Skip to content

Commit 58191c7

Browse files
committed
fix: align power rows and recipes across affected curves
1 parent 5b220e6 commit 58191c7

7 files changed

Lines changed: 40 additions & 11 deletions

File tree

‎inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/h200-fp8/agentx/agg-tp8-mtp-kvoffload.yaml‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -102,9 +102,6 @@ srun_options:
102102
# The custom benchmark installs its isolated AIPerf environment at runtime.
103103
container-remap-root: ""
104104

105-
telemetry:
106-
required: false
107-
108105
benchmark:
109106
env:
110107
AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"

‎inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/qwen3.8next/sglang/gb300-fp4-mtp/agentic.yaml‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -98,7 +98,7 @@ base:
9898
env:
9999
AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
100100
# Native single-node binding verifies these against the generated point.
101-
TOTAL_CPU_DRAM_GB: '225'
101+
TOTAL_CPU_DRAM_GB: '241'
102102
KV_OFFLOADING: dram
103103

104104
# Only client concurrency changes; serving/admission/graph/cache settings stay fixed.

‎inferencex-e2e/configs/nvidia-master.yaml‎

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6198,6 +6198,7 @@ kimik3-fp4-gb200-dynamo-vllm-agentic-dspark-mooncake-dcp16-agg:
61986198
scenarios:
61996199
agentic-coding:
62006200
- dram-utilization: 0.1065
6201+
require-power: true
62016202
search-space:
62026203
- spec-decoding: mtp
62036204
kv-offloading: dram
@@ -6227,6 +6228,7 @@ kimik3-fp4-gb200-dynamo-vllm-agentic-mooncake-dcp16-agg:
62276228
scenarios:
62286229
agentic-coding:
62296230
- dram-utilization: 0.1065
6231+
require-power: true
62306232
search-space:
62316233
- kv-offloading: dram
62326234
kv-offload-backend: { name: mooncake, version: "0.3.12.post1" }
@@ -6363,6 +6365,7 @@ kimik3-fp4-gb200-dynamo-vllm-agentic-dspark-mooncake-tp8pp2:
63636365
scenarios:
63646366
agentic-coding:
63656367
- dram-utilization: 0.8
6368+
require-power: true
63666369
search-space:
63676370
- spec-decoding: mtp
63686371
kv-offloading: dram
@@ -7889,6 +7892,7 @@ kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-disagg:
78897892
# 0.1775 resolves to 160 GB of host DRAM per 4-GPU server node on
78907893
# cluster:gb300-nv, matching the recipe's Mooncake segment size.
78917894
- dram-utilization: 0.1775
7895+
require-power: true
78927896
search-space:
78937897
- spec-decoding: mtp
78947898
kv-offloading: dram
@@ -7960,6 +7964,7 @@ kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-agg:
79607964
scenarios:
79617965
agentic-coding:
79627966
- dram-utilization: 0.1775
7967+
require-power: true
79637968
search-space:
79647969
- spec-decoding: mtp
79657970
kv-offloading: dram
@@ -9227,7 +9232,7 @@ qwen3.8next-fp4-gb300-sglang-agentic-mtp:
92279232
scenarios:
92289233
agentic-coding:
92299234
# Use the GB300 profile's proportional per-worker host DRAM allowance.
9230-
# Generated allowance: 225 GB. Recipe H224 reserves alignment headroom.
9235+
# Generated allowance: 241 GB. Recipe H224 retains its existing cache size.
92319236
- dram-utilization: 1.0
92329237
search-space:
92339238
- tp: 1

‎inferencex-e2e/infx/launch/drivers/srt/power.py‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -49,7 +49,6 @@ class PowerLane:
4949
PowerRule(
5050
Match(any_of("kimik3"), any_of("fp4"), any_of("dynamo-vllm"), agentic=True),
5151
agentx=True,
52-
require=True,
5352
),
5453
PowerRule(Match(frameworks=any_of("dynamo-sglang"), agentic=False), agentx=False),
5554
),
@@ -61,7 +60,6 @@ class PowerLane:
6160
PowerRule(
6261
Match(any_of("kimik3"), any_of("fp4"), any_of("dynamo-vllm"), agentic=True),
6362
agentx=True,
64-
require=True,
6563
),
6664
PowerRule(Match(frameworks=any_of("dynamo-sglang")), agentx=False),
6765
),

‎inferencex-e2e/infx/tests/launch/test_launch_policy.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -73,7 +73,7 @@ def test_power_eligibility_of_each_cluster_lane(
7373
assert (decision.dcgm, decision.agentx, decision.adapter) == (
7474
True, expected == AGENTX, expected == ADAPTER,
7575
) # fmt: skip
76-
assert decision.require is (prefix == "kimik3" and cluster in ("gb200-nv", "gb300-nv"))
76+
assert not decision.require
7777
assert decide_power(cluster, path, _multinode(**env, POWER="0")).dcgm is (
7878
cluster in power.NVL72_POWER_CLUSTERS
7979
)

‎inferencex-e2e/infx/tests/launch/test_nvl72_power_defaults.py‎

Lines changed: 8 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -47,13 +47,19 @@ def test_other_cluster_keeps_explicit_power_policy():
4747
assert not decide_power("h200-dgxc", LaunchPath.SRT_MULTI, request(POWER="0")).dcgm
4848

4949

50-
def test_existing_strict_agentx_rule_stays_strict():
50+
def test_existing_kimi_agentx_row_requires_explicit_flag():
5151
decision = decide_power(
5252
"gb200-nv",
5353
LaunchPath.SRT_MULTI,
5454
request(POWER="1", MODEL_PREFIX="kimik3", FRAMEWORK="dynamo-vllm"),
5555
)
56-
assert decision.dcgm and decision.agentx and decision.cpu and decision.require
56+
assert decision.dcgm and decision.agentx and decision.cpu and not decision.require
57+
strict = decide_power(
58+
"gb200-nv",
59+
LaunchPath.SRT_MULTI,
60+
request(REQUIRE_POWER="1", MODEL_PREFIX="kimik3", FRAMEWORK="dynamo-vllm"),
61+
)
62+
assert strict.require
5763

5864

5965
def test_explicit_requirement_uses_default_nvl72_collector_and_skips_evals():

‎inferencex-e2e/perf-changelog.yaml‎

Lines changed: 24 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9963,8 +9963,31 @@
99639963
- dsv41flash-fp4-gb200-sglang-agentic-dspark
99649964
- dsv41flash-fp4-gb300-vllm-agentic-dspark
99659965
- dsv41flash-fp4-gb300-sglang-agentic-dspark
9966+
- qwen3.5-fp8-gb300-dynamo-sglang-agentic-pp-pareto
9967+
- qwen3.5-fp8-gb300-dynamo-sglang-agentic-agg-pareto
9968+
- qwen3.5-fp8-gb300-dynamo-sglang-agentic-pp-midcurve
9969+
- qwen3.5-fp8-gb300-dynamo-sglang-agentic-agg-midcurve
9970+
- dsv4-fp4-gb200-dynamo-sglang-agentic-agg
9971+
- dsv4-fp4-gb200-dynamo-sglang-agentic-disagg
9972+
- dsv41flash-fp4-gb300-dynamo-sglang-agentic-agg
9973+
- dsv41flash-fp4-gb300-dynamo-sglang-agentic-disagg
9974+
- dsv41flash-fp4-gb300-dynamo-vllm-agentic-dspark-agg
9975+
- dsv41flash-fp4-gb300-dynamo-vllm-agentic-dspark-disagg
9976+
- dsv4-fp4-gb200-dynamo-vllm-agentic-dspark-agg
9977+
- dsv4-fp4-gb200-dynamo-vllm-agentic-dspark-disagg
9978+
- qwen3.8next-fp4-gb300-sglang-agentic-mtp
9979+
- kimik3-fp4-h200-vllm-agentic-latency
9980+
- kimik3-fp4-h200-vllm-agentic-balanced
9981+
- kimik3-fp4-h200-vllm-agentic-simple
9982+
- dsv4-fp8-h200-dynamo-sglang-agentic-agg
9983+
- glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-2p2d
9984+
- glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-1p1d-hicache
9985+
- glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-agg
9986+
- kimik3-fp4-b200-dynamo-vllm-agentic-dspark
9987+
- qwen3.5-fp8-b200-dynamo-sglang-agentic-disagg-mtp
99669988
description:
99679989
- "Enable best-effort noprof GPU and ACPI Grace CPU collection by default for GB200/GB300 benchmarks, including native single-node and AgentX, with the pinned cpu-power-exporter v2.51.1. Explicit required-power AgentX points fail on invalid GPU or expected CPU evidence; ordinary points retain valid benchmark results and independent GPU/CPU validity. Eval-only jobs omit collection."
99689990
- "Place benchmark clients and telemetry on the head clock without changing worker or allocation counts. Native single-node results retain their real topology and power audit. Placement changes require fresh performance results."
9969-
- "Require validated power on the 11 GLM-5.2 GB200 aggregate/disaggregated/MTP AgentX points and 24 DSV4 GB200 aggregate MTP TP8/DEP8 points, retaining their current images, selectors and topology."
9991+
- "Require validated power on the 11 GLM-5.2 GB200 aggregate/disaggregated/MTP AgentX points and 24 DSV4 GB200 aggregate MTP TP8/DEP8 points. Declare the five existing Kimi-K3 GB200/GB300 measured-power arms strict in their matrix rows instead of a cluster-wide Kimi rule. Retain their current images, selectors and topology."
9992+
- "Forward the selected required-power flag through existing H200/B200 AgentX clients, telemetry, and result collection. Align the Qwen3.8-Next GB300 recipe's declared host DRAM with the current runner's 241 GB allowance while retaining its H224 cache setting."
99709993
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3670

0 commit comments

Comments
 (0)