Skip to content

Commit 8a36b64

Browse files
nvpohanhnvpohanh
committed
fix(dsxe): migrate B300 AgentX recipes
将 B300 AgentX 配方迁移到重构后的 inferencex-e2e 布局，并采用 EFA 1.50.0 安装流程。
1 parent 5568982 commit 8a36b64

10 files changed

Lines changed: 979 additions & 1 deletion

File tree

‎inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh‎

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,15 @@
1+
#!/usr/bin/env bash
2+
set -eo pipefail
3+
4+
export EFA_VERSION=1.50.0
5+
curl --retry 3 --retry-delay 2 -fsSL -o aws-efa-installer-${EFA_VERSION}.tar.gz \
6+
https://efa-installer.amazonaws.com/aws-efa-installer-${EFA_VERSION}.tar.gz && \
7+
tar -xf aws-efa-installer-${EFA_VERSION}.tar.gz && \
8+
cd aws-efa-installer && \
9+
apt-get update && \
10+
./efa_installer.sh -y --skip-kmod --skip-limit-conf --no-verify && \
11+
cd .. && rm -rf aws-efa-installer* && \
12+
ldconfig
13+
14+
python3 -m pip uninstall --break-system-packages -y mooncake-transfer-engine-cuda13 mooncake-transfer-engine-efa-cuda13
15+
python3 -m pip install --break-system-packages --no-deps mooncake-transfer-engine-efa-cuda13==0.3.13.post1

‎inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c4-mtp.yaml‎

Lines changed: 128 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,128 @@
1+
schema: 2
2+
name: "agg-b300-tp4-c4-mtp"
3+
4+
# AgentX aggregate topology: one TP4 worker occupies half of one eight-GPU
5+
# B300 node and serves both prefill and decode with DSpark K=6.
6+
7+
model:
8+
path: "deepseek-v4-pro-0813"
9+
container: "dynamo-sglang"
10+
precision: "fp4"
11+
12+
identity:
13+
model:
14+
repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
15+
container:
16+
image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
17+
18+
dynamo:
19+
install: false
20+
21+
health_check:
22+
max_attempts: 1440
23+
interval_seconds: 10
24+
25+
resources:
26+
gpu_type: "b300"
27+
gpus_per_node: 8
28+
services:
29+
- name: etcd
30+
type: etcd
31+
placement:
32+
node: infra
33+
- name: nats
34+
type: nats
35+
placement:
36+
node: infra
37+
options:
38+
max_payload_mb: 32
39+
frontend:
40+
type: dynamo
41+
nginx_session_affinity: true
42+
nginx_session_affinity_header: X-Dynamo-Session-ID
43+
enable_multiple_frontends: false
44+
env:
45+
PIP_BREAK_SYSTEM_PACKAGES: "1"
46+
DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
47+
args:
48+
router-mode: "kv"
49+
router-session-affinity-ttl-secs: "3600"
50+
active-decode-blocks-threshold: "None"
51+
active-prefill-tokens-threshold: "None"
52+
active-prefill-tokens-threshold-frac: "None"
53+
54+
engine: sglang
55+
roles:
56+
agg:
57+
nodes: 1
58+
workers: 1
59+
gpus: 4
60+
61+
env:
62+
SGLANG_RAGGED_VERIFY_MODE: "static"
63+
SGLANG_DEFAULT_THINKING: "1"
64+
SGLANG_DSV4_REASONING_EFFORT: high
65+
PIP_BREAK_SYSTEM_PACKAGES: "1"
66+
SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
67+
SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
68+
SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
69+
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
70+
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
71+
SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
72+
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
73+
SGLANG_OPT_USE_JIT_NORM: "1"
74+
SGLANG_OPT_USE_TOPK_V2: "True"
75+
76+
args:
77+
served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
78+
enable-metrics: true
79+
enable-cache-report: true
80+
trust-remote-code: true
81+
weight-loader-prefetch-checkpoints: true
82+
stream-interval: 10
83+
watchdog-timeout: 1000000
84+
mem-fraction-static: 0.90
85+
page-size: 256
86+
chunked-prefill-size: 8192
87+
max-prefill-tokens: 8192
88+
moe-runner-backend: "flashinfer_mxfp4"
89+
enable-deepseek-v4-fp4-indexer: true
90+
disable-flashinfer-autotune: true
91+
swa-full-tokens-ratio: 0.1
92+
max-running-requests: 8
93+
cuda-graph-max-bs-decode: 8
94+
scheduler-recv-interval: 30
95+
dp-size: 1
96+
tp-size: 4
97+
ep-size: 1
98+
speculative-algorithm: DSPARK
99+
speculative-dspark-block-size: 6
100+
speculative-num-steps: 1
101+
speculative-eagle-topk: 1
102+
speculative-num-draft-tokens: 7
103+
104+
sbatch_directives:
105+
mem: "0"
106+
cpus-per-task: "192"
107+
exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-16"
108+
109+
srun_options:
110+
mem: "0"
111+
container-remap-root: ""
112+
113+
benchmark:
114+
type: custom
115+
command: bash /infmax-workspace/benchmarks/srt_agentic.sh
116+
env:
117+
INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
118+
RESULT_DIR: "/logs/agentic"
119+
PORT: "8000"
120+
IS_MULTINODE: "false"
121+
TP: "4"
122+
PP_SIZE: "1"
123+
PCP_SIZE: "1"
124+
AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
125+
AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
126+
AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
127+
AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
128+
HF_HUB_CACHE: "/hf_hub_cache"

‎inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c8-mtp.yaml‎

Lines changed: 128 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,128 @@
1+
schema: 2
2+
name: "agg-b300-tp4-c8-mtp"
3+
4+
# AgentX aggregate topology: one TP4 worker occupies half of one eight-GPU
5+
# B300 node and serves both prefill and decode with DSpark K=6.
6+
7+
model:
8+
path: "deepseek-v4-pro-0813"
9+
container: "dynamo-sglang"
10+
precision: "fp4"
11+
12+
identity:
13+
model:
14+
repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
15+
container:
16+
image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
17+
18+
dynamo:
19+
install: false
20+
21+
health_check:
22+
max_attempts: 1440
23+
interval_seconds: 10
24+
25+
resources:
26+
gpu_type: "b300"
27+
gpus_per_node: 8
28+
services:
29+
- name: etcd
30+
type: etcd
31+
placement:
32+
node: infra
33+
- name: nats
34+
type: nats
35+
placement:
36+
node: infra
37+
options:
38+
max_payload_mb: 32
39+
frontend:
40+
type: dynamo
41+
nginx_session_affinity: true
42+
nginx_session_affinity_header: X-Dynamo-Session-ID
43+
enable_multiple_frontends: false
44+
env:
45+
PIP_BREAK_SYSTEM_PACKAGES: "1"
46+
DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
47+
args:
48+
router-mode: "kv"
49+
router-session-affinity-ttl-secs: "3600"
50+
active-decode-blocks-threshold: "None"
51+
active-prefill-tokens-threshold: "None"
52+
active-prefill-tokens-threshold-frac: "None"
53+
54+
engine: sglang
55+
roles:
56+
agg:
57+
nodes: 1
58+
workers: 1
59+
gpus: 4
60+
61+
env:
62+
SGLANG_RAGGED_VERIFY_MODE: "static"
63+
SGLANG_DEFAULT_THINKING: "1"
64+
SGLANG_DSV4_REASONING_EFFORT: high
65+
PIP_BREAK_SYSTEM_PACKAGES: "1"
66+
SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
67+
SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
68+
SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
69+
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
70+
SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
71+
SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
72+
SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
73+
SGLANG_OPT_USE_JIT_NORM: "1"
74+
SGLANG_OPT_USE_TOPK_V2: "True"
75+
76+
args:
77+
served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
78+
enable-metrics: true
79+
enable-cache-report: true
80+
trust-remote-code: true
81+
weight-loader-prefetch-checkpoints: true
82+
stream-interval: 10
83+
watchdog-timeout: 1000000
84+
mem-fraction-static: 0.90
85+
page-size: 256
86+
chunked-prefill-size: 8192
87+
max-prefill-tokens: 8192
88+
moe-runner-backend: "flashinfer_mxfp4"
89+
enable-deepseek-v4-fp4-indexer: true
90+
disable-flashinfer-autotune: true
91+
swa-full-tokens-ratio: 0.1
92+
max-running-requests: 16
93+
cuda-graph-max-bs-decode: 16
94+
scheduler-recv-interval: 30
95+
dp-size: 1
96+
tp-size: 4
97+
ep-size: 1
98+
speculative-algorithm: DSPARK
99+
speculative-dspark-block-size: 6
100+
speculative-num-steps: 1
101+
speculative-eagle-topk: 1
102+
speculative-num-draft-tokens: 7
103+
104+
sbatch_directives:
105+
mem: "0"
106+
cpus-per-task: "192"
107+
exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-16"
108+
109+
srun_options:
110+
mem: "0"
111+
container-remap-root: ""
112+
113+
benchmark:
114+
type: custom
115+
command: bash /infmax-workspace/benchmarks/srt_agentic.sh
116+
env:
117+
INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
118+
RESULT_DIR: "/logs/agentic"
119+
PORT: "8000"
120+
IS_MULTINODE: "false"
121+
TP: "4"
122+
PP_SIZE: "1"
123+
PCP_SIZE: "1"
124+
AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
125+
AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
126+
AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
127+
AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
128+
HF_HUB_CACHE: "/hf_hub_cache"

0 commit comments

Comments
 (0)