[AMD][MI35X] Bump Qwen3.5 MXFP4 MI355X SGLang single-node image to v0.5.18

benchmarks/single_node/fixed_seq_len/qwen3.5_fp4_mi355x.sh

@@ -21,6 +21,7 @@ export SGLANG_USE_AITER=1
 export SGLANG_USE_AITER_UNIFIED_ATTN=1
 export AITER_FLYDSL_FORCE=1
 export SGLANG_MAMBA_SSM_DTYPE=bfloat16
+export ROCM_QUICK_REDUCE_QUANTIZATION=INT8
 
 SERVER_LOG=/workspace/server.log
 MEM_FRAC_STATIC=${MEM_FRAC_STATIC:-0.8}
@@ -41,8 +42,9 @@ python3 -m sglang.launch_server --model-path=$MODEL --trust-remote-code \
 --model-loader-extra-config '{"enable_multithread_load": true}' \
 --watchdog-timeout 1200  \
 --disable-radix-cache \
---enable-aiter-allreduce-fusion --max-running-requests $CONC \
+--max-running-requests $CONC \
 --page-size 16 \
+--kv-cache-dtype fp8_e4m3 \
 > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!

benchmarks/single_node/fixed_seq_len/qwen3.5_fp4_mi355x_mtp.sh

@@ -21,6 +21,7 @@ export SGLANG_USE_AITER=1
 export SGLANG_USE_AITER_UNIFIED_ATTN=1
 export AITER_FLYDSL_FORCE=1
 export SGLANG_MAMBA_SSM_DTYPE=bfloat16
+export ROCM_QUICK_REDUCE_QUANTIZATION=INT8
 
 SERVER_LOG=/workspace/server.log
 MEM_FRAC_STATIC=${MEM_FRAC_STATIC:-0.8}
@@ -41,12 +42,13 @@ python3 -m sglang.launch_server --model-path=$MODEL --trust-remote-code \
 --model-loader-extra-config '{"enable_multithread_load": true}' \
 --watchdog-timeout 1200  \
 --disable-radix-cache \
---enable-aiter-allreduce-fusion --max-running-requests $CONC \
+--max-running-requests $CONC \
 --page-size 16 \
 --speculative-algorithm EAGLE \
 --speculative-num-steps 3 \
 --speculative-eagle-topk 1 \
 --speculative-num-draft-tokens 4 \
+--kv-cache-dtype fp8_e4m3 \
 > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!

configs/amd-master.yaml

@@ -313,7 +313,7 @@ qwen3.5-fp8-mi355x-sglang-disagg:
           - "DECODE_MTP_SIZE=0"
 
 qwen3.5-fp4-mi355x-sglang:
-  image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730
+  image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260829
   model: amd/Qwen3.5-397B-A17B-MXFP4
   model-prefix: qwen3.5
   runner: mi355x
@@ -345,7 +345,7 @@ qwen3.5-fp4-mi355x-atom:
       - { tp: 4, conc-start: 4, conc-end: 16 }
 
 qwen3.5-fp4-mi355x-sglang-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730
+  image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260829
   model: amd/Qwen3.5-397B-A17B-MXFP4
   model-prefix: qwen3.5
   runner: mi355x

perf-changelog.yaml

@@ -6642,3 +6642,14 @@
     - "Capture the decode CUDA graph to min(2*CONC, 128) instead of min(CONC, 64). The replay keeps --max-running-requests 2*CONC in flight, so a CONC-sized graph dropped every decode batch above CONC onto the eager path. The 128 cap follows the sibling MI355X AgentX recipe dsv4_fp4_mi355x_sglang_mtp.sh."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2737
 
+- config-keys:
+    - qwen3.5-fp4-mi355x-sglang
+    - qwen3.5-fp4-mi355x-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Bump image from lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730 to lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260829"
+    - "Serve both arms with an fp8_e4m3 KV cache."
+    - "Route multi-GPU collectives through INT8-quantized ROCm quick all-reduce (ROCM_QUICK_REDUCE_QUANTIZATION=INT8) and drop --enable-aiter-allreduce-fusion, aligning both arms with the published SGLang cookbook recipe for MXFP4 on MI355X (https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.5). The two paths are mutually exclusive in SGLang: the AITER fused AR+RMSNorm path is gated on --enable-aiter-allreduce-fusion, and only with it off do collectives fall back to custom all-reduce where the quick-reduce regime applies. This supersedes the aiter allreduce fusion enabled for these two config keys in #1680, so throughput on this arm is not a like-for-like continuation of the previous MXFP4 series."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2640
+