[AMD][MI35X] Qwen3.5-fp8 SGLang single-node benchmark (#1669)

* [AMD] Switch Qwen3.5 FP8 MI355X benchmarks to aiter attention backend

* [AMD] Bump Qwen3.5 FP8 MI355X image to v0.5.12.post1-rocm720-mi35x-20260528

* Update change log

* Change mem-fraction-static back to 0.8

* Add context len args and limit batch to CONC size

* Fix typo

* Fix typo

* Change config to tp4

---------

Co-authored-by: jacky.cheng <yichiche@amd.com>

Author: 1am9trash

.github/configs/amd-master.yaml

@@ -261,7 +261,7 @@ qwen3.5-fp8-mi325x-sglang:
       - { tp: 8, conc-start: 4, conc-end: 64 }
 
 qwen3.5-fp8-mi355x-sglang:
-  image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260517
+  image: lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260528
   model: Qwen/Qwen3.5-397B-A17B-FP8
   model-prefix: qwen3.5
   runner: mi355x
@@ -273,17 +273,14 @@ qwen3.5-fp8-mi355x-sglang:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 8, ep: 1, conc-start: 4, conc-end: 32 }
-      - { tp: 8, ep: 8, conc-start: 64, conc-end: 256 }
-      - { tp: 2, ep: 2, conc-start: 128, conc-end: 256 }
+      - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 }
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 2, ep: 2, conc-start: 4, conc-end: 32 }
-      - { tp: 4, ep: 1, conc-start: 32, conc-end: 256 }
+      - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 }
 
 qwen3.5-fp8-mi355x-sglang-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260517
+  image: lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260528
   model: Qwen/Qwen3.5-397B-A17B-FP8
   model-prefix: qwen3.5
   runner: mi355x
@@ -295,14 +292,11 @@ qwen3.5-fp8-mi355x-sglang-mtp:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 8, ep: 1, conc-start: 4, conc-end: 32, spec-decoding: mtp }
-      - { tp: 8, ep: 8, conc-start: 64, conc-end: 256, spec-decoding: mtp }
-      - { tp: 2, ep: 2, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, ep: 1, conc-start: 4, conc-end: 256, spec-decoding: mtp }
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 2, ep: 2, conc-start: 4, conc-end: 32, spec-decoding: mtp }
-      - { tp: 4, ep: 1, conc-start: 32, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, ep: 1, conc-start: 4, conc-end: 256, spec-decoding: mtp }
 
 # Diverged from qwen3.5-fp8-mi355x-sglang (agentic-coding sibling). Metadata is
 # identical to origin/main's qwen3.5-fp8-mi355x-sglang; the split exists because this

benchmarks/single_node/fixed_seq_len/qwen3.5_fp8_mi355x.sh

@@ -18,9 +18,11 @@ fi
 
 if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
+export SGLANG_USE_AITER_UNIFIED_ATTN=1
+export SGLANG_USE_AITER=1
+
 SERVER_LOG=/workspace/server.log
 CONTEXT_LENGTH=$((ISL + OSL + 20))
-MAX_PREFILL_TOKENS=32768
 
 EVAL_CONTEXT_ARGS=""
 if [ "${EVAL_ONLY}" = "true" ]; then
@@ -32,7 +34,7 @@ fi
 start_gpu_monitor
 
 python3 -m sglang.launch_server \
-    --attention-backend triton \
+    --attention-backend aiter \
     --model-path $MODEL \
     --host=0.0.0.0 \
     --port $PORT \
@@ -41,11 +43,14 @@ python3 -m sglang.launch_server \
     --trust-remote-code \
     --tokenizer-worker-num 6 \
     --enable-aiter-allreduce-fusion \
+    --max-running-requests $CONC \
     --cuda-graph-max-bs $CONC \
     --disable-radix-cache \
-    --max-prefill-tokens $MAX_PREFILL_TOKENS \
+    --chunked-prefill-size 32768 \
     --scheduler-recv-interval 30 \
-    --mem-fraction-static 0.8 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+    --mem-fraction-static 0.8 \
+    --model-loader-extra-config '{"enable_multithread_load": true}' \
+    --page-size 16 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

benchmarks/single_node/fixed_seq_len/qwen3.5_fp8_mi355x_mtp.sh

@@ -18,9 +18,11 @@ fi
 
 if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
+export SGLANG_USE_AITER_UNIFIED_ATTN=1
+export SGLANG_USE_AITER=1
+
 SERVER_LOG=/workspace/server.log
 CONTEXT_LENGTH=$((ISL + OSL + 20))
-MAX_PREFILL_TOKENS=32768
 
 EVAL_CONTEXT_ARGS=""
 if [ "${EVAL_ONLY}" = "true" ]; then
@@ -32,7 +34,7 @@ fi
 start_gpu_monitor
 
 python3 -m sglang.launch_server \
-    --attention-backend triton \
+    --attention-backend aiter \
     --model-path $MODEL \
     --host=0.0.0.0 \
     --port $PORT \
@@ -41,11 +43,14 @@ python3 -m sglang.launch_server \
     --trust-remote-code \
     --tokenizer-worker-num 6 \
     --enable-aiter-allreduce-fusion \
+    --max-running-requests $CONC \
     --cuda-graph-max-bs $CONC \
     --disable-radix-cache \
-    --max-prefill-tokens $MAX_PREFILL_TOKENS \
+    --chunked-prefill-size 32768 \
     --scheduler-recv-interval 30 \
     --mem-fraction-static 0.8 \
+    --model-loader-extra-config '{"enable_multithread_load": true}' \
+    --page-size 16 \
     --speculative-algorithm EAGLE \
     --speculative-num-steps 3 \
     --speculative-eagle-topk 1 \

perf-changelog.yaml

@@ -3481,6 +3481,14 @@
     - "Set max-running-requests=256, chunked-prefill-size=16384, mem-fraction-static=0.8, cuda-graph-max-bs=CONC, and enable symm-mem"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1544
 
+- config-keys:
+    - qwen3.5-fp8-mi355x-sglang
+    - qwen3.5-fp8-mi355x-sglang-mtp
+  description:
+    - "Bump image to lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260528."
+    - "Update script for aiter attention backend from triton."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1669
+
 - config-keys:
     - minimaxm2.5-fp8-h200-vllm
   description: