Skip to content

Commit c171323

Browse files
Simplify Qwen H100 TEP sweep config
1 parent 29fb678 commit c171323

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
@@ -9237,14 +9237,12 @@ qwen3.5-fp8-h100-sglang:
92379237
osl: 1024
92389238
search-space:
92399239
- { tp: 8, ep: 1, conc-start: 1, conc-end: 8 }
9240-
- { tp: 8, ep: 8, conc-start: 16, conc-end: 64 }
9241-
- { tp: 8, ep: 8, conc-start: 128, conc-end: 256 }
9240+
- { tp: 8, ep: 8, conc-start: 16, conc-end: 256 }
92429241
- isl: 8192
92439242
osl: 1024
92449243
search-space:
92459244
- { tp: 8, ep: 1, conc-start: 1, conc-end: 8 }
9246-
- { tp: 8, ep: 8, conc-start: 16, conc-end: 64 }
9247-
- { tp: 8, ep: 8, conc-start: 128, conc-end: 256 }
9245+
- { tp: 8, ep: 8, conc-start: 16, conc-end: 256 }
92489246

92499247
qwen3.5-fp8-h100-sglang-mtp:
92509248
image: lmsysorg/sglang:v0.5.12-cu130

benchmarks/single_node/fixed_seq_len/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 "$0")/../../benchmark_lib.sh"
88
check_env_vars \
99
MODEL \
1010
TP \
11-
DP_ATTENTION \
1211
CONC \
1312
ISL \
1413
OSL \
@@ -37,42 +36,37 @@ if [ "${EP_SIZE}" -gt 1 ]; then
3736
fi
3837

3938
SCHEDULER_RECV_INTERVAL=
40-
if [ "${DP_ATTENTION}" != "true" ]; then
41-
case "$CONC" in
42-
1|2|4)
43-
SCHEDULER_RECV_INTERVAL=2
44-
;;
45-
8)
46-
SCHEDULER_RECV_INTERVAL=60
47-
;;
48-
16)
49-
SCHEDULER_RECV_INTERVAL=30
50-
;;
51-
32)
52-
SCHEDULER_RECV_INTERVAL=1200
53-
;;
54-
64)
55-
SCHEDULER_RECV_INTERVAL=600
56-
;;
57-
128|256)
58-
SCHEDULER_RECV_INTERVAL=1920
59-
;;
60-
*)
61-
echo "Unsupported CONC=$CONC for qwen3.5 FP8 H100 SGLang recipe" >&2
62-
exit 1
63-
;;
64-
esac
65-
fi
39+
case "$CONC" in
40+
1|2|4)
41+
SCHEDULER_RECV_INTERVAL=2
42+
;;
43+
8)
44+
SCHEDULER_RECV_INTERVAL=60
45+
;;
46+
16)
47+
SCHEDULER_RECV_INTERVAL=30
48+
;;
49+
32)
50+
SCHEDULER_RECV_INTERVAL=1200
51+
;;
52+
64)
53+
SCHEDULER_RECV_INTERVAL=600
54+
;;
55+
128|256)
56+
SCHEDULER_RECV_INTERVAL=1920
57+
;;
58+
*)
59+
echo "Unsupported CONC=$CONC for qwen3.5 FP8 H100 SGLang recipe" >&2
60+
exit 1
61+
;;
62+
esac
6663

6764
SCHEDULER_ARGS=()
6865
if [ -n "$SCHEDULER_RECV_INTERVAL" ]; then
6966
SCHEDULER_ARGS=(--scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL")
7067
fi
71-
if [ "${DP_ATTENTION}" = "true" ]; then
72-
PARALLEL_ARGS+=(--dp-size "$TP" --enable-dp-attention)
73-
fi
7468

75-
echo "TP: $TP, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION, CONC: $CONC, ISL: $ISL, OSL: $OSL, MAX_SEQ_LEN: $MAX_SEQ_LEN"
69+
echo "TP: $TP, EP_SIZE: $EP_SIZE, CONC: $CONC, ISL: $ISL, OSL: $OSL, MAX_SEQ_LEN: $MAX_SEQ_LEN"
7670
echo "SCHEDULER_RECV_INTERVAL: ${SCHEDULER_RECV_INTERVAL:-none}"
7771
echo "SCHEDULER_ARGS: ${SCHEDULER_ARGS[*]}"
7872

0 commit comments

Comments
 (0)