From 69acc81f12032a1206007947da80a5e2508e54ca Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 13:49:14 -0500
Subject: [PATCH 01/18] feat(b300): add DeepSeek V4.1 SGLang nightly AgentX
 recipe

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh | 134 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  17 +++
 perf-changelog.yaml                           |   9 ++
 runners/launch_b300-dsxe.sh                   |   2 +-
 4 files changed, 161 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
new file mode 100644
index 0000000000..44665f8cd9
--- /dev/null
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -0,0 +1,134 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# DeepSeek-V4.1-Flash AgentX on B300 with SGLang native DSpark, following the
+# cookbook's verified Blackwell TP4/EP4 low-latency cell. The KV cache is GPU-resident.
+# https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4_1
+source "$(dirname "$0")/../../benchmark_lib.sh"
+check_env_vars MODEL TP EP_SIZE CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+check_env_vars EVAL_ONLY
+require_agentic_kv_offload_none
+export GPU_COUNT="$TP"
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+# Complete/resume partial downloads instead of trusting nonempty directories.
+if [[ -n "${MODEL_PATH:-}" && "$MODEL_PATH" != "$MODEL" ]]; then
+    hf download "$MODEL" --local-dir "$MODEL_PATH"
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+nvidia-smi
+resolve_trace_source
+install_agentic_deps
+mkdir -p "$RESULT_DIR"
+SERVER_LOG="$RESULT_DIR/server.log"
+export PYTHONNOUSERSITE=1
+export PYTHONUNBUFFERED=1
+
+# Agentic warmup dispatches hundreds of large prompts at once and SGLang's
+# tokenizer can leave bytes unacknowledged past AIPerf's default 30 s
+# TCP_USER_TIMEOUT, so Linux aborts live localhost connections.
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+# Outlast AIPerf's pooled connections so an inter-turn idle gap cannot race
+# Uvicorn's five-second keep-alive closure.
+export SGLANG_TIMEOUT_KEEP_ALIVE=900
+
+# AgentX measures the thinking-on regime, which is also the committed golden-AL
+# curve. SGLang ships thinking off by default for this model.
+export SGLANG_DEFAULT_THINKING=1
+export SGLANG_DSV41_REASONING_EFFORT=high
+
+# One shared host copy of the two fp8 Engram tables instead of a row-sharded
+# copy per rank: the SGLang analogue of the vLLM arm's Engram CPU offload. It
+# frees ~46 GiB of HBM per GPU for the 1M-context prefill working set and the
+# KV pool, and output is bitwise unchanged (cookbook). The first sweep ran
+# with the tables on GPU and the server died on the first long AgentX prompts
+# (run 35304517453: c4 came up, then the server exited on the first two warmup prompts).
+export SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1
+
+# AgentX concurrency counts live session trees, not individual requests.
+# Allow subagent fan-out to exceed CONC without clipping request bursts, but
+# never let the pool exceed the decode graph batch: a DSpark verify step for a
+# batch above the captured 64 runs eagerly and allocates its attention
+# workspace on the fly, which OOMed the H200 eval at 128 running requests
+# (6.4 GiB allocation with 2 GiB free, run 35306704553). Batches within the
+# graph tier reuse the capture-time workspace instead.
+CUDA_GRAPH_MAX_BS=64
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+if (( MAX_RUNNING_REQUESTS > CUDA_GRAPH_MAX_BS )); then
+    MAX_RUNNING_REQUESTS=$CUDA_GRAPH_MAX_BS
+fi
+
+# Saturation arms carry a larger in-flight working set than the 30-minute
+# default warmup drain allows.
+if (( CONC >= 32 )); then
+    export AGENTIC_WARMUP_GRACE_PERIOD=3600
+fi
+
+# Pyxis shares the host network; port 8888 can already belong to a host service.
+select_available_server_port
+export AIPERF_SERVER_URL="http://localhost:${PORT}"
+export AIPERF_SERVER_METRICS_URLS="${AIPERF_SERVER_URL}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="sglang:"
+echo "Using SGLang endpoint ${AIPERF_SERVER_URL}"
+
+# DSpark is the checkpoint's own bundled draft: no EAGLE/MTP path and no
+# --speculative-num-steps knob; the block size is the only tunable. Golden AL:
+# golden_al_distribution/dsv41flash_dspark.yaml, thinking_on, five draft tokens.
+# Throughput fixes acceptance to AL 3.51; accuracy evals keep real verification.
+DSPARK_BLOCK_SIZE=5
+DSV41_GOLDEN_AL=3.51
+if [[ "${EVAL_ONLY}" != true ]]; then
+    export SGLANG_SIMULATE_ACC_LEN="$DSV41_GOLDEN_AL"
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+echo "DSpark block size: $DSPARK_BLOCK_SIZE, golden AL=$DSV41_GOLDEN_AL"
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH" --served-model-name "$MODEL"
+    --host 0.0.0.0 --port "$PORT"
+    --trust-remote-code
+    --tp "$TP" --ep-size "$EP_SIZE"
+    # Backends resolve automatically (dsv4 / flashinfer_mxfp4 / flashinfer_cutedsl
+    # on Blackwell); the cookbook warns that overriding them costs decode speed.
+    # 0.70 rather than the cookbook's 0.8, and a bounded prefill chunk: the
+    # sparse-attention indexer and DSpark prefill buffers scale with the chunk
+    # times the 1M context, and the default 16384 chunk exhausted HBM on the
+    # first 66k-99k-token AgentX prompts.
+    --mem-fraction-static 0.70
+    --chunked-prefill-size 4096
+    --speculative-algorithm DSPARK
+    --speculative-dspark-block-size "$DSPARK_BLOCK_SIZE"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
+    --reasoning-parser auto
+    --tool-call-parser auto
+    # Draft-token forward passes under long-context agentic load block the
+    # scheduler long enough to trip the 1800 s default watchdog mid-warmup.
+    --watchdog-timeout 3600
+    --enable-metrics
+)
+write_command "$RESULT_DIR/sglang_command.txt" "${SGLANG_CMD[@]}"
+{
+    echo "=== SGLANG_* env vars at launch ==="
+    env | grep -E '^SGLANG_' | sort
+    echo "==================================="
+} | tee "$SERVER_LOG"
+"${SGLANG_CMD[@]}" >> "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+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+=" --server-metrics ${AIPERF_SERVER_METRICS_URLS}"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 1215a1a5f7..cefe990429 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8394,3 +8394,20 @@ qwen3.5-fp8-b200-dynamo-sglang-agentic-disagg-mtp:
           dp-attn: false
         kv-offload-backend:
           name: hicache
+
+# Official 2026-09-21 CUDA 13 nightly, amd64 manifest pinned for B300.
+# Latest multiarch index: sha256:987c7e4bd26918647211a5dcad72a2bdf2a5f394ac1469eff730e7517fc139be.
+# Uses the cookbook's automatic backends and the existing B200 AgentX memory bounds.
+dsv41flash-fp4-b300-sglang-agentic-dspark:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260921-0f6761b5@sha256:18aed2cd75ed45932a9d409e167e75191e0596a47925de1f176d1cc03036cf8c
+  model: deepseek-ai/DeepSeek-V4.1-Flash
+  model-prefix: dsv41flash
+  runner: cluster:b300-dsxe
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9419a68f0b..795cae837c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8531,3 +8531,12 @@
     - "为 GB300 vLLM DeepSeek-V4.1-Flash AgentX 配方在现有 TP4 臂旁新增 TP2 臂，Engram 表继续通过 --engram-config cpu_offload 放在固定页主机 DRAM；每张 277 GiB GPU 的权重升至约 175 GiB"
     - "在 dsv41flash_fp4_vllm_mtp.sh 中将 B200 TP2 的上限推广到所有 TP2 臂：--max-num-batched-tokens 4096（上游 16384 时 indexer 的 [batched-tokens, 1M] fp8 缓冲区达 32 GiB）、--max-num-seqs 为并发的两倍（16-256）、CUDA graph 捕获上限 512，为每张 GPU 留出约 36 GiB KV；TP4 与 TP8 臂沿用上游默认值"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3321
+
+- config-keys:
+    - dsv41flash-fp4-b300-sglang-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals."
+    - "Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD
diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index d49b28b417..87581af4ec 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -336,7 +336,7 @@ else
     fi
 
     # Keep all new AgentX runtime directories outside /workspace.
-    if [[ "$MODEL_PREFIX" == "dsv41flash" && "$FRAMEWORK" == "vllm" ]]; then
+    if [[ "$MODEL_PREFIX" == "dsv41flash" && ( "$FRAMEWORK" == "vllm" || "$FRAMEWORK" == "sglang" ) ]]; then
         CONTAINER_MOUNT_DIR=/ix
         export INFMAX_CONTAINER_WORKSPACE=/ix
         export RESULT_DIR=/ix/results

From 7b29802813cc7a6e894ad4da6c157a91f458d27c Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 13:49:48 -0500
Subject: [PATCH 02/18] fix(b300): record recipe pull request provenance

---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 795cae837c..bf3bede6be 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8539,4 +8539,4 @@
   description:
     - "Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals."
     - "Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From 96daee34ce9343791884ccab56ecb3810eeabec2 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 14:11:56 -0500
Subject: [PATCH 03/18] feat(b300): add non-speculative SGLang AgentX arm

---
 .../agentic/dsv41flash_fp4_b300_sglang.sh     |  1 +
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh | 52 +++++++++++--------
 configs/nvidia-master.yaml                    | 15 ++++++
 perf-changelog.yaml                           |  9 ++++
 4 files changed, 54 insertions(+), 23 deletions(-)
 create mode 120000 benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh
new file mode 120000
index 0000000000..03357261bc
--- /dev/null
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh
@@ -0,0 +1 @@
+dsv41flash_fp4_b300_sglang_mtp.sh
\ No newline at end of file
diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index 44665f8cd9..c341b74672 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -1,12 +1,12 @@
 #!/usr/bin/env bash
 set -eo pipefail
 
-# DeepSeek-V4.1-Flash AgentX on B300 with SGLang native DSpark, following the
-# cookbook's verified Blackwell TP4/EP4 low-latency cell. The KV cache is GPU-resident.
+# DeepSeek-V4.1-Flash AgentX on B300 with SGLang, supporting STP and DSpark.
+# The KV cache is GPU-resident; caller SPEC_DECODING selects the serving mode.
 # https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4_1
 source "$(dirname "$0")/../../benchmark_lib.sh"
 check_env_vars MODEL TP EP_SIZE CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
-check_env_vars EVAL_ONLY
+check_env_vars EVAL_ONLY SPEC_DECODING
 require_agentic_kv_offload_none
 export GPU_COUNT="$TP"
 
@@ -43,13 +43,11 @@ export SGLANG_TIMEOUT_KEEP_ALIVE=900
 export SGLANG_DEFAULT_THINKING=1
 export SGLANG_DSV41_REASONING_EFFORT=high
 
-# One shared host copy of the two fp8 Engram tables instead of a row-sharded
-# copy per rank: the SGLang analogue of the vLLM arm's Engram CPU offload. It
-# frees ~46 GiB of HBM per GPU for the 1M-context prefill working set and the
-# KV pool, and output is bitwise unchanged (cookbook). The first sweep ran
-# with the tables on GPU and the server died on the first long AgentX prompts
-# (run 35304517453: c4 came up, then the server exited on the first two warmup prompts).
+# Keep Engram tables in host RAM to make room for long-context AgentX KV.
+# Per-rank anonymous mappings can use THP without requiring shared-memory THP
+# or host sysctl changes. The table payload remains native FP8.
 export SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1
+export SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT=per_rank
 
 # AgentX concurrency counts live session trees, not individual requests.
 # Allow subagent fan-out to exceed CONC without clipping request bursts, but
@@ -77,18 +75,27 @@ export AIPERF_SERVER_METRICS_URLS="${AIPERF_SERVER_URL}/metrics"
 export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="sglang:"
 echo "Using SGLang endpoint ${AIPERF_SERVER_URL}"
 
-# DSpark is the checkpoint's own bundled draft: no EAGLE/MTP path and no
-# --speculative-num-steps knob; the block size is the only tunable. Golden AL:
-# golden_al_distribution/dsv41flash_dspark.yaml, thinking_on, five draft tokens.
-# Throughput fixes acceptance to AL 3.51; accuracy evals keep real verification.
-DSPARK_BLOCK_SIZE=5
-DSV41_GOLDEN_AL=3.51
-if [[ "${EVAL_ONLY}" != true ]]; then
-    export SGLANG_SIMULATE_ACC_LEN="$DSV41_GOLDEN_AL"
-    export SGLANG_SIMULATE_ACC_METHOD=match-expected
-    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
-fi
-echo "DSpark block size: $DSPARK_BLOCK_SIZE, golden AL=$DSV41_GOLDEN_AL"
+# STP and accuracy evaluations must never inherit synthetic acceptance.
+unset SGLANG_SIMULATE_ACC_LEN SGLANG_SIMULATE_ACC_METHOD SGLANG_SIMULATE_ACC_TOKEN_MODE
+SPECULATIVE_ARGS=()
+case "$SPEC_DECODING" in
+    none)
+        echo "Non-speculative decoding; synthetic acceptance disabled"
+        ;;
+    mtp)
+        # Existing measured curve: dsv41flash_dspark.yaml, thinking_on, K5.
+        SPECULATIVE_ARGS=(--speculative-algorithm DSPARK --speculative-dspark-block-size 5)
+        if [[ "$EVAL_ONLY" != true ]]; then
+            export SGLANG_SIMULATE_ACC_LEN=3.51
+            export SGLANG_SIMULATE_ACC_METHOD=match-expected
+            export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+        fi
+        ;;
+    *)
+        echo "Unsupported SPEC_DECODING: $SPEC_DECODING" >&2
+        exit 1
+        ;;
+esac
 
 SGLANG_CMD=(
     python3 -m sglang.launch_server
@@ -104,8 +111,7 @@ SGLANG_CMD=(
     # first 66k-99k-token AgentX prompts.
     --mem-fraction-static 0.70
     --chunked-prefill-size 4096
-    --speculative-algorithm DSPARK
-    --speculative-dspark-block-size "$DSPARK_BLOCK_SIZE"
+    "${SPECULATIVE_ARGS[@]}"
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
     --reasoning-parser auto
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index cefe990429..cf0e84fcb8 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8411,3 +8411,18 @@ dsv41flash-fp4-b300-sglang-agentic-dspark:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
+
+# Non-speculative cookbook high-throughput path; no draft or synthetic acceptance.
+dsv41flash-fp4-b300-sglang-agentic:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260921-0f6761b5@sha256:18aed2cd75ed45932a9d409e167e75191e0596a47925de1f176d1cc03036cf8c
+  model: deepseek-ai/DeepSeek-V4.1-Flash
+  model-prefix: dsv41flash
+  runner: cluster:b300-dsxe
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: none, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2614f69540..e7829e42fc 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8551,3 +8551,12 @@
     - "Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals."
     - "Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342
+
+- config-keys:
+    - dsv41flash-fp4-b300-sglang-agentic
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, while native DSpark precision qualification remains blocked."
+    - "Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From b69b0641242400abef4027e478b3ea6a2e5a186d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 14:19:46 -0500
Subject: [PATCH 04/18] docs: consolidate performance changelog into one PR
 entry

---
 perf-changelog.yaml | 20 +++++++-------------
 1 file changed, 7 insertions(+), 13 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e7829e42fc..a990734188 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8544,19 +8544,13 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3321
 
 - config-keys:
-    - dsv41flash-fp4-b300-sglang-agentic-dspark
+  - dsv41flash-fp4-b300-sglang-agentic-dspark
+  - dsv41flash-fp4-b300-sglang-agentic
   scenario-type:
-    - agentic-coding
-  description:
-    - "Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals."
-    - "Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342
-
-- config-keys:
-    - dsv41flash-fp4-b300-sglang-agentic
-  scenario-type:
-    - agentic-coding
+  - agentic-coding
   description:
-    - "Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, while native DSpark precision qualification remains blocked."
-    - "Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection."
+  - Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals.
+  - Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix.
+  - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, while native DSpark precision qualification remains blocked.
+  - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From 0661c240faaa2f9a2fd00048fe42164c87860d07 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 14:40:05 -0500
Subject: [PATCH 05/18] fix(b300): preserve native DSpark FP8 draft projections

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh | 12 ++-
 .../agentic/patch_sglang_dsv41_native_wo_a.py | 66 ++++++++++++++++
 .../patches/sglang_dsv41_native_wo_a.patch    | 76 +++++++++++++++++++
 docs/configuration-procedures.md              |  8 ++
 docs/configuration-procedures_zh.md           |  7 ++
 perf-changelog.yaml                           |  1 +
 6 files changed, 169 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
 create mode 100644 benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index c341b74672..1631e5539a 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -19,7 +19,8 @@ if [[ -n "${MODEL_PATH:-}" && "$MODEL_PATH" != "$MODEL" ]]; then
     hf download "$MODEL" --local-dir "$MODEL_PATH"
 else
     hf download "$MODEL"
-    export MODEL_PATH="$MODEL"
+    MODEL_PATH=$(python3 -c 'from huggingface_hub import snapshot_download; import sys; print(snapshot_download(repo_id=sys.argv[1], local_files_only=True))' "$MODEL")
+    export MODEL_PATH
 fi
 
 nvidia-smi
@@ -30,6 +31,12 @@ SERVER_LOG="$RESULT_DIR/server.log"
 export PYTHONNOUSERSITE=1
 export PYTHONUNBUFFERED=1
 
+# Preserve native draft FP8 weights with the exact-nightly, hash-guarded fix.
+if [[ "$SPEC_DECODING" == mtp ]]; then
+    python3 "$(dirname "$0")/patch_sglang_dsv41_native_wo_a.py" \
+        | tee "$RESULT_DIR/native_wo_a_patch.txt"
+fi
+
 # Agentic warmup dispatches hundreds of large prompts at once and SGLang's
 # tokenizer can leave bytes unacknowledged past AIPerf's default 30 s
 # TCP_USER_TIMEOUT, so Linux aborts live localhost connections.
@@ -123,6 +130,9 @@ SGLANG_CMD=(
 )
 write_command "$RESULT_DIR/sglang_command.txt" "${SGLANG_CMD[@]}"
 {
+    if [[ "$SPEC_DECODING" == mtp ]]; then
+        cat "$RESULT_DIR/native_wo_a_patch.txt"
+    fi
     echo "=== SGLANG_* env vars at launch ==="
     env | grep -E '^SGLANG_' | sort
     echo "==================================="
diff --git a/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py b/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
new file mode 100644
index 0000000000..b462590fcd
--- /dev/null
+++ b/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
@@ -0,0 +1,66 @@
+"""Preserve native V4.1 DSpark FP8 WO_A weights in the official SGLang nightly.
+
+Applies only to the exact upstream source shipped at 0f6761b54facebb47f2068f87ecccd8f14da3a0e.
+The dense FP8 linear path supports the checkpoint's 32x32 scales; the upstream
+grouped projection instead dequantizes these draft weights permanently to BF16.
+"""
+
+import hashlib
+import importlib.util
+import subprocess
+from pathlib import Path
+
+BASE_SHA256 = "61dc79f075c9e1e5a68a466de5eb95a85ff1fa2e5a9f4d66c2fa69e956fd7b61"
+PATCHED_SHA256 = "f33b370d9e49f87c5909e23109c1f22968166690765a7df169144d149f0f3d87"
+HELPER_SHA256 = "9e0d5e2cb5afe0d729ffffa160f1b44ddcca163e44a675a393154a6148a610de"
+
+
+def digest(path: Path) -> str:
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def apply_native_wo_a_patch(package_root: Path) -> None:
+    model = package_root / "srt/models/deepseek_v4_dspark.py"
+    helper = package_root / "srt/models/dspark_projection.py"
+    actual = digest(model)
+    if (
+        actual == PATCHED_SHA256
+        and helper.is_file()
+        and digest(helper) == HELPER_SHA256
+    ):
+        print(f"Native DSpark WO_A patch already applied: sha256:{actual}")
+        return
+    if actual != BASE_SHA256 or helper.exists():
+        raise RuntimeError(
+            "Refusing to patch unexpected SGLang DSpark source: "
+            f"model sha256:{actual}; expected sha256:{BASE_SHA256}. "
+            "Revalidate the native FP8 fix when changing the serving image."
+        )
+    patch = Path(__file__).parent / "patches/sglang_dsv41_native_wo_a.patch"
+    command = [
+        "patch",
+        "--batch",
+        "--forward",
+        "--fuzz=0",
+        "-p3",
+        "-i",
+        str(patch.resolve()),
+    ]
+    subprocess.run([*command, "--dry-run"], cwd=package_root, check=True)
+    subprocess.run(command, cwd=package_root, check=True)
+    if digest(model) != PATCHED_SHA256 or digest(helper) != HELPER_SHA256:
+        raise RuntimeError(
+            "Patched DSpark source does not match the validated native FP8 fix"
+        )
+    print(f"Applied native DSpark WO_A patch: sha256:{PATCHED_SHA256}")
+
+
+def main() -> None:
+    spec = importlib.util.find_spec("sglang")
+    if spec is None or spec.origin is None:
+        raise RuntimeError("SGLang package is unavailable in the serving container")
+    apply_native_wo_a_patch(Path(spec.origin).parent)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch b/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
new file mode 100644
index 0000000000..21c94e0d58
--- /dev/null
+++ b/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
@@ -0,0 +1,76 @@
+diff --git a/python/sglang/srt/models/deepseek_v4_dspark.py b/python/sglang/srt/models/deepseek_v4_dspark.py
+index baebc2d..e2cd9f8 100644
+--- a/python/sglang/srt/models/deepseek_v4_dspark.py
++++ b/python/sglang/srt/models/deepseek_v4_dspark.py
+@@ -55,0 +56 @@ from sglang.srt.models.dspark import (
++from sglang.srt.models.dspark_projection import apply_grouped_wo_a
+@@ -106,0 +108,3 @@ class DSparkAttention(MqaAttentionBase):
++        native_wo_a = (
++            getattr(config, "model_type", None) == "deepseek_v41" and not _is_npu
++        )
+@@ -117 +121 @@ class DSparkAttention(MqaAttentionBase):
+-            wo_a_keeps_quant_config=False,
++            wo_a_keeps_quant_config=native_wo_a,
+@@ -123,0 +128 @@ class DSparkAttention(MqaAttentionBase):
++        self.native_wo_a = native_wo_a
+@@ -355,2 +360,8 @@ class DSparkAttention(MqaAttentionBase):
+-        wo_a = self.wo_a.weight.view(self.n_local_groups, self.o_lora_rank, -1)
+-        if self._use_fast_kernel:
++        if self.native_wo_a:
++            # V4.1 uses 32x32-scaled FP8 weights. The grouped absorb GEMM only
++            # supports 128x128 scales, so use the existing dense quant method
++            # and retain each group's own output. At TP8 there is one local
++            # group and this is a single ordinary FP8 linear projection.
++            o = apply_grouped_wo_a(self.wo_a, o, self.o_lora_rank)
++        elif self._use_fast_kernel:
++            wo_a = self.wo_a.weight.view(self.n_local_groups, self.o_lora_rank, -1)
+@@ -364,0 +376 @@ class DSparkAttention(MqaAttentionBase):
++            wo_a = self.wo_a.weight.view(self.n_local_groups, self.o_lora_rank, -1)
+@@ -1074 +1086,2 @@ class DeepseekV4ForCausalLMDSpark(nn.Module):
+-        weights = _dequant_fp8_wo_a_streaming(weights)
++        if getattr(self.config, "model_type", None) != "deepseek_v41" or _is_npu:
++            weights = _dequant_fp8_wo_a_streaming(weights)
+@@ -1145,0 +1159,17 @@ class DeepseekV4ForCausalLMDSpark(nn.Module):
++        if getattr(self.config, "model_type", None) == "deepseek_v41" and not _is_npu:
++            for stage_id, stage in enumerate(self.stages):
++                prefix = f"stages.{stage_id}.self_attn.wo_a"
++                required = {f"{prefix}.weight", f"{prefix}.weight_scale_inv"}
++                if not required.issubset(loaded_params):
++                    raise ValueError(
++                        f"Native FP8 DSpark projection missing {required - loaded_params}"
++                    )
++                if stage.self_attn.wo_a.weight.dtype != torch.float8_e4m3fn:
++                    raise ValueError(
++                        f"{prefix} must retain native FP8 checkpoint weights"
++                    )
++            logger.info(
++                "DSpark native FP8 WO_A weights and block scales loaded for %d stages",
++                len(self.stages),
++            )
++
+diff --git a/python/sglang/srt/models/dspark_projection.py b/python/sglang/srt/models/dspark_projection.py
+new file mode 100644
+index 0000000..89df415
+--- /dev/null
++++ b/python/sglang/srt/models/dspark_projection.py
+@@ -0,0 +1,20 @@
++from typing import Callable, Optional, Tuple
++
++import torch
++
++
++def apply_grouped_wo_a(
++    projection: Callable[[torch.Tensor], Tuple[torch.Tensor, Optional[torch.Tensor]]],
++    hidden_states: torch.Tensor,
++    output_rank: int,
++) -> torch.Tensor:
++    """Apply each local group's projection through its native linear method.
++
++    The linear owns all local groups' weight rows. Projecting each input group
++    through it and retaining the diagonal supports its existing quantized
++    kernels without changing weight storage. One local group needs one GEMM.
++    """
++    tokens, groups, width = hidden_states.shape
++    projected, _ = projection(hidden_states.reshape(tokens * groups, width))
++    projected = projected.view(tokens, groups, groups, output_rank)
++    return projected.diagonal(dim1=1, dim2=2).transpose(1, 2).contiguous()
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 06906129e9..70a7a0a3f4 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -405,6 +405,14 @@ multi-arch preview build `lmsysorg/sglang:dev-dsv41` and MI355X uses
 `lmsysorg/sglang:dev-dsv41-mi35x`. Both tags are mutable, so the master configs and the
 changelog record the digests they were validated against.
 
+B300 registers STP and DSpark on the pinned latest CUDA 13 nightly. Its DSpark
+path applies the hash-verified `patch_sglang_dsv41_native_wo_a.py` fix to preserve
+all three native FP8 WO_A weights and block scales; Markov weights remain native
+BF16. STP loads no draft. Both STP and evals clear synthetic acceptance. B300
+uses per-rank anonymous Engram host tables without host sysctl changes, and
+resolves downloaded Hugging Face snapshots locally before serving. Verify
+actual huge-page backing, native draft startup checks and full accuracy results.
+
 DSpark is the checkpoint's own bundled draft. SGLang exposes no EAGLE or MTP path and no
 `--speculative-num-steps` knob for it; the recipes pass `--speculative-algorithm DSPARK
 --speculative-dspark-block-size 5`. Throughput uses the same
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index f97f1737ce..117d4e416e 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -352,6 +352,13 @@ Maximum concurrency for 1,048,576 tokens per request: 6.70x
 `lmsysorg/sglang:dev-dsv41`，MI355X 使用 `lmsysorg/sglang:dev-dsv41-mi35x`。两个标签均可变，
 因此 master 配置与 changelog 记录了验证时的 digest。
 
+B300 在固定到摘要的最新 CUDA 13 nightly 上注册 STP 和 DSpark。DSpark 路径应用
+经过哈希校验的 `patch_sglang_dsv41_native_wo_a.py` 修复，保留三个原生 FP8 WO_A
+权重及分块 scale；Markov 权重保持原生 BF16。STP 不加载草稿模型，STP 与评估均清除
+合成接受率设置。B300 使用每个 rank 的匿名 Engram 主机表，无需修改主机 sysctl，
+并在服务启动前从本地解析已下载的 Hugging Face 快照。必须核实实际大页覆盖率、
+草稿原生精度启动检查及完整准确率结果。
+
 DSpark 是检查点自带的草稿模型。SGLang 对它不提供 EAGLE 或 MTP 路径，也没有
 `--speculative-num-steps` 参数；配方传入 `--speculative-algorithm DSPARK
 --speculative-dspark-block-size 5`。吞吐测试通过 `SGLANG_SIMULATE_ACC_LEN`（`match-expected`、
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a990734188..f0266b7eba 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8553,4 +8553,5 @@
   - Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix.
   - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, while native DSpark precision qualification remains blocked.
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
+  - Apply the hash-verified native FP8 DSpark WO_A preservation fix on the pinned nightly, retain native BF16 Markov weights, and resolve downloaded checkpoints locally before serving.
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From d09f8a66c9720e9f515d94b4e87bb0197617c396 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 14:46:19 -0500
Subject: [PATCH 06/18] fix(b300): normalize strided native FP8 projection
 inputs

---
 .../single_node/agentic/patch_sglang_dsv41_native_wo_a.py       | 2 +-
 .../single_node/agentic/patches/sglang_dsv41_native_wo_a.patch  | 2 +-
 perf-changelog.yaml                                             | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py b/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
index b462590fcd..75a8f9cd28 100644
--- a/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
+++ b/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
@@ -12,7 +12,7 @@
 
 BASE_SHA256 = "61dc79f075c9e1e5a68a466de5eb95a85ff1fa2e5a9f4d66c2fa69e956fd7b61"
 PATCHED_SHA256 = "f33b370d9e49f87c5909e23109c1f22968166690765a7df169144d149f0f3d87"
-HELPER_SHA256 = "9e0d5e2cb5afe0d729ffffa160f1b44ddcca163e44a675a393154a6148a610de"
+HELPER_SHA256 = "4d1eb84af8cd9813c17417d31bdcf1c01fff96e5eb281e8ea2282aa2bae41a0e"
 
 
 def digest(path: Path) -> str:
diff --git a/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch b/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
index 21c94e0d58..2e2265755f 100644
--- a/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
+++ b/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
@@ -71,6 +71,6 @@ index 0000000..89df415
 +    kernels without changing weight storage. One local group needs one GEMM.
 +    """
 +    tokens, groups, width = hidden_states.shape
-+    projected, _ = projection(hidden_states.reshape(tokens * groups, width))
++    projected, _ = projection(hidden_states.reshape(tokens * groups, width).contiguous())
 +    projected = projected.view(tokens, groups, groups, output_rank)
 +    return projected.diagonal(dim1=1, dim2=2).transpose(1, 2).contiguous()
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f0266b7eba..db8d50e589 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8553,5 +8553,5 @@
   - Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix.
   - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, while native DSpark precision qualification remains blocked.
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
-  - Apply the hash-verified native FP8 DSpark WO_A preservation fix on the pinned nightly, retain native BF16 Markov weights, and resolve downloaded checkpoints locally before serving.
+  - Apply the hash-verified native FP8 DSpark WO_A preservation fix on the pinned nightly, retain native BF16 Markov weights, normalize strided FP8 projection inputs for eager and CUDA-graph execution, and resolve downloaded checkpoints locally before serving.
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From 17e7d2539043e9f136bbb9569b63babd931c19dc Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 15:11:48 -0500
Subject: [PATCH 07/18] fix: restore default SGLang DSpark precision

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh |  9 +--
 .../agentic/patch_sglang_dsv41_native_wo_a.py | 66 ----------------
 .../patches/sglang_dsv41_native_wo_a.patch    | 76 -------------------
 docs/configuration-procedures.md              |  8 +-
 docs/configuration-procedures_zh.md           |  7 +-
 perf-changelog.yaml                           |  4 +-
 6 files changed, 5 insertions(+), 165 deletions(-)
 delete mode 100644 benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
 delete mode 100644 benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index 1631e5539a..7324d76d5f 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -31,11 +31,7 @@ SERVER_LOG="$RESULT_DIR/server.log"
 export PYTHONNOUSERSITE=1
 export PYTHONUNBUFFERED=1
 
-# Preserve native draft FP8 weights with the exact-nightly, hash-guarded fix.
-if [[ "$SPEC_DECODING" == mtp ]]; then
-    python3 "$(dirname "$0")/patch_sglang_dsv41_native_wo_a.py" \
-        | tee "$RESULT_DIR/native_wo_a_patch.txt"
-fi
+# Use the default DSpark precision shipped by the pinned SGLang nightly.
 
 # Agentic warmup dispatches hundreds of large prompts at once and SGLang's
 # tokenizer can leave bytes unacknowledged past AIPerf's default 30 s
@@ -130,9 +126,6 @@ SGLANG_CMD=(
 )
 write_command "$RESULT_DIR/sglang_command.txt" "${SGLANG_CMD[@]}"
 {
-    if [[ "$SPEC_DECODING" == mtp ]]; then
-        cat "$RESULT_DIR/native_wo_a_patch.txt"
-    fi
     echo "=== SGLANG_* env vars at launch ==="
     env | grep -E '^SGLANG_' | sort
     echo "==================================="
diff --git a/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py b/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
deleted file mode 100644
index 75a8f9cd28..0000000000
--- a/benchmarks/single_node/agentic/patch_sglang_dsv41_native_wo_a.py
+++ /dev/null
@@ -1,66 +0,0 @@
-"""Preserve native V4.1 DSpark FP8 WO_A weights in the official SGLang nightly.
-
-Applies only to the exact upstream source shipped at 0f6761b54facebb47f2068f87ecccd8f14da3a0e.
-The dense FP8 linear path supports the checkpoint's 32x32 scales; the upstream
-grouped projection instead dequantizes these draft weights permanently to BF16.
-"""
-
-import hashlib
-import importlib.util
-import subprocess
-from pathlib import Path
-
-BASE_SHA256 = "61dc79f075c9e1e5a68a466de5eb95a85ff1fa2e5a9f4d66c2fa69e956fd7b61"
-PATCHED_SHA256 = "f33b370d9e49f87c5909e23109c1f22968166690765a7df169144d149f0f3d87"
-HELPER_SHA256 = "4d1eb84af8cd9813c17417d31bdcf1c01fff96e5eb281e8ea2282aa2bae41a0e"
-
-
-def digest(path: Path) -> str:
-    return hashlib.sha256(path.read_bytes()).hexdigest()
-
-
-def apply_native_wo_a_patch(package_root: Path) -> None:
-    model = package_root / "srt/models/deepseek_v4_dspark.py"
-    helper = package_root / "srt/models/dspark_projection.py"
-    actual = digest(model)
-    if (
-        actual == PATCHED_SHA256
-        and helper.is_file()
-        and digest(helper) == HELPER_SHA256
-    ):
-        print(f"Native DSpark WO_A patch already applied: sha256:{actual}")
-        return
-    if actual != BASE_SHA256 or helper.exists():
-        raise RuntimeError(
-            "Refusing to patch unexpected SGLang DSpark source: "
-            f"model sha256:{actual}; expected sha256:{BASE_SHA256}. "
-            "Revalidate the native FP8 fix when changing the serving image."
-        )
-    patch = Path(__file__).parent / "patches/sglang_dsv41_native_wo_a.patch"
-    command = [
-        "patch",
-        "--batch",
-        "--forward",
-        "--fuzz=0",
-        "-p3",
-        "-i",
-        str(patch.resolve()),
-    ]
-    subprocess.run([*command, "--dry-run"], cwd=package_root, check=True)
-    subprocess.run(command, cwd=package_root, check=True)
-    if digest(model) != PATCHED_SHA256 or digest(helper) != HELPER_SHA256:
-        raise RuntimeError(
-            "Patched DSpark source does not match the validated native FP8 fix"
-        )
-    print(f"Applied native DSpark WO_A patch: sha256:{PATCHED_SHA256}")
-
-
-def main() -> None:
-    spec = importlib.util.find_spec("sglang")
-    if spec is None or spec.origin is None:
-        raise RuntimeError("SGLang package is unavailable in the serving container")
-    apply_native_wo_a_patch(Path(spec.origin).parent)
-
-
-if __name__ == "__main__":
-    main()
diff --git a/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch b/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
deleted file mode 100644
index 2e2265755f..0000000000
--- a/benchmarks/single_node/agentic/patches/sglang_dsv41_native_wo_a.patch
+++ /dev/null
@@ -1,76 +0,0 @@
-diff --git a/python/sglang/srt/models/deepseek_v4_dspark.py b/python/sglang/srt/models/deepseek_v4_dspark.py
-index baebc2d..e2cd9f8 100644
---- a/python/sglang/srt/models/deepseek_v4_dspark.py
-+++ b/python/sglang/srt/models/deepseek_v4_dspark.py
-@@ -55,0 +56 @@ from sglang.srt.models.dspark import (
-+from sglang.srt.models.dspark_projection import apply_grouped_wo_a
-@@ -106,0 +108,3 @@ class DSparkAttention(MqaAttentionBase):
-+        native_wo_a = (
-+            getattr(config, "model_type", None) == "deepseek_v41" and not _is_npu
-+        )
-@@ -117 +121 @@ class DSparkAttention(MqaAttentionBase):
--            wo_a_keeps_quant_config=False,
-+            wo_a_keeps_quant_config=native_wo_a,
-@@ -123,0 +128 @@ class DSparkAttention(MqaAttentionBase):
-+        self.native_wo_a = native_wo_a
-@@ -355,2 +360,8 @@ class DSparkAttention(MqaAttentionBase):
--        wo_a = self.wo_a.weight.view(self.n_local_groups, self.o_lora_rank, -1)
--        if self._use_fast_kernel:
-+        if self.native_wo_a:
-+            # V4.1 uses 32x32-scaled FP8 weights. The grouped absorb GEMM only
-+            # supports 128x128 scales, so use the existing dense quant method
-+            # and retain each group's own output. At TP8 there is one local
-+            # group and this is a single ordinary FP8 linear projection.
-+            o = apply_grouped_wo_a(self.wo_a, o, self.o_lora_rank)
-+        elif self._use_fast_kernel:
-+            wo_a = self.wo_a.weight.view(self.n_local_groups, self.o_lora_rank, -1)
-@@ -364,0 +376 @@ class DSparkAttention(MqaAttentionBase):
-+            wo_a = self.wo_a.weight.view(self.n_local_groups, self.o_lora_rank, -1)
-@@ -1074 +1086,2 @@ class DeepseekV4ForCausalLMDSpark(nn.Module):
--        weights = _dequant_fp8_wo_a_streaming(weights)
-+        if getattr(self.config, "model_type", None) != "deepseek_v41" or _is_npu:
-+            weights = _dequant_fp8_wo_a_streaming(weights)
-@@ -1145,0 +1159,17 @@ class DeepseekV4ForCausalLMDSpark(nn.Module):
-+        if getattr(self.config, "model_type", None) == "deepseek_v41" and not _is_npu:
-+            for stage_id, stage in enumerate(self.stages):
-+                prefix = f"stages.{stage_id}.self_attn.wo_a"
-+                required = {f"{prefix}.weight", f"{prefix}.weight_scale_inv"}
-+                if not required.issubset(loaded_params):
-+                    raise ValueError(
-+                        f"Native FP8 DSpark projection missing {required - loaded_params}"
-+                    )
-+                if stage.self_attn.wo_a.weight.dtype != torch.float8_e4m3fn:
-+                    raise ValueError(
-+                        f"{prefix} must retain native FP8 checkpoint weights"
-+                    )
-+            logger.info(
-+                "DSpark native FP8 WO_A weights and block scales loaded for %d stages",
-+                len(self.stages),
-+            )
-+
-diff --git a/python/sglang/srt/models/dspark_projection.py b/python/sglang/srt/models/dspark_projection.py
-new file mode 100644
-index 0000000..89df415
---- /dev/null
-+++ b/python/sglang/srt/models/dspark_projection.py
-@@ -0,0 +1,20 @@
-+from typing import Callable, Optional, Tuple
-+
-+import torch
-+
-+
-+def apply_grouped_wo_a(
-+    projection: Callable[[torch.Tensor], Tuple[torch.Tensor, Optional[torch.Tensor]]],
-+    hidden_states: torch.Tensor,
-+    output_rank: int,
-+) -> torch.Tensor:
-+    """Apply each local group's projection through its native linear method.
-+
-+    The linear owns all local groups' weight rows. Projecting each input group
-+    through it and retaining the diagonal supports its existing quantized
-+    kernels without changing weight storage. One local group needs one GEMM.
-+    """
-+    tokens, groups, width = hidden_states.shape
-+    projected, _ = projection(hidden_states.reshape(tokens * groups, width).contiguous())
-+    projected = projected.view(tokens, groups, groups, output_rank)
-+    return projected.diagonal(dim1=1, dim2=2).transpose(1, 2).contiguous()
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 70a7a0a3f4..6967f21d58 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -405,13 +405,7 @@ multi-arch preview build `lmsysorg/sglang:dev-dsv41` and MI355X uses
 `lmsysorg/sglang:dev-dsv41-mi35x`. Both tags are mutable, so the master configs and the
 changelog record the digests they were validated against.
 
-B300 registers STP and DSpark on the pinned latest CUDA 13 nightly. Its DSpark
-path applies the hash-verified `patch_sglang_dsv41_native_wo_a.py` fix to preserve
-all three native FP8 WO_A weights and block scales; Markov weights remain native
-BF16. STP loads no draft. Both STP and evals clear synthetic acceptance. B300
-uses per-rank anonymous Engram host tables without host sysctl changes, and
-resolves downloaded Hugging Face snapshots locally before serving. Verify
-actual huge-page backing, native draft startup checks and full accuracy results.
+DSpark uses the default precision shipped by the pinned official nightly, without custom draft quantization or precision patches. STP loads no draft; full accuracy and performance validation are still required.
 
 DSpark is the checkpoint's own bundled draft. SGLang exposes no EAGLE or MTP path and no
 `--speculative-num-steps` knob for it; the recipes pass `--speculative-algorithm DSPARK
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 117d4e416e..7bdf140336 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -352,12 +352,7 @@ Maximum concurrency for 1,048,576 tokens per request: 6.70x
 `lmsysorg/sglang:dev-dsv41`，MI355X 使用 `lmsysorg/sglang:dev-dsv41-mi35x`。两个标签均可变，
 因此 master 配置与 changelog 记录了验证时的 digest。
 
-B300 在固定到摘要的最新 CUDA 13 nightly 上注册 STP 和 DSpark。DSpark 路径应用
-经过哈希校验的 `patch_sglang_dsv41_native_wo_a.py` 修复，保留三个原生 FP8 WO_A
-权重及分块 scale；Markov 权重保持原生 BF16。STP 不加载草稿模型，STP 与评估均清除
-合成接受率设置。B300 使用每个 rank 的匿名 Engram 主机表，无需修改主机 sysctl，
-并在服务启动前从本地解析已下载的 Hugging Face 快照。必须核实实际大页覆盖率、
-草稿原生精度启动检查及完整准确率结果。
+DSpark 使用固定官方 nightly 默认提供的精度，不应用自定义草稿量化或精度补丁。STP 不加载草稿模型；完整准确率和性能验证仍然必需。
 
 DSpark 是检查点自带的草稿模型。SGLang 对它不提供 EAGLE 或 MTP 路径，也没有
 `--speculative-num-steps` 参数；配方传入 `--speculative-algorithm DSPARK
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index db8d50e589..87b897117c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8551,7 +8551,7 @@
   description:
   - Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals.
   - Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix.
-  - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, while native DSpark precision qualification remains blocked.
+  - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, alongside the default-precision DSpark arm.
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
-  - Apply the hash-verified native FP8 DSpark WO_A preservation fix on the pinned nightly, retain native BF16 Markov weights, normalize strided FP8 projection inputs for eager and CUDA-graph execution, and resolve downloaded checkpoints locally before serving.
+  - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From b06b4f832b99944e50b1038e5ddc3403aa8c3978 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 16:41:42 -0500
Subject: [PATCH 08/18] perf(b300): expand DSpark parallelism and interleave
 decode

---
 .../single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh       | 2 ++
 configs/nvidia-master.yaml                                      | 1 +
 perf-changelog.yaml                                             | 1 +
 3 files changed, 4 insertions(+)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index 7324d76d5f..28358f9edf 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -114,6 +114,8 @@ SGLANG_CMD=(
     # first 66k-99k-token AgentX prompts.
     --mem-fraction-static 0.70
     --chunked-prefill-size 4096
+    # Long AgentX prefills otherwise starve ready decode requests.
+    --prefill-decode-interval 16
     "${SPECULATIVE_ARGS[@]}"
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index cf0e84fcb8..e99d78a67d 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8411,6 +8411,7 @@ dsv41flash-fp4-b300-sglang-agentic-dspark:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
+      - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
 
 # Non-speculative cookbook high-throughput path; no draft or synthetic acceptance.
 dsv41flash-fp4-b300-sglang-agentic:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 87b897117c..00d3a2245c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8554,4 +8554,5 @@
   - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, alongside the default-precision DSpark arm.
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
+  - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From 0a68cc0e2aaee191b37bdcf9a08693422b73afdd Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 17:31:13 -0500
Subject: [PATCH 09/18] Launch B300 SGLang AgentX through an owned Slurm batch
 job

---
 perf-changelog.yaml         |  1 +
 runners/launch_b300-dsxe.sh | 53 ++++++++++++++++++++++++++++++++-----
 2 files changed, 48 insertions(+), 6 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 00d3a2245c..c9da37e86b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8555,4 +8555,5 @@
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
+  - "Submit this SGLang AgentX lane through a Slurm batch allocation and launch its steps from the compute node, avoiding failed interactive allocation notifications on the runner login host while preserving GPU reservations and terminal-status checks."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342
diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index f49ccec597..76c81630ed 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -13,6 +13,38 @@ source "$(dirname "${BASH_SOURCE[0]}")/slurm_utils.sh" || exit 1
 SLURM_PARTITION="batch_1"
 SLURM_ACCOUNT="benchmark"
 
+# This lane's interactive allocation notifications fail on login-02, while
+# batch submission and steps launched from the allocated node work. Keep the
+# workaround scoped to this recipe and use normal Slurm resource accounting.
+if [[ "$IS_MULTINODE" != true && "${MODEL_PREFIX:-}" == dsv41flash &&
+      "${FRAMEWORK:-}" == sglang && "${IS_AGENTIC:-}" == 1 &&
+      "${B300_AGENTX_BATCH:-}" != 1 ]]; then
+    check_env_vars GITHUB_WORKSPACE GPU_COUNT RUNNER_NAME
+    BATCH_SCRIPT=$(mktemp "${RUNNER_TEMP:-$GITHUB_WORKSPACE}/b300-agentx.XXXXXX.sh") || exit 1
+    BATCH_LOG="${BATCH_SCRIPT%.sh}.log"
+    {
+        printf '#!/usr/bin/env bash\nexport B300_AGENTX_BATCH=1\nexec bash '
+        printf '%q\n' "$GITHUB_WORKSPACE/runners/launch_b300-dsxe.sh"
+    } > "$BATCH_SCRIPT"
+    BATCH_ARGS=(--parsable --partition="$SLURM_PARTITION" --account="$SLURM_ACCOUNT"
+        --nodes=1 --ntasks=1 --gres="gpu:$GPU_COUNT" --exclusive --mem=0
+        --time="$SALLOC_TIME_LIMIT" --job-name="$RUNNER_NAME" --export=ALL
+        --chdir="$GITHUB_WORKSPACE" --output="$BATCH_LOG")
+    if [[ -n "${SALLOC_EXCLUDE:-}" ]]; then
+        BATCH_ARGS+=(--exclude="$SALLOC_EXCLUDE")
+    fi
+    JOB_ID=$(sbatch "${BATCH_ARGS[@]}" "$BATCH_SCRIPT") || { rm -f "$BATCH_SCRIPT"; exit 1; }
+    JOB_ID="${JOB_ID%%;*}"
+    [[ "$JOB_ID" =~ ^[0-9]+$ ]] || { echo 'ERROR: B300 batch allocation unavailable' >&2; rm -f "$BATCH_SCRIPT"; exit 1; }
+    trap 'rc=$?; scancel "$JOB_ID" 2>/dev/null || true; rm -f "$BATCH_SCRIPT"; exit "$rc"' EXIT
+    trap 'exit 130' INT
+    trap 'exit 143' TERM
+    echo "B300 AgentX batch job $JOB_ID; log: $BATCH_LOG"
+    stream_slurm_job_log "$JOB_ID" "$BATCH_LOG" || exit 1
+    verify_slurm_job_status "$JOB_ID"
+    exit $?
+fi
+
 # enroot squash images. Must be on storage every compute node mounts and writable
 # by the runner user (/data/squash is root-owned, hence the per-user default).
 SQUASH_DIR="/data/home/sa-gha-runner/squash"
@@ -67,8 +99,13 @@ import_squash_image() {
         return 0
     fi
 
-    srun -N 1 -A "$SLURM_ACCOUNT" -p "$SLURM_PARTITION" \
-        --time="${ENROOT_IMPORT_TIME_LIMIT}" bash -c "
+    local import_launcher=(srun -N 1 -A "$SLURM_ACCOUNT" -p "$SLURM_PARTITION"
+        --time="${ENROOT_IMPORT_TIME_LIMIT}")
+    if [[ "${B300_AGENTX_BATCH:-}" == 1 ]]; then
+        # Already inside our exclusive compute-node allocation.
+        import_launcher=()
+    fi
+    "${import_launcher[@]}" bash -c "
         set -eo pipefail
         exec 9>\"$lock\"
         flock -w 3600 9
@@ -364,13 +401,17 @@ else
         SALLOC_ARGS+=(--exclude="$SALLOC_EXCLUDE")
     fi
     # Capture this allocation's ID; a runner name can also match an older job.
-    JOB_ID=$(
+    if [[ "${B300_AGENTX_BATCH:-}" == 1 ]]; then
+        JOB_ID="${SLURM_JOB_ID:?B300 batch execution requires a Slurm allocation}"
+    else
+        JOB_ID=$(
         set -o pipefail
         LC_ALL=C salloc "${SALLOC_ARGS[@]}" 2>&1 | tee /dev/stderr |
             sed -n 's/.*Granted job allocation \([0-9][0-9]*\)$/\1/p'
-    ) || exit 1
-    [[ "$JOB_ID" =~ ^[0-9]+$ ]] || { echo 'ERROR: B300 allocation unavailable' >&2; exit 1; }
-    trap 'rc=$?; scancel "$JOB_ID" 2>/dev/null || true; exit "$rc"' EXIT
+        ) || exit 1
+        [[ "$JOB_ID" =~ ^[0-9]+$ ]] || { echo 'ERROR: B300 allocation unavailable' >&2; exit 1; }
+        trap 'rc=$?; scancel "$JOB_ID" 2>/dev/null || true; exit "$rc"' EXIT
+    fi
     if [[ "$MODEL_MOUNT_DIR" == "$MODEL_ROOT" ]]; then
         # MODEL_ROOT is node-local: probe the allocated compute node, not the login host.
         srun --jobid="$JOB_ID" test -r "$MODEL_PATH/config.json" || {

From e2b357a12ab90784b1b3e6e1e8baf12552ab4b4e Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 17:44:41 -0500
Subject: [PATCH 10/18] Match B300 batch PMI environment to non-MPI container
 step

---
 runners/launch_b300-dsxe.sh | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index 76c81630ed..59f7d12b8c 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -23,7 +23,9 @@ if [[ "$IS_MULTINODE" != true && "${MODEL_PREFIX:-}" == dsv41flash &&
     BATCH_SCRIPT=$(mktemp "${RUNNER_TEMP:-$GITHUB_WORKSPACE}/b300-agentx.XXXXXX.sh") || exit 1
     BATCH_LOG="${BATCH_SCRIPT%.sh}.log"
     {
-        printf '#!/usr/bin/env bash\nexport B300_AGENTX_BATCH=1\nexec bash '
+        # Enroot's PMI hook also inspects this variable. Without it, inherited
+        # batch PMIX variables trigger mounts absent from our --mpi=none step.
+        printf '#!/usr/bin/env bash\nexport B300_AGENTX_BATCH=1\nexport SLURM_MPI_TYPE=none\nexec bash '
         printf '%q\n' "$GITHUB_WORKSPACE/runners/launch_b300-dsxe.sh"
     } > "$BATCH_SCRIPT"
     BATCH_ARGS=(--parsable --partition="$SLURM_PARTITION" --account="$SLURM_ACCOUNT"

From e3803f030c504626c7ba7d32e1fb5e3f7de5a5cb Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 18:04:27 -0500
Subject: [PATCH 11/18] Use supported PMIx container steps for B300 AgentX
 batch jobs

---
 runners/launch_b300-dsxe.sh | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index 59f7d12b8c..da289a2f2e 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -23,9 +23,7 @@ if [[ "$IS_MULTINODE" != true && "${MODEL_PREFIX:-}" == dsv41flash &&
     BATCH_SCRIPT=$(mktemp "${RUNNER_TEMP:-$GITHUB_WORKSPACE}/b300-agentx.XXXXXX.sh") || exit 1
     BATCH_LOG="${BATCH_SCRIPT%.sh}.log"
     {
-        # Enroot's PMI hook also inspects this variable. Without it, inherited
-        # batch PMIX variables trigger mounts absent from our --mpi=none step.
-        printf '#!/usr/bin/env bash\nexport B300_AGENTX_BATCH=1\nexport SLURM_MPI_TYPE=none\nexec bash '
+        printf '#!/usr/bin/env bash\nexport B300_AGENTX_BATCH=1\nexec bash '
         printf '%q\n' "$GITHUB_WORKSPACE/runners/launch_b300-dsxe.sh"
     } > "$BATCH_SCRIPT"
     BATCH_ARGS=(--parsable --partition="$SLURM_PARTITION" --account="$SLURM_ACCOUNT"
@@ -438,8 +436,14 @@ else
     fi
     CONTAINER_MOUNTS_ARG=$(IFS=,; printf '%s' "${CONTAINER_MOUNTS[*]}")
 
+    B300_CONTAINER_MPI=none
+    if [[ "${B300_AGENTX_BATCH:-}" == 1 ]]; then
+        # The installed Enroot hook sees PMIx variables in batch jobs. Use the
+        # supported plugin so its required per-step mount directories exist.
+        B300_CONTAINER_MPI=pmix
+    fi
     srun --jobid="$JOB_ID" \
-        --mpi=none \
+        --mpi="$B300_CONTAINER_MPI" \
         --container-image="$SQUASH_FILE" \
         --container-mounts="$CONTAINER_MOUNTS_ARG" \
         --no-container-mount-home \

From a753192d9a9f2c341a6966d91ee5fb5491e33263 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 18:28:08 -0500
Subject: [PATCH 12/18] perf(b300): scope final AgentX sweep to DSpark
 candidates

---
 .../agentic/dsv41flash_fp4_b300_sglang.sh          |  1 -
 configs/nvidia-master.yaml                         | 14 --------------
 perf-changelog.yaml                                |  2 --
 3 files changed, 17 deletions(-)
 delete mode 120000 benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh
deleted file mode 120000
index 03357261bc..0000000000
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang.sh
+++ /dev/null
@@ -1 +0,0 @@
-dsv41flash_fp4_b300_sglang_mtp.sh
\ No newline at end of file
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index fd81aedb75..e7c5743a40 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8411,17 +8411,3 @@ dsv41flash-fp4-b300-sglang-agentic-dspark:
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
       - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
 
-# Non-speculative cookbook high-throughput path; no draft or synthetic acceptance.
-dsv41flash-fp4-b300-sglang-agentic:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260921-0f6761b5@sha256:18aed2cd75ed45932a9d409e167e75191e0596a47925de1f176d1cc03036cf8c
-  model: deepseek-ai/DeepSeek-V4.1-Flash
-  model-prefix: dsv41flash
-  runner: cluster:b300-dsxe
-  precision: fp4
-  framework: sglang
-  multinode: false
-  scenarios:
-    agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
-      - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: none, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c9da37e86b..401dc6cc29 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8545,13 +8545,11 @@
 
 - config-keys:
   - dsv41flash-fp4-b300-sglang-agentic-dspark
-  - dsv41flash-fp4-b300-sglang-agentic
   scenario-type:
   - agentic-coding
   description:
   - Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals.
   - Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix.
-  - Add non-speculative B300 DeepSeek-V4.1-Flash SGLang AgentX TP4/EP4 on the pinned latest nightly, with no draft and no synthetic acceptance, alongside the default-precision DSpark arm.
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."

From 72c41b899fb5b12f732f1690da55730b05cb1b67 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 20:16:02 -0500
Subject: [PATCH 13/18] Test B300 low-concurrency prefix retention and
 checkpoint prefetch

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh         | 11 +++++++++++
 perf-changelog.yaml                                   |  1 +
 2 files changed, 12 insertions(+)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index 28358f9edf..907d73653f 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -65,6 +65,14 @@ if (( MAX_RUNNING_REQUESTS > CUDA_GRAPH_MAX_BS )); then
     MAX_RUNNING_REQUESTS=$CUDA_GRAPH_MAX_BS
 fi
 
+# AgentX reuses long prefixes across turns even at low session concurrency.
+# Floor the nightly's four-tails-per-request default while preserving its
+# existing larger-request budget until the high-concurrency sweep is measured.
+SWA_PREFIX_TAILS=$((4 * MAX_RUNNING_REQUESTS))
+if (( SWA_PREFIX_TAILS < 128 )); then
+    SWA_PREFIX_TAILS=128
+fi
+
 # Saturation arms carry a larger in-flight working set than the 30-minute
 # default warmup drain allows.
 if (( CONC >= 32 )); then
@@ -105,6 +113,8 @@ SGLANG_CMD=(
     --model-path "$MODEL_PATH" --served-model-name "$MODEL"
     --host 0.0.0.0 --port "$PORT"
     --trust-remote-code
+    # Feed mmap weight copies sequentially from shared Lustre storage.
+    --weight-loader-prefetch-checkpoints
     --tp "$TP" --ep-size "$EP_SIZE"
     # Backends resolve automatically (dsv4 / flashinfer_mxfp4 / flashinfer_cutedsl
     # on Blackwell); the cookbook warns that overriding them costs decode speed.
@@ -118,6 +128,7 @@ SGLANG_CMD=(
     --prefill-decode-interval 16
     "${SPECULATIVE_ARGS[@]}"
     --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --swa-prefix-tails "$SWA_PREFIX_TAILS"
     --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
     --reasoning-parser auto
     --tool-call-parser auto
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b1eead9f5d..830b059e60 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8565,4 +8565,5 @@
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
   - "Submit this SGLang AgentX lane through a Slurm batch allocation and launch its steps from the compute node, avoiding failed interactive allocation notifications on the runner login host while preserving GPU reservations and terminal-status checks."
+  - "Test a 128-tail minimum for low-concurrency prefix reuse while preserving the larger-concurrency budget and 0.70 static memory fraction; prefetch checkpoints for shared Lustre loading."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From 72a8341750f3a1562c7eb0efc86ebf6b648ac032 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 20:32:03 -0500
Subject: [PATCH 14/18] Scale B300 TP4 prefix retention within measured KV
 budget

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh          | 10 ++++++++--
 perf-changelog.yaml                                    |  2 +-
 2 files changed, 9 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index 907d73653f..cd23f1eba3 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -66,9 +66,15 @@ if (( MAX_RUNNING_REQUESTS > CUDA_GRAPH_MAX_BS )); then
 fi
 
 # AgentX reuses long prefixes across turns even at low session concurrency.
-# Floor the nightly's four-tails-per-request default while preserving its
-# existing larger-request budget until the high-concurrency sweep is measured.
+# TP4's measured 110.55 GiB KV budget leaves room to retain session prefixes
+# across subagent turns. TP2 keeps its conservative budget until measured.
 SWA_PREFIX_TAILS=$((4 * MAX_RUNNING_REQUESTS))
+if (( TP == 4 )); then
+    SWA_PREFIX_TAILS=$((64 * CONC))
+    if (( SWA_PREFIX_TAILS > 4096 )); then
+        SWA_PREFIX_TAILS=4096
+    fi
+fi
 if (( SWA_PREFIX_TAILS < 128 )); then
     SWA_PREFIX_TAILS=128
 fi
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 830b059e60..debf876524 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8565,5 +8565,5 @@
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
   - "Submit this SGLang AgentX lane through a Slurm batch allocation and launch its steps from the compute node, avoiding failed interactive allocation notifications on the runner login host while preserving GPU reservations and terminal-status checks."
-  - "Test a 128-tail minimum for low-concurrency prefix reuse while preserving the larger-concurrency budget and 0.70 static memory fraction; prefetch checkpoints for shared Lustre loading."
+  - "Test a 128-tail minimum for low-concurrency prefix reuse and scale TP4 tails with concurrency up to 4096, keeping TP2's larger-concurrency budget and the 0.70 static memory fraction; prefetch checkpoints for shared Lustre loading."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From efd7f6621441ebcfb89922eb59179bc78c93dd4d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 20:36:43 -0500
Subject: [PATCH 15/18] Pin B300 to the September 22 SGLang nightly

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

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 83f54c5797..87f2c24658 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8579,7 +8579,7 @@ qwen3.5-fp8-b300-dynamo-sglang-agentic-disagg:
 # Latest multiarch index: sha256:987c7e4bd26918647211a5dcad72a2bdf2a5f394ac1469eff730e7517fc139be.
 # Uses the cookbook's automatic backends and the existing B200 AgentX memory bounds.
 dsv41flash-fp4-b300-sglang-agentic-dspark:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260921-0f6761b5@sha256:18aed2cd75ed45932a9d409e167e75191e0596a47925de1f176d1cc03036cf8c
+  image: lmsysorg/sglang:nightly-dev-cu13-20260922-582389ce@sha256:35ea4d321b0735051dcce2599fc495853ad1ef30f6a1908f0a75e74204362d14
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:b300-dsxe
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index debf876524..4f13f665b7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8559,7 +8559,7 @@
   scenario-type:
   - agentic-coding
   description:
-  - Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260921-0f6761b5), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals.
+  - Add B300 SGLang DeepSeek-V4.1-Flash AgentX TP4/EP4 with the newest official CUDA 13 nightly (20260922-582389ce), pinned amd64 digest, native DSpark5, measured thinking-on AL 3.51 for throughput and real verification for evals.
   - Keep automatic backends, host Engram tables, GPU-resident KV, bounded prefill and decode graphs; route the B300 SGLang runtime through /ix.
   - Use per-rank anonymous host Engram tables to allow transparent huge pages without changing host settings; preserve native FP8 table payload and automatic backend selection.
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."

From c684a14148ab8ff4224531054bcdd86e7068f75d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 20:54:49 -0500
Subject: [PATCH 16/18] Test larger B300 TP2 KV budget for high concurrency

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh | 27 ++++++++++---------
 perf-changelog.yaml                           |  2 +-
 2 files changed, 15 insertions(+), 14 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index cd23f1eba3..0a114c9d6a 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -66,14 +66,16 @@ if (( MAX_RUNNING_REQUESTS > CUDA_GRAPH_MAX_BS )); then
 fi
 
 # AgentX reuses long prefixes across turns even at low session concurrency.
-# TP4's measured 110.55 GiB KV budget leaves room to retain session prefixes
-# across subagent turns. TP2 keeps its conservative budget until measured.
-SWA_PREFIX_TAILS=$((4 * MAX_RUNNING_REQUESTS))
-if (( TP == 4 )); then
-    SWA_PREFIX_TAILS=$((64 * CONC))
-    if (( SWA_PREFIX_TAILS > 4096 )); then
-        SWA_PREFIX_TAILS=4096
-    fi
+# At memory fraction 0.70, TP4 has 110.55 GiB of KV budget but TP2 has only
+# 38.83 GiB. Give TP2's larger working sets more KV space before retaining
+# additional SWA tails; keep the conservative low-concurrency allocation.
+MEM_FRACTION_STATIC=0.70
+if (( TP == 2 && CONC >= 16 )); then
+    MEM_FRACTION_STATIC=0.80
+fi
+SWA_PREFIX_TAILS=$((64 * CONC))
+if (( SWA_PREFIX_TAILS > 4096 )); then
+    SWA_PREFIX_TAILS=4096
 fi
 if (( SWA_PREFIX_TAILS < 128 )); then
     SWA_PREFIX_TAILS=128
@@ -124,11 +126,10 @@ SGLANG_CMD=(
     --tp "$TP" --ep-size "$EP_SIZE"
     # Backends resolve automatically (dsv4 / flashinfer_mxfp4 / flashinfer_cutedsl
     # on Blackwell); the cookbook warns that overriding them costs decode speed.
-    # 0.70 rather than the cookbook's 0.8, and a bounded prefill chunk: the
-    # sparse-attention indexer and DSpark prefill buffers scale with the chunk
-    # times the 1M context, and the default 16384 chunk exhausted HBM on the
-    # first 66k-99k-token AgentX prompts.
-    --mem-fraction-static 0.70
+    # Bound transient prefill allocations: the sparse-attention indexer and
+    # DSpark buffers scale with the chunk times the 1M context. Static KV
+    # memory is selected above from the measured TP2/TP4 weight footprints.
+    --mem-fraction-static "$MEM_FRACTION_STATIC"
     --chunked-prefill-size 4096
     # Long AgentX prefills otherwise starve ready decode requests.
     --prefill-decode-interval 16
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index dafcd96283..fc835d4a68 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8599,5 +8599,5 @@
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
   - "Submit this SGLang AgentX lane through a Slurm batch allocation and launch its steps from the compute node, avoiding failed interactive allocation notifications on the runner login host while preserving GPU reservations and terminal-status checks."
-  - "Test a 128-tail minimum for low-concurrency prefix reuse and scale TP4 tails with concurrency up to 4096, keeping TP2's larger-concurrency budget and the 0.70 static memory fraction; prefetch checkpoints for shared Lustre loading."
+  - "Test retained SWA tails at 64 times concurrency, bounded to 128–4096; use static memory 0.80 for TP2 at concurrency 16 and above after measuring its smaller KV budget, otherwise retain 0.70. Prefetch checkpoints for shared Lustre loading."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From 252105c44688a9144c4bc7d342953f7113e74225 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 22:27:10 -0500
Subject: [PATCH 17/18] Increase high-concurrency TP2 cache memory on B300

---
 .../single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh   | 6 +++++-
 perf-changelog.yaml                                         | 2 +-
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index 0a114c9d6a..e00d9df036 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -70,7 +70,11 @@ fi
 # 38.83 GiB. Give TP2's larger working sets more KV space before retaining
 # additional SWA tails; keep the conservative low-concurrency allocation.
 MEM_FRACTION_STATIC=0.70
-if (( TP == 2 && CONC >= 16 )); then
+if (( TP == 2 && CONC >= 32 )); then
+    # At C64, 0.80 left 43.59 GiB after graphs but only 18.29M full tokens.
+    # Retain more long prefixes while leaving room for transient prefills.
+    MEM_FRACTION_STATIC=0.85
+elif (( TP == 2 && CONC >= 16 )); then
     MEM_FRACTION_STATIC=0.80
 fi
 SWA_PREFIX_TAILS=$((64 * CONC))
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index fc835d4a68..d1c2e6a3cd 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8599,5 +8599,5 @@
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
   - "Submit this SGLang AgentX lane through a Slurm batch allocation and launch its steps from the compute node, avoiding failed interactive allocation notifications on the runner login host while preserving GPU reservations and terminal-status checks."
-  - "Test retained SWA tails at 64 times concurrency, bounded to 128–4096; use static memory 0.80 for TP2 at concurrency 16 and above after measuring its smaller KV budget, otherwise retain 0.70. Prefetch checkpoints for shared Lustre loading."
+  - "Test retained SWA tails at 64 times concurrency, bounded to 128–4096; use static memory 0.80 for TP2 C16 and 0.85 at C32 and above after measuring its smaller KV budget and 43.59 GiB of post-graph headroom at C64, otherwise retain 0.70. Prefetch checkpoints for shared Lustre loading."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342

From cd702aefaa9e3732c16bf27fd51f02ffacdd3c82 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 21 Sep 2026 23:33:20 -0500
Subject: [PATCH 18/18] Screen higher prefill duty for B300 TP2 DSpark

---
 .../agentic/dsv41flash_fp4_b300_sglang_mtp.sh            | 9 ++++++++-
 perf-changelog.yaml                                      | 2 +-
 2 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
index e00d9df036..dfb7056385 100644
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_b300_sglang_mtp.sh
@@ -120,6 +120,13 @@ case "$SPEC_DECODING" in
         ;;
 esac
 
+# At high TP2 concurrency, test more prefill duty against the matched C64
+# baseline. Keep the latency-oriented cadence on low-C and TP4 points.
+PREFILL_DECODE_INTERVAL=16
+if (( TP == 2 && CONC >= 32 )); then
+    PREFILL_DECODE_INTERVAL=4
+fi
+
 SGLANG_CMD=(
     python3 -m sglang.launch_server
     --model-path "$MODEL_PATH" --served-model-name "$MODEL"
@@ -136,7 +143,7 @@ SGLANG_CMD=(
     --mem-fraction-static "$MEM_FRACTION_STATIC"
     --chunked-prefill-size 4096
     # Long AgentX prefills otherwise starve ready decode requests.
-    --prefill-decode-interval 16
+    --prefill-decode-interval "$PREFILL_DECODE_INTERVAL"
     "${SPECULATIVE_ARGS[@]}"
     --max-running-requests "$MAX_RUNNING_REQUESTS"
     --swa-prefix-tails "$SWA_PREFIX_TAILS"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d1c2e6a3cd..988294821a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8599,5 +8599,5 @@
   - "Use the default MTP/DSpark precision shipped by the official nightly, without custom draft quantization or precision patches."
   - "Add DSpark TP2/EP2 alongside TP4/EP4 to explore the vLLM parallelism frontier, keeping native host Engram tables and GPU KV; interleave 16 decode batches between prefill chunks to test long-context responsiveness."
   - "Submit this SGLang AgentX lane through a Slurm batch allocation and launch its steps from the compute node, avoiding failed interactive allocation notifications on the runner login host while preserving GPU reservations and terminal-status checks."
-  - "Test retained SWA tails at 64 times concurrency, bounded to 128–4096; use static memory 0.80 for TP2 C16 and 0.85 at C32 and above after measuring its smaller KV budget and 43.59 GiB of post-graph headroom at C64, otherwise retain 0.70. Prefetch checkpoints for shared Lustre loading."
+  - "Test retained SWA tails at 64 times concurrency, bounded to 128–4096; use static memory 0.80 for TP2 C16 and 0.85 at C32 and above after measuring its smaller KV budget and 43.59 GiB of post-graph headroom at C64, otherwise retain 0.70. Prefetch checkpoints for shared Lustre loading. Test prefill/decode interval 4 at TP2 C32 and above against the interval-16 baseline; keep other points at 16."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3342