diff --git a/.github/unittest/linux/scripts/run_all.sh b/.github/unittest/linux/scripts/run_all.sh
index 0e6b738ae6d..4fb11ce4508 100755
--- a/.github/unittest/linux/scripts/run_all.sh
+++ b/.github/unittest/linux/scripts/run_all.sh
@@ -130,7 +130,6 @@ uv_pip_install \
   wandb \
   mlflow \
   av \
-  torchcodec \
   coverage \
   transformers \
   ninja \
@@ -240,6 +239,17 @@ else
   uv_pip_install --no-deps tensordict
 fi
 
+# install torchcodec (from source for nightly to match PyTorch ABI)
+if [[ "$RELEASE" == 0 ]]; then
+  export BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1
+  export Python3_ROOT_DIR="$(python -c 'import sys; print(sys.base_prefix)')"
+  uv_pip_install --no-build-isolation git+https://github.com/meta-pytorch/torchcodec.git
+  unset BUILD_AGAINST_ALL_FFMPEG_FROM_S3
+  unset Python3_ROOT_DIR
+else
+  uv_pip_install torchcodec
+fi
+
 printf "* Installing torchrl\n"
 if [[ "$RELEASE" == 0 ]]; then
   uv_pip_install -e . --no-build-isolation --no-deps
diff --git a/.github/workflows/docs.yml b/.github/workflows/docs.yml
index d4c88a010b4..2ed7fc7222a 100644
--- a/.github/workflows/docs.yml
+++ b/.github/workflows/docs.yml
@@ -44,7 +44,8 @@ jobs:
         # Install system dependencies (Ubuntu-based PyTorch image)
         apt-get update
         apt-get install -y libglfw3 libglfw3-dev libgl1-mesa-glx libgl1-mesa-dev libegl1-mesa-dev \
-          freeglut3-dev libglu1-mesa libegl1 mesa-utils xvfb wget git cmake
+          freeglut3-dev libglu1-mesa libegl1 mesa-utils xvfb wget git cmake ffmpeg \
+          libavcodec-dev libavformat-dev libavutil-dev libswscale-dev libswresample-dev
         
         # 2. upgrade pip, ninja and packaging
         python -m pip install --upgrade pip
@@ -65,10 +66,12 @@ jobs:
           # Use stable PyTorch for releases (tags and release/* branches)
           python -m pip install torch torchvision --quiet --root-user-action=ignore
           python -m pip install tensordict --quiet --root-user-action=ignore
+          python -m pip install torchcodec --quiet --root-user-action=ignore
         else
           # Use nightly PyTorch for main, nightly, and PRs
           python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cpu -U --quiet --root-user-action=ignore
           python -m pip install git+https://github.com/pytorch/tensordict.git --quiet --root-user-action=ignore
+          BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1 python -m pip install --no-build-isolation git+https://github.com/meta-pytorch/torchcodec.git --quiet --root-user-action=ignore
         fi
         
         # 6. Install doc requirements BEFORE TorchRL to ensure gymnasium is available
diff --git a/test/test_loggers.py b/test/test_loggers.py
index e9a82c6fcff..ebec8a2d23f 100644
--- a/test/test_loggers.py
+++ b/test/test_loggers.py
@@ -404,9 +404,18 @@ def test_log_video(self, steps, mlflow_fixture):
             videos_dir = client.download_artifacts(run_id, "videos", artifacts_dir)
             for i, video_name in enumerate(os.listdir(videos_dir)):
                 video_path = os.path.join(videos_dir, video_name)
-                loaded_video, _, _ = torchvision.io.read_video(
-                    video_path, pts_unit="sec", output_format="TCHW"
-                )
+                if _has_torchcodec:
+                    from torchcodec.decoders import VideoDecoder
+
+                    loaded_video = (
+                        VideoDecoder(video_path)
+                        .get_frames_in_range(start=0, stop=128)
+                        .data
+                    )
+                else:
+                    loaded_video, _, _ = torchvision.io.read_video(
+                        video_path, pts_unit="sec", output_format="TCHW"
+                    )
                 if steps:
                     assert torch.allclose(loaded_video.int(), videos[i].int(), rtol=0.1)
                 else:
diff --git a/torchrl/record/loggers/common.py b/torchrl/record/loggers/common.py
index 80f7dbe3ac9..b0ad414c5e8 100644
--- a/torchrl/record/loggers/common.py
+++ b/torchrl/record/loggers/common.py
@@ -57,8 +57,8 @@ def _write_video(filename, video_array, **kwargs):  # noqa: F811
     preset = options.pop("preset", None)
     pixel_format = options.pop("pixel_format", None)
 
-    # VideoEncoder expects (N, C, H, W); callers pass (T, H, W, C)
-    video_array = video_array.permute(0, 3, 1, 2).contiguous()
+    # VideoEncoder expects (N, C, H, W) uint8; callers pass (T, H, W, C)
+    video_array = video_array.permute(0, 3, 1, 2).contiguous().to(torch.uint8)
 
     to_file_kwargs = {}
     if video_codec is not None: