Skip to content

[AMD][MI30X]Update Qwen3.5 perf #986

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
chunfangamd merged 6 commits into from
Apr 15, 2026
Merged

[AMD][MI30X]Update Qwen3.5 perf #986

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
a9e3221
Update Qwen3.5 benchmark configurations and scripts for improved perf…
Apr 9, 2026
f05c916
Update Qwen3.5 benchmark PR link in perf-changelog.yaml
Apr 9, 2026
c647a86
Remove --ep-size parameter from Qwen3.5 benchmark scripts for BF16 an…
Apr 9, 2026
2c17191
Update Qwen3.5 benchmark scripts to use 'aiter' as the attention back…
Apr 9, 2026
970742b
Update Qwen3.5 benchmark configurations to use stable Docker image ve…
Apr 9, 2026
a1c288e
Merge branch 'main' into todd/qwen35-mi30x
chunfangamd Apr 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions .github/configs/amd-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,7 +132,7 @@ qwen3.5-bf16-mi355x-sglang:
- { tp: 8, conc-start: 4, conc-end: 64 }

qwen3.5-bf16-mi300x-sglang:
image: lmsysorg/sglang:v0.5.9-rocm720-mi30x
image: lmsysorg/sglang:v0.5.10-rocm720-mi30x
model: Qwen/Qwen3.5-397B-A17B
model-prefix: qwen3.5
runner: mi300x
Expand All @@ -150,7 +150,7 @@ qwen3.5-bf16-mi300x-sglang:
- { tp: 8, conc-start: 4, conc-end: 64 }

qwen3.5-bf16-mi325x-sglang:
image: lmsysorg/sglang:v0.5.9-rocm720-mi30x
image: lmsysorg/sglang:v0.5.10-rocm720-mi30x
model: Qwen/Qwen3.5-397B-A17B
model-prefix: qwen3.5
runner: mi325x
Expand All @@ -168,7 +168,7 @@ qwen3.5-bf16-mi325x-sglang:
- { tp: 8, conc-start: 4, conc-end: 64 }

qwen3.5-fp8-mi325x-sglang:
image: lmsysorg/sglang:v0.5.9-rocm720-mi30x
image: lmsysorg/sglang:v0.5.10-rocm720-mi30x
model: Qwen/Qwen3.5-397B-A17B-FP8
model-prefix: qwen3.5
runner: mi325x
Expand Down Expand Up @@ -224,7 +224,7 @@ qwen3.5-fp4-mi355x-sglang:
- { tp: 4, conc-start: 4, conc-end: 32 }

qwen3.5-fp8-mi300x-sglang:
image: lmsysorg/sglang:v0.5.9-rocm720-mi30x
image: lmsysorg/sglang:v0.5.10-rocm720-mi30x
model: Qwen/Qwen3.5-397B-A17B-FP8
model-prefix: qwen3.5
runner: mi300x
Expand Down
15 changes: 12 additions & 3 deletions benchmarks/single_node/qwen3.5_bf16_mi300x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,25 +19,34 @@ hf download "$MODEL"

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
CONTEXT_LENGTH=$((ISL + OSL + 20))
Comment thread
zhentaocc marked this conversation as resolved.
MAX_PREFILL_TOKENS=32768
Comment thread
zhentaocc marked this conversation as resolved.

EVAL_CONTEXT_ARGS=""
if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
fi
# Start GPU monitoring (power, temperature, clocks every second)
start_gpu_monitor

# following Andy Luo linkedin's recipe https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/
python3 -m sglang.launch_server \
--attention-backend triton \
--attention-backend aiter \
Comment thread
chunfangamd marked this conversation as resolved.
--model-path $MODEL \
--host=0.0.0.0 \
--port $PORT \
--tensor-parallel-size $TP \
Comment thread
zhentaocc marked this conversation as resolved.
--data-parallel-size 1 \
Comment thread
zhentaocc marked this conversation as resolved.
--trust-remote-code \
--mem-fraction-static 0.8 \
--disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
--tokenizer-worker-num 6 \
Comment thread
zhentaocc marked this conversation as resolved.
--enable-aiter-allreduce-fusion \
Comment thread
zhentaocc marked this conversation as resolved.
--cuda-graph-max-bs $CONC \
--disable-radix-cache \
--max-prefill-tokens $MAX_PREFILL_TOKENS \
--scheduler-recv-interval 30 \
--mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
Comment thread
zhentaocc marked this conversation as resolved.

SERVER_PID=$!

Expand Down
15 changes: 12 additions & 3 deletions benchmarks/single_node/qwen3.5_bf16_mi325x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,25 +19,34 @@ hf download "$MODEL"

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
CONTEXT_LENGTH=$((ISL + OSL + 20))
MAX_PREFILL_TOKENS=32768

EVAL_CONTEXT_ARGS=""
if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
fi
# Start GPU monitoring (power, temperature, clocks every second)
start_gpu_monitor

# following Andy Luo linkedin's recipe https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/
python3 -m sglang.launch_server \
--attention-backend triton \
--attention-backend aiter \
--model-path $MODEL \
--host=0.0.0.0 \
--port $PORT \
--tensor-parallel-size $TP \
--data-parallel-size 1 \
--trust-remote-code \
--mem-fraction-static 0.8 \
--disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
--tokenizer-worker-num 6 \
--enable-aiter-allreduce-fusion \
Comment thread
zhentaocc marked this conversation as resolved.
--cuda-graph-max-bs $CONC \
--disable-radix-cache \
--max-prefill-tokens $MAX_PREFILL_TOKENS \
--scheduler-recv-interval 30 \
--mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &

SERVER_PID=$!

Expand Down
16 changes: 14 additions & 2 deletions benchmarks/single_node/qwen3.5_bf16_mi355x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,8 @@ check_env_vars \
ISL \
OSL \
RANDOM_RANGE_RATIO \
RESULT_FILENAME
RESULT_FILENAME \
EP_SIZE

if [[ -n "$SLURM_JOB_ID" ]]; then
echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
Expand All @@ -19,11 +20,14 @@ hf download "$MODEL"

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
CONTEXT_LENGTH=$((ISL + OSL + 20))
MAX_PREFILL_TOKENS=32768

EVAL_CONTEXT_ARGS=""
if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
fi
# Start GPU monitoring (power, temperature, clocks every second)
start_gpu_monitor
Expand All @@ -34,8 +38,16 @@ python3 -m sglang.launch_server \
--host=0.0.0.0 \
--port $PORT \
--tensor-parallel-size $TP \
--ep-size $EP_SIZE \
--data-parallel-size 1 \
--trust-remote-code \
--mem-fraction-static 0.8 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
--tokenizer-worker-num 6 \
--enable-aiter-allreduce-fusion \
--cuda-graph-max-bs $CONC \
--disable-radix-cache \
--max-prefill-tokens $MAX_PREFILL_TOKENS \
--scheduler-recv-interval 30 \
--mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &

SERVER_PID=$!

Expand Down
15 changes: 12 additions & 3 deletions benchmarks/single_node/qwen3.5_fp8_mi300x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,26 +19,35 @@ hf download "$MODEL"

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
CONTEXT_LENGTH=$((ISL + OSL + 20))
MAX_PREFILL_TOKENS=32768

EVAL_CONTEXT_ARGS=""
if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
fi
# Start GPU monitoring (power, temperature, clocks every second)
start_gpu_monitor

# following AMD Andy linkedin's recipe
# https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/
python3 -m sglang.launch_server \
--attention-backend triton \
--attention-backend aiter \
--model-path $MODEL \
--host=0.0.0.0 \
--port $PORT \
--tensor-parallel-size $TP \
--data-parallel-size 1 \
--trust-remote-code \
--mem-fraction-static 0.8 \
--disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
--tokenizer-worker-num 6 \
--enable-aiter-allreduce-fusion \
--cuda-graph-max-bs $CONC \
--disable-radix-cache \
--max-prefill-tokens $MAX_PREFILL_TOKENS \
--scheduler-recv-interval 30 \
--mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &

SERVER_PID=$!

Expand Down
15 changes: 12 additions & 3 deletions benchmarks/single_node/qwen3.5_fp8_mi325x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,26 +19,35 @@ hf download "$MODEL"

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
CONTEXT_LENGTH=$((ISL + OSL + 20))
MAX_PREFILL_TOKENS=32768

EVAL_CONTEXT_ARGS=""
if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
fi
# Start GPU monitoring (power, temperature, clocks every second)
start_gpu_monitor

# following AMD Andy linkedin's recipe
# https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/
python3 -m sglang.launch_server \
--attention-backend triton \
--attention-backend aiter \
--model-path $MODEL \
--host=0.0.0.0 \
--port $PORT \
--tensor-parallel-size $TP \
--data-parallel-size 1 \
--trust-remote-code \
--mem-fraction-static 0.8 \
--disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
--tokenizer-worker-num 6 \
--enable-aiter-allreduce-fusion \
--cuda-graph-max-bs $CONC \
--disable-radix-cache \
--max-prefill-tokens $MAX_PREFILL_TOKENS \
--scheduler-recv-interval 30 \
--mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &

SERVER_PID=$!

Expand Down
16 changes: 14 additions & 2 deletions benchmarks/single_node/qwen3.5_fp8_mi355x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,8 @@ check_env_vars \
ISL \
OSL \
RANDOM_RANGE_RATIO \
RESULT_FILENAME
RESULT_FILENAME \
EP_SIZE

if [[ -n "$SLURM_JOB_ID" ]]; then
echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
Expand All @@ -19,11 +20,14 @@ hf download "$MODEL"

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
CONTEXT_LENGTH=$((ISL + OSL + 20))
MAX_PREFILL_TOKENS=32768

EVAL_CONTEXT_ARGS=""
if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH"
fi
# Start GPU monitoring (power, temperature, clocks every second)
start_gpu_monitor
Expand All @@ -34,8 +38,16 @@ python3 -m sglang.launch_server \
--host=0.0.0.0 \
--port $PORT \
--tensor-parallel-size $TP \
--ep-size $EP_SIZE \
--data-parallel-size 1 \
--trust-remote-code \
--mem-fraction-static 0.8 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
--tokenizer-worker-num 6 \
--enable-aiter-allreduce-fusion \
--cuda-graph-max-bs $CONC \
--disable-radix-cache \
--max-prefill-tokens $MAX_PREFILL_TOKENS \
--scheduler-recv-interval 30 \
--mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &

SERVER_PID=$!

Expand Down
10 changes: 10 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1308,6 +1308,16 @@
- "Configs: 1k1k (TP4 conc 4-128), 8k1k (TP4 conc 4-128)"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/820

- config-keys:
- qwen3.5-bf16-mi300x-sglang
- qwen3.5-bf16-mi325x-sglang
- qwen3.5-fp8-mi300x-sglang
- qwen3.5-fp8-mi325x-sglang
description:
- "Update cli args of Qwen3.5 FP8 and BF16 SGLang benchmarks for MI300X and MI325X to achieve better performance"
- "Use lmsysorg/sglang:v0.5.10-rocm720-mi30x"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/986

- config-keys:
- glm5-fp4-b200-sglang
description:
Expand Down
Loading