Skip to content

Add aggregate mode support to AIDynamo workload #982

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

Open
saivishal1999 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add aggregate mode support to AIDynamo workload #982

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
13 changes: 10 additions & 3 deletions src/cloudai/workloads/ai_dynamo/ai_dynamo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,6 +161,7 @@ class AIDynamoArgs(BaseModel):

model: str = "Qwen/Qwen3-0.6B"
backend: Literal["vllm", "sglang", "sglang_dsr1"] = "vllm"
mode: Literal["aggregate", "disaggregate"] = "disaggregate"
endpoint: str = Field(default="v1/chat/completions")
connector: Optional[str | list[str]] = None

Expand Down Expand Up @@ -235,12 +236,14 @@ def populate_prefill_decode_args(self) -> "AIDynamoArgs":
"""Populate prefill/decode args."""
if self.backend.lower() == "vllm":
self.prefill_worker.args.model = self.model
self.decode_worker.args.model = self.model
if self.mode == "disaggregate":
self.decode_worker.args.model = self.model
elif self.backend.lower() in ["sglang", "sglang_dsr1"]:
self.prefill_worker.args.model_path = self.model
self.decode_worker.args.model_path = self.model
self.prefill_worker.args.served_model_name = self.model
self.decode_worker.args.served_model_name = self.model
if self.mode == "disaggregate":
self.decode_worker.args.model_path = self.model
self.decode_worker.args.served_model_name = self.model
else:
raise ValueError(f"Invalid backend: {self.backend}")

Expand Down Expand Up @@ -577,6 +580,10 @@ def was_run_successful(self, tr: TestRun) -> JobStatusResult:
return JobStatusResult(workloads_successful and accuracy_successful)

def constraint_check(self, tr: TestRun, system: Optional[System]) -> bool:
if tr.test.cmd_args.dynamo.mode == "aggregate":
logging.info("constraint_check skipped: aggregate mode has no prefill/decode split")
return True

Comment on lines +583 to +586

@coderabbitai coderabbitai Bot Jul 28, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Retain prefill GPU-capacity validation in aggregate mode.

This return also skips tp_times_pp_le_gpus_per_node for the prefill worker. An aggregate run with TP × PP exceeding available GPUs is accepted, then the runtime allocation can compute zero workers per node. Skip only decode/split-specific checks; continue validating prefill capacity.

This is evidenced by the allocation path in src/cloudai/workloads/ai_dynamo/ai_dynamo.sh Lines 376-390.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/cloudai/workloads/ai_dynamo/ai_dynamo.py` around lines 583 - 586, Update
the aggregate-mode branch in the constraint-check function so it skips only
decode/split-specific validation while still executing prefill GPU-capacity
validation, including tp_times_pp_le_gpus_per_node. Remove the unconditional
early return and preserve the existing aggregate-mode handling for checks that
do not apply.

prefill_worker = tr.test.cmd_args.dynamo.prefill_worker
decode_worker = tr.test.cmd_args.dynamo.decode_worker

Expand Down
50 changes: 36 additions & 14 deletions src/cloudai/workloads/ai_dynamo/ai_dynamo.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -307,7 +307,11 @@ _apply_connector_settings() {

_patch_dynamo_args() {
if [[ -z "${dynamo_args["frontend-node"]}" ]]; then
dynamo_args["frontend-node"]=$(echo "${decode_config["node-list"]}" | cut -d',' -f1)
if [[ -n "${decode_config["node-list"]:-}" ]]; then
dynamo_args["frontend-node"]=$(echo "${decode_config["node-list"]}" | cut -d',' -f1)
else
dynamo_args["frontend-node"]=$(echo "${prefill_config["node-list"]}" | cut -d',' -f1)
fi
fi

dynamo_args["url"]="http://${dynamo_args["frontend-node"]}:${dynamo_args["port"]}"
Expand Down Expand Up @@ -342,13 +346,20 @@ _compute_worker_allocation_vllm() {
fi

prefill_config["gpus-per-worker"]=$(( prefill_args["--tensor-parallel-size"] * prefill_args["--pipeline-parallel-size"] ))
decode_config["gpus-per-worker"]=$(( decode_args["--tensor-parallel-size"] * decode_args["--pipeline-parallel-size"] ))

if [[ ${prefill_config["gpus-per-worker"]} -eq 0 ]] || [[ ${decode_config["gpus-per-worker"]} -eq 0 ]]; then
if [[ ${prefill_config["gpus-per-worker"]} -eq 0 ]]; then
log "ERROR: Invalid TP/PP configuration"
exit 1
fi

if [[ "${decode_config["num-nodes"]:-0}" -gt 0 ]]; then
decode_config["gpus-per-worker"]=$(( decode_args["--tensor-parallel-size"] * decode_args["--pipeline-parallel-size"] ))
if [[ ${decode_config["gpus-per-worker"]} -eq 0 ]]; then
log "ERROR: Invalid decode TP/PP configuration"
exit 1
fi
fi

decode_config["gpu-offset"]=0
prefill_config["gpu-offset"]=0

Expand All @@ -362,16 +373,21 @@ _compute_worker_allocation_vllm() {
prefill_config["workers-per-node"]=1
prefill_config["gpu-offset"]=${decode_config["gpus-per-worker"]}
else
if [[ "${prefill_config["multiple-workers-per-node"]}" != "true" ]]; then
if [[ "${prefill_config["multiple-workers-per-node"],,}" != "true" ]]; then
prefill_config["gpus-per-worker"]=$num_gpus
fi

if [[ "${decode_config["multiple-workers-per-node"]}" != "true" ]]; then
decode_config["gpus-per-worker"]=$num_gpus
fi

prefill_config["workers-per-node"]=$(( num_gpus / prefill_config["gpus-per-worker"] ))
decode_config["workers-per-node"]=$(( num_gpus / decode_config["gpus-per-worker"] ))

if [[ "${decode_config["num-nodes"]:-0}" -gt 0 ]]; then
if [[ "${decode_config["multiple-workers-per-node"],,}" != "true" ]]; then
decode_config["gpus-per-worker"]=$num_gpus
fi
decode_config["workers-per-node"]=$(( num_gpus / decode_config["gpus-per-worker"] ))
else
decode_config["gpus-per-worker"]=0
decode_config["workers-per-node"]=0
fi
fi

log "DECODE: num GPUs: $num_gpus, GPUs per worker: ${decode_config["gpus-per-worker"]}"
Expand Down Expand Up @@ -1124,7 +1140,7 @@ _resolve_aiperf_server_metrics_urls() {
done
done

if [[ "${dynamo_args["dcgm-exporter-enabled"]}" == "True" || "${dynamo_args["dcgm-exporter-enabled"]}" == "true" ]]; then
if [[ "${dynamo_args["dcgm-exporter-enabled"],,}" == "true" ]]; then
local dcgm_nodes="${decode_config["node-list"]:-},${prefill_config["node-list"]:-}"
for node in $dcgm_nodes; do
[[ -z "$node" ]] && continue
Expand Down Expand Up @@ -1393,12 +1409,14 @@ function main()
launch_etcd &
launch_nats &
wait_for_etcd
launch_ingress
if _is_sglang_dsr1; then
launch_sgl_http_server
fi
fi

# Workers launch BEFORE the ingress: launch_ingress blocks in
# wait_for_router, and the router only becomes ready once a worker
# registers — on a combined frontend+worker node the old order serialized
# the whole ROUTER_START_TIMEOUT (120 s of failing readiness curls) in
# front of every worker start. Workers only need etcd/nats (waited above)
# and the lmcache config from setup_lmcache; they never talk to the router.
if _is_decode_node; then
log "Node ID: $SLURM_NODEID, Role: decode"
log_node_role "$(_current_node_name)" "decode"
Expand All @@ -1412,6 +1430,10 @@ function main()
fi

if _is_frontend_node; then
launch_ingress
if _is_sglang_dsr1; then
launch_sgl_http_server
fi
sleep 10

launch_workloads &
Expand Down
21 changes: 15 additions & 6 deletions src/cloudai/workloads/ai_dynamo/slurm_command_gen_strategy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -400,13 +400,17 @@ def _gen_script_args(self, td: AIDynamoTestDefinition) -> List[str]:
args.append('--dynamo-dcgm-exporter-enabled "True"')
args.append(f'--dynamo-dcgm-exporter-port "{td.cmd_args.dynamo.dcgm_exporter.port}"')

is_aggregate = td.cmd_args.dynamo.mode == "aggregate"

if td.cmd_args.dynamo.prefill_worker:
args.extend(self._get_nested_toml_args(td.cmd_args.dynamo.prefill_worker, "--prefill-", exclude=["nodes"]))
if td.cmd_args.dynamo.prefill_worker.nodes:
args.append(f"--prefill-node-list {shlex.quote(td.cmd_args.dynamo.prefill_worker.nodes)}")
args.extend(self._get_nested_toml_args(td.cmd_args.dynamo.decode_worker, "--decode-", exclude=["nodes"]))
if td.cmd_args.dynamo.decode_worker.nodes:
args.append(f"--decode-node-list {shlex.quote(td.cmd_args.dynamo.decode_worker.nodes)}")

if not is_aggregate:
args.extend(self._get_nested_toml_args(td.cmd_args.dynamo.decode_worker, "--decode-", exclude=["nodes"]))
if td.cmd_args.dynamo.decode_worker.nodes:
args.append(f"--decode-node-list {shlex.quote(td.cmd_args.dynamo.decode_worker.nodes)}")

args.extend(self._get_nested_toml_args(td.cmd_args.genai_perf, "--genai_perf-"))
if aiperf_script:
Expand Down Expand Up @@ -608,12 +612,16 @@ def get_cached_nodes_spec(self) -> tuple[int, list[str]]:
if cache_key in self._node_spec_cache:
return self._node_spec_cache[cache_key]

is_aggregate = self.td.cmd_args.dynamo.mode == "aggregate"

prefill_n, prefill_nodes = 0, ""
if self.td.cmd_args.dynamo.prefill_worker:
prefill_n = cast(int, self.td.cmd_args.dynamo.prefill_worker.num_nodes)
prefill_nodes = self.td.cmd_args.dynamo.prefill_worker.nodes
decode_n = self.td.cmd_args.dynamo.decode_worker.num_nodes
decode_nodes = self.td.cmd_args.dynamo.decode_worker.nodes

# In aggregate mode there is no separate decode worker; all nodes run the prefill worker.
decode_n = 0 if is_aggregate else cast(int, self.td.cmd_args.dynamo.decode_worker.num_nodes)
decode_nodes = None if is_aggregate else self.td.cmd_args.dynamo.decode_worker.nodes

assert isinstance(prefill_n, int), "prefill_worker.num_nodes must be an integer"
assert isinstance(decode_n, int), "decode_worker.num_nodes must be an integer"
Expand Down Expand Up @@ -649,7 +657,8 @@ def get_cached_nodes_spec(self) -> tuple[int, list[str]]:

if prefill_nodes or decode_nodes:
self._validate_worker_nodes(node_list, prefill_nodes, prefill_n, "prefill")
self._validate_worker_nodes(node_list, decode_nodes, decode_n, "decode")
if not is_aggregate:
self._validate_worker_nodes(node_list, decode_nodes, decode_n, "decode")
if self._worker_nodes_overlap(prefill_nodes, decode_nodes):
unique_worker_nodes = self._unique_nodes(
self._split_node_list(prefill_nodes) + self._split_node_list(decode_nodes)
Expand Down