update SP+EP - #209
Open
sfc-gh-mhidayetoglu wants to merge 87 commits into
Open
Conversation
…ffix_draft_token_ids` (#195)
Co-authored-by: Aurick Qiao <aurick.qiao@snowflake.com>
Shift Parallelism paper link: https://arxiv.org/abs/2509.16495
This reverts commit 170082f.
Co-authored-by: Mert Hidayetoglu <mert.hidayetoglu@snowflake.com> Co-authored-by: Aurick Qiao <aurick.qiao@snowflake.com>
Co-authored-by: Ye Wang <ye.wang@snowflake.com>
sfc-gh-mhidayetoglu
requested review from
sfc-gh-aqiao,
sfc-gh-goliaro,
sfc-gh-jrasley and
sfc-gh-yewang
as code owners
October 15, 2025 00:31
Signed-off-by: Stas Bekman <stas.bekman@snowflake.com>
Signed-off-by: Stas Bekman <stas.bekman@snowflake.com>
…-eng/ArcticInference-internal into tunji/verl_integration
… (#76) Co-authored-by: Ye Wang <ye.wang@snowflake.com> Co-authored-by: Jeff Rasley <jeff.rasley@snowflake.com> Co-authored-by: 151130f5470be3da5eea21fd02baa3_snow <bb9ecf6d788c5c0e03a13a6ec00039@snowflake.com> Co-authored-by: Michael Wyatt <michael.wyatt@snowflake.com> Co-authored-by: Cursor <cursoragent@cursor.com>
Compare sender HF names against the receiver model's expected set and fail loud on mismatch so Qwen3-32B colocated runs catch architecture drift early. Co-authored-by: Cursor <cursoragent@cursor.com>
- receiver: raise on incomplete CUDA IPC load (tensor-count mismatch), validate received parameter names on NCCL/CPU paths, and accept either serialized bytes or (name, tensor) pairs in load_weights_from_cpu - replica_pool: raise RuntimeError when any worker's sync_weights status is not "done" instead of swallowing the failure - api/multi_model: add /spec_weights_info and /sync_spec_weights endpoints plus Driver forwarding for multi-model mode Co-authored-by: Cursor <cursoragent@cursor.com>
Add load_weights_cuda_ipc_chunk to the weight-sync receiver, which loads one param (or small chunk) at a time from CUDA IPC handles instead of requiring the whole model in one payload. It raises on any GPU missing a handle and runs full param-name validation on the final chunk (context=cuda_ipc_stream). Add the matching InferenceWorker collective_rpc wrapper. This bounds peak extra GPU memory during sync to one full param per GPU. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Stas Bekman <stas.bekman@snowflake.com>
…Arctic Platform (#263)
Arctic rl/integration
Co-authored-by: Stas Bekman <stas.bekman@snowflake.com>
![arcticowl-ai-review[bot]](https://avatars.githubusercontent.com/in/1382583?s=60&v=4){kind=small}
Comment on lines
+100
to
+123
| @app.post("/init") | ||
| async def init_endpoint(request: InitRequest): | ||
| try: | ||
| n = await backend.initialize( | ||
| request.config, model_id=request.model_id, | ||
| num_replicas=request.num_replicas, | ||
| ) | ||
| return {"status": "ready", "model_id": request.model_id, "num_replicas": n} | ||
| except Exception as e: | ||
| raise HTTPException(status_code=500, detail=str(e)) | ||
|
|
||
|
|
||
| @app.post("/generate") | ||
| async def generate_endpoint(request: GenerateRequest): | ||
| try: | ||
| results = await backend.generate(**request.model_dump()) | ||
| return {"results": results} | ||
| except RuntimeError as e: | ||
| msg = str(e).lower() | ||
| if "paused" in msg or "cancelled" in msg: | ||
| raise HTTPException(status_code=503, detail=str(e)) | ||
| raise HTTPException(status_code=500, detail=str(e)) | ||
| except Exception as e: | ||
| raise HTTPException(status_code=500, detail=str(e)) |
There was a problem hiding this comment.
🚨 Critical blocker
🔴 HIGH · FastAPI inference server exposes model init / generation / weight-sync endpoints with no authentication [NEW] · CWE-306
How to override this blocker
Resolve the finding and push, or override it:
- On this comment: add a 👎 reaction and post an inline reply disputing it.
- Add the
critical_blocker_force_overridelabel to the PR to trigger re-evaluation.
Every Critical blocker on the PR must be disputed this way — the label on its own does not override a blocker that has no 👎 and reply.
Comment on lines
+136
to
+139
| data = path.read_bytes() | ||
| weights: list[tuple[str, torch.Tensor]] = torch.load( | ||
| io.BytesIO(data), map_location="cpu", weights_only=True, | ||
| ) |
There was a problem hiding this comment.
🟠 MEDIUM · load_weights_from_shm uses torch.load on /dev/shm file (weights_only=True but attacker-writable path) [NEW] · CWE-502
Comment on lines
+569
to
+572
| def load_weights_from_shm_path(self, path: str) -> dict: | ||
| """Load weights from a shared memory file path.""" | ||
| weights = torch.load(path, map_location="cpu", weights_only=True) | ||
| return self._load_cpu_weights_into_model(weights) |
There was a problem hiding this comment.
🟠 MEDIUM · load_weights_from_shm_path loads weights from a caller-supplied filesystem path via torch.load [NEW] · CWE-22
Comment on lines
+473
to
+474
| handles_list = pickle.loads( | ||
| base64.b64decode(ipc_payload["ipc_handles_pickled"])) |
There was a problem hiding this comment.
🚨 Critical blocker
🔴 HIGH · load_weights_cuda_ipc deserializes attacker-influenced pickled IPC handles via pickle.loads [NEW] · CWE-502
How to override this blocker
Resolve the finding and push, or override it:
- On this comment: add a 👎 reaction and post an inline reply disputing it.
- Add the
critical_blocker_force_overridelabel to the PR to trigger re-evaluation.
Every Critical blocker on the PR must be disputed this way — the label on its own does not override a blocker that has no 👎 and reply.
Comment on lines
+55
to
+70
| if is_server: | ||
| listen_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM) | ||
| listen_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1) | ||
| listen_socket.bind(("0.0.0.0", master_port)) | ||
| listen_socket.listen() | ||
| listen_fd = listen_socket.fileno() | ||
|
|
||
| store = TCPStore( | ||
| host_name=master_addr, | ||
| port=master_port, | ||
| world_size=world_size, | ||
| is_master=is_server, | ||
| timeout=timedelta(seconds=300), | ||
| use_libuv=False, | ||
| master_listen_fd=listen_fd, | ||
| ) |
There was a problem hiding this comment.
🟠 MEDIUM · stateless_init_nccl binds NCCL TCP rendezvous listener to 0.0.0.0 with no auth [NEW] · CWE-306
Comment on lines
+24
to
+30
| class GenerateRequest(BaseModel): | ||
| model_id: str | None = None | ||
| prompts: list[str | list[int]] | ||
| sampling_params: dict[str, Any] = Field(default_factory=dict) | ||
| routing_key: str | list[str | None] | None = None | ||
| strict: bool = False | ||
|
|
There was a problem hiding this comment.
🟠 MEDIUM · GenerateRequest forwards arbitrary sampling_params dict to SamplingParams(**...) without validation · CWE-20
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
13 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.