From 4bde51e8d4103551c30dd099466c56b4e74a3aae Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 06:43:45 +0000
Subject: [PATCH 1/7] feat: add B300 TP2 DeepSeek V4 Flash serving sweep
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

新增 B300 TP2 DeepSeek V4 Flash SGLang 8k1k serving 扫描，使用原生 EAGLE 三步 MTP。
---
 .../dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml   | 68 ++++++++++++++++
 benchmarks/single_node/srt_fixed_sequence.sh  |  2 +-
 configs/nvidia-master.yaml                    | 15 ++++
 docs/configuration-procedures.md              |  3 +
 docs/configuration-procedures_zh.md           |  3 +
 .../test_dsv4_fixed_sequence_client.py        | 79 +++++++++++++++++++
 perf-changelog.yaml                           |  9 +++
 7 files changed, 178 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
 create mode 100644 infx/tests/srt_slurm/test_dsv4_fixed_sequence_client.py

diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
new file mode 100644
index 0000000000..01d6a0c111
--- /dev/null
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
@@ -0,0 +1,68 @@
+base:
+  schema: 2
+  name: dsv4flash-fp4-b300-sglang-mtp-8k1k
+  model:
+    path: hf:deepseek-ai/DeepSeek-V4-Flash
+    container: lmsysorg/sglang:nightly-dev-cu13-20260925-8ca82118@sha256:baea7ec5ea86412ce59837f17011028b940f9777fa15129254de02bd6a842559
+    precision: fp4
+  resources:
+    gpu_type: b300
+    gpus_per_node: 8
+  frontend:
+    type: sglang
+    enable_multiple_frontends: false
+  observability:
+    enabled: false
+    tachometer:
+      enabled: false
+  engine: sglang
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      gpus: 2
+      args:
+        served-model-name: deepseek-ai/DeepSeek-V4-Flash
+        trust-remote-code: true
+        tensor-parallel-size: 2
+        data-parallel-size: 1
+        expert-parallel-size: 1
+        moe-runner-backend: flashinfer_mxfp4
+        disable-flashinfer-autotune: true
+        disable-radix-cache: true
+        mem-fraction-static: 0.85
+        swa-full-tokens-ratio: 0.1
+        chunked-prefill-size: 8192
+        context-length: 9236
+        reasoning-parser: deepseek-v4
+        tool-call-parser: deepseekv4
+        speculative-algorithm: EAGLE
+        speculative-num-steps: 3
+        speculative-eagle-topk: 1
+        speculative-num-draft-tokens: 4
+        watchdog-timeout: 3600
+        enable-metrics: true
+      env:
+        PYTHONNOUSERSITE: '1'
+        PYTHONUNBUFFERED: '1'
+        SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '0'
+  benchmark:
+    type: custom
+    command: bash /infmax-workspace/benchmarks/single_node/srt_fixed_sequence.sh --dsv4
+    env:
+      MODEL: deepseek-ai/DeepSeek-V4-Flash
+      ISL: '8192'
+      OSL: '1024'
+      RANDOM_RANGE_RATIO: '0.8'
+      USE_CHAT_TEMPLATE: 'true'
+      KV_OFFLOADING: none
+
+zip_override_tp2:
+  roles:
+    agg:
+      args:
+        max-running-requests: [1, 2, 4, 8, 16, 32, 64, 128]
+        cuda-graph-max-bs-decode: [1, 2, 4, 8, 16, 32, 64, 128]
+  benchmark:
+    env:
+      CONC: ['1', '2', '4', '8', '16', '32', '64', '128']
diff --git a/benchmarks/single_node/srt_fixed_sequence.sh b/benchmarks/single_node/srt_fixed_sequence.sh
index 0bc79b9ab4..0a4c511444 100644
--- a/benchmarks/single_node/srt_fixed_sequence.sh
+++ b/benchmarks/single_node/srt_fixed_sequence.sh
@@ -20,7 +20,7 @@ SRT_MONITOR_INTERVAL="$GPU_MONITOR_INTERVAL"
 CLIENT_ARGS=()
 for argument in "$@"; do
     case "$argument" in
-        --trust-remote-code) CLIENT_ARGS+=("$argument") ;;
+        --trust-remote-code|--dsv4) CLIENT_ARGS+=("$argument") ;;
         *) echo "ERROR: unsupported fixed-sequence argument: $argument" >&2; exit 1 ;;
     esac
 done
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index b0fc9ef451..6dc1079d88 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8928,3 +8928,18 @@ dsv41flash-fp4-gb300-sglang-agentic-dspark:
       search-space:
       - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4-mtp/agentic.yaml }
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4-mtp/agentic.yaml }
+
+dsv4flash-fp4-b300-sglang-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260925-8ca82118@sha256:baea7ec5ea86412ce59837f17011028b940f9777fa15129254de02bd6a842559
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  runner: cluster:b300-dsxe
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 2, ep: 1, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml }
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index d480a4bfd7..cae948c781 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -44,6 +44,9 @@ router image does not require changing the worker image. TRT-LLM recipes use nat
 `engine.served_model_name`, without duplicating that flag in `roles.agg.extra_args`.
 The former fork's direct ATOM frontend is not required.
 
+DeepSeek-V4 fixed-sequence recipes pass `--dsv4` to `srt_fixed_sequence.sh` and set
+`USE_CHAT_TEMPLATE: 'true'` to use the shared DeepSeek-V4 encoder rather than a missing HF chat template.
+
 ### Cluster profiles
 
 Launchers that use srt-slurm keep their cluster configuration in
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 47812b4a67..32408f27fd 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -42,6 +42,9 @@ frontend、一个聚合 worker，并设置 `enable_multiple_frontends: false`。
 镜像。TRT-LLM 配方使用原生 `engine.served_model_name`，不再通过 `roles.agg.extra_args`
 重复传入该参数。不再依赖此前分叉中的 ATOM 直连 frontend。
 
+DeepSeek-V4 固定序列长度配方向 `srt_fixed_sequence.sh` 传入 `--dsv4`，并设置
+`USE_CHAT_TEMPLATE: 'true'`，使用共享 DeepSeek-V4 编码器，而不是 checkpoint 中缺失的 HF chat template。
+
 ## 规程索引
 
 1. [准备 worktree](#准备-worktree)
diff --git a/infx/tests/srt_slurm/test_dsv4_fixed_sequence_client.py b/infx/tests/srt_slurm/test_dsv4_fixed_sequence_client.py
new file mode 100644
index 0000000000..bacb33261e
--- /dev/null
+++ b/infx/tests/srt_slurm/test_dsv4_fixed_sequence_client.py
@@ -0,0 +1,79 @@
+"""Exercise the fixed-sequence client's DeepSeek-V4 encoder forwarding."""
+
+import os
+import subprocess
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[3]
+
+
+def test_dsv4_client_forwards_encoder_and_workload(tmp_path: Path) -> None:
+    harness = tmp_path / "harness.sh"
+    harness.write_text(
+        """
+source() {
+    if [[ "$1" == */benchmark_lib.sh && "$2" != --validation-only ]]; then
+        start_gpu_monitor() { :; }
+        stop_gpu_monitor() { :; }
+        run_benchmark_serving() { printf '%s\\n' "$@" > "$CAPTURE"; }
+    else
+        builtin source "$@"
+    fi
+}
+pip3() { :; }
+"""
+    )
+    capture = tmp_path / "arguments"
+    result = subprocess.run(
+        ["bash", str(ROOT / "benchmarks/single_node/srt_fixed_sequence.sh"), "--dsv4"],
+        env={
+            **os.environ,
+            "BASH_ENV": str(harness),
+            "CAPTURE": str(capture),
+            "INFERENCEX_REPO_ROOT": str(ROOT),
+            "MODEL": "test/model",
+            "FRAMEWORK": "sglang",
+            "CONC": "2",
+            "ISL": "256",
+            "OSL": "64",
+            "RANDOM_RANGE_RATIO": "0.8",
+            "RESULT_FILENAME": "point",
+            "RESULT_DIR": str(tmp_path),
+            "SRT_FRONTEND_HOST": "127.0.0.1",
+            "SRT_FRONTEND_PORT": "8000",
+            "RUN_EVAL": "false",
+            "EVAL_ONLY": "false",
+            "GPU_MONITOR_INTERVAL": "3",
+            "USE_CHAT_TEMPLATE": "true",
+        },
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stdout + result.stderr
+    assert capture.read_text().splitlines() == [
+        "--model",
+        "test/model",
+        "--port",
+        "8000",
+        "--base-url",
+        "http://127.0.0.1:8000",
+        "--backend",
+        "vllm",
+        "--input-len",
+        "256",
+        "--output-len",
+        "64",
+        "--random-range-ratio",
+        "0.8",
+        "--num-prompts",
+        "20",
+        "--max-concurrency",
+        "2",
+        "--result-filename",
+        "point",
+        "--result-dir",
+        str(tmp_path),
+        "--dsv4",
+        "--use-chat-template",
+    ]
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 352c61fc0d..7ab7ed8fcf 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8986,3 +8986,12 @@
     - "Hopper 使用 FA3，Blackwell FA4 fp8 descale 问题不适用，EAGLE3 draft 保持 attention_backend FLASH_ATTN。"
     - "This change does not alter the EAGLE3 draft model data type. The draft loads unmodified from the published Inferact/MiniMax-M3-EAGLE3-GQA checkpoint via --speculative-config (method=eagle3). kv-cache-dtype fp8 sets KV-cache storage precision, not the draft weights, and no flag overrides or re-quantizes the draft weights."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3503
+
+- config-keys:
+    - dsv4flash-fp4-b300-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (3 steps, top-k 1, 4 draft tokens), real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder."
+    - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（3 steps、top-k 1、4 draft tokens），采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From db0df130b7ee3f2961ade612487d67a769bfbe81 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 06:44:06 +0000
Subject: [PATCH 2/7] docs: link V4 Flash sweep changelog to PR 3516
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将新增 V4 Flash 扫描记录关联到 PR #3516。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7ab7ed8fcf..b40e7036c6 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8994,4 +8994,4 @@
   description:
     - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (3 steps, top-k 1, 4 draft tokens), real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder."
     - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（3 steps、top-k 1、4 draft tokens），采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516

From 8766bfd1e1219c484a2f11f47ecbbbd56c61dc43 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 06:48:52 +0000
Subject: [PATCH 3/7] fix: match V4 Flash recipe to fixed-sequence CI metadata
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

移除仅适用于 AgentX 的 KV_OFFLOADING 元数据，修复固定序列 CI 配方选择失败。
---
 .../dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml              | 1 -
 perf-changelog.yaml                                      | 9 +++++++++
 2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
index 01d6a0c111..fba3f44057 100644
--- a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
@@ -55,7 +55,6 @@ base:
       OSL: '1024'
       RANDOM_RANGE_RATIO: '0.8'
       USE_CHAT_TEMPLATE: 'true'
-      KV_OFFLOADING: none
 
 zip_override_tp2:
   roles:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b40e7036c6..35b5e9a25a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8995,3 +8995,12 @@
     - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (3 steps, top-k 1, 4 draft tokens), real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder."
     - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（3 steps、top-k 1、4 draft tokens），采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
+
+- config-keys:
+    - dsv4flash-fp4-b300-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Remove AgentX-only KV_OFFLOADING metadata from the fixed-sequence recipe: CI passes an empty value for this scenario, and the explicit 'none' value rejected recipe selection before server launch. Serving settings are unchanged."
+    - "移除固定序列配方中仅适用于 AgentX 的 KV_OFFLOADING 元数据：该场景的 CI 传入空值，显式设置 'none' 导致配方选择在服务启动前失败。Serving 设置保持不变。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516

From cc1c168bc3401ebbade872106c7d0dd243104f45 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 06:54:30 +0000
Subject: [PATCH 4/7] fix: download V4 Flash into B300 runner home before
 submission
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

提交前将 V4 Flash 下载到 B300 runner 的 home 可写目录，保留其他模型路径，并测试下载失败时停止提交。
---
 docs/configuration-procedures.md             |  5 +++
 docs/configuration-procedures_zh.md          |  4 +++
 infx/tests/srt_slurm/test_srt_single_node.py | 34 ++++++++++++++++++--
 perf-changelog.yaml                          |  9 ++++++
 runners/launch_b300-dsxe.sh                  |  5 +++
 5 files changed, 54 insertions(+), 3 deletions(-)

diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index cae948c781..1971249cc1 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -195,6 +195,11 @@ Concurrent cells serialize draft staging with a per-model lock. Each cell lets
 `hf download` validate or resume the existing cache before serving; a nonempty
 directory is not a completion signal.
 
+The B300 native single-node launcher downloads `deepseek-ai/DeepSeek-V4-Flash`
+with `hf download --local-dir` into `/data/home/sa-gha-runner/models/DeepSeek-V4-Flash`
+before SRT submission. Every launch validates or resumes the local download; a failure
+stops submission. This checkpoint does not require a pre-staged `/data/models` copy.
+
 ## Native TileRT power
 
 TileRT's shared importer preserves Docker Hub image names and converts explicit registries such as `ghcr.io/team/image:tag` to Enroot's `docker://ghcr.io#team/image:tag` syntax. Existing `#` references are preserved. Valid cached squash images are reused without importing; a cache hit does not validate the registry import path. Invalid cached images are removed under the import lock before retrying the import.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 32408f27fd..5e5709de15 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -144,6 +144,10 @@ B300 DSXE 的 Kimi-K3 AgentX 路径在 `/scratch/models` 下挂载预置目标
 并发任务通过模型专用锁串行准备草稿权重。每个任务在启动服务前由 `hf download`
 校验或续传现有缓存；目录非空不代表下载完成。
 
+B300 原生单节点启动器在 SRT 提交前，通过 `hf download --local-dir` 将
+`deepseek-ai/DeepSeek-V4-Flash` 下载至 `/data/home/sa-gha-runner/models/DeepSeek-V4-Flash`。
+每次启动都会验证或续传本地下载，下载失败则停止提交，不要求 `/data/models` 中存在预置副本。
+
 ## TileRT 原生功耗
 
 TileRT 的共享导入器保留 Docker Hub 镜像名称，并将 `ghcr.io/team/image:tag` 等显式仓库地址转换为 Enroot 的 `docker://ghcr.io#team/image:tag` 格式。已有的 `#` 地址保持不变。有效的缓存 squash 镜像会直接复用；命中缓存不能证明仓库导入路径有效。无效的缓存镜像会在持有导入锁时删除，再重新导入。
diff --git a/infx/tests/srt_slurm/test_srt_single_node.py b/infx/tests/srt_slurm/test_srt_single_node.py
index d0b160dd5f..c705c64bfa 100644
--- a/infx/tests/srt_slurm/test_srt_single_node.py
+++ b/infx/tests/srt_slurm/test_srt_single_node.py
@@ -260,6 +260,7 @@ def test_submission_manifest(tmp_path, record, expected):
     ("b200-cw", "none"), ("b200-nb", "none"), ("b200-nscale-slurm", "none"),
     ("b200-nscale-slurm", "agentic"),
     ("b300-dsxe", "none"),
+    ("b300-dsxe", "download"), ("b300-dsxe", "download-failure"),
     ("mi300x-amd", "none"), ("mi325x-amds", "none"), ("mi355x-amds", "none"),
 ] + [(pool, "missing-recipe") for pool in (
     "b200-cw", "b200-nb", "b200-nscale-slurm", "b300-dsxe", "h100-cw",
@@ -267,7 +268,12 @@ def test_submission_manifest(tmp_path, record, expected):
     "mi325x-amds", "mi355x-amds",
 )])
 def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path, pool, failure):
-    path, _, point_env = point
+    path, recipe, point_env = point
+    if failure.startswith("download"):
+        point_env = {**point_env, "MODEL": "deepseek-ai/DeepSeek-V4-Flash"}
+        recipe["model"]["path"] = "hf:deepseek-ai/DeepSeek-V4-Flash"
+        recipe["benchmark"]["env"]["MODEL"] = "deepseek-ai/DeepSeek-V4-Flash"
+        path.write_text(yaml.safe_dump({"base": recipe}))
     binaries = tmp_path / "bin"
     binaries.mkdir()
     model = tmp_path / "model"
@@ -279,7 +285,15 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
     # setup/profile/acceptance helpers, binder, and artifact collection.
     scripts = {
         "git": 'if [[ " $* " == *" clone "* ]]; then mkdir -p "${@: -1}/configs"; else echo test-commit; fi',
-        "uv": 'if [[ "$1" == venv ]]; then mkdir -p .venv/bin; echo ":" > .venv/bin/activate; fi',
+        "uv": """
+if [[ "$1" == tool ]]; then
+    printf '%s\\n' "$@" > "$DOWNLOAD_CAPTURE"
+    [[ "$TEST_FAILURE" != download-failure ]] || exit 17
+elif [[ "$1" == venv ]]; then
+    mkdir -p .venv/bin
+    echo ":" > .venv/bin/activate
+fi
+""",
         "make": '[[ "$TEST_FAILURE" == bootstrap ]] && exit 13; mkdir -p bin; touch bin/uv',
         "squeue": '[[ "$TEST_FAILURE" == submission || "$TEST_FAILURE" == agentic ]] && echo "42"; exit 0',
         "salloc": 'echo "Granted job allocation 42"',
@@ -325,6 +339,7 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         "B300_HF_CACHE_CONTAINER_DIR": "/hf", "ENROOT_IMPORT_TIME_LIMIT": "10",
         "INFERENCEX_RUNTIME_ENV_VARS": "REQUIRE_POWER",
         "TEST_FAILURE": failure, "CANCEL_CAPTURE": str(capture),
+        "DOWNLOAD_CAPTURE": str(tmp_path / "download-args"),
         "SRUN_CAPTURE": str(tmp_path / "srun.jsonl"),
         "KEEP_LOGS": "0",
     }
@@ -340,7 +355,16 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         ["bash", str(ROOT / f"runners/launch_{pool}.sh")], cwd=tmp_path,
         env=env, capture_output=True, text=True, timeout=30,
     )
-    assert result.returncode == {"none": 0, "allocation": 1, "submission": 7, "bootstrap": 13, "missing-recipe": 1, "agentic": 0}[failure], result.stderr
+    assert result.returncode == {"none": 0, "allocation": 1, "submission": 7, "bootstrap": 13, "missing-recipe": 1, "agentic": 0, "download": 0, "download-failure": 1}[failure], result.stderr
+    if failure.startswith("download"):
+        assert Path(env["DOWNLOAD_CAPTURE"]).read_text().splitlines() == [
+            "tool", "run", "--from", "huggingface-hub>=0.34,<2", "hf", "download",
+            "deepseek-ai/DeepSeek-V4-Flash", "--local-dir",
+            "/data/home/sa-gha-runner/models/DeepSeek-V4-Flash",
+        ]
+    if failure == "download-failure":
+        assert not (tmp_path / "srt-single-node-submission.json").exists()
+        return
     if failure == "agentic":
         calls = [json.loads(line) for line in Path(env["SRUN_CAPTURE"]).read_text().splitlines()]
         assert calls[-1][-2:] == ["bash", "benchmarks/single_node/agentic/fixture_fp8_b200.sh"]
@@ -364,6 +388,10 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
     cluster_config = yaml.safe_load(next(tmp_path.glob("srt-single.*/checkout/srtslurm.yaml")).read_text())
     assert cluster_config["containers"]["test:tag"] == "test:tag"
     assert cluster_config["use_exclusive_sbatch_directive"] is True
+    if failure == "download":
+        assert cluster_config["model_paths"]["hf:deepseek-ai/DeepSeek-V4-Flash"] == (
+            "/data/home/sa-gha-runner/models/DeepSeek-V4-Flash"
+        )
     assert (capture.read_text() if capture.exists() else "") == ("42\n" if failure == "submission" else "")
 
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 35b5e9a25a..0f9b70cb10 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -9004,3 +9004,12 @@
     - "Remove AgentX-only KV_OFFLOADING metadata from the fixed-sequence recipe: CI passes an empty value for this scenario, and the explicit 'none' value rejected recipe selection before server launch. Serving settings are unchanged."
     - "移除固定序列配方中仅适用于 AgentX 的 KV_OFFLOADING 元数据：该场景的 CI 传入空值，显式设置 'none' 导致配方选择在服务启动前失败。Serving 设置保持不变。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
+
+- config-keys:
+    - dsv4flash-fp4-b300-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Download DeepSeek-V4-Flash into the B300 runner's writable home models directory before native SRT submission instead of requiring a missing /data/models copy. Resume or validate the download on every launch and stop before submission on download failure."
+    - "在原生 SRT 提交前，将 DeepSeek-V4-Flash 下载到 B300 runner 的 home 可写模型目录，不再要求不存在的 /data/models 副本。每次启动均续传或验证下载，下载失败时停止提交。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index efd0506131..137096e38e 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -137,6 +137,11 @@ if [[ "$EXECUTION_PATH" == native-single-node ]]; then
     SRT_MODEL_PATH="$MODEL_ROOT/${MODEL##*/}"
     if [[ "$MODEL" == nvidia/DeepSeek-R1-0528-FP4-V2 ]]; then
         SRT_MODEL_PATH="$MODEL_ROOT/DeepSeek-R1-0528-NVFP4-v2"
+    elif [[ "$MODEL" == deepseek-ai/DeepSeek-V4-Flash ]]; then
+        SRT_MODEL_PATH="$WRITABLE_MODELS_DIR/${MODEL##*/}"
+        # Resume and verify the download before SRT checks the local model path.
+        uv tool run --from 'huggingface-hub>=0.34,<2' hf download "$MODEL" \
+            --local-dir "$SRT_MODEL_PATH" || exit 1
     elif [[ " ${STAGED_MODELS[*]} " != *" ${MODEL##*/} "* || "${MODEL##*/}" == DeepSeek-V4-Pro-0813 ]]; then
         # Not staged on every node's NVMe; read the shared copy.
         SRT_MODEL_PATH="$SHARED_MODEL_ROOT/${MODEL##*/}"

From 32720374252a81f7e4923569f4cc010982a7c847 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 07:03:10 +0000
Subject: [PATCH 5/7] fix: isolate host model download cache from container
 paths
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

主机下载显式使用 home 下的 HF 和 Xet 缓存，避免继承容器路径导致权限错误。
---
 docs/configuration-procedures.md             | 3 +++
 docs/configuration-procedures_zh.md          | 3 +++
 infx/tests/srt_slurm/test_srt_single_node.py | 8 ++++++++
 perf-changelog.yaml                          | 9 +++++++++
 runners/launch_b300-dsxe.sh                  | 3 +++
 5 files changed, 26 insertions(+)

diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 1971249cc1..c7778e80ae 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -199,6 +199,9 @@ The B300 native single-node launcher downloads `deepseek-ai/DeepSeek-V4-Flash`
 with `hf download --local-dir` into `/data/home/sa-gha-runner/models/DeepSeek-V4-Flash`
 before SRT submission. Every launch validates or resumes the local download; a failure
 stops submission. This checkpoint does not require a pre-staged `/data/models` copy.
+The host download explicitly sets `HF_HOME`, `HF_HUB_CACHE`, and `HF_XET_CACHE`
+under `B300_HF_CACHE_HOST_DIR`; container-only cache paths must not leak into the
+host downloader. These overrides are command-scoped so SRT retains its container cache mount.
 
 ## Native TileRT power
 
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 5e5709de15..0a605e0e6f 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -147,6 +147,9 @@ B300 DSXE 的 Kimi-K3 AgentX 路径在 `/scratch/models` 下挂载预置目标
 B300 原生单节点启动器在 SRT 提交前，通过 `hf download --local-dir` 将
 `deepseek-ai/DeepSeek-V4-Flash` 下载至 `/data/home/sa-gha-runner/models/DeepSeek-V4-Flash`。
 每次启动都会验证或续传本地下载，下载失败则停止提交，不要求 `/data/models` 中存在预置副本。
+主机下载命令显式将 `HF_HOME`、`HF_HUB_CACHE` 和 `HF_XET_CACHE` 设置到
+`B300_HF_CACHE_HOST_DIR` 下，防止继承仅适用于容器的缓存路径。这些覆盖仅作用于下载命令，
+不会改变 SRT 的容器缓存挂载。
 
 ## TileRT 原生功耗
 
diff --git a/infx/tests/srt_slurm/test_srt_single_node.py b/infx/tests/srt_slurm/test_srt_single_node.py
index c705c64bfa..f0daf014ef 100644
--- a/infx/tests/srt_slurm/test_srt_single_node.py
+++ b/infx/tests/srt_slurm/test_srt_single_node.py
@@ -288,6 +288,7 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         "uv": """
 if [[ "$1" == tool ]]; then
     printf '%s\\n' "$@" > "$DOWNLOAD_CAPTURE"
+    printf '%s\\n' "$HF_HOME" "$HF_HUB_CACHE" "$HF_XET_CACHE" > "$DOWNLOAD_ENV_CAPTURE"
     [[ "$TEST_FAILURE" != download-failure ]] || exit 17
 elif [[ "$1" == venv ]]; then
     mkdir -p .venv/bin
@@ -340,12 +341,15 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         "INFERENCEX_RUNTIME_ENV_VARS": "REQUIRE_POWER",
         "TEST_FAILURE": failure, "CANCEL_CAPTURE": str(capture),
         "DOWNLOAD_CAPTURE": str(tmp_path / "download-args"),
+        "DOWNLOAD_ENV_CAPTURE": str(tmp_path / "download-env"),
         "SRUN_CAPTURE": str(tmp_path / "srun.jsonl"),
         "KEEP_LOGS": "0",
     }
     env.pop("AIPERF_DRAIN_TIMEOUT_SECONDS", None)
     env.pop("AIPERF_DRAIN_POLL_SECONDS", None)
     env.pop("BENCH_SCRIPT_OVERRIDE", None)
+    if failure.startswith("download"):
+        env.update(HF_HOME="/data/models", HF_HUB_CACHE="/mnt/hf_hub_cache", HF_XET_CACHE="/data/models/xet")
     if failure == "missing-recipe":
         env.pop("SRT_RECIPE")
     if failure == "agentic":
@@ -357,6 +361,9 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
     )
     assert result.returncode == {"none": 0, "allocation": 1, "submission": 7, "bootstrap": 13, "missing-recipe": 1, "agentic": 0, "download": 0, "download-failure": 1}[failure], result.stderr
     if failure.startswith("download"):
+        assert Path(env["DOWNLOAD_ENV_CAPTURE"]).read_text().splitlines() == [
+            str(tmp_path), str(tmp_path / "hub"), str(tmp_path / "xet"),
+        ]
         assert Path(env["DOWNLOAD_CAPTURE"]).read_text().splitlines() == [
             "tool", "run", "--from", "huggingface-hub>=0.34,<2", "hf", "download",
             "deepseek-ai/DeepSeek-V4-Flash", "--local-dir",
@@ -392,6 +399,7 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         assert cluster_config["model_paths"]["hf:deepseek-ai/DeepSeek-V4-Flash"] == (
             "/data/home/sa-gha-runner/models/DeepSeek-V4-Flash"
         )
+        assert cluster_config["default_mounts"][str(tmp_path / "hub")] == "/mnt/hf_hub_cache"
     assert (capture.read_text() if capture.exists() else "") == ("42\n" if failure == "submission" else "")
 
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0f9b70cb10..35315178d1 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -9013,3 +9013,12 @@
     - "Download DeepSeek-V4-Flash into the B300 runner's writable home models directory before native SRT submission instead of requiring a missing /data/models copy. Resume or validate the download on every launch and stop before submission on download failure."
     - "在原生 SRT 提交前，将 DeepSeek-V4-Flash 下载到 B300 runner 的 home 可写模型目录，不再要求不存在的 /data/models 副本。每次启动均续传或验证下载，下载失败时停止提交。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
+
+- config-keys:
+    - dsv4flash-fp4-b300-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Scope HF_HOME, HF_HUB_CACHE, and HF_XET_CACHE to the runner's writable home cache during the host-side V4 Flash download. Preserve the container cache environment and mounts for subsequent SRT execution."
+    - "在主机下载 V4 Flash 时，将 HF_HOME、HF_HUB_CACHE 和 HF_XET_CACHE 限定到 runner 的 home 可写缓存；保留后续 SRT 执行所需的容器缓存环境和挂载。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index 137096e38e..9f7214a411 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -140,6 +140,9 @@ if [[ "$EXECUTION_PATH" == native-single-node ]]; then
     elif [[ "$MODEL" == deepseek-ai/DeepSeek-V4-Flash ]]; then
         SRT_MODEL_PATH="$WRITABLE_MODELS_DIR/${MODEL##*/}"
         # Resume and verify the download before SRT checks the local model path.
+        HF_HOME="$B300_HF_CACHE_HOST_DIR" \
+        HF_HUB_CACHE="$HF_HUB_CACHE_MOUNT" \
+        HF_XET_CACHE="$B300_HF_CACHE_HOST_DIR/xet" \
         uv tool run --from 'huggingface-hub>=0.34,<2' hf download "$MODEL" \
             --local-dir "$SRT_MODEL_PATH" || exit 1
     elif [[ " ${STAGED_MODELS[*]} " != *" ${MODEL##*/} "* || "${MODEL##*/}" == DeepSeek-V4-Pro-0813 ]]; then

From e98dc7dec3a6030c32ed9d8f3542e72416f7d84f Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sun, 27 Sep 2026 18:44:30 +0800
Subject: [PATCH 6/7] fix: load V4 Flash from node-local NVMe and use EAGLE
 2/1/3

The shared Lustre copy is read through mmap page faults at ~10 MB/s, so
TP2 weight loading outlasted the 1800 s health timeout. DeepSeek-V4-Flash
is now staged under /scratch/models on every B300 node; drop the host-side
download path and list it in STAGED_MODELS. Switch MTP to 2 steps, top-k 1,
3 draft tokens.
---
 .../dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml   |  4 +-
 docs/configuration-procedures.md              |  8 ----
 docs/configuration-procedures_zh.md           |  7 ----
 infx/tests/srt_slurm/test_srt_single_node.py  | 42 ++-----------------
 perf-changelog.yaml                           | 31 +-------------
 runners/launch_b300-dsxe.sh                   |  9 +---
 6 files changed, 8 insertions(+), 93 deletions(-)

diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
index fba3f44057..d86ac85958 100644
--- a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
@@ -37,9 +37,9 @@ base:
         reasoning-parser: deepseek-v4
         tool-call-parser: deepseekv4
         speculative-algorithm: EAGLE
-        speculative-num-steps: 3
+        speculative-num-steps: 2
         speculative-eagle-topk: 1
-        speculative-num-draft-tokens: 4
+        speculative-num-draft-tokens: 3
         watchdog-timeout: 3600
         enable-metrics: true
       env:
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index c7778e80ae..cae948c781 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -195,14 +195,6 @@ Concurrent cells serialize draft staging with a per-model lock. Each cell lets
 `hf download` validate or resume the existing cache before serving; a nonempty
 directory is not a completion signal.
 
-The B300 native single-node launcher downloads `deepseek-ai/DeepSeek-V4-Flash`
-with `hf download --local-dir` into `/data/home/sa-gha-runner/models/DeepSeek-V4-Flash`
-before SRT submission. Every launch validates or resumes the local download; a failure
-stops submission. This checkpoint does not require a pre-staged `/data/models` copy.
-The host download explicitly sets `HF_HOME`, `HF_HUB_CACHE`, and `HF_XET_CACHE`
-under `B300_HF_CACHE_HOST_DIR`; container-only cache paths must not leak into the
-host downloader. These overrides are command-scoped so SRT retains its container cache mount.
-
 ## Native TileRT power
 
 TileRT's shared importer preserves Docker Hub image names and converts explicit registries such as `ghcr.io/team/image:tag` to Enroot's `docker://ghcr.io#team/image:tag` syntax. Existing `#` references are preserved. Valid cached squash images are reused without importing; a cache hit does not validate the registry import path. Invalid cached images are removed under the import lock before retrying the import.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 0a605e0e6f..32408f27fd 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -144,13 +144,6 @@ B300 DSXE 的 Kimi-K3 AgentX 路径在 `/scratch/models` 下挂载预置目标
 并发任务通过模型专用锁串行准备草稿权重。每个任务在启动服务前由 `hf download`
 校验或续传现有缓存；目录非空不代表下载完成。
 
-B300 原生单节点启动器在 SRT 提交前，通过 `hf download --local-dir` 将
-`deepseek-ai/DeepSeek-V4-Flash` 下载至 `/data/home/sa-gha-runner/models/DeepSeek-V4-Flash`。
-每次启动都会验证或续传本地下载，下载失败则停止提交，不要求 `/data/models` 中存在预置副本。
-主机下载命令显式将 `HF_HOME`、`HF_HUB_CACHE` 和 `HF_XET_CACHE` 设置到
-`B300_HF_CACHE_HOST_DIR` 下，防止继承仅适用于容器的缓存路径。这些覆盖仅作用于下载命令，
-不会改变 SRT 的容器缓存挂载。
-
 ## TileRT 原生功耗
 
 TileRT 的共享导入器保留 Docker Hub 镜像名称，并将 `ghcr.io/team/image:tag` 等显式仓库地址转换为 Enroot 的 `docker://ghcr.io#team/image:tag` 格式。已有的 `#` 地址保持不变。有效的缓存 squash 镜像会直接复用；命中缓存不能证明仓库导入路径有效。无效的缓存镜像会在持有导入锁时删除，再重新导入。
diff --git a/infx/tests/srt_slurm/test_srt_single_node.py b/infx/tests/srt_slurm/test_srt_single_node.py
index f0daf014ef..d0b160dd5f 100644
--- a/infx/tests/srt_slurm/test_srt_single_node.py
+++ b/infx/tests/srt_slurm/test_srt_single_node.py
@@ -260,7 +260,6 @@ def test_submission_manifest(tmp_path, record, expected):
     ("b200-cw", "none"), ("b200-nb", "none"), ("b200-nscale-slurm", "none"),
     ("b200-nscale-slurm", "agentic"),
     ("b300-dsxe", "none"),
-    ("b300-dsxe", "download"), ("b300-dsxe", "download-failure"),
     ("mi300x-amd", "none"), ("mi325x-amds", "none"), ("mi355x-amds", "none"),
 ] + [(pool, "missing-recipe") for pool in (
     "b200-cw", "b200-nb", "b200-nscale-slurm", "b300-dsxe", "h100-cw",
@@ -268,12 +267,7 @@ def test_submission_manifest(tmp_path, record, expected):
     "mi325x-amds", "mi355x-amds",
 )])
 def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path, pool, failure):
-    path, recipe, point_env = point
-    if failure.startswith("download"):
-        point_env = {**point_env, "MODEL": "deepseek-ai/DeepSeek-V4-Flash"}
-        recipe["model"]["path"] = "hf:deepseek-ai/DeepSeek-V4-Flash"
-        recipe["benchmark"]["env"]["MODEL"] = "deepseek-ai/DeepSeek-V4-Flash"
-        path.write_text(yaml.safe_dump({"base": recipe}))
+    path, _, point_env = point
     binaries = tmp_path / "bin"
     binaries.mkdir()
     model = tmp_path / "model"
@@ -285,16 +279,7 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
     # setup/profile/acceptance helpers, binder, and artifact collection.
     scripts = {
         "git": 'if [[ " $* " == *" clone "* ]]; then mkdir -p "${@: -1}/configs"; else echo test-commit; fi',
-        "uv": """
-if [[ "$1" == tool ]]; then
-    printf '%s\\n' "$@" > "$DOWNLOAD_CAPTURE"
-    printf '%s\\n' "$HF_HOME" "$HF_HUB_CACHE" "$HF_XET_CACHE" > "$DOWNLOAD_ENV_CAPTURE"
-    [[ "$TEST_FAILURE" != download-failure ]] || exit 17
-elif [[ "$1" == venv ]]; then
-    mkdir -p .venv/bin
-    echo ":" > .venv/bin/activate
-fi
-""",
+        "uv": 'if [[ "$1" == venv ]]; then mkdir -p .venv/bin; echo ":" > .venv/bin/activate; fi',
         "make": '[[ "$TEST_FAILURE" == bootstrap ]] && exit 13; mkdir -p bin; touch bin/uv',
         "squeue": '[[ "$TEST_FAILURE" == submission || "$TEST_FAILURE" == agentic ]] && echo "42"; exit 0',
         "salloc": 'echo "Granted job allocation 42"',
@@ -340,16 +325,12 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         "B300_HF_CACHE_CONTAINER_DIR": "/hf", "ENROOT_IMPORT_TIME_LIMIT": "10",
         "INFERENCEX_RUNTIME_ENV_VARS": "REQUIRE_POWER",
         "TEST_FAILURE": failure, "CANCEL_CAPTURE": str(capture),
-        "DOWNLOAD_CAPTURE": str(tmp_path / "download-args"),
-        "DOWNLOAD_ENV_CAPTURE": str(tmp_path / "download-env"),
         "SRUN_CAPTURE": str(tmp_path / "srun.jsonl"),
         "KEEP_LOGS": "0",
     }
     env.pop("AIPERF_DRAIN_TIMEOUT_SECONDS", None)
     env.pop("AIPERF_DRAIN_POLL_SECONDS", None)
     env.pop("BENCH_SCRIPT_OVERRIDE", None)
-    if failure.startswith("download"):
-        env.update(HF_HOME="/data/models", HF_HUB_CACHE="/mnt/hf_hub_cache", HF_XET_CACHE="/data/models/xet")
     if failure == "missing-recipe":
         env.pop("SRT_RECIPE")
     if failure == "agentic":
@@ -359,19 +340,7 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         ["bash", str(ROOT / f"runners/launch_{pool}.sh")], cwd=tmp_path,
         env=env, capture_output=True, text=True, timeout=30,
     )
-    assert result.returncode == {"none": 0, "allocation": 1, "submission": 7, "bootstrap": 13, "missing-recipe": 1, "agentic": 0, "download": 0, "download-failure": 1}[failure], result.stderr
-    if failure.startswith("download"):
-        assert Path(env["DOWNLOAD_ENV_CAPTURE"]).read_text().splitlines() == [
-            str(tmp_path), str(tmp_path / "hub"), str(tmp_path / "xet"),
-        ]
-        assert Path(env["DOWNLOAD_CAPTURE"]).read_text().splitlines() == [
-            "tool", "run", "--from", "huggingface-hub>=0.34,<2", "hf", "download",
-            "deepseek-ai/DeepSeek-V4-Flash", "--local-dir",
-            "/data/home/sa-gha-runner/models/DeepSeek-V4-Flash",
-        ]
-    if failure == "download-failure":
-        assert not (tmp_path / "srt-single-node-submission.json").exists()
-        return
+    assert result.returncode == {"none": 0, "allocation": 1, "submission": 7, "bootstrap": 13, "missing-recipe": 1, "agentic": 0}[failure], result.stderr
     if failure == "agentic":
         calls = [json.loads(line) for line in Path(env["SRUN_CAPTURE"]).read_text().splitlines()]
         assert calls[-1][-2:] == ["bash", "benchmarks/single_node/agentic/fixture_fp8_b200.sh"]
@@ -395,11 +364,6 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
     cluster_config = yaml.safe_load(next(tmp_path.glob("srt-single.*/checkout/srtslurm.yaml")).read_text())
     assert cluster_config["containers"]["test:tag"] == "test:tag"
     assert cluster_config["use_exclusive_sbatch_directive"] is True
-    if failure == "download":
-        assert cluster_config["model_paths"]["hf:deepseek-ai/DeepSeek-V4-Flash"] == (
-            "/data/home/sa-gha-runner/models/DeepSeek-V4-Flash"
-        )
-        assert cluster_config["default_mounts"][str(tmp_path / "hub")] == "/mnt/hf_hub_cache"
     assert (capture.read_text() if capture.exists() else "") == ("42\n" if failure == "submission" else "")
 
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 35315178d1..2a32a37af1 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8992,33 +8992,6 @@
   scenario-type:
     - fixed-seq-len
   description:
-    - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (3 steps, top-k 1, 4 draft tokens), real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder."
-    - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（3 steps、top-k 1、4 draft tokens），采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
-
-- config-keys:
-    - dsv4flash-fp4-b300-sglang-mtp
-  scenario-type:
-    - fixed-seq-len
-  description:
-    - "Remove AgentX-only KV_OFFLOADING metadata from the fixed-sequence recipe: CI passes an empty value for this scenario, and the explicit 'none' value rejected recipe selection before server launch. Serving settings are unchanged."
-    - "移除固定序列配方中仅适用于 AgentX 的 KV_OFFLOADING 元数据：该场景的 CI 传入空值，显式设置 'none' 导致配方选择在服务启动前失败。Serving 设置保持不变。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
-
-- config-keys:
-    - dsv4flash-fp4-b300-sglang-mtp
-  scenario-type:
-    - fixed-seq-len
-  description:
-    - "Download DeepSeek-V4-Flash into the B300 runner's writable home models directory before native SRT submission instead of requiring a missing /data/models copy. Resume or validate the download on every launch and stop before submission on download failure."
-    - "在原生 SRT 提交前，将 DeepSeek-V4-Flash 下载到 B300 runner 的 home 可写模型目录，不再要求不存在的 /data/models 副本。每次启动均续传或验证下载，下载失败时停止提交。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
-
-- config-keys:
-    - dsv4flash-fp4-b300-sglang-mtp
-  scenario-type:
-    - fixed-seq-len
-  description:
-    - "Scope HF_HOME, HF_HUB_CACHE, and HF_XET_CACHE to the runner's writable home cache during the host-side V4 Flash download. Preserve the container cache environment and mounts for subsequent SRT execution."
-    - "在主机下载 V4 Flash 时，将 HF_HOME、HF_HUB_CACHE 和 HF_XET_CACHE 限定到 runner 的 home 可写缓存；保留后续 SRT 执行所需的容器缓存环境和挂载。"
+    - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (2 steps, top-k 1, 3 draft tokens), real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder. Weights load from the node-local /scratch/models copy; mmap page faults on the shared /data copy read at ~10 MB/s and outlast the 1800 s health timeout."
+    - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（2 steps、top-k 1、3 draft tokens），采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。权重从节点本地 /scratch/models 副本加载；共享 /data 副本的 mmap 缺页读取仅约 10 MB/s，会超过 1800 s 健康检查超时。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516
diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index 9f7214a411..6769a3c1aa 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -61,6 +61,7 @@ WRITABLE_MODELS_DIR="/data/home/sa-gha-runner/models"
 STAGED_MODELS=(
     DeepSeek-R1-0528
     DeepSeek-R1-0528-NVFP4-v2
+    DeepSeek-V4-Flash
     DeepSeek-V4-Pro
     DeepSeek-V4-Pro-0813
     DeepSeek-V4-Pro-NVFP4
@@ -137,14 +138,6 @@ if [[ "$EXECUTION_PATH" == native-single-node ]]; then
     SRT_MODEL_PATH="$MODEL_ROOT/${MODEL##*/}"
     if [[ "$MODEL" == nvidia/DeepSeek-R1-0528-FP4-V2 ]]; then
         SRT_MODEL_PATH="$MODEL_ROOT/DeepSeek-R1-0528-NVFP4-v2"
-    elif [[ "$MODEL" == deepseek-ai/DeepSeek-V4-Flash ]]; then
-        SRT_MODEL_PATH="$WRITABLE_MODELS_DIR/${MODEL##*/}"
-        # Resume and verify the download before SRT checks the local model path.
-        HF_HOME="$B300_HF_CACHE_HOST_DIR" \
-        HF_HUB_CACHE="$HF_HUB_CACHE_MOUNT" \
-        HF_XET_CACHE="$B300_HF_CACHE_HOST_DIR/xet" \
-        uv tool run --from 'huggingface-hub>=0.34,<2' hf download "$MODEL" \
-            --local-dir "$SRT_MODEL_PATH" || exit 1
     elif [[ " ${STAGED_MODELS[*]} " != *" ${MODEL##*/} "* || "${MODEL##*/}" == DeepSeek-V4-Pro-0813 ]]; then
         # Not staged on every node's NVMe; read the shared copy.
         SRT_MODEL_PATH="$SHARED_MODEL_ROOT/${MODEL##*/}"

From c9bb03d9a0346f633f08ac80f34d54fe307ec93a Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sun, 27 Sep 2026 23:10:00 +0800
Subject: [PATCH 7/7] perf: run V4 Flash decode and speculative steps without
 graphs

Set cuda-graph-backend-{decode,prefill}=disabled, which also skips target
verify and draft capture, and drop the per-concurrency decode graph batch.
---
 .../srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml | 4 +++-
 perf-changelog.yaml                                           | 4 ++--
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
index d86ac85958..ed7f590936 100644
--- a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b300-fp4-mtp/8k1k.yaml
@@ -40,6 +40,9 @@ base:
         speculative-num-steps: 2
         speculative-eagle-topk: 1
         speculative-num-draft-tokens: 3
+        # Eager decode and speculative verify/draft; no CUDA graph capture.
+        cuda-graph-backend-decode: disabled
+        cuda-graph-backend-prefill: disabled
         watchdog-timeout: 3600
         enable-metrics: true
       env:
@@ -61,7 +64,6 @@ zip_override_tp2:
     agg:
       args:
         max-running-requests: [1, 2, 4, 8, 16, 32, 64, 128]
-        cuda-graph-max-bs-decode: [1, 2, 4, 8, 16, 32, 64, 128]
   benchmark:
     env:
       CONC: ['1', '2', '4', '8', '16', '32', '64', '128']
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2a32a37af1..2f6218f838 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8992,6 +8992,6 @@
   scenario-type:
     - fixed-seq-len
   description:
-    - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (2 steps, top-k 1, 3 draft tokens), real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder. Weights load from the node-local /scratch/models copy; mmap page faults on the shared /data copy read at ~10 MB/s and outlast the 1800 s health timeout."
-    - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（2 steps、top-k 1、3 draft tokens），采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。权重从节点本地 /scratch/models 副本加载；共享 /data 副本的 mmap 缺页读取仅约 10 MB/s，会超过 1800 s 健康检查超时。"
+    - "Add DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled MTP via EAGLE (2 steps, top-k 1, 3 draft tokens) with CUDA/HIP graphs disabled for decode, prefill, and speculative verify/draft, real verification, GPU-resident weights/KV, and the DeepSeek-V4 chat encoder. Weights load from the node-local /scratch/models copy; mmap page faults on the shared /data copy read at ~10 MB/s and outlast the 1800 s health timeout."
+    - "新增 DeepSeek-V4-Flash B300 TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128；通过 EAGLE 使用原生 MTP（2 steps、top-k 1、3 draft tokens），decode、prefill 及投机验证/draft 均关闭 CUDA/HIP graph，采用真实验证、GPU 常驻权重/KV 和 DeepSeek-V4 chat 编码器。权重从节点本地 /scratch/models 副本加载；共享 /data 副本的 mmap 缺页读取仅约 10 MB/s，会超过 1800 s 健康检查超时。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3516