Skip to content

Commit ddcabb1

Browse files
YukioZzzYukioZzz
committed
feat(amd): add qwen3.5-fp4 MI355X SGLang PD-disaggregation
Based on chun-chang/sglang-disagg-qwen3.5 (successful FP8 disagg path). Mirror qwen3.5-fp8-mi355x-sglang-disagg topology; MXFP4 model; MoRI overlay.
1 parent c4e397d commit ddcabb1

4 files changed

Lines changed: 177 additions & 0 deletions

File tree

‎.github/configs/amd-master.yaml‎

Lines changed: 55 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -348,6 +348,61 @@ qwen3.5-fp4-mi355x-atom:
348348
- { tp: 2, conc-start: 4, conc-end: 256 }
349349
- { tp: 4, conc-start: 4, conc-end: 16 }
350350

351+
qwen3.5-fp4-mi355x-sglang-disagg:
352+
image: lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260523
353+
model: amd/Qwen3.5-397B-A17B-MXFP4
354+
model-prefix: qwen3.5
355+
runner: mi355x-disagg
356+
precision: fp4
357+
framework: sglang-disagg
358+
multinode: true
359+
disagg: true
360+
scenarios:
361+
fixed-seq-len:
362+
- isl: 1024
363+
osl: 1024
364+
search-space:
365+
# Mirrors qwen3.5-fp8-mi355x-sglang-disagg (chun-chang/sglang-disagg-qwen3.5 @ 688ebe60).
366+
# Same image, 1P1D TP8/EP1, dp-attn false; MoRI conn.py overlay via job.slurm.
367+
- spec-decoding: "none"
368+
conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
369+
prefill:
370+
num-worker: 1
371+
tp: 8
372+
ep: 1
373+
dp-attn: false
374+
additional-settings:
375+
- "PREFILL_NODES=1"
376+
decode:
377+
num-worker: 1
378+
tp: 8
379+
ep: 1
380+
dp-attn: false
381+
additional-settings:
382+
- "DECODE_NODES=1"
383+
- "DECODE_MTP_SIZE=0"
384+
385+
- isl: 8192
386+
osl: 1024
387+
search-space:
388+
- spec-decoding: "none"
389+
conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
390+
prefill:
391+
num-worker: 1
392+
tp: 8
393+
ep: 1
394+
dp-attn: false
395+
additional-settings:
396+
- "PREFILL_NODES=1"
397+
decode:
398+
num-worker: 1
399+
tp: 8
400+
ep: 1
401+
dp-attn: false
402+
additional-settings:
403+
- "DECODE_NODES=1"
404+
- "DECODE_MTP_SIZE=0"
405+
351406
qwen3.5-fp8-mi355x-sglang-disagg:
352407
image: lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260523
353408
model: Qwen/Qwen3.5-397B-A17B-FP8

‎benchmarks/multi_node/amd_utils/models.yaml‎

Lines changed: 31 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -161,6 +161,37 @@ DeepSeek-R1-0528:
161161
chunked_prefill_size: 262144
162162
cuda_graph_bs_range: "1-128"
163163

164+
Qwen3.5-397B-A17B-MXFP4:
165+
base_flags: "--decode-log-interval 1000 --log-level warning --watchdog-timeout 3600 --load-balance-method round_robin --kv-cache-dtype fp8_e4m3 --attention-backend aiter --disaggregation-transfer-backend mori --moe-dense-tp-size 1"
166+
mtp_flags: ""
167+
dp_flags: "--moe-a2a-backend mori --enable-dp-attention --enable-dp-lm-head"
168+
prefill:
169+
mem_fraction_static: 0.8
170+
disable_radix_cache: true
171+
dp:
172+
max_running_requests: 24
173+
chunked_prefill_size: "MORI_MAX_DISPATCH_TOKENS_PREFILL * PREFILL_TP_SIZE"
174+
cuda_graph_bs: "1 2 3"
175+
no_dp:
176+
max_running_requests: 128
177+
chunked_prefill_size: 262144
178+
cuda_graph_bs_range: "1-128"
179+
decode:
180+
mem_fraction_static: 0.85
181+
prefill_round_robin_balance: true
182+
dp:
183+
max_running_requests: 4096
184+
chunked_prefill_size: "MORI_MAX_DISPATCH_TOKENS_DECODE * DECODE_TP_SIZE"
185+
cuda_graph_bs_range: "1-160"
186+
ep_only:
187+
max_running_requests: 256
188+
chunked_prefill_size: 262144
189+
cuda_graph_bs_range: "1-256"
190+
no_dp:
191+
max_running_requests: 128
192+
chunked_prefill_size: 262144
193+
cuda_graph_bs_range: "1-128"
194+
164195
Qwen3.5-397B-A17B-FP8:
165196
base_flags: "--decode-log-interval 1000 --log-level warning --watchdog-timeout 3600 --load-balance-method round_robin --kv-cache-dtype fp8_e4m3 --attention-backend aiter --disaggregation-transfer-backend mori --moe-dense-tp-size 1"
166197
mtp_flags: ""

‎benchmarks/multi_node/qwen3.5_fp4_mi355x_sglang-disagg.sh‎

Lines changed: 83 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,83 @@
1+
#!/usr/bin/env bash
2+
3+
source "$(dirname "$0")/../benchmark_lib.sh"
4+
5+
check_env_vars \
6+
CONC_LIST \
7+
ISL \
8+
OSL \
9+
IMAGE \
10+
SPEC_DECODING \
11+
MODEL_PATH \
12+
PREFILL_NUM_WORKERS \
13+
PREFILL_TP \
14+
PREFILL_EP \
15+
PREFILL_DP_ATTN \
16+
DECODE_NUM_WORKERS \
17+
DECODE_TP \
18+
DECODE_EP \
19+
DECODE_DP_ATTN \
20+
PREFILL_NODES \
21+
DECODE_NODES \
22+
RANDOM_RANGE_RATIO
23+
24+
if [[ -n "$SLURM_JOB_ID" ]]; then
25+
echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
26+
fi
27+
28+
set -x
29+
30+
# Use upstreamed multi_node scripts (no external clone needed)
31+
cd "$GITHUB_WORKSPACE/benchmarks/multi_node/amd_utils" || exit 1
32+
33+
# Set up SGL launch script-specific environment variables
34+
export TIME_LIMIT="08:00:00"
35+
export MODEL_PATH=$MODEL_PATH
36+
export MODEL_NAME=$MODEL_NAME
37+
export CONTAINER_IMAGE=$IMAGE
38+
39+
if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
40+
export PREFILL_ENABLE_EP=false
41+
else
42+
export PREFILL_ENABLE_EP=true
43+
fi
44+
45+
if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
46+
export PREFILL_ENABLE_DP=true
47+
else
48+
export PREFILL_ENABLE_DP=false
49+
fi
50+
51+
if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
52+
export DECODE_ENABLE_EP=false
53+
else
54+
export DECODE_ENABLE_EP=true
55+
fi
56+
57+
if [[ "$DECODE_DP_ATTN" == "true" ]]; then
58+
export DECODE_ENABLE_DP=true
59+
else
60+
export DECODE_ENABLE_DP=false
61+
fi
62+
63+
# Launch jobs based on ISL/OSL
64+
# Replace ' ' in CONC_LIST with 'x' such that the concurrency list is represented
65+
# by a list of numbers delimited by 'x'. This is because of how the underlying launch script
66+
# expects the concurrencies.
67+
JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
68+
$PREFILL_NUM_WORKERS \
69+
$DECODE_NODES \
70+
$DECODE_NUM_WORKERS \
71+
$ISL $OSL "${CONC_LIST// /x}" inf \
72+
${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
73+
${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
74+
${PREFILL_TP} ${DECODE_TP} \
75+
${RANDOM_RANGE_RATIO} \
76+
${NODE_LIST:-})
77+
78+
if [[ $? -ne 0 ]]; then
79+
echo "Failed to submit job" >&2
80+
exit 1
81+
fi
82+
83+
echo "$JOB_ID"

‎perf-changelog.yaml‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2307,3 +2307,11 @@
23072307
- "Tune DSv4 FP4 MI355X SGLang runtime envs: enable aiter MHC pre/post, and enable triton swa prepare kernel."
23082308
- "Add --context-length. Add --enable-prefill-delayer for dp config"
23092309
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1300
2310+
2311+
- config-keys:
2312+
- qwen3.5-fp4-mi355x-sglang-disagg
2313+
description:
2314+
- "Add Qwen3.5-397B-A17B-MXFP4 MI355X SGLang PD-disaggregation (base: chun-chang/sglang-disagg-qwen3.5)"
2315+
- "Mirror qwen3.5-fp8-mi355x-sglang-disagg: v0.5.12.post1 image, 1P1D TP8/EP1, dp-attn false, conc [8..512]"
2316+
- "MoRI conn.py overlay (48e459bd) via job.slurm; launcher qwen3.5_fp4_mi355x_sglang-disagg.sh"
2317+
pr-link: XXX

0 commit comments

Comments
 (0)