From f3dc0f8b90da39bffd630ededa0c81c0837b4bbd Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 13 Aug 2026 11:15:08 -0700
Subject: [PATCH 1/6] feat(config): add H200 MTP
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

添加 H200 MTP 配置。
---
 .../agentic/qwen3.5_fp8_h200_sglang_mtp.sh    | 160 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  16 ++
 perf-changelog.yaml                           |   9 +
 3 files changed, 185 insertions(+)
 create mode 100755 benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh
new file mode 100755
index 000000000..f0fee4946
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh
@@ -0,0 +1,160 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# Agentic trace replay for Qwen3.5 FP8 on H200 with native NEXTN MTP and
+# DRAM HiCache. Throughput pins the committed golden acceptance length; evals
+# retain target-model verification.
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL MODEL_PREFIX TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR \
+    DURATION EP_SIZE EVAL_ONLY SPEC_DECODING
+
+if ! require_agentic_kv_offload_backend hicache; then
+    echo "Error: this H200 MTP recipe requires KV_OFFLOADING=dram with HiCache" >&2
+    exit 1
+fi
+
+if [ "$TP" -ne 8 ] || [ "$EP_SIZE" -ne 1 ]; then
+    echo "Error: this H200 MTP recipe requires TP=8 and EP_SIZE=1" >&2
+    exit 1
+fi
+
+if [ "$SPEC_DECODING" != "mtp" ]; then
+    echo "Error: this launcher requires SPEC_DECODING=mtp" >&2
+    exit 1
+fi
+
+if [ "$TOTAL_CPU_DRAM_GB" -lt 1200 ]; then
+    echo "Error: TP8 requires at least 1200 GB host DRAM; generated budget is ${TOTAL_CPU_DRAM_GB} GB" >&2
+    exit 1
+fi
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${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="$MODEL"
+fi
+nvidia-smi
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_with_subagents_256k
+resolve_trace_source
+install_agentic_deps
+
+# Pin the published FlashInfer 0.6.15 CUDA 13 wheels used by this launcher.
+agentic_pip_install --no-deps --force-reinstall flashinfer_python==0.6.15
+agentic_pip_install \
+    --no-deps --force-reinstall flashinfer-cubin==0.6.15 \
+    --index-url https://flashinfer.ai/whl
+agentic_pip_install \
+    --no-deps --force-reinstall flashinfer-jit-cache==0.6.15+cu130 \
+    --index-url https://flashinfer.ai/whl/cu130
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+export TORCH_CUDA_ARCH_LIST=9.0
+export PYTHONNOUSERSITE=1
+export PYTHONUNBUFFERED=1
+export FLASHINFER_DISABLE_VERSION_CHECK=1
+export FLASHINFER_WORKSPACE_BASE=/tmp/flashinfer-cache
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
+export SGLANG_ENABLE_SPEC_V2=1
+
+# Three speculative tokens per verification step. For agentic throughput, pin
+# the committed Qwen3.5 MTP golden AL (thinking_on, K=3); evals use real
+# verification because simulated acceptance does not produce accuracy data.
+if [ "$EVAL_ONLY" != "true" ]; then
+    export SGLANG_SIMULATE_ACC_LEN=3.39
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    --tensor-parallel-size "$TP"
+    --data-parallel-size 1
+    --expert-parallel-size "$EP_SIZE"
+    --quantization fp8
+    --kv-cache-dtype fp8_e4m3
+    --mamba-ssm-dtype bfloat16
+    --mamba-scheduler-strategy extra_buffer
+    --mamba-track-interval 8192
+    --mamba-max-states-per-path 1
+    --attention-backend flashinfer
+    --cuda-graph-max-bs 32
+    --max-running-requests 128
+    --max-prefill-tokens 16384
+    --chunked-prefill-size 16384
+    --mem-fraction-static 0.78
+    --max-mamba-cache-size 360
+    --allow-auto-truncate
+    --stream-interval 50
+    --scheduler-recv-interval 10
+    --tokenizer-worker-num 6
+    --enable-cache-report
+    --enable-symm-mem
+    --enable-metrics
+    --page-size 64
+    --enable-hierarchical-cache
+    --hicache-ratio 0.9
+    --hicache-io-backend kernel
+    --hicache-mem-layout page_first_direct
+    --hicache-write-policy write_back
+    --speculative-algorithm NEXTN
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+
+SERVER_PID=""
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "SGLang server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+{
+    echo "=== SGLANG_* env vars at launch ==="
+    env | grep -E '^SGLANG_' | sort
+    echo "==================================="
+} > "$SERVER_LOG"
+
+"${SGLANG_CMD[@]}" >> "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+echo "Server PID: $SERVER_PID"
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [ "$EVAL_ONLY" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --apply-chat-template"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index b95f59618..6544e66c7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7151,6 +7151,22 @@ qwen3.5-fp8-h200-sglang-agentic-mtp:
       - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [4, 8, 12, 16] }
       - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [4, 8, 12, 16] }
 
+# H200 AgentX MTP frontier with DRAM HiCache. This is intentionally an MTP-only
+# submission; the model's non-speculative AgentX arm is not included.
+qwen3.5-fp8-h200-sglang-agentic-hicache-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260716-b0b2dfbd
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: cluster:h200-dgxc
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [2, 4, 8, 10, 12, 16, 20, 24] }
+
 
 qwen3.5-fp8-h100-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6082f2b86..656994384 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5918,3 +5918,12 @@
     - "Add a TEP2 arm (tp 2, ep 2) to the qwen3.5-fp4-b200-sglang-mtp 8k/1k sweep at concurrency 16, 32, and 64"
     - "Rides on the NVFP4-V2 checkpoint switch from #2205"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2550
+
+- config-keys:
+    - qwen3.5-fp8-h200-sglang-agentic-hicache-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add Qwen3.5 FP8 H200 SGLang AgentX MTP points with DRAM HiCache"
+    - "Use TP8/EP1 concurrency 2 through 24, published FlashInfer 0.6.15 CUDA 13 wheels, and golden MTP acceptance length 3.39"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From 9faefbbd5aa9d8f751cf2f9bd3159cbe04389425 Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 13 Aug 2026 11:15:42 -0700
Subject: [PATCH 2/6] chore(changelog): link PR
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

补充性能变更日志中的 PR 链接。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 656994384..12d2df312 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5596,7 +5596,7 @@
     - "Image ghcr.io/tile-ai/tilert:0.1.5 (tilert 0.1.5.post2 installed at container start); commands aligned to TileRT README Topology A -- NIXL KV transfer, --kv-cache-dtype fp8_ds_mla (prefill) <-> fp8 (decode), max-seq-len 202752; MTP speculative-config wired via spec-decoding=mtp"
     - "Topology: 1 prefill node (TP8) + 1 decode node (TP8), each 8xB200 exclusive; TileRT decode is bs=1 only so conc-list is a single point [1], ISL 1k/8k OSL 1k"
     - "Runner: launch_b200-dgxc.sh tilert early-return branch (zero impact on the dynamo path); tilert_utils/submit.sh issues two srun --ntasks=1, one per role, because prefill and decode need different container images; roles are dispatched by the TILERT_ROLE it exports, and torn down across nodes via a sentinel file on the shared /workspace"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2592
 
 - config-keys:
     - qwen3.5-fp8-b200-sglang-agentic-mtp

From e222ad48f9594c087f8333010aef153f0b8d33fe Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 13 Aug 2026 11:30:57 -0700
Subject: [PATCH 3/6] Revert "chore(changelog): link PR"

This reverts commit 9faefbbd5aa9d8f751cf2f9bd3159cbe04389425.
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 12d2df312..656994384 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5596,7 +5596,7 @@
     - "Image ghcr.io/tile-ai/tilert:0.1.5 (tilert 0.1.5.post2 installed at container start); commands aligned to TileRT README Topology A -- NIXL KV transfer, --kv-cache-dtype fp8_ds_mla (prefill) <-> fp8 (decode), max-seq-len 202752; MTP speculative-config wired via spec-decoding=mtp"
     - "Topology: 1 prefill node (TP8) + 1 decode node (TP8), each 8xB200 exclusive; TileRT decode is bs=1 only so conc-list is a single point [1], ISL 1k/8k OSL 1k"
     - "Runner: launch_b200-dgxc.sh tilert early-return branch (zero impact on the dynamo path); tilert_utils/submit.sh issues two srun --ntasks=1, one per role, because prefill and decode need different container images; roles are dispatched by the TILERT_ROLE it exports, and torn down across nodes via a sentinel file on the shared /workspace"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2592
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
 
 - config-keys:
     - qwen3.5-fp8-b200-sglang-agentic-mtp

From 15b43f211cb339c72e33b2242983d9635cfe026a Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 13 Aug 2026 14:20:42 -0700
Subject: [PATCH 4/6] fix(h200): update cache path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

更新 H200 缓存路径。
---
 .../agentic/qwen3.5_fp8_h200_mtp.sh           | 209 ++++++++----------
 .../agentic/qwen3.5_fp8_h200_sglang_mtp.sh    | 160 --------------
 runners/launch_h200-dgxc-slurm.sh             |   2 +-
 3 files changed, 92 insertions(+), 279 deletions(-)
 delete mode 100755 benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh
index 62982a397..1ed9ef9ca 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh
@@ -2,44 +2,34 @@
 set -euo pipefail
 set -x
 
-# Agentic trace replay benchmark for Qwen3.5 FP8 on H200 using SGLang with MTP
-# speculative decoding. First Qwen3.5 AgentX recipe on H200; it is spec-decode
-# only, per the AgentX policy that new agentic arms ship with speculative
-# decoding enabled rather than as an STP/MTP A/B (MODELS.md).
-#
-# Structure follows the proven agentic/qwen3.5_fp8_h100.sh replay path
-# (HiCache host-DRAM offload, the multi_tokenizer cached_tokens_details patch,
-# aiperf-driven trace replay). H200's 141 GB HBM3e is roomier than H100's 80 GB,
-# so --mem-fraction-static is 0.8 rather than 0.75, matching
-# fixed_seq_len/qwen3.5_fp8_h200_mtp.sh. Attention backend stays flashinfer
-# (sm_90); the trtllm_mha path is Blackwell-only.
-#
-# Speculative decoding mirrors fixed_seq_len/qwen3.5_fp8_h100_mtp.sh:
-# SGLANG_ENABLE_SPEC_V2=1 with --speculative-algorithm EAGLE, 3 steps, eagle-topk
-# 1 and 4 draft tokens, i.e. 3 speculative tokens per verification step.
-#
-# Throughput runs pin acceptance to the committed golden AL through SGLang's
-# simulated-acceptance path; the EVAL_ONLY accuracy run leaves it off and keeps
-# real verification. See the SGLANG_SIMULATE_ACC_* block.
-#
-# Required env vars:
-#   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
-#
-# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=hicache.
-
-source "$(dirname "$0")/../../benchmark_lib.sh"
-
-check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE
-
-SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-10}
+# Agentic trace replay benchmark for Qwen3.5 FP8 on H200 with SGLang MTP.
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL MODEL_PREFIX TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR \
+    DURATION EP_SIZE EVAL_ONLY SPEC_DECODING
+
+if ! require_agentic_kv_offload_backend hicache; then
+    echo "Error: the H200 MTP recipe requires KV_OFFLOADING=dram with HiCache" >&2
+    exit 1
+fi
+
+if [ "$TP" -ne 8 ] || [ "$EP_SIZE" -ne 1 ] || [ "$SPEC_DECODING" != "mtp" ]; then
+    echo "Error: the H200 MTP recipe requires TP=8, EP_SIZE=1, and SPEC_DECODING=mtp" >&2
+    exit 1
+fi
+
+if [ "$TOTAL_CPU_DRAM_GB" -lt 1200 ]; then
+    echo "Error: TP8 requires at least 1200 GB host DRAM; generated budget is ${TOTAL_CPU_DRAM_GB} GB" >&2
+    exit 1
+fi
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-# `hf download` creates the target dir if missing and is itself idempotent.
-# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
-# Either way, MODEL_PATH is what the server is launched with.
 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"
@@ -50,134 +40,117 @@ else
 fi
 nvidia-smi
 
-# ---- Resolve traces and install deps ----------------------------------------
-# Keep the 256k-capped with-subagents corpus the H100 Qwen3.5 AgentX recipe
-# uses (470 traces, max in+out <= 256k). The unfiltered corpus has requests up
-# to ~1M proxy tokens that the server would reject; H200's extra HBM raises the
-# context ceiling but not past 256k for this model at TP8.
 export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_with_subagents_256k
-
 resolve_trace_source
 install_agentic_deps
 
-# ---- Server config ----------------------------------------------------------
+agentic_pip_install --no-deps --force-reinstall flashinfer_python==0.6.15
+agentic_pip_install \
+    --no-deps --force-reinstall flashinfer-cubin==0.6.15 \
+    --index-url https://flashinfer.ai/whl
+agentic_pip_install \
+    --no-deps --force-reinstall flashinfer-jit-cache==0.6.15+cu130 \
+    --index-url https://flashinfer.ai/whl/cu130
+
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
-CACHE_ARGS=()
-if require_agentic_kv_offload_backend hicache; then
-    # HiCache extends RadixAttention, so do not pass --disable-radix-cache.
-    # Hybrid GDN/Mamba allocates one KV and one Mamba host pool per rank.
-    REQUESTED_HICACHE_TOTAL_GB="${HICACHE_TOTAL_CPU_DRAM_GB:-$TOTAL_CPU_DRAM_GB}"
-    if [ "$REQUESTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
-        echo "Error: requested HiCache pool ${REQUESTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
-        exit 1
-    fi
-    TOTAL_CPU_DRAM_GB="$REQUESTED_HICACHE_TOTAL_GB"
-    HICACHE_HOST_POOL_COUNT="${HICACHE_HOST_POOL_COUNT:-2}"
-    HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through_selective}"
-    MAX_HICACHE_SIZE_GB=$((TOTAL_CPU_DRAM_GB / TP / HICACHE_HOST_POOL_COUNT))
-    HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-$MAX_HICACHE_SIZE_GB}"
-    if [ "$HICACHE_SIZE_GB" -gt "$MAX_HICACHE_SIZE_GB" ]; then
-        echo "Error: HICACHE_SIZE_GB=$HICACHE_SIZE_GB exceeds configured per-pool limit $MAX_HICACHE_SIZE_GB" >&2
-        exit 1
-    fi
-    if [ "$HICACHE_SIZE_GB" -lt 1 ]; then
-        echo "Error: computed HICACHE_SIZE_GB=$HICACHE_SIZE_GB from TOTAL_CPU_DRAM_GB=$TOTAL_CPU_DRAM_GB, TP=$TP, HICACHE_HOST_POOL_COUNT=$HICACHE_HOST_POOL_COUNT" >&2
-        exit 1
-    fi
-    echo "HiCache CPU pool: ${HICACHE_SIZE_GB} GB per rank per host pool across TP=${TP}, host_pool_count=${HICACHE_HOST_POOL_COUNT}"
-    CACHE_ARGS=(
-        --page-size 64
-        --enable-hierarchical-cache
-        --hicache-size "$HICACHE_SIZE_GB"
-        --hicache-io-backend kernel
-        --hicache-mem-layout page_first
-        --hicache-write-policy "$HICACHE_WRITE_POLICY"
-    )
-fi
-
-echo "Starting SGLang server..."
+export TORCH_CUDA_ARCH_LIST=9.0
 export PYTHONNOUSERSITE=1
+export PYTHONUNBUFFERED=1
+export FLASHINFER_DISABLE_VERSION_CHECK=1
+export FLASHINFER_WORKSPACE_BASE=/tmp/flashinfer-cache
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
 export SGLANG_ENABLE_SPEC_V2=1
 
-# 3 speculative tokens per step (num-steps 3, eagle-topk 1, 4 draft tokens),
-# the same MTP shape as the fixed-seq-len Qwen3.5 recipes.
 SPEC_ARGS=(
-    --speculative-algorithm EAGLE
+    --speculative-algorithm NEXTN
     --speculative-num-steps 3
     --speculative-eagle-topk 1
     --speculative-num-draft-tokens 4
 )
 
-# AgentX pins acceptance to the committed golden AL so submissions are compared
-# on system performance at a fixed acceptance target rather than on draft-head
-# quality (golden_al_distribution/README.md). 3.39 is the Qwen3.5 MTP curve at
-# num_speculative_tokens=3, thinking_on (golden_al_distribution/qwen3.5_mtp.yaml)
-# -- the same value the GB300 Qwen3.5 AgentX srt-slurm recipes pin.
-# SGLANG_SIMULATE_ACC_TOKEN_MODE landed in SGLang v0.5.16, which is why this
-# recipe pins that image rather than the non-MTP agentic sibling's v0.5.12.
-#
-# EVAL_ONLY leaves simulated acceptance off: it commits drafted tokens
-# regardless of the target logits, so generated text is wrong and the eval would
-# score ~0.
-if [ "${EVAL_ONLY:-false}" != "true" ]; then
+if [ "$EVAL_ONLY" != "true" ]; then
     export SGLANG_SIMULATE_ACC_LEN=3.39
     export SGLANG_SIMULATE_ACC_METHOD=match-expected
     export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
 fi
 
-SGLANG_MULTI_TOKENIZER=/sgl-workspace/sglang/python/sglang/srt/managers/multi_tokenizer_mixin.py
-if ! sed -n '/elif isinstance(output, BatchStrOutput):/,/input_token_logprobs_val=_extract_field_by_index/p' "$SGLANG_MULTI_TOKENIZER" \
-    | grep -q 'cached_tokens_details=_extract_field_by_index'; then
-    sed -i '/elif isinstance(output, BatchStrOutput):/,/input_token_logprobs_val=_extract_field_by_index/ {
-        /cached_tokens=_extract_field_by_index(output, "cached_tokens", i),/a\
-            cached_tokens_details=_extract_field_by_index(\
-                output, "cached_tokens_details", i\
-            ),
-    }' "$SGLANG_MULTI_TOKENIZER"
-fi
+CACHE_ARGS=(
+    --page-size 64
+    --enable-hierarchical-cache
+    --hicache-ratio 0.9
+    --hicache-io-backend kernel
+    --hicache-mem-layout page_first_direct
+    --hicache-write-policy write_back
+)
 
-{ set +x; } 2>/dev/null
 SGLANG_CMD=(
     python3 -m sglang.launch_server
-    --model-path="$MODEL_PATH" --served-model-name="$MODEL"
-    --host=0.0.0.0
-    --port="$PORT"
-    --served-model-name "Qwen/Qwen3.5-397B-A17B-FP8"
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
     --trust-remote-code
-    --tensor-parallel-size="$TP"
-    --data-parallel-size=1
-    --expert-parallel-size="$EP_SIZE"
+    --tensor-parallel-size "$TP"
+    --data-parallel-size 1
+    --expert-parallel-size "$EP_SIZE"
     --quantization fp8
     --kv-cache-dtype fp8_e4m3
     --mamba-ssm-dtype bfloat16
+    --mamba-scheduler-strategy extra_buffer
+    --mamba-track-interval 8192
     --attention-backend flashinfer
-    --enable-flashinfer-allreduce-fusion
-    # --cuda-graph-max-bs "$CONC"
-    # --max-running-requests "$CONC"
-    # --max-prefill-tokens 8192
-    # --chunked-prefill-size 8192
-    --mem-fraction-static 0.8
+    --cuda-graph-max-bs 32
+    --max-running-requests 128
+    --max-prefill-tokens 16384
+    --chunked-prefill-size 16384
+    --mem-fraction-static 0.78
+    --max-mamba-cache-size 360
+    --allow-auto-truncate
     --stream-interval 50
-    --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
+    --scheduler-recv-interval 10
     --tokenizer-worker-num 6
-    --tokenizer-path "$MODEL"
+    --enable-cache-report
+    --enable-symm-mem
     --enable-metrics
-    "${SPEC_ARGS[@]}"
     "${CACHE_ARGS[@]}"
+    "${SPEC_ARGS[@]}"
 )
+
 printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
 printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
-"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=""
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "SGLang server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+{
+    echo "=== SGLANG_* env vars at launch ==="
+    env | grep -E '^SGLANG_' | sort
+    echo "==================================="
+} > "$SERVER_LOG"
+
+"${SGLANG_CMD[@]}" >> "$SERVER_LOG" 2>&1 &
 SERVER_PID=$!
 echo "Server PID: $SERVER_PID"
 
 wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
 
-if [ "${EVAL_ONLY}" = "true" ]; then
+if [ "$EVAL_ONLY" = "true" ]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --apply-chat-template"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
 fi
diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh
deleted file mode 100755
index f0fee4946..000000000
--- a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_sglang_mtp.sh
+++ /dev/null
@@ -1,160 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-set -x
-
-# Agentic trace replay for Qwen3.5 FP8 on H200 with native NEXTN MTP and
-# DRAM HiCache. Throughput pins the committed golden acceptance length; evals
-# retain target-model verification.
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-source "$SCRIPT_DIR/../../benchmark_lib.sh"
-
-check_env_vars \
-    MODEL MODEL_PREFIX TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR \
-    DURATION EP_SIZE EVAL_ONLY SPEC_DECODING
-
-if ! require_agentic_kv_offload_backend hicache; then
-    echo "Error: this H200 MTP recipe requires KV_OFFLOADING=dram with HiCache" >&2
-    exit 1
-fi
-
-if [ "$TP" -ne 8 ] || [ "$EP_SIZE" -ne 1 ]; then
-    echo "Error: this H200 MTP recipe requires TP=8 and EP_SIZE=1" >&2
-    exit 1
-fi
-
-if [ "$SPEC_DECODING" != "mtp" ]; then
-    echo "Error: this launcher requires SPEC_DECODING=mtp" >&2
-    exit 1
-fi
-
-if [ "$TOTAL_CPU_DRAM_GB" -lt 1200 ]; then
-    echo "Error: TP8 requires at least 1200 GB host DRAM; generated budget is ${TOTAL_CPU_DRAM_GB} GB" >&2
-    exit 1
-fi
-
-if [[ -n "${SLURM_JOB_ID:-}" ]]; then
-    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
-fi
-
-if [[ -n "${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="$MODEL"
-fi
-nvidia-smi
-
-export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_with_subagents_256k
-resolve_trace_source
-install_agentic_deps
-
-# Pin the published FlashInfer 0.6.15 CUDA 13 wheels used by this launcher.
-agentic_pip_install --no-deps --force-reinstall flashinfer_python==0.6.15
-agentic_pip_install \
-    --no-deps --force-reinstall flashinfer-cubin==0.6.15 \
-    --index-url https://flashinfer.ai/whl
-agentic_pip_install \
-    --no-deps --force-reinstall flashinfer-jit-cache==0.6.15+cu130 \
-    --index-url https://flashinfer.ai/whl/cu130
-
-SERVER_LOG="$RESULT_DIR/server.log"
-mkdir -p "$RESULT_DIR"
-
-export TORCH_CUDA_ARCH_LIST=9.0
-export PYTHONNOUSERSITE=1
-export PYTHONUNBUFFERED=1
-export FLASHINFER_DISABLE_VERSION_CHECK=1
-export FLASHINFER_WORKSPACE_BASE=/tmp/flashinfer-cache
-export SGL_ENABLE_JIT_DEEPGEMM=false
-export SGLANG_ENABLE_FLASHINFER_GEMM=true
-export SGLANG_ENABLE_SPEC_V2=1
-
-# Three speculative tokens per verification step. For agentic throughput, pin
-# the committed Qwen3.5 MTP golden AL (thinking_on, K=3); evals use real
-# verification because simulated acceptance does not produce accuracy data.
-if [ "$EVAL_ONLY" != "true" ]; then
-    export SGLANG_SIMULATE_ACC_LEN=3.39
-    export SGLANG_SIMULATE_ACC_METHOD=match-expected
-    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
-fi
-
-SGLANG_CMD=(
-    python3 -m sglang.launch_server
-    --model-path "$MODEL_PATH"
-    --served-model-name "$MODEL"
-    --host 0.0.0.0
-    --port "$PORT"
-    --trust-remote-code
-    --tensor-parallel-size "$TP"
-    --data-parallel-size 1
-    --expert-parallel-size "$EP_SIZE"
-    --quantization fp8
-    --kv-cache-dtype fp8_e4m3
-    --mamba-ssm-dtype bfloat16
-    --mamba-scheduler-strategy extra_buffer
-    --mamba-track-interval 8192
-    --mamba-max-states-per-path 1
-    --attention-backend flashinfer
-    --cuda-graph-max-bs 32
-    --max-running-requests 128
-    --max-prefill-tokens 16384
-    --chunked-prefill-size 16384
-    --mem-fraction-static 0.78
-    --max-mamba-cache-size 360
-    --allow-auto-truncate
-    --stream-interval 50
-    --scheduler-recv-interval 10
-    --tokenizer-worker-num 6
-    --enable-cache-report
-    --enable-symm-mem
-    --enable-metrics
-    --page-size 64
-    --enable-hierarchical-cache
-    --hicache-ratio 0.9
-    --hicache-io-backend kernel
-    --hicache-mem-layout page_first_direct
-    --hicache-write-policy write_back
-    --speculative-algorithm NEXTN
-    --speculative-num-steps 3
-    --speculative-eagle-topk 1
-    --speculative-num-draft-tokens 4
-)
-
-printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
-printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
-
-SERVER_PID=""
-cleanup_agentic_services() {
-    local exit_code=$?
-    trap - EXIT INT TERM
-    set +e
-    stop_background_process_tree "$SERVER_PID" "SGLang server" 60
-    exit "$exit_code"
-}
-trap cleanup_agentic_services EXIT
-trap 'exit 130' INT
-trap 'exit 143' TERM
-
-{
-    echo "=== SGLANG_* env vars at launch ==="
-    env | grep -E '^SGLANG_' | sort
-    echo "==================================="
-} > "$SERVER_LOG"
-
-"${SGLANG_CMD[@]}" >> "$SERVER_LOG" 2>&1 &
-SERVER_PID=$!
-echo "Server PID: $SERVER_PID"
-
-wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
-
-if [ "$EVAL_ONLY" = "true" ]; then
-    run_eval --port "$PORT"
-else
-    build_replay_cmd "$RESULT_DIR"
-    REPLAY_CMD+=" --apply-chat-template"
-    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
-    run_agentic_replay_and_write_outputs "$RESULT_DIR"
-fi
diff --git a/runners/launch_h200-dgxc-slurm.sh b/runners/launch_h200-dgxc-slurm.sh
index f904d8bc2..a0ff8b14b 100755
--- a/runners/launch_h200-dgxc-slurm.sh
+++ b/runners/launch_h200-dgxc-slurm.sh
@@ -350,7 +350,7 @@ EOF
     find . -name '.nfs*' -delete 2>/dev/null || true
 
 else
-    SQUASH_FILE="/data/gharunners/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    SQUASH_FILE="/data/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
 
     # Convert pyxis image format (nvcr.io#path) to docker format (nvcr.io/path) for enroot import
     DOCKER_IMAGE=$(echo "$IMAGE" | sed 's/#/\//g')

From 52074c3bc818fc96eedebcca32cd6c5ede662b0f Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 13 Aug 2026 14:44:36 -0700
Subject: [PATCH 5/6] fix(config): refresh sglang image
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

更新 SGLang 镜像。
---
 configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 40bce0904..f0f770e07 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7189,7 +7189,7 @@ qwen3.5-fp8-h200-sglang-agentic-mtp:
 # H200 AgentX MTP frontier with DRAM HiCache. This is intentionally an MTP-only
 # submission; the model's non-speculative AgentX arm is not included.
 qwen3.5-fp8-h200-sglang-agentic-hicache-mtp:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260716-b0b2dfbd
+  image: lmsysorg/sglang:nightly-dev-cu13-20260813-273d978b
   model: Qwen/Qwen3.5-397B-A17B-FP8
   model-prefix: qwen3.5
   runner: cluster:h200-dgxc

From 73faadad4b57badff2e7ef1363def15357c5fa94 Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 13 Aug 2026 15:04:01 -0700
Subject: [PATCH 6/6] fix(agentic): update flashinfer
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

更新 FlashInfer 版本。
---
 benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh | 6 +++---
 perf-changelog.yaml                                    | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh
index 1ed9ef9ca..b6f6d7eeb 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh
@@ -44,12 +44,12 @@ export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_with_subagents_256k
 resolve_trace_source
 install_agentic_deps
 
-agentic_pip_install --no-deps --force-reinstall flashinfer_python==0.6.15
+agentic_pip_install --no-deps --force-reinstall flashinfer_python==0.6.17
 agentic_pip_install \
-    --no-deps --force-reinstall flashinfer-cubin==0.6.15 \
+    --no-deps --force-reinstall flashinfer-cubin==0.6.17 \
     --index-url https://flashinfer.ai/whl
 agentic_pip_install \
-    --no-deps --force-reinstall flashinfer-jit-cache==0.6.15+cu130 \
+    --no-deps --force-reinstall flashinfer-jit-cache==0.6.17+cu130 \
     --index-url https://flashinfer.ai/whl/cu130
 
 SERVER_LOG="$RESULT_DIR/server.log"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index af1c6a759..1bd6cf219 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5960,5 +5960,5 @@
     - agentic-coding
   description:
     - "Add Qwen3.5 FP8 H200 SGLang AgentX MTP points with DRAM HiCache"
-    - "Use TP8/EP1 concurrency 2 through 24, published FlashInfer 0.6.15 CUDA 13 wheels, and golden MTP acceptance length 3.39"
+    - "Use TP8/EP1 concurrency 2 through 24, published FlashInfer 0.6.17 CUDA 13 wheels, and golden MTP acceptance length 3.39"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX