
+                                             1, z_dim, 1, 1, 1))
+            latents_std = (1.0 / torch.tensor(
+                vae.latents_std, device=latents.device,
+                dtype=latents.dtype).view(1, z_dim, 1, 1, 1))
+            latents = latents / latents_std + latents_mean
+
+        # Decode using VAE
+        with torch.no_grad():
+            video = vae.decode(latents.float(), return_dict=False)[0]
+            # VAE.decode returns tensor directly (not tuple)
+
+            # Postprocess video: convert from [-1, 1] to [0, 1]
+            # FastVideo VAE typically outputs in [-1, 1] range
+            video = (video / 2 + 0.5).clamp(0, 1)
+    else:
+        video = latents
+
+    return video, all_latents, all_log_probs, all_kl, prompt_ids

fastvideo/training/training_pipeline.py

@@ -174,17 +174,18 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 last_epoch=self.init_steps - 1,
             )
 
-        self.train_dataset, self.train_dataloader = build_parquet_map_style_dataloader(
-            training_args.data_path,
-            training_args.train_batch_size,
-            parquet_schema=self.train_dataset_schema,
-            num_data_workers=training_args.dataloader_num_workers,
-            cfg_rate=training_args.training_cfg_rate,
-            drop_last=True,
-            text_padding_length=training_args.pipeline_config.
-            text_encoder_configs[0].arch_config.
-            text_len,  # type: ignore[attr-defined]
-            seed=self.seed)
+        if not self.training_args.rl_args.rl_mode:
+            self.train_dataset, self.train_dataloader = build_parquet_map_style_dataloader(
+                training_args.data_path,
+                training_args.train_batch_size,
+                parquet_schema=self.train_dataset_schema,
+                num_data_workers=training_args.dataloader_num_workers,
+                cfg_rate=training_args.training_cfg_rate,
+                drop_last=True,
+                text_padding_length=training_args.pipeline_config.
+                text_encoder_configs[0].arch_config.
+                text_len,  # type: ignore[attr-defined]
+                seed=self.seed)
 
         self.noise_scheduler = noise_scheduler
         if self.training_args.boundary_ratio is not None:
@@ -192,19 +193,21 @@ class TrainingPipeline(LoRAPipeline, ABC):
         else:
             self.boundary_timestep = None
 
-        logger.info("train_dataloader length: %s", len(self.train_dataloader))
+        if not self.training_args.rl_args.rl_mode:
+            logger.info("train_dataloader length: %s", len(self.train_dataloader))
         logger.info("train_sp_batch_size: %s",
                     training_args.train_sp_batch_size)
         logger.info("gradient_accumulation_steps: %s",
                     training_args.gradient_accumulation_steps)
         logger.info("sp_size: %s", training_args.sp_size)
 
-        self.num_update_steps_per_epoch = math.ceil(
-            len(self.train_dataloader) /
-            training_args.gradient_accumulation_steps * training_args.sp_size /
-            training_args.train_sp_batch_size)
-        self.num_train_epochs = math.ceil(training_args.max_train_steps /
-                                          self.num_update_steps_per_epoch)
+        if not self.training_args.rl_args.rl_mode:
+            self.num_update_steps_per_epoch = math.ceil(
+                len(self.train_dataloader) /
+                training_args.gradient_accumulation_steps * training_args.sp_size /
+                training_args.train_sp_batch_size)
+            self.num_train_epochs = math.ceil(training_args.max_train_steps /
+                                            self.num_update_steps_per_epoch)
 
         # TODO(will): is there a cleaner way to track epochs?
         self.current_epoch = 0
@@ -575,8 +578,12 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 round(num_trainable_params / 1e9, 3))
 
         # Set random seeds for deterministic training
-        self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
-            self.seed)
+        if not self.training_args.rl_args.rl_mode:
+            self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
+                self.seed)
+        else:
+            self.noise_random_generator = torch.Generator(device=self.device).manual_seed(
+                self.seed)
         self.noise_gen_cuda = torch.Generator(
             device=current_platform.device_name).manual_seed(self.seed)
         self.validation_random_generator = torch.Generator(

fastvideo/training/wan_rl_training_pipeline.py

@@ -0,0 +1,100 @@
+# SPDX-License-Identifier: Apache-2.0
+import sys
+from copy import deepcopy
+
+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_pipeline import WanPipeline
+from fastvideo.training.rl.rl_pipeline import RLPipeline
+from fastvideo.utils import is_vsa_available
+
+vsa_available = is_vsa_available()
+
+logger = init_logger(__name__)
+
+
+class WanRLTrainingPipeline(RLPipeline):
+    """
+    A training pipeline for Wan with RL/GRPO support.
+    
+    This pipeline extends RLPipeline with Wan-specific initialization.
+    """
+    _required_config_modules = [
+        "scheduler", "transformer", "vae", "text_encoder", "tokenizer"
+    ]
+
+    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 _create_inference_pipeline(self, training_args: TrainingArgs,
+                                   dit_cpu_offload: bool):
+        args_copy = deepcopy(training_args)
+        args_copy.inference_mode = True
+        loaded_modules = {
+            "transformer": self.get_module("transformer"),
+        }
+        transformer_2 = self.get_module("transformer_2", None)
+        if transformer_2 is not None:
+            loaded_modules["transformer_2"] = transformer_2
+        text_encoder = self.get_module("text_encoder", None)
+        if text_encoder is not None:
+            loaded_modules["text_encoder"] = text_encoder
+        tokenizer = self.get_module("tokenizer", None)
+        if tokenizer is not None:
+            loaded_modules["tokenizer"] = tokenizer
+        vae = self.get_module("vae", None)
+        if vae is not None:
+            loaded_modules["vae"] = vae
+
+        return WanPipeline.from_pretrained(
+            training_args.model_path,
+            args=args_copy,  # type: ignore
+            inference_mode=True,
+            loaded_modules=loaded_modules,
+            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=dit_cpu_offload)
+
+    def initialize_validation_pipeline(self, training_args: TrainingArgs):
+        logger.info("Initializing validation pipeline...")
+        self.validation_pipeline = self._create_inference_pipeline(
+            training_args, dit_cpu_offload=True)
+
+    def _build_sampling_pipeline(self, training_args: TrainingArgs):
+        return self._create_inference_pipeline(training_args,
+                                               dit_cpu_offload=False)
+
+
+def main(args) -> None:
+    logger.info("Starting RL training pipeline...")
+
+    pipeline = WanRLTrainingPipeline.from_pretrained(
+        args.pretrained_model_name_or_path, args=args)
+    args = pipeline.training_args
+    pipeline.train()
+    logger.info("RL 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
+    # Enable RL mode
+    args.rl_mode = True
+    main(args)

fastvideo/utils.py

@@ -22,7 +22,7 @@ import threading
 import traceback
 from collections.abc import Callable
 from dataclasses import dataclass, fields, is_dataclass
-from functools import lru_cache, partial, wraps
+from functools import lru_cache, partial, wraps, cache
 from pathlib import Path
 from typing import Any, TextIO, TypeVar, cast
 
@@ -1193,3 +1193,30 @@ def decorate_logs(process_name: str | None = None) -> None:
     pid = os.getpid()
     _add_prefix(sys.stdout, process_name, pid)
     _add_prefix(sys.stderr, process_name, pid)
+
+
+def _probe_pin_memory() -> bool:
+    from fastvideo.platforms import current_platform
+    if current_platform.is_cpu() or current_platform.is_mps(
+    ) or current_platform.is_npu():
+        return False
+
+    try:
+        if torch.cuda.is_available():
+            torch.cuda.current_device()
+        _ = torch.empty(1024, device="cpu").pin_memory()
+        _ = torch.empty(1024, device="cpu", pin_memory=True)
+    except Exception as exc:
+        logger.warning("Pinned memory is unavailable: %s", exc)
+        return False
+
+    return True
+
+
+@cache
+def _cached_pin_memory_available(pid: int) -> bool:
+    return _probe_pin_memory()
+
+
+def is_pin_memory_available() -> bool:
+    return _cached_pin_memory_available(os.getpid())

pyproject.toml

@@ -63,7 +63,7 @@ dependencies = [
     "remote-pdb",
 
     # Kernel & Packaging
-    "fastvideo-kernel==0.2.2",
+    "fastvideo-kernel==0.2.4",
     "wheel",
 
     # Training Dependencies
@@ -111,7 +111,7 @@ lint = [
 ]
 
 test = [
-    "av==14.3.0",
+    "av",
     "pytorch-msssim==1.0.0",
     "pytest",
 ]

pyproject_other.toml

@@ -63,7 +63,7 @@ dependencies = [
     "remote-pdb",
 
     # Kernel & Packaging
-    "fastvideo-kernel==0.2.2",
+    "fastvideo-kernel==0.2.4",
     "wheel",
 
     # Training Dependencies
@@ -90,7 +90,7 @@ lint = [
 ]
 
 test = [
-    "av==14.3.0",
+    "av",
     "pytorch-msssim==1.0.0",
     "pytest",
 ]

scripts/checkpoint_conversion/pt_to_safetensors.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+"""Convert a PyTorch checkpoint (.pt) to a safetensors file."""
+
+import argparse
+from pathlib import Path
+
+import torch
+from safetensors.torch import save_file
+
+
+def convert_pt_to_safetensors(
+    input_path: str,
+    output_path: str,
+    key: str | None = None,
+    force: bool = False,
+    skip_patterns: list[str] | None = None,
+):
+    input_path = Path(input_path)
+    output_path = Path(output_path)
+
+    if not input_path.exists():
+        raise FileNotFoundError(f"Input file not found: {input_path}")
+
+    if output_path.exists() and not force:
+        raise FileExistsError(
+            f"Output file already exists: {output_path}. Use --force to overwrite."
+        )
+
+    checkpoint = torch.load(input_path, map_location="cpu")
+
+    state_dict: dict[str, torch.Tensor]
+    if isinstance(checkpoint, dict):
+        if key is not None:
+            if key not in checkpoint:
+                raise KeyError(f"Key {key!r} not found in checkpoint.")
+            state_dict = checkpoint[key]
+        else:
+            for k in ("state_dict", "model_state_dict", "model", "ema"):
+                if k in checkpoint:
+                    state_dict = checkpoint[k]
+                    break
+            else:
+                state_dict = checkpoint
+    else:
+        state_dict = checkpoint
+
+    if not isinstance(state_dict, dict):
+        raise TypeError(f"Expected a dict state_dict, got {type(state_dict)}")
+
+    if skip_patterns:
+        state_dict = {
+            k: v
+            for k, v in state_dict.items()
+            if not any(pat in k for pat in skip_patterns)
+        }
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    save_file(state_dict, str(output_path))
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert a PyTorch checkpoint (.pt) to safetensors."
+    )
+    parser.add_argument(
+        "input",
+        type=str,
+        help="Path to input .pt checkpoint file"
+    )
+    parser.add_argument(
+        "output",
+        type=str,
+        help="Path to output .safetensors file"
+    )
+    parser.add_argument(
+        "--key",
+        type=str,
+        default=None,
+        help="Optional key to extract from checkpoint dict (e.g. 'model_state_dict')"
+    )
+    parser.add_argument(
+        "--force",
+        action="store_true",
+        help="Overwrite output file if it exists"
+    )
+    parser.add_argument(
+        "--skip-pattern",
+        action="append",
+        dest="skip_patterns",
+        help="Parameter name patterns to skip (can be used multiple times)"
+    )
+
+    args = parser.parse_args()
+
+    convert_pt_to_safetensors(
+        args.input,
+        args.output,
+        args.key,
+        args.force,
+        args.skip_patterns
+    )
+
+
+if __name__ == "__main__":
+    main()
+