From d18fd7d172373b5baca4c6efd6b814d72d7b7c18 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 14 Jul 2026 18:01:32 +0900
Subject: [PATCH 01/29] [AMD][MI355X] DSv4 SGLang: add agentic MTP+HiCache
 recipe, exclude g12

- Add dsv4_fp4_mi355x_sglang_mtp.sh agentic launcher with EAGLE spec decoding
- Add dsv4-fp4-mi355x-sglang-agentic-hicache-mtp config in amd-master.yaml
- Sweep TP8 +/- DPA +/- HiCache with MTP (conc 1-96)
- Exclude mia1-p01-g12 from agentic salloc
- Update perf-changelog

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../agentic/dsv4_fp4_mi355x_sglang_mtp.sh     | 209 ++++++++++++++++++
 configs/amd-master.yaml                       |  18 +-
 perf-changelog.yaml                           |   9 +
 runners/launch_mi355x-amds.sh                 |   2 +-
 4 files changed, 235 insertions(+), 3 deletions(-)
 create mode 100755 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
new file mode 100755
index 0000000000..de891dd71c
--- /dev/null
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -0,0 +1,209 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# Agentic trace replay benchmark for DeepSeek-V4-Pro FP4 on MI355X using SGLang.
+#
+# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=hicache.
+#
+# Required env vars:
+#   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
+#
+# KV_OFFLOADING=dram requires one of these. 
+#   KV_OFFLOAD_BACKEND=hicache.
+
+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 ${SLURMD_NODENAME:-unknown}"
+fi
+
+# ROCR/HIP visibility under slurm cgroups.
+if [ -n "${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="$MODEL"
+fi
+rocm-smi || true
+amd-smi || true
+
+# ---- Resolve traces and install deps ----------------------------------------
+resolve_trace_source
+install_agentic_deps
+
+# ---- Server config ----------------------------------------------------------
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
+export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
+
+CACHE_ARGS=()
+if agentic_kv_offload_enabled; then
+    # HiCache config — https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+    case "${KV_OFFLOAD_BACKEND:-}" in
+        hicache)
+            HICACHE_RATIO=4
+            HICACHE_WRITE_POLICY="write_through"
+            HICACHE_IO_BACKEND="direct"
+            HICACHE_MEM_LAYOUT="page_first_direct"
+            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"
+            )
+            echo "HiCache DSv4 CPU tier: ratio=$HICACHE_RATIO, write_policy=$HICACHE_WRITE_POLICY, io_backend=$HICACHE_IO_BACKEND, mem_layout=$HICACHE_MEM_LAYOUT"
+            ;;
+        *)
+            echo "Error: unsupported KV_OFFLOAD_BACKEND '${KV_OFFLOAD_BACKEND:-}' (expected: hicache)" >&2
+            exit 1
+            ;;
+    esac
+fi
+
+# ---- LLM server config ----------------------------------------------------------
+USE_SGLANG_ROUTER=false
+SGLANG_BACKEND_PORT="$PORT"
+ROUTER_LOG="$RESULT_DIR/router.log"
+MEM_FRACTION_STATIC=0.95
+CHUNKED_PREFILL_SIZE=16384
+if [ "$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)
+
+    export SGLANG_SHARED_EXPERT_TP1=1
+    export SGLANG_DP_SHARED_EXPERT_LOCAL=1
+    export SGLANG_DP_USE_GATHERV=1
+    export SGLANG_DP_USE_REDUCE_SCATTER=1
+    export GPU_MAX_HW_QUEUES=5
+
+    CHUNKED_PREFILL_SIZE=$((8192 * TP))
+    PARALLEL_ARGS+=(
+        --dp "$TP"
+        --enable-dp-attention
+        --enable-prefill-delayer
+        --enable-two-batch-overlap
+    )
+fi
+
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    PARALLEL_ARGS+=(--ep-size "$EP_SIZE")
+fi
+
+# AgentX concurrency counts live session trees, not individual requests.
+# Allow subagent fan-out to exceed CONC without clipping request bursts.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+CUDA_GRAPH_MAX_BS=$CONC
+[ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
+
+export SGLANG_DEFAULT_THINKING=1
+export SGLANG_DSV4_REASONING_EFFORT=max
+export SGLANG_USE_ROCM700A=0
+export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
+export AITER_BF16_FP8_MOE_BOUND=0
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP")
+METRICS_ARGS=(--enable-metrics)
+SPEC_ARGS=(
+    --speculative-algorithm EAGLE
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+)
+
+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[@]}"
+    --attention-backend compressed
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --mem-fraction-static "$MEM_FRACTION_STATIC"
+    --swa-full-tokens-ratio 0.10
+    --page-size 256
+    --kv-cache-dtype fp8_e4m3
+    --chunked-prefill-size "$CHUNKED_PREFILL_SIZE"
+    --disable-shared-experts-fusion
+    --tool-call-parser deepseekv4
+    --reasoning-parser deepseek-v4
+    --chat-template "$(dirname "$0")/../chat_templates/deepseek_v4_thinking.jinja"
+    --watchdog-timeout 1800
+    "${METRICS_ARGS[@]}"
+    "${SPEC_ARGS[@]}"
+    "${CACHE_ARGS[@]}"
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+
+{
+    echo "=== SGLANG_* env vars at launch ==="
+    env | grep -E '^SGLANG_' | sort
+    echo "==================================="
+} | tee "$SERVER_LOG"
+
+echo "Starting SGLang server for MI355X..."
+"${SGLANG_CMD[@]}" >> "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+echo "Server PID: $SERVER_PID"
+
+capture_cache_metrics() {
+    {
+        echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
+        curl -fsS "http://localhost:$SGLANG_BACKEND_PORT/metrics" 2>/dev/null \
+            | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|hicache_host_used_tokens|hicache_host_total_tokens|token_usage|num_requests_running|num_requests_waiting))' \
+            || true
+        echo "============================================================"
+    } >> "$SERVER_LOG"
+}
+
+wait_for_server_ready --port "$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 [ "${#METRICS_ARGS[@]}" -gt 0 ]; then
+    capture_cache_metrics
+    trap capture_cache_metrics EXIT
+fi
+
+# ---- Run benchmark ----------------------------------------------------------
+build_replay_cmd "$RESULT_DIR"
+REPLAY_CMD+=" --server-metrics http://localhost:$SGLANG_BACKEND_PORT/metrics"
+
+run_agentic_replay_and_write_outputs "$RESULT_DIR"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 4b1ea4aed3..2781c48485 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2351,8 +2351,6 @@ dsv4-fp4-mi355x-sglang-agentic:
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds
-  precision: fp4
-  framework: sglang
   multinode: false
   scenarios:
     agentic-coding:
@@ -2360,6 +2358,22 @@ dsv4-fp4-mi355x-sglang-agentic:
       - { tp: 8, kv-offloading: none, conc-list: [16, 32, 64] }
       - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [64, 128, 256] }
 
+dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
+  image: lmsysorg/sglang-rocm:v0.5.15-rocm700-mi35x-20260713
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80, 96], spec-decoding: mtp }
+
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
 # MXFP8 runs from TP=4 on gfx950; block size 128 is mandatory for MSA.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e05c2ca1c4..9e82f7533a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4765,3 +4765,12 @@
     - "Bump image to lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708"
     - "Clean the export envs"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2198
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
+  description:
+    - "Add DSv4 FP4 MI355X SGLang agentic MTP recipe with HiCache support"
+    - "Image: lmsysorg/sglang-rocm:v0.5.15-rocm700-mi35x-20260713"
+    - "EAGLE speculative decoding: 3 steps, topk 1, 4 draft tokens"
+    - "Sweep TP8 +/- DPA +/- HiCache with MTP across conc 1-96"
+    - "Exclude mia1-p01-g12 from agentic salloc"
diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh
index 8cb92b7a16..ec99d1698d 100644
--- a/runners/launch_mi355x-amds.sh
+++ b/runners/launch_mi355x-amds.sh
@@ -212,7 +212,7 @@ else
     #   mia1-p01-g09: pyxis broken (persistently fails to create container filesystem)
     #   mia1-p01-g11: docker.sock permissions denied (cluster-cleanup step fails)
     # Both have been root-caused via #1431/#1432/#1440/#1441/#1443 sweep failures.
-    salloc --partition=$PARTITION --exclude=mia1-p01-g09,mia1-p01-g11 --gres=gpu:$GPU_COUNT --exclusive --cpus-per-task=128 --time=500 --no-shell --job-name="$RUNNER_NAME"
+    salloc --partition=$PARTITION --exclude=mia1-p01-g09,mia1-p01-g11,mia1-p01-g12 --gres=gpu:$GPU_COUNT --exclusive --cpus-per-task=128 --time=500 --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID bash -c "docker stop \$(docker ps -a -q)"

From d8bf2b3d3545cda9d292e08c871e2576d032145f Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 14 Jul 2026 18:04:37 +0900
Subject: [PATCH 02/29] [AMD][MI355X] DSv4 SGLang: restore precision and
 framework fields to agentic config

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 configs/amd-master.yaml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 2781c48485..56a761d332 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2351,6 +2351,8 @@ dsv4-fp4-mi355x-sglang-agentic:
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds
+  precision: fp4
+  framework: sglang
   multinode: false
   scenarios:
     agentic-coding:

From 0b0cc49c7ca9707162e043cf10b9c02ecf7c895b Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 14 Jul 2026 18:11:37 +0900
Subject: [PATCH 03/29] [AMD][MI355X] DSv4 SGLang: add PR link to
 perf-changelog MTP entry

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 perf-changelog.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9e82f7533a..77485ea470 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4774,3 +4774,4 @@
     - "EAGLE speculative decoding: 3 steps, topk 1, 4 draft tokens"
     - "Sweep TP8 +/- DPA +/- HiCache with MTP across conc 1-96"
     - "Exclude mia1-p01-g12 from agentic salloc"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200

From a0f7911f0a15d2fbf136afa12743f4a2dbcd6df5 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 14 Jul 2026 18:12:57 +0900
Subject: [PATCH 04/29] [AMD][MI355X] DSv4 SGLang MTP: move PARALLEL_ARGS init
 before DPA block

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index de891dd71c..923253c7cb 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -78,6 +78,7 @@ SGLANG_BACKEND_PORT="$PORT"
 ROUTER_LOG="$RESULT_DIR/router.log"
 MEM_FRACTION_STATIC=0.95
 CHUNKED_PREFILL_SIZE=16384
+PARALLEL_ARGS=(--tensor-parallel-size "$TP")
 if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true
     export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
@@ -116,7 +117,6 @@ export SGLANG_USE_ROCM700A=0
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export AITER_BF16_FP8_MOE_BOUND=0
 
-PARALLEL_ARGS=(--tensor-parallel-size "$TP")
 METRICS_ARGS=(--enable-metrics)
 SPEC_ARGS=(
     --speculative-algorithm EAGLE

From f773bb0ca7f057591b03097cb1bdc0b1f8360859 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 14 Jul 2026 21:41:00 +0900
Subject: [PATCH 05/29] [AMD][MI355X] DSv4 SGLang: revert agentic-hicache-mtp
 image to v0.5.14

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 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 56a761d332..eb0970787b 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2361,7 +2361,7 @@ dsv4-fp4-mi355x-sglang-agentic:
       - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [64, 128, 256] }
 
 dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.15-rocm700-mi35x-20260713
+  image: lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260710
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds

From 36b1f54c96a7d556887187b2dd8de3dbf640e7e2 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 14 Jul 2026 22:11:37 +0900
Subject: [PATCH 06/29] [AMD][MI355X] DSv4 SGLang MTP: revert
 mem-fraction-static to 0.90 and chunked-prefill to 8192

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 923253c7cb..ec2ab7b935 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -76,8 +76,8 @@ fi
 USE_SGLANG_ROUTER=false
 SGLANG_BACKEND_PORT="$PORT"
 ROUTER_LOG="$RESULT_DIR/router.log"
-MEM_FRACTION_STATIC=0.95
-CHUNKED_PREFILL_SIZE=16384
+MEM_FRACTION_STATIC=0.90
+CHUNKED_PREFILL_SIZE=8192
 PARALLEL_ARGS=(--tensor-parallel-size "$TP")
 if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true

From 53c2c7738dc14fb0c20e6e593c41def0573bbaa5 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 14 Jul 2026 15:28:55 -0500
Subject: [PATCH 07/29] fix(agentx): require explicit benchmark environment

---
 .../agentic/dsv4_fp4_mi355x_sglang_mtp.sh        | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index ec2ab7b935..9b73977ea0 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-set -euo pipefail
+set -eo pipefail
 set -x
 
 # Agentic trace replay benchmark for DeepSeek-V4-Pro FP4 on MI355X using SGLang.
@@ -16,16 +16,16 @@ 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 ${SLURMD_NODENAME:-unknown}"
+if [[ -n "$SLURM_JOB_ID" ]]; then
+    echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
 fi
 
 # ROCR/HIP visibility under slurm cgroups.
-if [ -n "${ROCR_VISIBLE_DEVICES:-}" ]; then
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
     export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
 fi
 
-if [[ -n "${MODEL_PATH:-}" ]]; then
+if [[ -n "$MODEL_PATH" ]]; then
     if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
         hf download "$MODEL" --local-dir "$MODEL_PATH"
     fi
@@ -50,7 +50,7 @@ export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
 CACHE_ARGS=()
 if agentic_kv_offload_enabled; then
     # HiCache config — https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4
-    case "${KV_OFFLOAD_BACKEND:-}" in
+    case "$KV_OFFLOAD_BACKEND" in
         hicache)
             HICACHE_RATIO=4
             HICACHE_WRITE_POLICY="write_through"
@@ -66,7 +66,7 @@ if agentic_kv_offload_enabled; then
             echo "HiCache DSv4 CPU tier: ratio=$HICACHE_RATIO, write_policy=$HICACHE_WRITE_POLICY, io_backend=$HICACHE_IO_BACKEND, mem_layout=$HICACHE_MEM_LAYOUT"
             ;;
         *)
-            echo "Error: unsupported KV_OFFLOAD_BACKEND '${KV_OFFLOAD_BACKEND:-}' (expected: hicache)" >&2
+            echo "Error: unsupported KV_OFFLOAD_BACKEND '$KV_OFFLOAD_BACKEND' (expected: hicache)" >&2
             exit 1
             ;;
     esac
@@ -101,7 +101,7 @@ if [ "$DP_ATTENTION" = "true" ]; then
     )
 fi
 
-if [ "${EP_SIZE:-1}" -gt 1 ]; then
+if [ "$EP_SIZE" -gt 1 ]; then
     PARALLEL_ARGS+=(--ep-size "$EP_SIZE")
 fi
 

From 8850665b706497bfa1a270115d65482200c68d03 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 14 Jul 2026 15:34:01 -0500
Subject: [PATCH 08/29] fix(agentx): use golden MTP acceptance length

---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 9b73977ea0..284a12f62c 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -113,6 +113,10 @@ CUDA_GRAPH_MAX_BS=$CONC
 
 export SGLANG_DEFAULT_THINKING=1
 export SGLANG_DSV4_REASONING_EFFORT=max
+# DeepSeek-V4-Pro thinking-on golden AL for three speculative tokens.
+export SGLANG_SIMULATE_ACC_LEN=2.49
+export SGLANG_SIMULATE_ACC_METHOD=match-expected
+export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
 export SGLANG_USE_ROCM700A=0
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export AITER_BF16_FP8_MOE_BOUND=0

From 25f137d27f275451e51d960f4c74383f14a129a1 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 14 Jul 2026 15:44:00 -0500
Subject: [PATCH 09/29] fix(agentx): match golden AL reasoning effort

---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 284a12f62c..a09a929d98 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -112,7 +112,7 @@ CUDA_GRAPH_MAX_BS=$CONC
 [ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
 
 export SGLANG_DEFAULT_THINKING=1
-export SGLANG_DSV4_REASONING_EFFORT=max
+export SGLANG_DSV4_REASONING_EFFORT=high
 # DeepSeek-V4-Pro thinking-on golden AL for three speculative tokens.
 export SGLANG_SIMULATE_ACC_LEN=2.49
 export SGLANG_SIMULATE_ACC_METHOD=match-expected

From 9139777f25b697939c5d0b0d8fc18cd91cab1e8d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 14 Jul 2026 15:50:38 -0500
Subject: [PATCH 10/29] style(agentx): group simulated AL settings
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Place the thinking and simulated acceptance-length environment variables in a dedicated, clearly labeled block.

中文：将思考模式与模拟接受长度（AL）环境变量整理到独立且标注清晰的配置块中。
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index a09a929d98..723ea4f731 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -111,12 +111,13 @@ MAX_RUNNING_REQUESTS=$((2 * CONC))
 CUDA_GRAPH_MAX_BS=$CONC
 [ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
 
+# Simulated acceptance-length (AL) settings.
 export SGLANG_DEFAULT_THINKING=1
 export SGLANG_DSV4_REASONING_EFFORT=high
-# DeepSeek-V4-Pro thinking-on golden AL for three speculative tokens.
 export SGLANG_SIMULATE_ACC_LEN=2.49
 export SGLANG_SIMULATE_ACC_METHOD=match-expected
 export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+
 export SGLANG_USE_ROCM700A=0
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export AITER_BF16_FP8_MOE_BOUND=0

From 369e49b829049d7456f657ba2a3019f6389bb5ad Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 14 Jul 2026 15:58:59 -0500
Subject: [PATCH 11/29] chore(agentx): remove cache metric snapshots
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove the manual metrics scrape and EXIT trap while retaining the metrics endpoint used by the replay client.

中文：移除手动指标抓取与 EXIT trap，同时保留回放客户端使用的 metrics endpoint。
---
 .../agentic/dsv4_fp4_mi355x_sglang_mtp.sh         | 15 ---------------
 1 file changed, 15 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 723ea4f731..a601daa15f 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -170,16 +170,6 @@ echo "Starting SGLang server for MI355X..."
 SERVER_PID=$!
 echo "Server PID: $SERVER_PID"
 
-capture_cache_metrics() {
-    {
-        echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
-        curl -fsS "http://localhost:$SGLANG_BACKEND_PORT/metrics" 2>/dev/null \
-            | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|hicache_host_used_tokens|hicache_host_total_tokens|token_usage|num_requests_running|num_requests_waiting))' \
-            || true
-        echo "============================================================"
-    } >> "$SERVER_LOG"
-}
-
 wait_for_server_ready --port "$SGLANG_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
 
 if [ "$USE_SGLANG_ROUTER" = "true" ]; then
@@ -202,11 +192,6 @@ if [ "$USE_SGLANG_ROUTER" = "true" ]; then
     wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
 fi
 
-if [ "${#METRICS_ARGS[@]}" -gt 0 ]; then
-    capture_cache_metrics
-    trap capture_cache_metrics EXIT
-fi
-
 # ---- Run benchmark ----------------------------------------------------------
 build_replay_cmd "$RESULT_DIR"
 REPLAY_CMD+=" --server-metrics http://localhost:$SGLANG_BACKEND_PORT/metrics"

From a016c0eb858ab3e2871aedb46f5a25daed9c66f0 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 14 Jul 2026 16:31:35 -0500
Subject: [PATCH 12/29] fix(runner): remove MI355X node exclusions

---
 perf-changelog.yaml           | 1 -
 runners/launch_mi355x-amds.sh | 6 +-----
 2 files changed, 1 insertion(+), 6 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 77485ea470..602dee8aec 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4773,5 +4773,4 @@
     - "Image: lmsysorg/sglang-rocm:v0.5.15-rocm700-mi35x-20260713"
     - "EAGLE speculative decoding: 3 steps, topk 1, 4 draft tokens"
     - "Sweep TP8 +/- DPA +/- HiCache with MTP across conc 1-96"
-    - "Exclude mia1-p01-g12 from agentic salloc"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200
diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh
index ec99d1698d..d7007aaf09 100644
--- a/runners/launch_mi355x-amds.sh
+++ b/runners/launch_mi355x-amds.sh
@@ -208,11 +208,7 @@ else
     export GPU_COUNT="${GPU_COUNT:-${TP:?TP must be set}}"
 
     set -x
-    # Exclude known-bad mi355x compute nodes (KLAUD_DEBUG §5.1 / §5.2):
-    #   mia1-p01-g09: pyxis broken (persistently fails to create container filesystem)
-    #   mia1-p01-g11: docker.sock permissions denied (cluster-cleanup step fails)
-    # Both have been root-caused via #1431/#1432/#1440/#1441/#1443 sweep failures.
-    salloc --partition=$PARTITION --exclude=mia1-p01-g09,mia1-p01-g11,mia1-p01-g12 --gres=gpu:$GPU_COUNT --exclusive --cpus-per-task=128 --time=500 --no-shell --job-name="$RUNNER_NAME"
+    salloc --partition=$PARTITION --gres=gpu:$GPU_COUNT --exclusive --cpus-per-task=128 --time=500 --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID bash -c "docker stop \$(docker ps -a -q)"

From 024a184d0cb8c87784134ae8d5384ca673cba479 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Wed, 15 Jul 2026 11:53:57 +0900
Subject: [PATCH 13/29] [AMD][MI355X] DSv4 SGLang MTP: drop two-batch-overlap,
 add TCP user timeout, clean up var defaults

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index a601daa15f..bbfd40c242 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -71,6 +71,8 @@ if agentic_kv_offload_enabled; then
             ;;
     esac
 fi
+# ---- Client config ----------------------------------------------------------
+export AIPERF_HTTP_TCP_USER_TIMEOUT=1000000 
 
 # ---- LLM server config ----------------------------------------------------------
 USE_SGLANG_ROUTER=false
@@ -97,7 +99,6 @@ if [ "$DP_ATTENTION" = "true" ]; then
         --dp "$TP"
         --enable-dp-attention
         --enable-prefill-delayer
-        --enable-two-batch-overlap
     )
 fi
 

From 75584f7dd2e3713fdeadba019bde2f7ff2d49818 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Wed, 15 Jul 2026 15:41:15 +0900
Subject: [PATCH 14/29] [AMD][MI355X] DSv4 SGLang MTP: cap cuda-graph-max-bs at
 64, set reasoning-effort to max, narrow HiCache sweep to c96

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh      | 4 ++--
 configs/amd-master.yaml                                    | 7 ++++---
 2 files changed, 6 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index bbfd40c242..3d68fc8b15 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -110,11 +110,11 @@ fi
 # Allow subagent fan-out to exceed CONC without clipping request bursts.
 MAX_RUNNING_REQUESTS=$((2 * CONC))
 CUDA_GRAPH_MAX_BS=$CONC
-[ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
+[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
 
 # Simulated acceptance-length (AL) settings.
 export SGLANG_DEFAULT_THINKING=1
-export SGLANG_DSV4_REASONING_EFFORT=high
+export SGLANG_DSV4_REASONING_EFFORT=max
 export SGLANG_SIMULATE_ACC_LEN=2.49
 export SGLANG_SIMULATE_ACC_METHOD=match-expected
 export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index eb0970787b..b6530f0063 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2372,9 +2372,10 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80, 96], spec-decoding: mtp }
+      #- { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
+      #- { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
+      #- { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80, 96], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [96], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5

From 6ba055961d548246e81edf9d97326fd8cb1de14e Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Wed, 15 Jul 2026 17:38:17 +0900
Subject: [PATCH 15/29] [AMD][MI355X] DSv4 SGLang MTP: revert cuda-graph-max-bs
 cap and reasoning-effort, lower HiCache sweep to c64

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 4 ++--
 configs/amd-master.yaml                                      | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 3d68fc8b15..bbfd40c242 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -110,11 +110,11 @@ fi
 # Allow subagent fan-out to exceed CONC without clipping request bursts.
 MAX_RUNNING_REQUESTS=$((2 * CONC))
 CUDA_GRAPH_MAX_BS=$CONC
-[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+[ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
 
 # Simulated acceptance-length (AL) settings.
 export SGLANG_DEFAULT_THINKING=1
-export SGLANG_DSV4_REASONING_EFFORT=max
+export SGLANG_DSV4_REASONING_EFFORT=high
 export SGLANG_SIMULATE_ACC_LEN=2.49
 export SGLANG_SIMULATE_ACC_METHOD=match-expected
 export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index b6530f0063..ad1542ebb1 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2375,7 +2375,7 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
       #- { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
       #- { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
       #- { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80, 96], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [96], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5

From 1c7e030face82c4e0cc57968d357a5eb566e552f Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Wed, 15 Jul 2026 18:31:02 +0900
Subject: [PATCH 16/29] [AMD][MI355X] DSv4 SGLang MTP: switch HiCache c64 to no
 KV offloading

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 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 ad1542ebb1..69177a2092 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2375,7 +2375,7 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
       #- { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
       #- { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
       #- { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80, 96], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [64], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5

From 82039f0514a5e40581288d3c34d9c31667035698 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Wed, 15 Jul 2026 20:32:06 +0900
Subject: [PATCH 17/29] [AMD][MI355X] DSv4 SGLang MTP: use
 cuda-graph-max-bs-decode flag and restore full sweep to c72

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh       | 2 +-
 configs/amd-master.yaml                                     | 6 ++----
 2 files changed, 3 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index bbfd40c242..9c84e55249 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -140,7 +140,7 @@ SGLANG_CMD=(
     --trust-remote-code
     "${PARALLEL_ARGS[@]}"
     --attention-backend compressed
-    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --mem-fraction-static "$MEM_FRACTION_STATIC"
     --swa-full-tokens-ratio 0.10
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 69177a2092..27fce274dc 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2372,10 +2372,8 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      #- { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
-      #- { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
-      #- { tp: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80, 96], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [64], spec-decoding: mtp }
+      - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64, 72], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5

From 3e52d17293f422c4f7227c12bcd2868faa67ef57 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Wed, 15 Jul 2026 21:04:58 +0900
Subject: [PATCH 18/29] [AMD][MI355X] DSv4 SGLang MTP: drop c72 from DPA
 no-offload sweep

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 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 27fce274dc..e8e801a23d 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2373,7 +2373,7 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64, 72], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5

From fe0c5b598f6af8eac3888b7469d4224524fca6cc Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 16 Jul 2026 00:11:29 +0900
Subject: [PATCH 19/29] [AMD][MI355X] DSv4 SGLang MTP: disable
 unified-radix-tree per sgl-project/sglang#30339, drop GPU_MAX_HW_QUEUES, set
 max-running-requests to CUDA_GRAPH_MAX_BS

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../agentic/dsv4_fp4_mi355x_sglang_mtp.sh            | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 9c84e55249..965ff843a0 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -44,9 +44,6 @@ install_agentic_deps
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
-export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
-export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
-
 CACHE_ARGS=()
 if agentic_kv_offload_enabled; then
     # HiCache config — https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4
@@ -92,7 +89,6 @@ if [ "$DP_ATTENTION" = "true" ]; then
     export SGLANG_DP_SHARED_EXPERT_LOCAL=1
     export SGLANG_DP_USE_GATHERV=1
     export SGLANG_DP_USE_REDUCE_SCATTER=1
-    export GPU_MAX_HW_QUEUES=5
 
     CHUNKED_PREFILL_SIZE=$((8192 * TP))
     PARALLEL_ARGS+=(
@@ -108,7 +104,6 @@ fi
 
 # AgentX concurrency counts live session trees, not individual requests.
 # Allow subagent fan-out to exceed CONC without clipping request bursts.
-MAX_RUNNING_REQUESTS=$((2 * CONC))
 CUDA_GRAPH_MAX_BS=$CONC
 [ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
 
@@ -123,6 +118,11 @@ export SGLANG_USE_ROCM700A=0
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export AITER_BF16_FP8_MOE_BOUND=0
 
+# sglang kv cache
+# https://github.com/sgl-project/sglang/pull/30339
+#export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
+export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
+
 METRICS_ARGS=(--enable-metrics)
 SPEC_ARGS=(
     --speculative-algorithm EAGLE
@@ -141,7 +141,7 @@ SGLANG_CMD=(
     "${PARALLEL_ARGS[@]}"
     --attention-backend compressed
     --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
-    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --max-running-requests "$CUDA_GRAPH_MAX_BS"
     --mem-fraction-static "$MEM_FRACTION_STATIC"
     --swa-full-tokens-ratio 0.10
     --page-size 256

From df554692f37452eda80fd6887a85e271b67d73bd Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 16 Jul 2026 01:37:45 +0900
Subject: [PATCH 20/29] [AMD][MI355X] DSv4 SGLang MTP: also disable
 unified-cache-free-out-of-window-slots when hicache off

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 965ff843a0..41656ec95b 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -118,10 +118,10 @@ export SGLANG_USE_ROCM700A=0
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export AITER_BF16_FP8_MOE_BOUND=0
 
-# sglang kv cache
+# sglang kv cache, when hicache is disabled
 # https://github.com/sgl-project/sglang/pull/30339
 #export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
-export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
+#export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
 
 METRICS_ARGS=(--enable-metrics)
 SPEC_ARGS=(

From 57aa3c6ff81267875e64c96f8d4ea4d7daad0ad7 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 12:51:42 -0500
Subject: [PATCH 21/29] fix(mi355x): bump agentic MTP SGLang image

---
 configs/amd-master.yaml | 2 +-
 perf-changelog.yaml     | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index e8e801a23d..683d64ebb0 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2361,7 +2361,7 @@ dsv4-fp4-mi355x-sglang-agentic:
       - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [64, 128, 256] }
 
 dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260710
+  image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260714
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 271e20d7c2..8c7c3784e8 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4786,7 +4786,7 @@
     - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
   description:
     - "Add DSv4 FP4 MI355X SGLang agentic MTP recipe with HiCache support"
-    - "Image: lmsysorg/sglang-rocm:v0.5.15-rocm700-mi35x-20260713"
+    - "Image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260714"
     - "EAGLE speculative decoding: 3 steps, topk 1, 4 draft tokens"
     - "Sweep TP8 +/- DPA +/- HiCache with MTP across conc 1-96"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200

From 48b08c6f03a4cd6a1b83e66724e5c3f0730e353f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 12:54:12 -0500
Subject: [PATCH 22/29] fix(changelog): append MTP image bump

---
 perf-changelog.yaml | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 8c7c3784e8..08c501fac3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4786,7 +4786,7 @@
     - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
   description:
     - "Add DSv4 FP4 MI355X SGLang agentic MTP recipe with HiCache support"
-    - "Image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260714"
+    - "Image: lmsysorg/sglang-rocm:v0.5.15-rocm700-mi35x-20260713"
     - "EAGLE speculative decoding: 3 steps, topk 1, 4 draft tokens"
     - "Sweep TP8 +/- DPA +/- HiCache with MTP across conc 1-96"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200
@@ -4804,3 +4804,10 @@
     - "Add EAGLE3 speculative-decoding arm for the Kimi K2.6 NVFP4 B300 AgentX recipe (draft lightseekorg/kimi-k2.6-eagle3-mla, TOKENSPEED_MLA attention backend with TRT-LLM ragged MLA kernel)."
     - "TP8/TP4 GPU-only KV points plus a TP4 native CPU-offload ladder via SimpleCPUOffloadConnector with lazy_offload off; TP4/DCP4 high-concurrency points (conc 32/64) using num_speculative_tokens=3 and synthetic_acceptance_length=2.88."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2222
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
+  description:
+    - "Bump image to lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260714"
+    - "Fix HSA out-of-resources crashes during DPA agentic MTP warmup"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200

From 8aeac8a7351745c511a1083705d4a828760a0afb Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 13:38:55 -0500
Subject: [PATCH 23/29] fix(mi355x): scale MTP DPA limits per engine

---
 .../agentic/dsv4_fp4_mi355x_sglang_mtp.sh           | 13 ++++++++++---
 perf-changelog.yaml                                 |  7 +++++++
 2 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 41656ec95b..a4d6045c50 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -102,9 +102,16 @@ if [ "$EP_SIZE" -gt 1 ]; then
     PARALLEL_ARGS+=(--ep-size "$EP_SIZE")
 fi
 
-# AgentX concurrency counts live session trees, not individual requests.
-# Allow subagent fan-out to exceed CONC without clipping request bursts.
+# AgentX concurrency counts live session trees, not individual requests, so
+# retain 2x aggregate request headroom for subagent fan-out. SGLang applies
+# both limits per engine; with DP attention, divide the aggregate limits over
+# the DP ranks instead of granting the full global limit to every rank.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
 CUDA_GRAPH_MAX_BS=$CONC
+if [ "$DP_ATTENTION" = "true" ]; then
+    MAX_RUNNING_REQUESTS=$(( (MAX_RUNNING_REQUESTS + TP - 1) / TP ))
+    CUDA_GRAPH_MAX_BS=$(( (CUDA_GRAPH_MAX_BS + TP - 1) / TP ))
+fi
 [ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
 
 # Simulated acceptance-length (AL) settings.
@@ -141,7 +148,7 @@ SGLANG_CMD=(
     "${PARALLEL_ARGS[@]}"
     --attention-backend compressed
     --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
-    --max-running-requests "$CUDA_GRAPH_MAX_BS"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
     --mem-fraction-static "$MEM_FRACTION_STATIC"
     --swa-full-tokens-ratio 0.10
     --page-size 256
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b96383e9a2..c7469793ce 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4804,3 +4804,10 @@
     - "Bump image to lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260714"
     - "Fix HSA out-of-resources crashes during DPA agentic MTP warmup"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
+  description:
+    - "Scale SGLang request and CUDA-graph limits per DP-attention engine while preserving aggregate AgentX fan-out headroom"
+    - "Prevent per-rank HSA out-of-resources crashes during accelerated warmup"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200

From ea8caa92c9539899eec5dd9d0c847e5b1ffcd99d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 14:04:23 -0500
Subject: [PATCH 24/29] fix(mi355x): reserve DPA MTP runtime headroom

---
 .../agentic/dsv4_fp4_mi355x_sglang_mtp.sh           | 13 +++++++------
 perf-changelog.yaml                                 |  7 +++++++
 2 files changed, 14 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index a4d6045c50..d102b32049 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -80,6 +80,9 @@ CHUNKED_PREFILL_SIZE=8192
 PARALLEL_ARGS=(--tensor-parallel-size "$TP")
 if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true
+    # DPA + MTP needs additional runtime headroom for speculative decode and
+    # communication buffers beyond SGLang's static KV pool.
+    MEM_FRACTION_STATIC=0.80
     export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
     SGLANG_BACKEND_PORT=$((PORT + 1))
     SGLANG_ROUTER_METRICS_PORT=$((PORT + 10000))
@@ -102,14 +105,12 @@ if [ "$EP_SIZE" -gt 1 ]; then
     PARALLEL_ARGS+=(--ep-size "$EP_SIZE")
 fi
 
-# AgentX concurrency counts live session trees, not individual requests, so
-# retain 2x aggregate request headroom for subagent fan-out. SGLang applies
-# both limits per engine; with DP attention, divide the aggregate limits over
-# the DP ranks instead of granting the full global limit to every rank.
-MAX_RUNNING_REQUESTS=$((2 * CONC))
+# SGLang treats max-running-requests as a global DPA limit and partitions it
+# internally. CUDA graph capture is per scheduler, so only its batch size is
+# divided across DP ranks.
+MAX_RUNNING_REQUESTS=$CONC
 CUDA_GRAPH_MAX_BS=$CONC
 if [ "$DP_ATTENTION" = "true" ]; then
-    MAX_RUNNING_REQUESTS=$(( (MAX_RUNNING_REQUESTS + TP - 1) / TP ))
     CUDA_GRAPH_MAX_BS=$(( (CUDA_GRAPH_MAX_BS + TP - 1) / TP ))
 fi
 [ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c7469793ce..6543010f64 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4811,3 +4811,10 @@
     - "Scale SGLang request and CUDA-graph limits per DP-attention engine while preserving aggregate AgentX fan-out headroom"
     - "Prevent per-rank HSA out-of-resources crashes during accelerated warmup"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
+  description:
+    - "Keep max-running-requests global for SGLang DP attention and scale only per-scheduler CUDA graph capture"
+    - "Reserve DPA MTP runtime headroom with mem-fraction-static 0.80"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200

From 53fb3a58c63a09b595cd30c669e8358fe8858c79 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 14:43:02 -0500
Subject: [PATCH 25/29] fix(mi355x): increase DPA MTP memory headroom

---
 .../single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh       | 2 +-
 perf-changelog.yaml                                         | 6 ++++++
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index d102b32049..4def38ec90 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -82,7 +82,7 @@ if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true
     # DPA + MTP needs additional runtime headroom for speculative decode and
     # communication buffers beyond SGLang's static KV pool.
-    MEM_FRACTION_STATIC=0.80
+    MEM_FRACTION_STATIC=0.75
     export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
     SGLANG_BACKEND_PORT=$((PORT + 1))
     SGLANG_ROUTER_METRICS_PORT=$((PORT + 10000))
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6543010f64..cd24eb1392 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4818,3 +4818,9 @@
     - "Keep max-running-requests global for SGLang DP attention and scale only per-scheduler CUDA graph capture"
     - "Reserve DPA MTP runtime headroom with mem-fraction-static 0.80"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
+  description:
+    - "Increase DPA MTP runtime headroom to mem-fraction-static 0.75 after c64 exhausted VRAM at 0.80 during regular warmup"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200

From 0360b6c48ad4f8b34b24f90e81024bf7d928cc4b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 15:44:11 -0500
Subject: [PATCH 26/29] fix(mi355x): increase DPA agentic prefill chunk

---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 4def38ec90..86f4a8f7f9 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -93,7 +93,10 @@ if [ "$DP_ATTENTION" = "true" ]; then
     export SGLANG_DP_USE_GATHERV=1
     export SGLANG_DP_USE_REDUCE_SCATTER=1
 
-    CHUNKED_PREFILL_SIZE=$((8192 * TP))
+    # SGLang divides the configured chunk across DP schedulers. Use a 16K
+    # per-scheduler chunk so long agentic prefill tails drain within the
+    # standard 600-second warmup grace period.
+    CHUNKED_PREFILL_SIZE=$((16384 * TP))
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention

From ff083efe84fcd1841c5db3b14de81fea489c1fbd Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 15:47:20 -0500
Subject: [PATCH 27/29] style(mi355x): clean MTP launcher whitespace

---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 86f4a8f7f9..c52787b9a1 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -9,7 +9,7 @@ set -x
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# KV_OFFLOADING=dram requires one of these. 
+# KV_OFFLOADING=dram requires one of these.
 #   KV_OFFLOAD_BACKEND=hicache.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
@@ -69,7 +69,7 @@ if agentic_kv_offload_enabled; then
     esac
 fi
 # ---- Client config ----------------------------------------------------------
-export AIPERF_HTTP_TCP_USER_TIMEOUT=1000000 
+export AIPERF_HTTP_TCP_USER_TIMEOUT=1000000
 
 # ---- LLM server config ----------------------------------------------------------
 USE_SGLANG_ROUTER=false

From b53fc3a301d026e1bbb96636c23b0cfe1dd1e582 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 16 Jul 2026 10:48:20 +0900
Subject: [PATCH 28/29] [AMD][MI355X] DSv4 SGLang MTP: re-enable
 unified-cache-free-out-of-window-slots, drop conc=64 from hicache MTP

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh | 2 +-
 configs/amd-master.yaml                                      | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index c52787b9a1..95945c409d 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -132,7 +132,7 @@ export AITER_BF16_FP8_MOE_BOUND=0
 # sglang kv cache, when hicache is disabled
 # https://github.com/sgl-project/sglang/pull/30339
 #export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
-#export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
+export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
 
 METRICS_ARGS=(--enable-metrics)
 SPEC_ARGS=(
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 683d64ebb0..d1b1d69029 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2373,7 +2373,7 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5

From bd4f4e1a12fafd9d794b9df55af2b78291c48462 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 15 Jul 2026 22:21:11 -0500
Subject: [PATCH 29/29] fix(mi355x): disable DPA prefill delayer

---
 .../single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh      | 1 -
 configs/amd-master.yaml                                    | 2 +-
 perf-changelog.yaml                                        | 7 +++++++
 3 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
index 95945c409d..ce3c2ad882 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
@@ -100,7 +100,6 @@ if [ "$DP_ATTENTION" = "true" ]; then
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention
-        --enable-prefill-delayer
     )
 fi
 
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0c787be1d7..24c4e0e7b3 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2428,7 +2428,7 @@ dsv4-fp4-mi355x-sglang-agentic-hicache-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 64], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 42ea007d07..a6059e1589 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4876,3 +4876,10 @@
     - "Image: lmsysorg/sglang:v0.5.13.post1-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2123
   
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-hicache-mtp
+  description:
+    - "Disable the SGLang prefill delayer so DP-attention MTP c64 warmup drains within the standard 600-second grace period"
+    - "Restore DP-attention concurrency 64 to the full sweep"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2200
+