Add bulk-buffer registration leases for IBGDA (#3414) - #3414
Open
snarayankh wants to merge 4 commits into
Open
Conversation
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@snarayankh has exported this pull request. If you are a Meta employee, you can view the originating Diff in D114554503. |
meta-codesync
Bot
changed the title
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 5, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
2 times, most recently
from
August 5, 2026 19:14
7ed75c4 to
8f9a63a
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 5, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Differential Revision: D114554503
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 5, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Differential Revision: D114554503
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 5, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 5, 2026 20:06
8f9a63a to
c3b82c8
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 6, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 6, 2026 07:04
c3b82c8 to
d28941e
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 6, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 9, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 9, 2026 19:13
d28941e to
6af6bf3
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 9, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 9, 2026 19:27
6af6bf3 to
9b86c7a
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 9, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 9, 2026 19:54
9b86c7a to
18d7a5a
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 11, 2026
Summary: Pull Request resolved: meta-pytorch#3414 Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 11, 2026 00:00
18d7a5a to
8709d31
Compare
snarayankh
pushed a commit
to snarayankh/torchcomms
that referenced
this pull request
Aug 11, 2026
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 11, 2026 00:48
8709d31 to
31a56e7
Compare
added 4 commits
August 10, 2026 22:19
Summary: Add resumable and blocking registered-source send APIs that let IBGDA read directly from a registered caller buffer. Track posting separately from local NIC completion so callers can safely reuse the source only after an explicit drain. Preserve the existing staged protocol cursor, receiver staging, signaling, credits, and default behavior. Reviewed By: rmahidhar Differential Revision: D114552007
Summary: Cover two protocol interactions for registered-source IBGDA sends: sharing the normal send cursor with staged sends, and maintaining backpressure and data integrity across repeated staging-slot wrap. This is test-only hardening for the registered-source transport primitive. Reviewed By: rmahidhar Differential Revision: D114552006
Summary: Add an opt-in `--ibgda_sendrecv_enable_registered` mode to the IBGDA send/receive benchmark so registered-source sends can be measured against the existing staged path. Keep staged sends as the default, use the same one-block benchmark geometry for both paths, and explicitly drain registered sends before their source storage can be reused. Reviewed By: rmahidhar Differential Revision: D114552005
Summary: Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges. Expose the lease operations through `MultiPeerTransport` for the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree. Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope. Reviewed By: rmahidhar Differential Revision: D114554503
snarayankh
force-pushed
the
export-D114554503
branch
from
August 11, 2026 05:20
31a56e7 to
d0c36c5
Compare
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
1 participant
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.
Summary:
Add move-only bulk-buffer registration leases with stable generations and non-owning contained-range views. Bulk registrations request PCIe Relaxed Ordering, share the existing per-allocation MR cache, and reject invalid, overflowing, stale, or released ranges.
Expose the lease operations through
MultiPeerTransportfor the IBGDA zero-copy AllReduce path. This change does not alter collective behavior or enable zero-copy Ring or Tree.Add focused lifetime, containment, overlap, ordering, and re-registration coverage. IBRC exercises the shared transport-base implementation as regression coverage; IBGDA remains the product scope.
Reviewed By: rmahidhar
Differential Revision: D114554503