Skip to content

CollectiveX: run the MI300X and MI325X pools on the vLLM ROCm image / CollectiveX：在 vLLM ROCm 镜像上运行 MI300X 与 MI325X 资源池 #4581

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Oseltamivir merged 13 commits into from
Oct 28, 2026
Merged

CollectiveX: run the MI300X and MI325X pools on the vLLM ROCm image / CollectiveX：在 vLLM ROCm 镜像上运行 MI300X 与 MI325X 资源池 #4581

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
b0c8597
CollectiveX: run EP16 and low-latency on the Barite MI300X and MI325X…
Oseltamivir Oct 27, 2026
ae8335c
CollectiveX: give MoRI's bootstrap the registry socket interface
Oseltamivir Oct 27, 2026
e030aed
CollectiveX: bind Barite's Broadcom verbs provider into multi-node sh…
Oseltamivir Oct 27, 2026
51b0573
CollectiveX: install Barite's Broadcom verbs provider inside the cont…
Oseltamivir Oct 27, 2026
1e34553
CollectiveX: keep an image's own verbs provider when it already works
Oseltamivir Oct 27, 2026
bce4551
CollectiveX: give each run.sh invocation its own shard directory
Oseltamivir Oct 27, 2026
a6983c1
CollectiveX: stop waiting on an allocation whose job already ended
Oseltamivir Oct 27, 2026
8be7bc5
CollectiveX: run mi300x and mi325x EP in the vLLM v0.32.0 ROCm image
Oseltamivir Oct 27, 2026
63f4b54
CollectiveX: publish the hand-run mi300x and mi325x MoRI results
Oseltamivir Oct 27, 2026
c11826b
CollectiveX: publish the hand-run mi300x and mi325x UCCL-EP and swap-…
Oseltamivir Oct 27, 2026
f573b99
CollectiveX: replace the noisy mi300x MoRI EP16 fp8 prefill result
Oseltamivir Oct 27, 2026
c3a422b
CollectiveX: stop a hand run whose checkout changes, and keep transie…
Oseltamivir Oct 28, 2026
292b6f0
CollectiveX: describe the Broadcom pools by SKU in runtime comments
Oseltamivir Oct 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
32 changes: 22 additions & 10 deletions collectivex/configs/platform_config.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,17 +133,23 @@
"mi300x": {
"arch": "gfx942",
"product": "mi300x",
"image": "rocm/sgl-dev:sglang-0.5.14-rocm720-mi35x-mori-0701",
"image": "vllm/vllm-openai-rocm:v0.32.0",
"image_platform": "linux/amd64",
"gpus_per_node": 8,
"scale_up_domain": 8,
"scale_up_transport": "xgmi",
"launcher": "mi-amds",
"backends": {"mori": [8], "uccl-ep": [8]},
"ll_backends": {},
"backends": {"mori": [8, 16], "uccl-ep": [8, 16]},
"ll_backends": {"mori": [16], "uccl-ep": [16]},
"fabric": {
"nic": "n/a (single-node scale-up)",
"switch": "n/a (single-node scale-up)"
"nic": "Broadcom 400GbE RoCE v2 (bnxt_re), 8 rails",
"switch": "routed L3 RoCE fabric"
},
"network": {
"socket_ifname": "ens50f1np1",
"rdma_devices": "bnxt_re_bond0,bnxt_re_bond1,bnxt_re_bond2,bnxt_re_bond3,bnxt_re_bond4,bnxt_re_bond5,bnxt_re_bond6,bnxt_re_bond7",
"ib_gid_index": 3,
"host_verbs_provider": "/usr/local/lib/x86_64-linux-gnu/libbnxt_re-rdmav34.so"
},
"operator": {
"partition": "MI300X-UBUNTU",
Expand All @@ -153,18 +159,24 @@
"mi325x": {
"arch": "gfx942",
"product": "mi325x",
"image": "rocm/sgl-dev:sglang-0.5.14-rocm720-mi35x-mori-0701",
"image": "vllm/vllm-openai-rocm:v0.32.0",
"image_platform": "linux/amd64",
"gpus_per_node": 8,
"scale_up_domain": 8,
"scale_up_transport": "xgmi",
"launcher": "mi-amds",
"runner_label": "cluster:mi325x-amds",
"backends": {"mori": [8], "uccl-ep": [8]},
"ll_backends": {},
"backends": {"mori": [8, 16], "uccl-ep": [8, 16]},
"ll_backends": {"mori": [16], "uccl-ep": [16]},
"fabric": {
"nic": "n/a (single-node scale-up)",
"switch": "n/a (single-node scale-up)"
"nic": "Broadcom 400GbE RoCE v2 (bnxt_re), 8 rails",
"switch": "routed L3 RoCE fabric"
},
"network": {
"socket_ifname": "enp193s0f0np0",
"rdma_devices": "bnxt_re4,bnxt_re5,bnxt_re1,bnxt_re0,bnxt_re10,bnxt_re11,bnxt_re7,bnxt_re6",
"ib_gid_index": 3,
"host_verbs_provider": "/usr/local/lib/x86_64-linux-gnu/libbnxt_re-rdmav34.so"
},
"operator": {
"partition": "MI325X-UBUNTU",
Expand Down
1 change: 1 addition & 0 deletions collectivex/launchers/launch_mi-amds.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,7 @@ for allocation_attempt in 1 2 3; do
done
unset COLLX_SALLOC_ATTEMPT COLLX_NETWORK_VALIDATION_ATTEMPT
CONTAINER_MOUNTS="$MOUNT_SRC:$MOUNT_DIR$DEVICE_MOUNTS"
collx_add_host_verbs_provider_mount
# uccl-ep builds from source, so it gets the cross-allocation backend cache single-slurm uses.
# Its parent is COLLX_STAGE_DIR, not SQUASH_DIR: this SKU's squash dir is node-local and
# root-owned (/var/lib/squash), so the submit-side mkdir fails and the path is not what compute
Expand Down
9 changes: 8 additions & 1 deletion collectivex/launchers/launch_single-slurm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,7 @@ collx_select_image "$IMAGE"
collx_stage_with_backend_cache
# Multi-node shards on the EFA pools (b300, h100-dgxc) carry the host EFA user space themselves.
collx_add_efa_host_mounts
collx_add_host_verbs_provider_mount

command -v salloc >/dev/null || collx_die "salloc not found on this runner"
allocation=(--partition="$COLLX_PARTITION" --nodes="$NODES" --gres=gpu:"$GPN"
Expand All @@ -81,7 +82,13 @@ for allocation_attempt in 1 2 3; do
[ -z "$excluded_nodes" ] || attempt_allocation+=(--exclude="$excluded_nodes")
export COLLX_SALLOC_ATTEMPT="$allocation_attempt"
export COLLX_NETWORK_VALIDATION_ATTEMPT="$allocation_attempt"
collx_salloc_jobid "${attempt_allocation[@]}"
JOB_ID=""
if ! collx_salloc_jobid "${attempt_allocation[@]}"; then
# A job that dies before its grant (a failed node prolog) is worth another allocation.
[ "$allocation_attempt" != 3 ] || collx_die "scheduler allocation failed on every attempt"
collx_log "allocation attempt $allocation_attempt failed; retrying"
continue
fi
[ -n "$JOB_ID" ] || collx_die "could not resolve allocated JOB_ID from salloc"
if ! collx_validate_network_profile_on_job "$JOB_ID" "$NODES" "$COLLX_TRANSPORT"; then
validation_failure=network
Expand Down
Loading
Loading