Repository navigation
GIN proxy on AINIC (Pollara): IB-CAST RMA backend, DMA-BUF for cuMem, spread over UDMA sets (#4480) - #4480
Open
JinghanHuang wants to merge 2 commits into
Open
GIN proxy on AINIC (Pollara): IB-CAST RMA backend, DMA-BUF for cuMem, spread over UDMA sets (#4480)#4480JinghanHuang wants to merge 2 commits into
JinghanHuang wants to merge 2 commits into
Conversation
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JinghanHuang has exported this pull request. If you are a Meta employee, you can view the originating Diff in D125311946. |
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 19, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Differential Revision: D125311946
JinghanHuang
force-pushed
the
export-D125311946
branch
from
October 19, 2026 19:11
6bafdf2 to
980c133
Compare
meta-codesync
Bot
changed the title
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 19, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 19, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
JinghanHuang
force-pushed
the
export-D125311946
branch
from
October 19, 2026 20:54
980c133 to
227d9ba
Compare
Summary: `RCCL_EP_HYB_Y_LOCAL=1` keeps only the node's own sources in a hybrid window's `y`/`cw` and moves their base down by `node_base * T` layout rows, so `slot(s, i)` still indexes them. A layout row is `num_topk` rows of `hidden` unless the window is `grouped_only`, but the hybrid combine is grouped only and writes and reads one row per slot. So on a window without `grouped_combine_only`, every node but node 0 addressed `(num_topk - 1) * node_base * T` rows below the region and faulted (memory access fault on all 8 ranks of node 1, then node 0 hangs). A `y_local` window now always gets one `y` row per slot, which is all the hybrid combine uses. Windows that set `grouped_combine_only` together with `Y_LOCAL` are unaffected. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Differential Revision: D125364154
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 21, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 21, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 21, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
JinghanHuang
force-pushed
the
export-D125311946
branch
from
October 21, 2026 21:19
227d9ba to
81b5695
Compare
JinghanHuang
added a commit
to JinghanHuang/torchcomms
that referenced
this pull request
Oct 21, 2026
… spread over UDMA sets (meta-pytorch#4480) Summary: Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
… spread over UDMA sets (meta-pytorch#4480) Summary: Pull Request resolved: meta-pytorch#4480 Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches: - On AINIC autodetect, GIN proxy's RMA backend is generic `net_ib` instead of IB-CAST. `plugin/rma.cc` only honors a literal `NCCL_NET=IB-CAST`; AMD's ROCM-27070 rule (ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom: `async fatal event on QP ... local access violation work queue error` in `ncclCommInitRank`. The fix mirrors `plugin/net.cc`. - GIN registers cuMem memory, which peermem cannot pin. With `NCCL_DMABUF_ENABLE=0` (as some cluster job environments set), devComm creation fails with `ibv_reg_mr ... Cannot allocate memory`. GIN proxy now uses DMA-BUF whenever cuMem is on, as `rocmwrap.cc` already does for hipMem. - On AINIC, a QP's channel id picks its UDMA engine set. Every IB-CAST GIN context was created with channel 0, so all GIN traffic shared one set, capping it near 30 GB/s per NIC. Each context now takes its own id from the top of the range, so consecutive contexts alternate sets. Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects; `NCCL_CROSS_NIC=0` avoids it. The same change is made in both the RCCL and the RCCLX trees, with their `patches/` mirrors. Reviewed By: srinathb-meta Differential Revision: D125311946
JinghanHuang
force-pushed
the
export-D125311946
branch
from
October 21, 2026 21:22
81b5695 to
3e6b685
Compare
This branch has not been deployed
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.
Summary:
Two-host RCCL GIN (the proxy backend) fails on MI355X hosts with Pollara NICs and, once it runs, is slow. Three RCCL bugs, fixed here as carry patches:
net_ibinstead of IB-CAST.plugin/rma.cconly honors a literalNCCL_NET=IB-CAST; AMD's ROCM-27070 rule (ROCM-27070: honor NCCL_GIN_ENABLE=0 and select ib-cast GIN backend on AINIC ROCm/rocm-systems#9085) was lost in the 2.30.7 GIN/RMA split, and develop still has the bug. Symptom:async fatal event on QP ... local access violation work queue errorinncclCommInitRank. The fix mirrorsplugin/net.cc.NCCL_DMABUF_ENABLE=0(as some cluster job environments set), devComm creation fails withibv_reg_mr ... Cannot allocate memory. GIN proxy now uses DMA-BUF whenever cuMem is on, asrocmwrap.ccalready does for hipMem.Out of scope: RCCL's own symmetric-kernel GIN connects a 16-rank ring that crosses NICs within a host, which a rail-only fabric rejects;
NCCL_CROSS_NIC=0avoids it.The same change is made in both the RCCL and the RCCLX trees, with their
patches/mirrors.Reviewed By: srinathb-meta
Differential Revision: D125311946