Update abstractions.md

Signed-off-by: Huy Vu <[email protected]>

Author: huvunvidia

docs/about/concepts/video/abstractions.md

@@ -49,13 +49,11 @@ Each processing stage:
 
 1. Inherits from `ProcessingStage`
 2. Declares a stable `name` and `resources: Resources` (CPU cores, GPU memory, optional NVDEC/NVENC, or more than one GPU)
-3. Defines `inputs()`/`outputs()` to document required attributes and produced attributes on tasks
+3. Defines `inputs()`/`outputs()` to document that requires attributes or produces attributes on tasks
 4. Implements `setup(worker_metadata)` for model initialization and `process(task)` to transform tasks
 
 This design enables map-style execution with executor-managed fault tolerance and dynamic scaling per stage. Stages can optionally provide `process_batch()` to support vectorized batch processing.
 
-Composite stages provide a user-facing convenience API and decompose into one or more execution stages at build time.
-
 ```python
 class MyStage(ProcessingStage[X, Y]):
     name: str = "..."
@@ -68,6 +66,8 @@ class MyStage(ProcessingStage[X, Y]):
     def process(self, task: X) -> Y | list[Y]: ...
 ```
 
+Composite stages provide a user-facing convenience API and decompose into one or more execution stages at build time.
+
 Refer to the stage base and resources definitions in Curator for full details.
 
 ### Resource Semantics
@@ -89,8 +89,8 @@ Video pipelines operate on task types defined in Curator:
 - `Video`: Holds decoded metadata, frames, and lists of `Clip`
 - `Clip`: Holds buffers, extracted frames, embeddings, and caption windows
 
-Stages transform tasks stage by stage (for example, `VideoReader` populates `Video`, splitting stages create `Clip` objects, embedding and captioning stages annotate clips, and writer stages persist outputs).
+Stages transform tasks stage by stage, and might run with different tasks on different stages (for example, `VideoReader` populates `Video`, splitting stages create `Clip` objects, embedding and captioning stages annotate clips, and writer stages persist outputs).
 
 ## Executors
 
-Executors run pipelines on a backend. Curator uses [`XennaExecutor`](https://github.com/nvidia-cosmos/cosmos-xenna) to translate `ProcessingStage` definitions into Cosmos-Xenna stage specifications and run them on Ray with automatic scaling. Execution modes include streaming (default) and batch.
+Executors run pipelines on a backend. Curator uses [`XennaExecutor`](https://github.com/nvidia-cosmos/cosmos-xenna) to translate `ProcessingStage` definitions into Cosmos-Xenna stage specifications and run them on Ray with automatic scaling. Execution modes include streaming (by default) and batch.