Purge InternVideo2 - #1451
Conversation
Signed-off-by: Ao Tang <aot@nvidia.com>
|
/ok to test 2f16c57 |
Greptile OverviewGreptile SummaryThis PR purges InternVideo2 support across the video curation stack: it removes the InternVideo2 model/stages/tests and drops the CI + Docker steps that cloned/patched the upstream InternVideo repo. Documentation is updated to position Cosmos-Embed1 as the sole embedding backend, and Main concern: several video docs/tutorials still reference the removed InternVideo2 output directories and helpers ( Confidence Score: 3/5
Important Files Changed
Sequence DiagramsequenceDiagram
participant User as User/Caller
participant Pipe as Pipeline
participant Writer as ClipWriterStage
participant Storage as Storage (local/S3)
User->>Pipe: run()
Pipe->>Writer: setup(worker_metadata)
Pipe->>Writer: process(VideoTask)
loop for each clip
Writer->>Writer: _write_clip_embedding_to_buffer(clip)
end
par clip-level writes (thread pool)
loop for each clip
Writer->>Storage: write clip mp4 (optional)
Writer->>Storage: write preview webp (optional)
Writer->>Storage: write embedding pickle (Cosmos-Embed1 only)
Writer->>Storage: write clip metadata JSON
end
and filtered clip writes
loop for each filtered clip
Writer->>Storage: write filtered clip mp4 (optional)
Writer->>Storage: write filtered clip metadata JSON
end
end
Writer->>Storage: write CE1 parquet batch (ce1_embd_parquet/)
Writer->>Storage: write video metadata JSON + chunk metadata JSON
Writer->>Writer: cleanup in-memory buffers on clips/windows
Pipe-->>User: completed
|
Additional Comments (6)
Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
|
Signed-off-by: Ao Tang <aot@nvidia.com>
|
/ok to test 7e0ad4d |
praateekmahajan
left a comment
praateekmahajan
left a comment
There was a problem hiding this comment.
LGTM but let's wait for @thomasdhc to also TAL
Additional Comments (12)
|
|
/ok to test e28aa0e |
…embedding in total size calculation Signed-off-by: Ao Tang <aot@nvidia.com>
|
/ok to test b3afc96 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Additional Comments (2)
|
Signed-off-by: Ao Tang <aot@nvidia.com>
|
/ok to test ffe3801 |
Additional Comments (12)
|
|
/ok to test 4acdd10 |
| def setup(self, worker_metadata: WorkerMetadata | None = None) -> None: # noqa: ARG002 | ||
| self._iv2_embedding_buffer: list[dict[str, Any]] = [] | ||
| self._ce1_embedding_buffer: list[dict[str, Any]] = [] | ||
|
|
There was a problem hiding this comment.
[P0] ClipWriterStage.setup() no longer initializes the removed IV2 buffer, but process() can call _write_clip_embedding_to_buffer() before setup() depending on how the stage lifecycle is used. Since _ce1_embedding_buffer is created in setup(), this would raise AttributeError if a caller invokes process() without calling setup() first. If the framework guarantees setup() is always called, it may be fine, but previously setup() was also where IV2 buffer was created and tests explicitly call it.
If you want this stage to be robust to direct use, consider initializing _ce1_embedding_buffer in __post_init__ or as a field default.
Additional Comments (1)
Also appears in: |
4 participants
Description
Usage
# Add snippet demonstrating usageChecklist