Fixups for NVIDIA fabric support#534
Open
artulab wants to merge 1 commit into
Open
Conversation
github-actions
Bot
added
in-progress
Contributor
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
This PR addresses follow-up fixes for NVIDIA fabric support across Iris topology discovery, memory import paths, and benchmark/example launch behavior.
Changes:
- Add NVIDIA CUDA driver hooks for exporting arbitrary VMM-backed pointer handles and querying allocation address ranges.
- Improve topology discovery on NVIDIA by probing fabric connectivity via CUDA fabric handle import/export when NVML fabric UUIDs are missing.
- Update benchmarks/examples and AG+MM auto-config to work under externally launched multi-node
torchrunand NVIDIA environments.
Reviewed changes
Copilot reviewed 7 out of 7 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| iris/ops/all_gather_matmul_hbm_buffer.py | Makes a ROCm-specific launch kwarg conditional to avoid applying it on NVIDIA. |
| iris/host/memory/allocators/vmem_chunked_allocator.py | Adds fallback behavior for external tensor import on single-rank local CUDA jobs and improves DriverNotSupported handling. |
| iris/host/distributed/topology.py | Adds an NVIDIA fabric connectivity probe and uses it to synthesize fabric domains when NVML fabric info is missing. |
| iris/drivers/local/nvidia.py | Adds CUDA driver calls to export retained allocation handles and query pointer address ranges, plus context management. |
| iris/bench/_runner.py | Treats an already-launched torchrun job as the benchmark rank instead of spawning a nested elastic job. |
| examples/14_all_gather_gemm/example_run_pull.py | Supports torchrun-style environment launches, local-rank device selection, and optional topology printing. |
| benchmark/ops/all_gather_matmul/auto_config.py | Adds NVIDIA arch detection and heuristic config fallback for NVIDIA (and currently for any non-AMD arch). |
Comment on lines
+547
to
+554
| if arch not in ("mi300x", "mi355x"): | ||
| heuristic_config, heuristic_hbm = _apply_heuristic(M, N, K, arch=arch) | ||
| return AutoConfigResult( | ||
| enabled=True, | ||
| config_params=heuristic_config, | ||
| hbm_buffer_params=heuristic_hbm, | ||
| source=f"Heuristic fallback for {arch} (no tuned configs available)", | ||
| ) |
Comment on lines
+389
to
+404
| local_row = [False] * world_size | ||
| for record in records: | ||
| if not record or not record.get("ok"): | ||
| continue | ||
| peer_rank = int(record["rank"]) | ||
| if peer_rank == rank: | ||
| local_row[peer_rank] = True | ||
| continue | ||
| if driver is None: | ||
| continue | ||
| try: | ||
| mapping = driver.import_and_map(peer_rank, record["handle"], int(record["size"])) | ||
| imported_mappings.append(mapping) | ||
| local_row[peer_rank] = True | ||
| except Exception as exc: | ||
| logger.debug("[Rank %d] CUDA fabric handle import from rank %d failed: %s", rank, peer_rank, exc) |
Comment on lines
+558
to
+560
| # If launched by torchrun/srun for a multi-node job, do not spawn another | ||
| # local elastic job. The current process is already one benchmark rank. | ||
| if all(key in os.environ for key in ("RANK", "LOCAL_RANK", "WORLD_SIZE")): |
Comment on lines
+451
to
+452
| def _can_return_external_tensor_alias(self) -> bool: | ||
| return self.num_ranks == 1 and self.driver.__class__.__name__ == "LocalCudaDriver" |
Comment on lines
+472
to
+474
| def export_pointer_handle(self, ptr: int, size: int) -> bytes: | ||
| """Export the VMM allocation containing ptr as a 4-byte native-endian POSIX FD.""" | ||
| self._check_initialized() |
mawad-amd
approved these changes
Jun 11, 2026
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
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.
Motivation
Fix follow-up issues needed for NVIDIA fabric support, including runtime setup, topology handling, and benchmark/example compatibility.
Technical Details
This PR adds NVIDIA CUDA address-range/export hooks, improves VMM chunked external tensor import behavior, supports externally launched multinode
torchruniniris.bench, and updates all-gather GEMM benchmark/example paths for NVIDIA environments.Test Plan
Tested the benchmark and example iris program on the clusters.
Test Result
PASS
Submission Checklist