From 264f2aed68f3d1debe4e8f4faa3f23835fbca4a4 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sat, 18 Jul 2026 14:51:09 +0800
Subject: [PATCH 1/7] feat: add GLM-5.2 NVFP4 B300 SGLang single-node agentic
 benchmarks

Add glm5.2-fp4-b300-sglang-agentic from the SGLang cookbook B300 NVFP4
single-node recipes (STP only, no spec decoding):

- Low-latency arm: TP8 with fp8 KV cache and cutedsl bf16 GEMM backend,
  conc [1, 2, 4, 8, 16, 32]
- High-throughput arm: TP8/DP8 attention-DP behind sglang-router
  consistent hashing for session affinity, conc [48-512]
- Image: lmsysorg/sglang:v0.5.15.post1-cu130 (ships sglang-router 0.3.2)
- benchmark_lib.sh: glm5.2* added to the unfiltered agentic corpus
  branch (GLM-5.2 is a 1M-context model)

Both arms validated on b300-nv: LL conc16 8k1k 10,043 total tok/s
(TPOT 12.1 ms); HT conc256 34,429 total tok/s (TPOT 48.9 ms). Weights
pre-staged at /data/models/GLM-5.2-NVFP4.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 benchmarks/benchmark_lib.sh                   |   2 +-
 .../agentic/glm5.2_fp4_b300_sglang.sh         | 150 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  22 +++
 perf-changelog.yaml                           |  10 ++
 4 files changed, 183 insertions(+), 1 deletion(-)
 create mode 100755 benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh

diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index b4c947932b..96492072c8 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -1660,7 +1660,7 @@ resolve_trace_source() {
     # WEKA_LOADER_OVERRIDE.
     local default_loader
     case "${MODEL_PREFIX:-}" in
-        dsv4*|minimaxm3*)
+        dsv4*|glm5.2*|minimaxm3*)
             default_loader="semianalysis_cc_traces_weka_062126"
             ;;
         *)
diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
new file mode 100755
index 0000000000..3ab74818cd
--- /dev/null
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
@@ -0,0 +1,150 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# Agentic trace replay benchmark for GLM-5.2 NVFP4 on B300 using SGLang.
+#
+# Server flags follow the SGLang cookbook B300 NVFP4 single-node recipes
+# (https://docs.sglang.io/cookbook/autoregressive/GLM/GLM-5.2), STP only:
+# the cookbook's EAGLE MTP variants are intentionally not wired up yet.
+#   DP_ATTENTION=false -> low-latency arm (TP8, fp8 KV, cutedsl bf16 GEMM)
+#   DP_ATTENTION=true  -> high-throughput arm (TP8 + DP8 attention-DP)
+#
+# Required env vars:
+#   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR, DURATION,
+#   EP_SIZE, DP_ATTENTION
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+if [[ "$KV_OFFLOADING" != "none" ]]; then
+    echo "Error: KV_OFFLOADING=$KV_OFFLOADING is not supported by this recipe" >&2
+    exit 1
+fi
+
+if [[ -n "${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 "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+nvidia-smi
+
+resolve_trace_source
+install_agentic_deps
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+# With attention-DP, front the DP ranks with sglang-router using consistent
+# hashing on the AIPerf correlation id so multi-turn sessions stay on the DP
+# rank that holds their radix-cache prefix.
+USE_SGLANG_ROUTER=false
+SGLANG_BACKEND_PORT="$PORT"
+ROUTER_LOG="$RESULT_DIR/router.log"
+if [ "$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))
+fi
+
+PARALLEL_ARGS=(--tp "$TP" --ep-size "$EP_SIZE")
+if [ "$DP_ATTENTION" = "true" ]; then
+    PARALLEL_ARGS+=(
+        --dp "$TP"
+        --enable-dp-attention
+        --tokenizer-worker-num "$TP"
+        --dist-init-addr "127.0.0.1:$((PORT + 2000))"
+    )
+else
+    # Cookbook low-latency levers; the DP-attention cell omits them.
+    PARALLEL_ARGS+=(
+        --kv-cache-dtype fp8_e4m3
+        --bf16-gemm-backend cutedsl
+        --max-prefill-tokens 8192
+    )
+fi
+
+# AgentX concurrency counts live session trees, not individual requests.
+# Allow subagent fan-out to exceed CONC without clipping request bursts.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+GRAPH_ARGS=()
+if [ "$DP_ATTENTION" != "true" ]; then
+    # Cookbook low-latency captures graphs up to its request cap; the
+    # DP-attention cell leaves the CUDA-graph batch list at SGLang defaults.
+    CUDA_GRAPH_MAX_BS=$MAX_RUNNING_REQUESTS
+    [ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+    GRAPH_ARGS=(--cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS")
+fi
+
+export PYTHONNOUSERSITE=1
+export TORCH_CUDA_ARCH_LIST=10.0
+# Agentic warmup dispatches hundreds of large prompts at once; allow up to
+# 15 minutes of TCP progress before AIPerf declares a connection dead.
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$SGLANG_BACKEND_PORT"
+    --trust-remote-code
+    "${PARALLEL_ARGS[@]}"
+    --quantization modelopt_fp4
+    --chunked-prefill-size 8192
+    --mem-fraction-static 0.85
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    "${GRAPH_ARGS[@]}"
+    --watchdog-timeout 1800
+    --enable-metrics
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+
+echo "Starting SGLang server for B300..."
+"${SGLANG_CMD[@]}" > "$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 $TP DP ranks..."
+    python3 -m sglang_router.launch_router \
+        --worker-urls "http://localhost:$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 [ "${EVAL_ONLY}" = "true" ]; then
+    run_eval --port "$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
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 7d7af13236..853a375295 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -9891,3 +9891,25 @@ qwen3.5-fp8-gb300-dynamo-sglang:
           tp: 16
           ep: 16
           dp-attn: true
+
+# GLM-5.2 B300 NVFP4 AgentX frontier from the SGLang cookbook B300 NVFP4
+# single-node recipes (https://docs.sglang.io/cookbook/autoregressive/GLM/GLM-5.2),
+# STP only (the cookbook's EAGLE MTP variants are deferred). The TP8 arm is the
+# cookbook low-latency recipe (fp8 KV, cutedsl bf16 GEMM) and covers the
+# interactivity end; the TP8/DP8 attention-DP arm is the cookbook
+# high-throughput recipe behind sglang-router consistent hashing for session
+# affinity. Conc lists are disjoint between arms so exp-names stay unique.
+glm5.2-fp4-b300-sglang-agentic:
+  image: lmsysorg/sglang:v0.5.15.post1-cu130
+  model: nvidia/GLM-5.2-NVFP4
+  model-prefix: glm5.2
+  runner: cluster:b300-nv
+  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, 16, 32] }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [48, 64, 96, 128, 192, 256, 512], router: { name: sglang-router, version: "0.3.2" } }
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b358f10bf1..f1607a872b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4933,3 +4933,13 @@
   description:
     - "Add MiniMax M3 NVFP4 B300 Dynamo-vLLM disaggregated EAGLE3 recipes"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2182
+
+- config-keys:
+    - glm5.2-fp4-b300-sglang-agentic
+  description:
+    - "Add GLM-5.2 NVFP4 B300 SGLang single-node agentic benchmarks from the SGLang cookbook B300 NVFP4 recipes (STP only, no spec decoding)"
+    - "Image: lmsysorg/sglang:v0.5.15.post1-cu130; model nvidia/GLM-5.2-NVFP4"
+    - "Low-latency arm: TP8 (fp8 KV cache, cutedsl bf16 GEMM backend) at conc [1, 2, 4, 8, 16, 32]"
+    - "High-throughput arm: TP8/DP8 attention-DP behind sglang-router consistent hashing at conc [48, 64, 96, 128, 192, 256, 512]"
+    - "New 1M-context model prefix glm5.2 added to the unfiltered agentic corpus branch in benchmark_lib.sh"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From e46fae94ce6fe859d52c502dd76d116ab3d7c087 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sat, 18 Jul 2026 14:52:07 +0800
Subject: [PATCH 2/7] chore: fill perf-changelog pr-link for glm5.2 b300
 agentic entry

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f1607a872b..279b06c9c9 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4942,4 +4942,4 @@
     - "Low-latency arm: TP8 (fp8 KV cache, cutedsl bf16 GEMM backend) at conc [1, 2, 4, 8, 16, 32]"
     - "High-throughput arm: TP8/DP8 attention-DP behind sglang-router consistent hashing at conc [48, 64, 96, 128, 192, 256, 512]"
     - "New 1M-context model prefix glm5.2 added to the unfiltered agentic corpus branch in benchmark_lib.sh"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2268

From 8933d94c85229820857e0fe65c364737d8d97972 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sat, 18 Jul 2026 15:20:44 +0800
Subject: [PATCH 3/7] fix: outlast AIPerf per-session keep-alive in glm5.2 b300
 agentic recipe

One warmup request hit ECONNRESET (uvicorn closes idle connections after
SGLANG_TIMEOUT_KEEP_ALIVE=5s while AIPerf reuses one pooled connection per
session with a 300s client keep-alive across inter-turn gaps of up to 10s),
and any terminal warmup failure aborts the AIPerf run by design. Set
SGLANG_TIMEOUT_KEEP_ALIVE=900 so the server outlasts the client pool.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
index 3ab74818cd..8a274a86e3 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
@@ -93,6 +93,12 @@ export TORCH_CUDA_ARCH_LIST=10.0
 # 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 (capped at 10s) can reuse a socket exactly as the
+# server closes it -> ECONNRESET -> terminal warmup failure. Outlast the
+# client pool so the race cannot occur.
+export SGLANG_TIMEOUT_KEEP_ALIVE=900
 
 SGLANG_CMD=(
     python3 -m sglang.launch_server

From 3643eb214d2e1d0d5dfd0d26d707ca9767bd8936 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sat, 18 Jul 2026 16:27:34 +0800
Subject: [PATCH 4/7] fix: widen DP-attention chunked prefill for agentic
 corpus in glm5.2 b300 recipe

The cookbook HT cell's --chunked-prefill-size 8192 is a whole-engine budget:
under dp8 each rank prefills 1,024 tokens/step, which starves prefill on the
1M-context agentic corpus. A conc-256 warmup timed out after AIPerf's 1800s
drain grace period with 15 giant sessions still prefilling while KV usage sat
at ~0.01 (prefill-rate-bound, not memory-bound). Use the cookbook's own dp8
lever from the B200 cells: 32768 total = ~4096 tokens/rank/step. The TP arm
keeps 8192 (full budget per step, passed warmup fine).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../single_node/agentic/glm5.2_fp4_b300_sglang.sh      | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
index 8a274a86e3..480e435a43 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
@@ -60,7 +60,15 @@ if [ "$DP_ATTENTION" = "true" ]; then
 fi
 
 PARALLEL_ARGS=(--tp "$TP" --ep-size "$EP_SIZE")
+CHUNKED_PREFILL_SIZE=8192
 if [ "$DP_ATTENTION" = "true" ]; then
+    # chunked-prefill-size is a whole-engine budget split across DP ranks:
+    # the cookbook HT cell's 8192 becomes 1,024 tokens/rank/step under dp8,
+    # which starves prefill on the 1M-context agentic corpus (observed: a
+    # conc-256 warmup could not drain within AIPerf's 1800s grace period
+    # while KV usage sat at ~0.01). Use the cookbook's own dp8 lever from
+    # the B200 cells (32768 = ~4096/rank).
+    CHUNKED_PREFILL_SIZE=32768
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention
@@ -109,7 +117,7 @@ SGLANG_CMD=(
     --trust-remote-code
     "${PARALLEL_ARGS[@]}"
     --quantization modelopt_fp4
-    --chunked-prefill-size 8192
+    --chunked-prefill-size "$CHUNKED_PREFILL_SIZE"
     --mem-fraction-static 0.85
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     "${GRAPH_ARGS[@]}"

From 21d9de2dd7cefb14d04cf6ae3b9abe2892cf5597 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sat, 18 Jul 2026 17:12:32 +0800
Subject: [PATCH 5/7] fix: enable glm47 tool-call and glm45 reasoning parsers
 for SWE-bench evals

The agentic SWE-bench eval drives mini-swe-agent through native tool
calling. Without --tool-call-parser, GLM-5.2's tool-call tokens stay as
raw text in message.content, every instance dies with RepeatedFormatError
("No tool calls found in the response"), 287/300 patches come back empty,
and swebench_lite scores 0.0000 against the 0.50 default threshold.

Per the GLM-5.2 cookbook: the model emits the GLM-4.7-style
<tool_call>/<arg_key>/<arg_value> format, so it needs the glm47 parser
(glm45 does not parse it); the glm45 reasoning parser separates hybrid
thinking into reasoning_content. Neither flag affects trace-replay
throughput (pre-canned replay discards live responses).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../single_node/agentic/glm5.2_fp4_b300_sglang.sh      | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
index 480e435a43..175a0316fb 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
@@ -117,6 +117,16 @@ SGLANG_CMD=(
     --trust-remote-code
     "${PARALLEL_ARGS[@]}"
     --quantization modelopt_fp4
+    # GLM-5.2 emits the GLM-4.7-style <tool_call>/<arg_key>/<arg_value> format;
+    # the glm47 parser is required for structured message.tool_calls (glm45
+    # leaves calls as raw text). Without it the SWE-bench mini-swe-agent eval
+    # dies with RepeatedFormatError ("No tool calls found in the response") on
+    # every instance and scores 0. Reasoning parser keeps hybrid-thinking
+    # output in reasoning_content instead of polluting content. Neither flag
+    # affects trace-replay throughput (pre-canned replay discards live
+    # responses).
+    --tool-call-parser glm47
+    --reasoning-parser glm45
     --chunked-prefill-size "$CHUNKED_PREFILL_SIZE"
     --mem-fraction-static 0.85
     --max-running-requests "$MAX_RUNNING_REQUESTS"

From c21ff08ba79636353d43d2280f463fbcdaf3a257 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sat, 18 Jul 2026 18:21:08 +0800
Subject: [PATCH 6/7] fix: overridable agentic warmup grace; 3600s for glm5.2
 b300 DPA arm

The DPA conc-512 warmup drain converges healthily (~0.45 req/s, zero
errors) but needs ~2500s end to end - the shared 1800s grace cuts it off
with ~300 requests in flight. Make the grace period overridable via
AGENTIC_WARMUP_GRACE_PERIOD (default unchanged at 1800) and set 3600 for
the glm5.2 DP-attention arm. Grace is a maximum wait, not a fixed sleep,
so lower-conc points are unaffected.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 benchmarks/benchmark_lib.sh                              | 5 ++++-
 benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh | 5 +++++
 2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index 96492072c8..2ba3d8923d 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -1780,8 +1780,11 @@ build_replay_cmd() {
     # only after those requests drain and resumes from the resulting live state.
     REPLAY_CMD+=" --agentic-cache-warmup-duration 600"
     # Give long-context warmup requests up to 30 minutes to drain before
+    # declaring warmup failed. Recipes whose saturation arms carry a larger
+    # in-flight working set may override via AGENTIC_WARMUP_GRACE_PERIOD
+    # (grace is a maximum wait, not a fixed sleep — drain exits when done).
     # cancelling any remaining requests and starting profiling.
-    REPLAY_CMD+=" --warmup-grace-period 1800"
+    REPLAY_CMD+=" --warmup-grace-period ${AGENTIC_WARMUP_GRACE_PERIOD:-1800}"
     # Use server-reported usage fields (prompt_tokens / completion_tokens) for
     # ISL/OSL instead of client-side tokenizer.encode(). Auto-enables
     # stream_options.include_usage on the OpenAI chat endpoint. Skips the
diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
index 175a0316fb..d570fbc972 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
@@ -69,6 +69,11 @@ if [ "$DP_ATTENTION" = "true" ]; then
     # while KV usage sat at ~0.01). Use the cookbook's own dp8 lever from
     # the B200 cells (32768 = ~4096/rank).
     CHUNKED_PREFILL_SIZE=32768
+    # At conc 512 the saturation working set outlives the default 1800s
+    # warmup drain grace: the drain converges healthily (~0.45 req/s, zero
+    # errors) but needs ~2500s end to end. 3600 is a maximum wait, not a
+    # fixed sleep — lower-conc DPA points still finish as fast as they drain.
+    export AGENTIC_WARMUP_GRACE_PERIOD=3600
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention

From 53fa0bb01b61906cbfe2e5a3c5da68912ff0ac24 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sun, 19 Jul 2026 00:04:17 +0800
Subject: [PATCH 7/7] fix: double SWE-bench step budget for glm5.2 b300 agentic
 evals

With the glm47 parser fix in, the agentic SWE-bench eval produces real
work: 10/11 submitted trajectories resolved (0.91 precision), but 12/23
exited LimitsExceeded before submitting - GLM-5.2's chat template
defaults to reasoning_effort=Max when the client passes no
chat_template_kwargs, and the heavy thinking exhausts the default
75-step budget. Score landed at 0.4348 vs the 0.50 floor. Export
SWEBENCH_AGENT_STEP_LIMIT=150 in the recipe's eval path only; the
shared default stays 75.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
index d570fbc972..9f4b9e0871 100755
--- a/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
@@ -171,6 +171,13 @@ if [ "$USE_SGLANG_ROUTER" = "true" ]; then
 fi
 
 if [ "${EVAL_ONLY}" = "true" ]; then
+    # GLM-5.2's chat template defaults to reasoning_effort=Max when the
+    # client passes no chat_template_kwargs (mini-swe-agent doesn't), and the
+    # heavy thinking burns the default 75-step budget: on the 23-instance
+    # slice, 12/23 trajectories exited LimitsExceeded unsubmitted while 10 of
+    # the 11 that submitted resolved. Double the step budget for this recipe;
+    # other recipes keep the shared 75 default.
+    export SWEBENCH_AGENT_STEP_LIMIT=150
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"