From d5cd452a5aecd3f4b0e95355e464cf844d1cf0bc Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Tue, 4 Aug 2026 15:19:26 +0200
Subject: [PATCH 01/15] [AMD] [WIP] [AGENTX] Adding GLM5.2 MI355X Support -
 Agentx

Signed-off-by: Giovanni Guasti <giovanni.guasti@amd.com>
---
 .../agentic/glm5.2_fp4_mi355x_sglang_mtp.sh   | 252 ++++++++++++++++++
 configs/amd-master.yaml                       |  18 ++
 perf-changelog.yaml                           |   7 +
 3 files changed, 277 insertions(+)
 create mode 100644 benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh

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
new file mode 100644
index 0000000000..6a4e642f94
--- /dev/null
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -0,0 +1,252 @@
+#!/usr/bin/env bash
+set -eo pipefail
+set -x
+
+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
+
+
+# A server killed on this node minutes earlier (previous job, crashed run)
+# can still be draining its ~1.4 TB of HBM: KFD reclaim takes minutes, and
+# booting into a half-drained node fails RCCL init with HIP 'unhandled cuda
+# error' / 'invalid argument' (observed as the mooncake-c64 CI failure).
+# Wait for the GPUs to come back before launching.
+# Per-GPU threshold: idle nodes hold a small driver/firmware VRAM baseline
+# (observed up to ~4%/GPU, node-dependent), while a draining or occupied
+# GPU sits at 50-90%. Require every GPU <= 10%.
+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
+# Agentic warmup dispatches hundreds of large prompts at once; allow up to
+# 15 minutes of TCP progress before AIPerf declares a connection dead.
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+# AIPerf pins one pooled keep-alive connection per session (client-side
+# keep-alive 300s) while uvicorn's default SGLANG_TIMEOUT_KEEP_ALIVE is 5s;
+# inter-turn idle gaps can reuse a socket exactly as the server closes it.
+# Outlast the client pool so the race cannot occur.
+export SGLANG_TIMEOUT_KEEP_ALIVE=900
+# The DSA indexer's top-k v2 kernel (default since v0.5.14) is JIT-compiled
+# from CUDA-only source (cooperative_groups.h) and cannot build for gfx950;
+# v1 dispatches to the precompiled HIP op in sgl-kernel (upstream MI355X CI
+# runs DSA models the same way).
+export SGLANG_OPT_USE_TOPK_V2=false
+
+# HiCache L2 + Mooncake L3 on every point (sizing rationale in the header).
+# Per-arm L2 ratio, both measured on-node. TP arm (182.7 GB/rank device
+# pool): the working set oversubscribes the device pool ~3x at conc 32, so
+# the host tier is what carries the radix hits - ratio 1.5 (~2.9 TB pinned
+# incl. sidecars) validates through the conc-24 long-context storm. The
+# DP-attention arm (159.4 GB/rank) only runs at conc >= 32, where each DP
+# rank's ~8 sessions nearly fit in its own device pool (~1.5-1.6M of 1.7M
+# tokens at conc 64) and the host tier just absorbs overflow - ratio 1.5
+# boots but the host OOM killer takes the server mid-storm at conc 48, so
+# it runs ratio 0.5 (~1.2 TB pinned, ~1.8 TB of load headroom) at
+# negligible hit-rate cost.
+CACHE_ARGS=()
+if require_agentic_kv_offload_backend mooncake; 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}"
+    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 "$HICACHE_MEM_LAYOUT"
+        --hicache-storage-backend mooncake
+        --hicache-storage-prefetch-policy wait_complete
+    )
+fi
+
+# Arm selection. TP arm keeps the FP8 sibling's cookbook batch-shaping
+# bands.
+#
+# NOTE: the DP-attention path below is currently DORMANT (no dp-attn arms
+# in amd-master.yaml): DSA + dp-attention hangs a collective under
+# long-context prefill on ROCm v0.5.14 (watchdog kills the scheduler with
+# zero completions; reproduced with and without HiCache, with and without
+# the DSv4 DP collective envs; short prompts are fine). Re-enable the
+# config arm once upstream fixes the DSA DP prefill path.
+#
+# When active, the DP-attention (DEP) arm fronts the DP ranks with sglang-router
+# using consistent hashing on the AIPerf correlation id so multi-turn
+# sessions stay on the DP rank holding their radix/hicache prefix, and
+# widens chunked-prefill (whole-engine, /dp ranks) like the B300 sibling.
+USE_SGLANG_ROUTER=false
+SGLANG_BACKEND_PORT="$PORT"
+PARALLEL_ARGS=(--tp "$TP" --ep-size "$EP_SIZE")
+MEM_FRACTION_STATIC=0.85
+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
+    # Swap the DP gather collectives to gatherv/reduce-scatter on ROCm
+    # (dsv4_fp4_mi355x_sglang.sh precedent - the only green DP-attention
+    # config on this cluster/image): with the defaults the DSA DP path
+    # hangs a collective under long-context prefill load until the
+    # watchdog kills the scheduler (0/96 storm completions, twice).
+    export SGLANG_DP_USE_GATHERV=1
+    export SGLANG_DP_USE_REDUCE_SCATTER=1
+    export GPU_MAX_HW_QUEUES=5
+elif [ "$CONC" -le 16 ]; then
+    # A full 131072-token prefill chunk needs ~7 GiB/rank of activation
+    # headroom on top of the static pool; pair it with mem-fraction 0.80
+    # like the FP8 sibling's low-conc band (0.85 OOMs the device mid-replay:
+    # "Tried to allocate 6.86 GiB ... 5.15 GiB is free", run 29751563205).
+    CHUNKED_PREFILL_SIZE=131072
+    MEM_FRACTION_STATIC=0.80
+else
+    CHUNKED_PREFILL_SIZE=32768
+    export AGENTIC_WARMUP_GRACE_PERIOD=3600
+fi
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+[ "$MAX_RUNNING_REQUESTS" -gt 256 ] && MAX_RUNNING_REQUESTS=256
+CUDA_GRAPH_MAX_BS=$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[@]}"
+    --dsa-prefill-backend tilelang
+    --dsa-decode-backend tilelang
+    # GLM-5.2 emits the GLM-4.7-style tool-call format; glm47 is required for
+    # structured message.tool_calls (SWE-bench agentic evals die without it).
+    # The glm45 reasoning parser keeps hybrid thinking in reasoning_content.
+    --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"
+    --speculative-algorithm EAGLE \
+    --speculative-num-steps 5 \
+    --speculative-eagle-topk 1 \
+    --speculative-num-draft-tokens 6 \    
+    "${CACHE_ARGS[@]}"
+    --watchdog-timeout 1800
+    --enable-metrics
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+
+echo "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 [ "${EVAL_ONLY}" = "true" ]; then
+    # GLM-5.2's chat template defaults to reasoning_effort=Max when the
+    # client passes no chat_template_kwargs (mini-swe-agent doesn't), and the
+    # heavy thinking burns the default 75-step budget before submission.
+    # Double the step budget for this recipe; others keep the shared default.
+    export SWEBENCH_AGENT_STEP_LIMIT=150
+    # Pin eval agent parallelism to the proven-green level: workers default
+    # to CONC, and at 64 concurrent Modal sandboxes the cluster's egress
+    # collapses (18k "Cannot connect to *.modal.host" errors crippled the
+    # trajectories in run 29764760177) while 32 ran clean. The serving
+    # config is unchanged - only the agent's session fan-out is capped.
+    export SWEBENCH_AGENT_WORKERS="${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
\ No newline at end of file
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0d46afe2e1..45739a3d4e 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2361,3 +2361,21 @@ 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 FP4 agentic-coding benchmark on MI355X via SGLang with MTP speculative
+# decoding. TP=4 EP=4 with KV offloading to DRAM (hicache backend) to support
+# long agentic context windows. Concurrency sweep [1, 2, 4, 8, 10].
+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: 0.8
+      search-space:
+      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4, 8, 10], spec-decoding: mtp }
+
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b5aa7a3857..823195b4c7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5425,3 +5425,10 @@
   description:
     - "Extend the search space to include the TP2EP1 configuration for fixed seq len 8k1k"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2446
+
+- config-keys:
+    - glm5.2-fp4-mi355x-sglang-agentic-mtp
+  description:
+    - "Add GLM-5.2 FP4 SGLANG Single Node Agentic Support"
+    - "Image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728"
+  pr-link: TO-BE-ADDED

From b7e7b895f2b0a97f7c5dc9aeac691b5716cc36a9 Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Tue, 4 Aug 2026 16:06:18 +0200
Subject: [PATCH 02/15] [AMD] [WIP] [AGENTX] Adding GLM5.2 MI355X Support -
 Agentx

Signed-off-by: Giovanni Guasti <giovanni.guasti@amd.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 823195b4c7..d7139ec2f5 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5431,4 +5431,4 @@
   description:
     - "Add GLM-5.2 FP4 SGLANG Single Node Agentic Support"
     - "Image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728"
-  pr-link: TO-BE-ADDED
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2463

From ab95e49b730fcdbb37b4ab3e4eaaf9e5476d14b9 Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Tue, 4 Aug 2026 16:27:51 +0200
Subject: [PATCH 03/15] [AMD] [WIP] [AGENTX] Adding GLM5.2 MI355X Support -
 Agentx

Signed-off-by: Giovanni Guasti <giovanni.guasti@amd.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d7139ec2f5..823195b4c7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5431,4 +5431,4 @@
   description:
     - "Add GLM-5.2 FP4 SGLANG Single Node Agentic Support"
     - "Image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2463
+  pr-link: TO-BE-ADDED

From 8bf03a2844b8f375e7cd7d3414724c8098b569b6 Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Tue, 4 Aug 2026 16:31:37 +0200
Subject: [PATCH 04/15] chore: update pr-link for
 glm5.2-fp4-mi355x-sglang-agentic-mtp
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：更新 glm5.2-fp4-mi355x-sglang-agentic-mtp 配置的 pr-link 至 PR #2488
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 823195b4c7..2b197680c2 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5431,4 +5431,4 @@
   description:
     - "Add GLM-5.2 FP4 SGLANG Single Node Agentic Support"
     - "Image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728"
-  pr-link: TO-BE-ADDED
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2488

From a11fcf2fa3c25bb4815415df3c727401f0a4d33a Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Wed, 5 Aug 2026 12:34:00 +0530
Subject: [PATCH 05/15] [AMD] [WIP] [AGENTX] GLM 5.2 - Update
 glm5.2_fp4_mi355x_sglang_mtp.sh

---
 .../single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh  | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

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
index 6a4e642f94..993bfc7b3c 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -4,6 +4,8 @@ set -x
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
+export EVAL_FRAMEWORK="lm-eval"
+
 check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
 
 if [[ -n "$SLURM_JOB_ID" ]]; then
@@ -232,6 +234,11 @@ if [ "$USE_SGLANG_ROUTER" = "true" ]; then
     wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
 fi
 
+if [ "${#METRICS_ARGS[@]}" -gt 0 ]; then
+    capture_cache_metrics
+    trap capture_cache_metrics EXIT
+fi
+
 if [ "${EVAL_ONLY}" = "true" ]; then
     # GLM-5.2's chat template defaults to reasoning_effort=Max when the
     # client passes no chat_template_kwargs (mini-swe-agent doesn't), and the
@@ -249,4 +256,4 @@ 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
\ No newline at end of file
+fi

From 6fa806ba4cea3e09731596597602384a19d98f8c Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Wed, 5 Aug 2026 12:40:38 +0530
Subject: [PATCH 06/15] [AMD] [WIP] [AGENTX] GLM 5.2 - Update
 perf-changelog.yaml

---
 perf-changelog.yaml | 1 -
 1 file changed, 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2cfe348bdd..dbc8c9a340 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5456,4 +5456,3 @@
     - "Add GLM-5.2 FP4 SGLANG Single Node Agentic Support"
     - "Image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2488
-  
\ No newline at end of file

From 194f975834649b71cd4e61dc9b2e9d04bf67a1db Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Wed, 5 Aug 2026 13:15:24 +0530
Subject: [PATCH 07/15] [AMD] [WIP] [AGENTX] GLM 5.2 - Update
 glm5.2_fp4_mi355x_sglang_mtp.sh

---
 .../agentic/glm5.2_fp4_mi355x_sglang_mtp.sh   | 46 -------------------
 1 file changed, 46 deletions(-)

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
index 993bfc7b3c..e76aab99e7 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -29,7 +29,6 @@ fi
 rocm-smi || true
 amd-smi || true
 
-
 # A server killed on this node minutes earlier (previous job, crashed run)
 # can still be draining its ~1.4 TB of HBM: KFD reclaim takes minutes, and
 # booting into a half-drained node fails RCCL init with HIP 'unhandled cuda
@@ -80,51 +79,6 @@ export SGLANG_OPT_USE_TOPK_V2=false
 # it runs ratio 0.5 (~1.2 TB pinned, ~1.8 TB of load headroom) at
 # negligible hit-rate cost.
 CACHE_ARGS=()
-if require_agentic_kv_offload_backend mooncake; 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}"
-    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
-    )
-fi
 
 # Arm selection. TP arm keeps the FP8 sibling's cookbook batch-shaping
 # bands.

From 0175c275f5d3a208cb44a862582783418197e451 Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Wed, 5 Aug 2026 10:14:11 +0200
Subject: [PATCH 08/15] fix(glm5.2-fp4-mi355x-sglang-mtp): remove trailing
 backslashes from SGLANG_CMD array
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Backslashes inside bash arrays are treated as literal characters, causing
SGLang to receive arguments like 'EAGLE \' instead of 'EAGLE', which
results in 'unrecognized arguments' errors.

中文：修复 SGLANG_CMD 数组中多余的行继续符（反斜杠），避免 SGLang 将其作为字面参数解析导致启动失败
---
 .../single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh   | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

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
index e76aab99e7..5198c2797c 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -149,10 +149,10 @@ SGLANG_CMD=(
     --mem-fraction-static "$MEM_FRACTION_STATIC"
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
-    --speculative-algorithm EAGLE \
-    --speculative-num-steps 5 \
-    --speculative-eagle-topk 1 \
-    --speculative-num-draft-tokens 6 \    
+    --speculative-algorithm EAGLE
+    --speculative-num-steps 5
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 6
     "${CACHE_ARGS[@]}"
     --watchdog-timeout 1800
     --enable-metrics

From 6d65f91579cb5451e4efab0f98e590718203a9ce Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Wed, 5 Aug 2026 11:13:13 +0200
Subject: [PATCH 09/15] fix(glm5.2-fp4-mi355x-sglang-mtp): increase
 watchdog-timeout to 3600s
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

MTP draft-token forward passes under long-context agentic load block the
SGLang scheduler long enough to trip the 1800s watchdog, killing the
server mid-warmup. Matches the value used by multi-node AMD disagg recipes.

中文：将 watchdog-timeout 从 1800s 增加至 3600s，避免 MTP 投机解码在长上下文推理负载下触发 SGLang 调度器看门狗超时导致服务崩溃
---
 benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

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
index 5198c2797c..a4d2b3515f 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -154,7 +154,7 @@ SGLANG_CMD=(
     --speculative-eagle-topk 1
     --speculative-num-draft-tokens 6
     "${CACHE_ARGS[@]}"
-    --watchdog-timeout 1800
+    --watchdog-timeout 3600
     --enable-metrics
 )
 

From 95dd7450dda16b0961da97f061a188a91390d694 Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Wed, 5 Aug 2026 15:11:28 +0530
Subject: [PATCH 10/15] [AMD] [WIP] [AGENTX] GLM 5.2 - Update
 glm5.2_fp4_mi355x_sglang_mtp.sh

---
 .../agentic/glm5.2_fp4_mi355x_sglang_mtp.sh   | 151 ++++++++++++------
 1 file changed, 102 insertions(+), 49 deletions(-)

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
index a4d2b3515f..8106332b1e 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -1,23 +1,23 @@
 #!/usr/bin/env bash
 set -eo pipefail
 set -x
-
+ 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
-export EVAL_FRAMEWORK="lm-eval"
-
+ export EVAL_FRAMEWORK="lm-eval"
+ 
 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"
@@ -28,7 +28,7 @@ else
 fi
 rocm-smi || true
 amd-smi || true
-
+  
 # A server killed on this node minutes earlier (previous job, crashed run)
 # can still be draining its ~1.4 TB of HBM: KFD reclaim takes minutes, and
 # booting into a half-drained node fails RCCL init with HIP 'unhandled cuda
@@ -44,14 +44,14 @@ for i in $(seq 1 90); do
     echo "waiting for prior-job GPU memory reclaim: vram%max=$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
 # Agentic warmup dispatches hundreds of large prompts at once; allow up to
 # 15 minutes of TCP progress before AIPerf declares a connection dead.
@@ -66,20 +66,88 @@ export SGLANG_TIMEOUT_KEEP_ALIVE=900
 # v1 dispatches to the precompiled HIP op in sgl-kernel (upstream MI355X CI
 # runs DSA models the same way).
 export SGLANG_OPT_USE_TOPK_V2=false
-
-# HiCache L2 + Mooncake L3 on every point (sizing rationale in the header).
-# Per-arm L2 ratio, both measured on-node. TP arm (182.7 GB/rank device
-# pool): the working set oversubscribes the device pool ~3x at conc 32, so
-# the host tier is what carries the radix hits - ratio 1.5 (~2.9 TB pinned
-# incl. sidecars) validates through the conc-24 long-context storm. The
-# DP-attention arm (159.4 GB/rank) only runs at conc >= 32, where each DP
-# rank's ~8 sessions nearly fit in its own device pool (~1.5-1.6M of 1.7M
-# tokens at conc 64) and the host tier just absorbs overflow - ratio 1.5
-# boots but the host OOM killer takes the server mid-storm at conc 48, so
-# it runs ratio 0.5 (~1.2 TB pinned, ~1.8 TB of load headroom) at
-# negligible hit-rate cost.
+ 
+# HiCache L2 (host DRAM), optionally extended with Mooncake L3.
+# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=hicache or mooncake.
+#
+# Per-arm L2 ratio (sizing rationale below) applies to both backends unless
+# overridden via HICACHE_RATIO. TP arm (182.7 GB/rank device pool): the
+# working set oversubscribes the device pool ~3x at conc 32, so the host
+# tier is what carries the radix hits - ratio 1.5 (~2.9 TB pinned incl.
+# sidecars) validates through the conc-24 long-context storm for the
+# mooncake arm. The DP-attention arm (159.4 GB/rank) only runs at conc >=
+# 32, where each DP rank's ~8 sessions nearly fit in its own device pool
+# (~1.5-1.6M of 1.7M tokens at conc 64) and the host tier just absorbs
+# overflow - ratio 1.5 boots but the host OOM killer takes the server
+# mid-storm at conc 48, so it runs ratio 0.5 (~1.2 TB pinned, ~1.8 TB of
+# load headroom) at negligible hit-rate cost. The hicache-only arm has no
+# L3 to fall back on, so these ratios are unvalidated there - override with
+# HICACHE_RATIO if the host OOMs or hit-rate is poor.
 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 '$KV_OFFLOAD_BACKEND' (expected: hicache or mooncake)" >&2
+            exit 1
+            ;;
+    esac
+fi
+ 
 # Arm selection. TP arm keeps the FP8 sibling's cookbook batch-shaping
 # bands.
 #
@@ -121,15 +189,15 @@ elif [ "$CONC" -le 16 ]; then
     # like the FP8 sibling's low-conc band (0.85 OOMs the device mid-replay:
     # "Tried to allocate 6.86 GiB ... 5.15 GiB is free", run 29751563205).
     CHUNKED_PREFILL_SIZE=131072
-    MEM_FRACTION_STATIC=0.80
+    MEM_FRACTION_STATIC=0.85
 else
     CHUNKED_PREFILL_SIZE=32768
     export AGENTIC_WARMUP_GRACE_PERIOD=3600
 fi
-MAX_RUNNING_REQUESTS=$((2 * CONC))
+MAX_RUNNING_REQUESTS=$((1 * CONC))
 [ "$MAX_RUNNING_REQUESTS" -gt 256 ] && MAX_RUNNING_REQUESTS=256
 CUDA_GRAPH_MAX_BS=$MAX_RUNNING_REQUESTS
-
+ 
 SGLANG_CMD=(
     python3 -m sglang.launch_server
     --model-path "$MODEL_PATH"
@@ -138,6 +206,7 @@ SGLANG_CMD=(
     --port "$SGLANG_BACKEND_PORT"
     --trust-remote-code
     "${PARALLEL_ARGS[@]}"
+    --kv-cache-dtype fp8_e4m3
     --dsa-prefill-backend tilelang
     --dsa-decode-backend tilelang
     # GLM-5.2 emits the GLM-4.7-style tool-call format; glm47 is required for
@@ -154,20 +223,20 @@ SGLANG_CMD=(
     --speculative-eagle-topk 1
     --speculative-num-draft-tokens 6
     "${CACHE_ARGS[@]}"
-    --watchdog-timeout 3600
+    --watchdog-timeout 1800
     --enable-metrics
 )
-
+ 
 printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
 printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
-
+ 
 echo "Starting SGLang server for MI355X..."
 "${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[@]}" \
@@ -187,24 +256,8 @@ if [ "$USE_SGLANG_ROUTER" = "true" ]; then
     echo "Router PID: $ROUTER_PID"
     wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
 fi
-
-if [ "${#METRICS_ARGS[@]}" -gt 0 ]; then
-    capture_cache_metrics
-    trap capture_cache_metrics EXIT
-fi
-
+ 
 if [ "${EVAL_ONLY}" = "true" ]; then
-    # GLM-5.2's chat template defaults to reasoning_effort=Max when the
-    # client passes no chat_template_kwargs (mini-swe-agent doesn't), and the
-    # heavy thinking burns the default 75-step budget before submission.
-    # Double the step budget for this recipe; others keep the shared default.
-    export SWEBENCH_AGENT_STEP_LIMIT=150
-    # Pin eval agent parallelism to the proven-green level: workers default
-    # to CONC, and at 64 concurrent Modal sandboxes the cluster's egress
-    # collapses (18k "Cannot connect to *.modal.host" errors crippled the
-    # trajectories in run 29764760177) while 32 ran clean. The serving
-    # config is unchanged - only the agent's session fan-out is capped.
-    export SWEBENCH_AGENT_WORKERS="${SWEBENCH_AGENT_WORKERS:-32}"
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"

From 6eca9be001b10b304b86d83963e862e24e54a98c Mon Sep 17 00:00:00 2001
From: Giovanni Guasti <giovanni.guasti@amd.com>
Date: Wed, 5 Aug 2026 12:06:11 +0200
Subject: [PATCH 11/15] =?UTF-8?q?fix(glm5.2-fp4-mi355x-sglang-agentic-mtp)?=
 =?UTF-8?q?:=20set=20mem-fraction-static=200.80=20for=20conc<=3D16=20/=20?=
 =?UTF-8?q?=E4=BF=AE=E6=AD=A3=E4=BD=8E=E5=B9=B6=E5=8F=91=E5=9C=BA=E6=99=AF?=
 =?UTF-8?q?=20mem-fraction-static=20=E5=8F=82=E6=95=B0=E4=B8=BA=200.80?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

0.85 OOMs the device mid-replay on 131072-token prefill chunks
(observed: "Tried to allocate 6.86 GiB ... 5.15 GiB is free",
run 29751563205). The comment already documented 0.80 as the
correct value; the assignment was wrong.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

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
index 8106332b1e..014719415c 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -189,7 +189,7 @@ elif [ "$CONC" -le 16 ]; then
     # like the FP8 sibling's low-conc band (0.85 OOMs the device mid-replay:
     # "Tried to allocate 6.86 GiB ... 5.15 GiB is free", run 29751563205).
     CHUNKED_PREFILL_SIZE=131072
-    MEM_FRACTION_STATIC=0.85
+    MEM_FRACTION_STATIC=0.80
 else
     CHUNKED_PREFILL_SIZE=32768
     export AGENTIC_WARMUP_GRACE_PERIOD=3600

From 02c59f7f38f2bc55c1a049a7ce17c3f07393a27d Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 11 Aug 2026 11:03:54 +0900
Subject: [PATCH 12/15] [AMD][AgentX] glm5.2 fp4 mi355x sglang mtp: set
 SGLANG_SIMULATE_ACC_LEN=2.99 for non-eval runs

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh     | 6 ++++++
 1 file changed, 6 insertions(+)

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
index 014719415c..064a2d5724 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -197,6 +197,12 @@ fi
 MAX_RUNNING_REQUESTS=$((1 * CONC))
 [ "$MAX_RUNNING_REQUESTS" -gt 256 ] && MAX_RUNNING_REQUESTS=256
 CUDA_GRAPH_MAX_BS=$MAX_RUNNING_REQUESTS
+
+if [ "${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
  
 SGLANG_CMD=(
     python3 -m sglang.launch_server

From fff9004dca322afeff360c523d424aef6b016ee7 Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Tue, 11 Aug 2026 14:11:14 +0530
Subject: [PATCH 13/15] [AMD] [AGENTX] GLM 5.2 - Update
 glm5.2_fp4_mi355x_sglang_mtp.sh

---
 .../single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh       | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

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
index 064a2d5724..d7ed7a359d 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -225,9 +225,9 @@ SGLANG_CMD=(
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
     --speculative-algorithm EAGLE
-    --speculative-num-steps 5
+    --speculative-num-steps 3
     --speculative-eagle-topk 1
-    --speculative-num-draft-tokens 6
+    --speculative-num-draft-tokens 4
     "${CACHE_ARGS[@]}"
     --watchdog-timeout 1800
     --enable-metrics

From 532d32838b7a9472fe8aa32598cf1537de361892 Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Tue, 11 Aug 2026 14:13:57 +0530
Subject: [PATCH 14/15] [AMD] [AGENTX] GLM 5.2 - Update amd-master.yaml

---
 configs/amd-master.yaml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 3b1822dd62..67fdf2a041 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1631,5 +1631,6 @@ glm5.2-fp4-mi355x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.8
       search-space:
-      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4, 8, 10], spec-decoding: mtp }
+      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4, 8, 10, 12], spec-decoding: mtp }
+      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4], spec-decoding: mtp }
 

From d204ac0de29d506d5de9ad534495b6fe0ae21370 Mon Sep 17 00:00:00 2001
From: ajith-sirra-amd <122240613+ajith-sirra-amd@users.noreply.github.com>
Date: Tue, 11 Aug 2026 14:36:59 +0530
Subject: [PATCH 15/15] [AMD] [AGENTX] GLM 5.2 Update Search Space

---
 configs/amd-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 67fdf2a041..9bc224b161 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1631,6 +1631,6 @@ glm5.2-fp4-mi355x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.8
       search-space:
-      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4, 8, 10, 12], spec-decoding: mtp }
+      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4, 8, 10, 12, 16], spec-decoding: mtp }
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 2, 4], spec-decoding: mtp }