SemiAnalysisAI · ajith-sirra-amd · Aug 3, 2026 · Aug 3, 2026
diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -0,0 +1,278 @@
+#!/usr/bin/env bash
+set -eo pipefail
+set -x
+
+# Agentic trace replay benchmark for GLM-5.2 MXFP4 on MI355X using SGLang with
+# EAGLE/MTP speculative decoding.
+#
+# Spec-decode only, per the AgentX policy that agentic recipes are run and
+# published with speculative decoding enabled rather than as an STP/MTP A/B
+# (MODELS.md: GLM-5.2 agentic non-MTP is deprecated after 2026-08-03). The
+# non-MTP arm is neither wired into the master config nor kept as a separate
+# script.
+#
+# Sibling of agentic/glm5.2_fp4_b300_sglang_mtp.sh (NVFP4/B300): same nextn
+# head, same draft length, same golden AL; the serve flags below are the ROCm
+# ones (tilelang DSA prefill/decode, MXFP4 checkpoint, HiCache or Mooncake host
+# offload).
+#
+# Required env vars:
+#   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR, DURATION,
+#   EP_SIZE, DP_ATTENTION
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+    echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+# ROCR/HIP visibility under slurm cgroups.
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+
+if [[ -n "$MODEL_PATH" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+rocm-smi || true
+amd-smi || true
+
+
+GPU_CLEAN=false
+for i in $(seq 1 90); do
+    VRAM_MAX=$(rocm-smi --showmemuse 2>/dev/null | grep -oE "GPU Memory Allocated \(VRAM%\): [0-9]+" | awk '{if ($NF > m) m = $NF} END {print m+0}')
+    if [ "${VRAM_MAX:-0}" -le 10 ]; then echo "GPUs clean (vram%max=$VRAM_MAX after $((i*10))s)"; GPU_CLEAN=true; break; fi
+    echo "waiting for prior-job GPU memory reclaim: vram%max=$VRAM_MAX"; sleep 10
+done
+[ "$GPU_CLEAN" = "true" ] || { echo "Error: GPUs still draining prior job's memory after 15min" >&2; exit 1; }
+
+resolve_trace_source
+install_agentic_deps
+
+SERVER_LOG="$RESULT_DIR/server.log"
+ROUTER_LOG="$RESULT_DIR/router.log"
+mkdir -p "$RESULT_DIR"
+
+export PYTHONNOUSERSITE=1
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+export SGLANG_TIMEOUT_KEEP_ALIVE=900
+export SGLANG_OPT_USE_TOPK_V2=false
+
+# AgentX pins acceptance to the committed golden AL so submissions are compared
+# on system performance at a fixed acceptance target rather than on draft-head
+# quality (golden_al_distribution/README.md). 2.99 is the GLM-5.2 curve at
+# num_speculative_tokens=3, thinking_on (SPEED-Bench coding, speedbench-al.yml
+# run 28058352479); the curve is committed as
+# golden_al_distribution/glm5.2_mtp.yaml by the B300 sibling recipe (#2447).
+# One curve per model: it was collected on the FP8 checkpoint, and the MXFP4
+# checkpoint ships the same nextn head.
+#
+# SGLANG_SIMULATE_ACC_TOKEN_MODE is only read from SGLang v0.5.16 onward --
+# ACC_LEN / ACC_METHOD exist further back, so an older image would silently
+# honor two thirds of the contract. The pinned image is v0.5.16-rocm720.
+#
+# EVAL_ONLY leaves simulated acceptance off: it commits drafted tokens
+# regardless of the target logits, so generated text is wrong and the eval
+# would score ~0.
+if [ "${EVAL_ONLY:-false}" != "true" ]; then
+    export SGLANG_SIMULATE_ACC_LEN=2.99
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+
+CACHE_ARGS=()
+if agentic_kv_offload_enabled; then
+    if [ "$DP_ATTENTION" = "true" ]; then
+        HICACHE_RATIO="${HICACHE_RATIO:-0.5}"
+    else
+        HICACHE_RATIO="${HICACHE_RATIO:-1.5}"
+    fi
+    HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through}"
+    HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
+    HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-page_first_direct}"
+    case "$KV_OFFLOAD_BACKEND" in
+        hicache)
+            echo "HiCache (GPU+host DRAM only): ratio=$HICACHE_RATIO, write_policy=$HICACHE_WRITE_POLICY, io_backend=$HICACHE_IO_BACKEND, mem_layout=$HICACHE_MEM_LAYOUT"
+            CACHE_ARGS=(
+                --enable-hierarchical-cache
+                --hicache-ratio "$HICACHE_RATIO"
+                --hicache-write-policy "$HICACHE_WRITE_POLICY"
+                --hicache-io-backend "$HICACHE_IO_BACKEND"
+                --hicache-mem-layout "$HICACHE_MEM_LAYOUT"
+            )
+            ;;
+        mooncake)
+            L3_PER_RANK_GB="${L3_PER_RANK_GB:-40}"
+            python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
+            MOONCAKE_MASTER_PORT=$((PORT + 12000))
+            MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
+            MOONCAKE_CONFIG_PATH="$RESULT_DIR/mooncake_config.json"
+            cat > "$MOONCAKE_CONFIG_PATH" <<EOF
+{
+  "local_hostname": "127.0.0.1",
+  "metadata_server": "P2PHANDSHAKE",
+  "master_server_address": "127.0.0.1:$MOONCAKE_MASTER_PORT",
+  "global_segment_size": "${L3_PER_RANK_GB}gb",
+  "local_buffer_size": "4gb",
+  "protocol": "tcp",
+  "device_name": ""
+}
+EOF
+            export SGLANG_HICACHE_MOONCAKE_CONFIG_PATH="$MOONCAKE_CONFIG_PATH"
+            mooncake_master --port "$MOONCAKE_MASTER_PORT" \
+                --default_kv_lease_ttl=120s \
+                --eviction_high_watermark_ratio=0.80 \
+                --eviction_ratio=0.10 > "$MOONCAKE_MASTER_LOG" 2>&1 &
+            MOONCAKE_MASTER_PID=$!
+            sleep 2
+            kill -0 "$MOONCAKE_MASTER_PID"
+            echo "HiCache+Mooncake: ratio=$HICACHE_RATIO, l3_per_rank=${L3_PER_RANK_GB} GB, dram_budget=${TOTAL_CPU_DRAM_GB} GB"
+            CACHE_ARGS=(
+                --enable-hierarchical-cache
+                --hicache-ratio "$HICACHE_RATIO"
+                --hicache-size 0
+                --hicache-write-policy "$HICACHE_WRITE_POLICY"
+                --hicache-io-backend "$HICACHE_IO_BACKEND"
+                --hicache-mem-layout "$HICACHE_MEM_LAYOUT"
+                --hicache-storage-backend mooncake
+                --hicache-storage-prefetch-policy wait_complete
+            )
+            ;;
+        *)
+            echo "Error: unsupported KV_OFFLOAD_BACKEND '$KV_OFFLOAD_BACKEND' (expected: hicache or mooncake)" >&2
+            exit 1
+            ;;
+    esac
+fi
+
+# MTP: GLM-5.2 ships its own nextn head, so SGLang EAGLE runs off the
+# checkpoint with no external draft model. num-steps 3 / eagle-topk 1 /
+# num-draft-tokens 4 is 3 speculative tokens per verification step -- the same
+# shape as the B300 sibling and the GLM-5.2 GB300 dynamo-sglang agentic
+# recipes, and the draft length whose golden AL is pinned above.
+#
+# The draft MoE backend is left to SGLang: GLM-5.2's nextn layer is unquantized
+# bf16, so it cannot inherit the target model's MXFP4 MoE runner once expert
+# parallelism puts an all-to-all in the path. Upstream's
+# _deepseek_spec_moe_resolution fixes that up automatically, and its hook is
+# gated on is_hip() -- i.e. it fires here, unlike on the CUDA sibling where the
+# same pair has to be passed explicitly.
+SPEC_ARGS=(
+    --speculative-algorithm EAGLE
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+)
+
+USE_SGLANG_ROUTER=false
+SGLANG_BACKEND_PORT="$PORT"
+PARALLEL_ARGS=(--tp "$TP" --ep-size "$EP_SIZE")
+# The nextn layer's weights and its own KV pool come out of the same static
+# budget as the target model, and CUDA-graph capture is over 4-token
+# verification batches rather than 1-token decodes. Hold the whole grid at the
+# conservative 0.80 instead of taking the +0.05 the non-speculative serve shape
+# could afford at conc <= 16: an OOM mid-warmup costs the whole sweep point,
+# and HiCache's host tier (ratio relative to the device pool) absorbs the
+# slightly smaller HBM KV pool.
+MEM_FRACTION_STATIC=0.80
+if [ "$DP_ATTENTION" = "true" ]; then
+    USE_SGLANG_ROUTER=true
+    export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
+    SGLANG_BACKEND_PORT=$((PORT + 1))
+    SGLANG_ROUTER_METRICS_PORT=$((PORT + 10000))
+    SGLANG_ROUTER_CMD=(python3 -m sglang_router.launch_router)
+    PARALLEL_ARGS+=(--dp "$TP" --enable-dp-attention)
+    CHUNKED_PREFILL_SIZE=32768
+    export AGENTIC_WARMUP_GRACE_PERIOD=3600
+    export SGLANG_DP_USE_GATHERV=1
+    export SGLANG_DP_USE_REDUCE_SCATTER=1
+    export GPU_MAX_HW_QUEUES=5
+elif [ "$CONC" -le 16 ]; then
+    CHUNKED_PREFILL_SIZE=131072
+else
+    CHUNKED_PREFILL_SIZE=32768
+    export AGENTIC_WARMUP_GRACE_PERIOD=3600
+fi
+MAX_RUNNING_REQUESTS=$((1 * CONC))
+[ "$MAX_RUNNING_REQUESTS" -gt 256 ] && MAX_RUNNING_REQUESTS=256
+# --cuda-graph-max-bs counts requests, not verification tokens: SGLang's
+# spec-decode graph runner scales each captured batch by
+# --speculative-num-draft-tokens itself.
+CUDA_GRAPH_MAX_BS=$MAX_RUNNING_REQUESTS
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$SGLANG_BACKEND_PORT"
+    --trust-remote-code
+    "${PARALLEL_ARGS[@]}"
+    --kv-cache-dtype fp8_e4m3
+    --dsa-prefill-backend tilelang
+    --dsa-decode-backend tilelang
+    --tool-call-parser glm47
+    --reasoning-parser glm45
+    --chunked-prefill-size "$CHUNKED_PREFILL_SIZE"
+    --mem-fraction-static "$MEM_FRACTION_STATIC"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    "${SPEC_ARGS[@]}"
+    "${CACHE_ARGS[@]}"
+    --watchdog-timeout 1800
+    --enable-metrics
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+
+{
+    echo "=== SGLANG_SIMULATE_ACC_* env vars at launch (empty => real verification) ==="
+    env | grep -E '^SGLANG_SIMULATE_ACC_' | sort || true
+    echo "============================================================================"
+} | tee "$SERVER_LOG"
+
+echo "Starting SGLang server for MI355X..."
+"${SGLANG_CMD[@]}" >> "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+echo "Server PID: $SERVER_PID"
+
+wait_for_server_ready --port "$SGLANG_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [ "$USE_SGLANG_ROUTER" = "true" ]; then
+    echo "Starting SGLang router on port $PORT for $TP DP ranks..."
+    "${SGLANG_ROUTER_CMD[@]}" \
+        --worker-urls "http://localhost:$SGLANG_BACKEND_PORT" \
+        --policy consistent_hashing \
+        --request-id-headers x-correlation-id \
+        --dp-aware \
+        --host 0.0.0.0 \
+        --port "$PORT" \
+        --prometheus-host 127.0.0.1 \
+        --prometheus-port "$SGLANG_ROUTER_METRICS_PORT" \
+        --connect-timeout-secs 900 \
+        --request-timeout-secs 14400 \
+        --disable-health-check \
+        --disable-retries > "$ROUTER_LOG" 2>&1 &
+    ROUTER_PID=$!
+    echo "Router PID: $ROUTER_PID"
+    wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
+fi
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    export SWEBENCH_AGENT_STEP_LIMIT=150
+    export SWEBENCH_AGENT_WORKERS="${SWEBENCH_AGENT_WORKERS:-32}"
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --server-metrics http://localhost:$SGLANG_BACKEND_PORT/metrics"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
@@ -2360,3 +2360,36 @@ dsv4-fp8-mi325x-vllm-mtp:
       # is 55.8%. conc128 passed cleanly on the memory-tightest SKU (MI300X);
       # cap 8k1k MTP at 128 (normal holds 256, 1k1k holds 512).
       - { tp: 8, conc-start: 4, conc-end: 128, spec-decoding: mtp }
+
+# GLM-5.2 MI355X MXFP4 AgentX with EAGLE/MTP speculative decoding. Spec-decode
+# only: agentic recipes are run and published with speculative decoding enabled
+# rather than as an STP/MTP A/B, and GLM-5.2 agentic non-MTP is deprecated after
+# 2026-08-03 (MODELS.md), so no non-MTP arm is onboarded here.
+#
+# SGLang EAGLE off GLM-5.2's built-in nextn head (num-steps 3, eagle-topk 1,
+# 4 draft tokens = 3 speculative tokens), with acceptance pinned to the golden
+# AL 2.99 (golden_al_distribution/glm5.2_mtp.yaml, thinking_on, K=3) through
+# SGLANG_SIMULATE_ACC_*, matching the B300 sibling glm5.2-fp4-b300-sglang-agentic-mtp.
+# v0.5.16-rocm720 is the first ROCm image that reads
+# SGLANG_SIMULATE_ACC_TOKEN_MODE.
+#
+# TP8 + EP8 with host-DRAM HiCache on the AgentX MTP concurrency grid
+# [1, 4, 8, 12, 16]: steps of at least 2 because single-step sampling cannot
+# separate configurations by more than run-to-run noise on the agentic corpus,
+# and a hard stop at conc 16. HiCache on every point -- without it the working
+# set outgrows HBM past conc 8 and the radix hit rate collapses. Mooncake is
+# deliberately not offered as the offload backend on this SKU (the HIP/TCP
+# transfer engine is not stable on gfx942/gfx950 under sustained agentic load).
+glm5.2-fp4-mi355x-sglang-agentic-mtp:
+  image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728
+  model: amd/GLM-5.2-MXFP4
+  model-prefix: glm5.2
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.85
+      search-space:
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 4, 8, 12, 16] }
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
@@ -5355,3 +5355,14 @@
     - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path"
     - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213
+
+- config-keys:
+    - glm5.2-fp4-mi355x-sglang-agentic-mtp
+  description:
+    - "Add the GLM-5.2 MXFP4 MI355X AgentX (agentic-coding) recipe with EAGLE/MTP speculative decoding, routed to benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh. Spec-decode only: agentic recipes are run and published with speculative decoding enabled rather than as an STP/MTP A/B, and GLM-5.2 agentic non-MTP is deprecated after 2026-08-03 (MODELS.md), so no non-MTP arm is onboarded for this SKU."
+    - "Speculative config is GLM-5.2's built-in nextn head via SGLang EAGLE (no external draft model): --speculative-algorithm EAGLE, --speculative-num-steps 3, --speculative-eagle-topk 1, --speculative-num-draft-tokens 4, i.e. 3 speculative tokens per verification step. Same shape as the B300 sibling glm5.2-fp4-b300-sglang-agentic-mtp. The draft MoE backend is left to SGLang: GLM-5.2's nextn layer is unquantized bf16 and cannot inherit the target model's MXFP4 MoE runner once EP8 puts an all-to-all in the path, and upstream's _deepseek_spec_moe_resolution hook is gated on is_hip(), so it fires on ROCm without the explicit backend pair the CUDA sibling needs."
+    - "Throughput runs pin SGLang simulated acceptance to the golden AL: SGLANG_SIMULATE_ACC_LEN=2.99 (GLM-5.2 MTP curve at K=3, thinking_on, SPEED-Bench coding, speedbench-al.yml run 28058352479), SGLANG_SIMULATE_ACC_METHOD=match-expected, SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token -- the triple golden_al_distribution/README.md prescribes for SGLang. The curve itself is committed as golden_al_distribution/glm5.2_mtp.yaml by the B300 sibling recipe (#2447); it was collected on the FP8 checkpoint and the MXFP4 checkpoint ships the same nextn head. EVAL_ONLY runs leave simulated acceptance off and keep real verification, since simulated acceptance commits drafted tokens regardless of the target logits and would score the eval at ~0."
+    - "Image lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728 (verified on Docker Hub, pushed 2026-07-28). v0.5.16 is the first release that reads SGLANG_SIMULATE_ACC_TOKEN_MODE -- ACC_LEN / ACC_METHOD exist further back, so an older ROCm image would silently ignore the token-mode half of the AgentX contract."
+    - "Serve shape is the ROCm GLM-5.2 low-latency arm: TP8 + EP8, --kv-cache-dtype fp8_e4m3, tilelang DSA prefill/decode backends, glm47 tool-call parser, glm45 reasoning parser, --chunked-prefill-size 131072 at conc <= 16, HiCache host-DRAM offload at hicache-ratio 1.5 / write_through / direct / page_first_direct. --mem-fraction-static is held at 0.80 across the whole grid rather than taking the +0.05 a non-speculative serve shape could afford at conc <= 16: the nextn layer's weights and its own KV pool come out of the same static budget and CUDA-graph capture is over 4-token verification batches, so an OOM mid-warmup would cost the sweep point, and HiCache's host tier (ratio relative to the device pool) absorbs the slightly smaller HBM KV pool."
+    - "Search space is one arm on the AgentX MTP concurrency grid: TP8 + EP8 + HiCache at conc [1, 4, 8, 12, 16]. Steps of at least 2 because single-step sampling cannot separate configurations by more than run-to-run noise on the agentic corpus, and a hard stop at conc 16. Mooncake is kept out of the master config as an offload backend on this SKU (the HIP/TCP transfer engine is not stable on gfx950 under sustained agentic load); the branch stays in the benchmark script for on-node experiments."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2459