Skip to content

Commit 213a1d2

Browse files
Simplify Qwen H100 TEP sweep config
1 parent 4587b6e commit 213a1d2

2 files changed

Lines changed: 27 additions & 35 deletions

File tree

.github/configs/nvidia-master.yaml

Lines changed: 2 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -9209,14 +9209,12 @@ qwen3.5-fp8-h100-sglang:
92099209
osl: 1024
92109210
search-space:
92119211
- { tp: 8, ep: 1, conc-start: 1, conc-end: 8 }
9212-
- { tp: 8, ep: 8, conc-start: 16, conc-end: 64 }
9213-
- { tp: 8, ep: 8, conc-start: 128, conc-end: 256 }
9212+
- { tp: 8, ep: 8, conc-start: 16, conc-end: 256 }
92149213
- isl: 8192
92159214
osl: 1024
92169215
search-space:
92179216
- { tp: 8, ep: 1, conc-start: 1, conc-end: 8 }
9218-
- { tp: 8, ep: 8, conc-start: 16, conc-end: 64 }
9219-
- { tp: 8, ep: 8, conc-start: 128, conc-end: 256 }
9217+
- { tp: 8, ep: 8, conc-start: 16, conc-end: 256 }
92209218

92219219
qwen3.5-fp8-h100-sglang-mtp:
92229220
image: lmsysorg/sglang:v0.5.12-cu130

benchmarks/single_node/qwen3.5_fp8_h100.sh

Lines changed: 25 additions & 31 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,6 @@ source "$(dirname "$0")/../benchmark_lib.sh"
88
check_env_vars \
99
MODEL \
1010
TP \
11-
DP_ATTENTION \
1211
CONC \
1312
ISL \
1413
OSL \
@@ -38,42 +37,37 @@ if [ "${EP_SIZE}" -gt 1 ]; then
3837
fi
3938

4039
SCHEDULER_RECV_INTERVAL=
41-
if [ "${DP_ATTENTION}" != "true" ]; then
42-
case "$CONC" in
43-
1|2|4)
44-
SCHEDULER_RECV_INTERVAL=2
45-
;;
46-
8)
47-
SCHEDULER_RECV_INTERVAL=60
48-
;;
49-
16)
50-
SCHEDULER_RECV_INTERVAL=30
51-
;;
52-
32)
53-
SCHEDULER_RECV_INTERVAL=1200
54-
;;
55-
64)
56-
SCHEDULER_RECV_INTERVAL=600
57-
;;
58-
128|256)
59-
SCHEDULER_RECV_INTERVAL=1920
60-
;;
61-
*)
62-
echo "Unsupported CONC=$CONC for qwen3.5 FP8 H100 SGLang recipe" >&2
63-
exit 1
64-
;;
65-
esac
66-
fi
40+
case "$CONC" in
41+
1|2|4)
42+
SCHEDULER_RECV_INTERVAL=2
43+
;;
44+
8)
45+
SCHEDULER_RECV_INTERVAL=60
46+
;;
47+
16)
48+
SCHEDULER_RECV_INTERVAL=30
49+
;;
50+
32)
51+
SCHEDULER_RECV_INTERVAL=1200
52+
;;
53+
64)
54+
SCHEDULER_RECV_INTERVAL=600
55+
;;
56+
128|256)
57+
SCHEDULER_RECV_INTERVAL=1920
58+
;;
59+
*)
60+
echo "Unsupported CONC=$CONC for qwen3.5 FP8 H100 SGLang recipe" >&2
61+
exit 1
62+
;;
63+
esac
6764

6865
SCHEDULER_ARGS=()
6966
if [ -n "$SCHEDULER_RECV_INTERVAL" ]; then
7067
SCHEDULER_ARGS=(--scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL")
7168
fi
72-
if [ "${DP_ATTENTION}" = "true" ]; then
73-
PARALLEL_ARGS+=(--dp-size "$TP" --enable-dp-attention)
74-
fi
7569

76-
echo "TP: $TP, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION, CONC: $CONC, ISL: $ISL, OSL: $OSL, MAX_SEQ_LEN: $MAX_SEQ_LEN"
70+
echo "TP: $TP, EP_SIZE: $EP_SIZE, CONC: $CONC, ISL: $ISL, OSL: $OSL, MAX_SEQ_LEN: $MAX_SEQ_LEN"
7771
echo "SCHEDULER_RECV_INTERVAL: ${SCHEDULER_RECV_INTERVAL:-none}"
7872
echo "SCHEDULER_ARGS: ${SCHEDULER_ARGS[*]}"
7973

0 commit comments

Comments
 (0)