Repository navigation
Add CUDA VMM multimodal feature transport - #33899
Merged
hanming-lu merged 2 commits intoAug 7, 2026
Merged
Conversation
oulgen
requested review from
CatherineSue,
JustinTong0323,
Ying1123,
hnyls2002,
ispobock,
merrymercy,
mickqian,
slin1237,
xiezhq-hermann,
yhyang201 and
yuan-luo
as code owners
August 6, 2026 19:26
Collaborator
Author
|
/tag-and-rerun-ci |
1 similar comment
Collaborator
Author
|
/tag-and-rerun-ci |
Collaborator
Author
|
cc: @yinghai @merrymercy |
Collaborator
|
/tag-and-rerun-ci |
Share multimodal feature tensors through a bounded VMM pool. Use CUDA FABRIC handles when available, fall back to POSIX file descriptors on one node, and use CPU transport when the pool is full. Co-authored-by: Yinghai Lu <yinghai@meta.com>
oulgen
force-pushed
the
cuda-vmm-mm-feature-transport
branch
from
August 7, 2026 04:08
0be7561 to
94e6e6f
Compare
Collaborator
Author
|
/tag-and-rerun-ci |
Collaborator
|
base-c-test-4-gpu-gb300 is flaky on main, merging. |
hanming-lu
approved these changes
Aug 7, 2026
This was referenced Aug 7, 2026
saturn-acc
pushed a commit
to saturn-acc/sglang
that referenced
this pull request
Aug 16, 2026
Co-authored-by: Yinghai Lu <yinghai@meta.com>
chencjcj
pushed a commit
to chencjcj/sglang
that referenced
this pull request
Aug 17, 2026
Co-authored-by: Yinghai Lu <yinghai@meta.com> (cherry picked from commit 7f6b4cb) Conflict resolution: - server_args.py: took upstream's mm_feature_transport declaration, which widens the Literal from {cpu, cuda_ipc} to include cuda_vmm and rewrites the help text. Strict superset of this base's version. - tokenizer_manager.py: pure adjacency conflict. This base carries init_request_preprocessor/run_in_request_preprocessor from sgl-project#30771 at the same insertion point as upstream's new _validate_cuda_vmm_feature_transport_support. The two are unrelated, so both were kept.
Atituiset
pushed a commit
to Atituiset/sglang
that referenced
this pull request
Sep 10, 2026
Co-authored-by: Yinghai Lu <yinghai@meta.com>
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
Large multimodal feature tensors can make CPU serialization and object broadcasts a bottleneck. This adds an explicit, model-gated GPU transport with bounded memory use and CPU fallback when capacity is unavailable.
Modifications
Accuracy Tests
No model forward or numerical code changes.
pytest -q test/registered/unit/multimodal/test_gpu_feature_transport.py: 23 passed.CUDA_VISIBLE_DEVICES=0,1,2 pytest -q test/registered/unit/multimodal/test_cuda_vmm_transport.py: 14 passed.Speed Tests and Profiling
Performance benchmark not run: no in-tree model opts into this transport.
Checklist
Review and Merge Process
CI States
Latest PR Test (Base): ❌ Run #31146495118
Latest PR Test (Extra): ❌ Run #31146494923