[AMD/ROCM] GLM5.1 FP4 (MXFP4) MI355X Support

.github/configs/amd-master.yaml

@@ -337,6 +337,26 @@ glm5-fp8-mi355x-atom:
     search-space:
     - { tp: 8, conc-start: 4, conc-end: 256 }
 
+glm5.1-fp4-mi355x-sglang:
+  image: lmsysorg/sglang-rocm:v0.5.10rc0-rocm720-mi35x-20260415
+  model: amd/GLM-5.1-MXFP4
+  model-prefix: glm5.1
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 2, conc-start: 4, conc-end: 256 }
+    - { tp: 4, conc-start: 4, conc-end: 16 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 2, conc-start: 4, conc-end: 256 }
+    - { tp: 4, conc-start: 4, conc-end: 16 }
+
 kimik2.5-int4-mi355x-vllm:
   image: vllm/vllm-openai-rocm:v0.18.0
   model: moonshotai/Kimi-K2.5

benchmarks/single_node/glm5.1_fp4_mi355x.sh

@@ -0,0 +1,83 @@
+#!/usr/bin/env bash
+set -x
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+# ROCm / SGLang performance tuning for MI355X
+export SGLANG_ROCM_FUSED_DECODE_MLA=0
+export ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export SAFETENSORS_FAST_GPU=1
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+CONTEXT_LENGTH=$((ISL + OSL + 32))
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+pip install -U transformers
+
+python3 -m sglang.launch_server \
+    --model-path $MODEL \
+    --host=0.0.0.0 \
+    --port $PORT \
+    --tensor-parallel-size $TP \
+    --trust-remote-code \
+    --cuda-graph-max-bs $CONC \
+    --context-length $CONTEXT_LENGTH \
+    --mem-fraction-static 0.85 \
+    --tool-call-parser glm47 \
+    --reasoning-parser glm45 \
+    --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 8}' \
+    --nsa-prefill-backend tilelang \
+    --nsa-decode-backend tilelang $EVAL_CONTEXT_ARGS  \
+    --kv-cache-dtype fp8_e4m3 \
+    --tokenizer-worker-num $((TP*2)) \
+    --disable-radix-cache> $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -1,3 +1,10 @@
+- config-keys:
+    - glm5.1-fp4-mi355x-sglang
+  description:
+    - "Add GLM5.1 MXFP4 (FP4) MI355X SGLang Support" 
+    - "Container : lmsysorg/sglang-rocm:v0.5.10rc0-rocm720-mi35x-20260415"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1098
+
 - config-keys:
     - kimik2.5-fp4-gb200-dynamo-trt
   description: