From d24d369ed6bddd0391a77db9185129125f8c4c51 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 07:08:13 +0000
Subject: [PATCH 1/6] feat: add B200 TP2 DeepSeek V4 Flash serving sweep
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

新增 B200 TP2 DeepSeek V4 Flash serving 扫描，在共享可写目录下载模型并隔离 HF/Xet 缓存。
---
 .../dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml   | 67 ++++++++++++++++
 benchmarks/single_node/srt_fixed_sequence.sh  |  2 +-
 configs/nvidia-master.yaml                    | 15 ++++
 docs/configuration-procedures.md              | 11 +++
 docs/configuration-procedures_zh.md           |  9 +++
 .../test_dsv4_fixed_sequence_client.py        | 79 +++++++++++++++++++
 infx/tests/srt_slurm/test_srt_single_node.py  | 43 +++++++++-
 perf-changelog.yaml                           |  9 +++
 runners/launch_b200-nscale-slurm.sh           |  9 +++
 9 files changed, 240 insertions(+), 4 deletions(-)
 create mode 100644 benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-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/b200-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml
new file mode 100644
index 0000000000..5d15a30a20
--- /dev/null
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml
@@ -0,0 +1,67 @@
+base:
+  schema: 2
+  name: dsv4flash-fp4-b200-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: b200
+    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'
+
+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..1410fbb7cd 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-b200-sglang-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260925-8ca82118@sha256:baea7ec5ea86412ce59837f17011028b940f9777fa15129254de02bd6a842559
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  runner: cluster:b200-nscale
+  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/b200-fp4-mtp/8k1k.yaml }
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index d480a4bfd7..8eee6d72c0 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -46,6 +46,17 @@ The former fork's direct ATOM frontend is not required.
 
 ### Cluster profiles
 
+The B200 Nscale fixed-sequence `deepseek-ai/DeepSeek-V4-Flash` launcher stages the
+checkpoint with `hf download --local-dir` under
+`/data/home/sa-shared/gharunners/models/DeepSeek-V4-Flash` before SRT submission.
+Each launch resumes or validates that download; failure stops submission. This
+shared writable location does not require a pre-staged `/scratch/models` copy.
+The download process uses explicit writable `HF_HOME`, `HF_HUB_CACHE`, and
+`HF_XET_CACHE` paths beneath the checkpoint's `.cache/huggingface` directory;
+the serving process retains its container cache settings.
+The recipe uses bundled MTP through `EAGLE` (3 steps, top-k 1, 4 draft tokens),
+with the DeepSeek-V4 chat encoder selected by the client's `--dsv4` option.
+
 Launchers that use srt-slurm keep their cluster configuration in
 [`runners/srt-slurm/<launcher>.yaml`](../runners/srt-slurm/). The native settings
 (GPU count, scheduling directives, aliases, and mounts) are separate from workload recipes.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 47812b4a67..eabc5c28c9 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -42,6 +42,15 @@ 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。
 
+B200 Nscale 的固定序列 `deepseek-ai/DeepSeek-V4-Flash` 启动器在提交 SRT 前，
+通过 `hf download --local-dir` 将 checkpoint 下载至
+`/data/home/sa-shared/gharunners/models/DeepSeek-V4-Flash`。
+每次启动均续传或验证下载，失败时停止提交；此共享可写目录不依赖
+`/scratch/models` 中的预置副本。下载进程显式将 `HF_HOME`、`HF_HUB_CACHE`
+和 `HF_XET_CACHE` 指向 checkpoint 的 `.cache/huggingface` 下的可写路径，
+服务进程仍保留容器缓存设置。配方通过 `EAGLE` 使用原生 MTP
+（3 steps、top-k 1、4 draft tokens），客户端以 `--dsv4` 选择 DeepSeek-V4 chat 编码器。
+
 ## 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/infx/tests/srt_slurm/test_srt_single_node.py b/infx/tests/srt_slurm/test_srt_single_node.py
index d0b160dd5f..3b9d265aba 100644
--- a/infx/tests/srt_slurm/test_srt_single_node.py
+++ b/infx/tests/srt_slurm/test_srt_single_node.py
@@ -259,6 +259,7 @@ def test_submission_manifest(tmp_path, record, expected):
     ("h200-cw", "none"), ("h100-cw", "none"), ("h100-dgxc-slurm", "none"),
     ("b200-cw", "none"), ("b200-nb", "none"), ("b200-nscale-slurm", "none"),
     ("b200-nscale-slurm", "agentic"),
+    ("b200-nscale-slurm", "download"), ("b200-nscale-slurm", "download-failure"),
     ("b300-dsxe", "none"),
     ("mi300x-amd", "none"), ("mi325x-amds", "none"), ("mi355x-amds", "none"),
 ] + [(pool, "missing-recipe") for pool in (
@@ -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 +285,16 @@ 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 [[ "$1" == tool ]]; then
+    printf '%s\\n' "$@" > "$DOWNLOAD_CAPTURE"
+    printf '%s\\n' "$HF_HOME" "$HF_HUB_CACHE" "$HF_XET_CACHE" > "$DOWNLOAD_CACHE_CAPTURE"
+    [[ "$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 +340,9 @@ 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_CACHE_CAPTURE": str(tmp_path / "download-cache"),
+        "HF_HOME": "/read-only/hf", "HF_XET_CACHE": "/read-only/xet",
         "SRUN_CAPTURE": str(tmp_path / "srun.jsonl"),
         "KEEP_LOGS": "0",
     }
@@ -340,7 +358,21 @@ 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-shared/gharunners/models/DeepSeek-V4-Flash",
+        ]
+        assert Path(env["DOWNLOAD_CACHE_CAPTURE"]).read_text().splitlines() == [
+            "/data/home/sa-shared/gharunners/models/DeepSeek-V4-Flash/.cache/huggingface",
+            "/data/home/sa-shared/gharunners/models/DeepSeek-V4-Flash/.cache/huggingface/hub",
+            "/data/home/sa-shared/gharunners/models/DeepSeek-V4-Flash/.cache/huggingface/xet",
+        ]
+    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 +396,11 @@ 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-shared/gharunners/models/DeepSeek-V4-Flash"
+        )
+        assert cluster_config["default_mounts"]["/data/home/sa-shared/gharunners/hf-hub-cache"] == "/hf"
     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 352c61fc0d..75c330e277 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-b200-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (3 steps, top-k 1, 4 draft tokens), real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
+    - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（3 steps、top-k 1、4 draft tokens）、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 1c6c7afc04..4f10c19f6b 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -90,6 +90,15 @@ if [[ "$LAUNCH_PATH" == "native-srt" ]]; then
             export SRT_SLURM_MODEL_PREFIX="glm5.1-fp8"
             ;;
     esac
+elif [[ "$LAUNCH_PATH" == native-single-node && "$MODEL" == deepseek-ai/DeepSeek-V4-Flash ]]; then
+    export MODEL_PATH="/data/home/sa-shared/gharunners/models/${MODEL##*/}"
+    # Stage on writable shared storage before SRT validates the local path.
+    # Host downloads must not inherit container-only or read-only cache paths.
+    HF_HOME="$MODEL_PATH/.cache/huggingface" \
+    HF_HUB_CACHE="$MODEL_PATH/.cache/huggingface/hub" \
+    HF_XET_CACHE="$MODEL_PATH/.cache/huggingface/xet" \
+    uv tool run --from 'huggingface-hub>=0.34,<2' hf download "$MODEL" \
+        --local-dir "$MODEL_PATH" || exit 1
 elif [[ "$MODEL_PREFIX" == "dsv41flash" && "$PRECISION" == "fp4" && ( "$FRAMEWORK" == "vllm" || "$FRAMEWORK" == "sglang" ) && "$IS_MULTINODE" != "true" ]]; then
     export MODEL_PATH="$MODEL"
     export HF_HUB_CACHE_HOST_PATH="/data/home/sa-shared/gharunners/hf-hub-cache"

From c359fcf6220d47e99c6c4a8f1250b4266bb66171 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 07:08:37 +0000
Subject: [PATCH 2/6] docs: link B200 sweep changelog to PR 3517
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 B200 扫描变更记录关联至 PR #3517。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 75c330e277..07f9100d20 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8994,4 +8994,4 @@
   description:
     - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (3 steps, top-k 1, 4 draft tokens), real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
     - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（3 steps、top-k 1、4 draft tokens）、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3517

From 840c69f1f8e22df939214c06b98ed4404db61f3f Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sun, 27 Sep 2026 07:24:10 +0000
Subject: [PATCH 3/6] fix: normalize B200 Flash image digest for Pyxis
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 B200 Flash 镜像 digest 转为 Pyxis/Enroot 原生语法，保留固定 digest 和有效 squash 缓存优先级。
---
 docs/configuration-procedures.md             |  3 ++
 docs/configuration-procedures_zh.md          |  3 ++
 infx/tests/srt_slurm/test_srt_single_node.py | 23 ++++++++++++---
 perf-changelog.yaml                          |  9 ++++++
 runners/launch_b200-nscale-slurm.sh          | 31 +++++++++++++-------
 5 files changed, 54 insertions(+), 15 deletions(-)

diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 8eee6d72c0..06785960ad 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -54,6 +54,9 @@ shared writable location does not require a pre-staged `/scratch/models` copy.
 The download process uses explicit writable `HF_HOME`, `HF_HUB_CACHE`, and
 `HF_XET_CACHE` paths beneath the checkpoint's `.cache/huggingface` directory;
 the serving process retains its container cache settings.
+For an uncached digest-pinned image, this lane maps Docker's `repo:tag@sha256:...`
+to Pyxis/Enroot's `repo:sha256:...` manifest reference. The recorded image and
+digest stay unchanged; a valid cached squash image still takes precedence.
 The recipe uses bundled MTP through `EAGLE` (3 steps, top-k 1, 4 draft tokens),
 with the DeepSeek-V4 chat encoder selected by the client's `--dsv4` option.
 
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index eabc5c28c9..f7894acf9e 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -50,6 +50,9 @@ B200 Nscale 的固定序列 `deepseek-ai/DeepSeek-V4-Flash` 启动器在提交 S
 和 `HF_XET_CACHE` 指向 checkpoint 的 `.cache/huggingface` 下的可写路径，
 服务进程仍保留容器缓存设置。配方通过 `EAGLE` 使用原生 MTP
 （3 steps、top-k 1、4 draft tokens），客户端以 `--dsv4` 选择 DeepSeek-V4 chat 编码器。
+对于尚未缓存的 digest 固定镜像，此路径将 Docker 的 `repo:tag@sha256:...`
+转换为 Pyxis/Enroot 的 `repo:sha256:...` manifest 引用。记录的镜像和 digest
+保持不变，已有有效 squash 镜像仍优先使用。
 
 ## 规程索引
 
diff --git a/infx/tests/srt_slurm/test_srt_single_node.py b/infx/tests/srt_slurm/test_srt_single_node.py
index 3b9d265aba..501de0eb20 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"),
     ("b200-nscale-slurm", "download"), ("b200-nscale-slurm", "download-failure"),
+    ("b200-nscale-slurm", "download-cached"),
     ("b300-dsxe", "none"),
     ("mi300x-amd", "none"), ("mi325x-amds", "none"), ("mi355x-amds", "none"),
 ] + [(pool, "missing-recipe") for pool in (
@@ -270,10 +271,16 @@ def test_submission_manifest(tmp_path, record, expected):
 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"}
+        point_env = {
+            **point_env, "MODEL": "deepseek-ai/DeepSeek-V4-Flash",
+            "IMAGE": f"example/server:nightly@sha256:{'a' * 64}",
+        }
         recipe["model"]["path"] = "hf:deepseek-ai/DeepSeek-V4-Flash"
+        recipe["model"]["container"] = point_env["IMAGE"]
         recipe["benchmark"]["env"]["MODEL"] = "deepseek-ai/DeepSeek-V4-Flash"
         path.write_text(yaml.safe_dump({"base": recipe}))
+    if failure == "download-cached":
+        (tmp_path / f"example_server_nightly_sha256_{'a' * 64}.sqsh").write_text("cache")
     binaries = tmp_path / "bin"
     binaries.mkdir()
     model = tmp_path / "model"
@@ -301,6 +308,7 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
         "sacct": 'if [[ "$TEST_FAILURE" == allocation ]]; then echo "FAILED|1:0"; else echo "COMPLETED|0:0"; fi',
         "scancel": 'printf "%s\\n" "$@" >> "$CANCEL_CAPTURE"',
         "tail": 'exit 0',
+        "unsquashfs": '[[ "$TEST_FAILURE" == download-cached ]]',
     }
     for name, script in scripts.items():
         binary = binaries / name
@@ -358,7 +366,7 @@ 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
+    assert result.returncode == {"none": 0, "allocation": 1, "submission": 7, "bootstrap": 13, "missing-recipe": 1, "agentic": 0, "download": 0, "download-failure": 1, "download-cached": 0}[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",
@@ -394,9 +402,16 @@ def test_pool_launcher_stages_artifacts_and_propagates_failure(point, tmp_path,
     assert json.loads((tmp_path / "gpu_metrics_context.json").read_text()) == {"device_count": 4}
     assert (tmp_path / "srt-single-node-logs.tar.gz").stat().st_size > 0
     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["containers"][point_env["IMAGE"]] == f"example/server:sha256:{'a' * 64}"
+    elif failure == "download-cached":
+        assert cluster_config["containers"][point_env["IMAGE"]] == str(
+            tmp_path / f"example_server_nightly_sha256_{'a' * 64}.sqsh"
+        )
+    else:
+        assert cluster_config["containers"]["test:tag"] == "test:tag"
+    assert cluster_config["use_exclusive_sbatch_directive"] is True
+    if failure in {"download", "download-cached"}:
         assert cluster_config["model_paths"]["hf:deepseek-ai/DeepSeek-V4-Flash"] == (
             "/data/home/sa-shared/gharunners/models/DeepSeek-V4-Flash"
         )
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 07f9100d20..cb7cf9087e 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8995,3 +8995,12 @@
     - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (3 steps, top-k 1, 4 draft tokens), real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
     - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（3 steps、top-k 1、4 draft tokens）、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3517
+
+- config-keys:
+    - dsv4flash-fp4-b200-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Fix B200 V4 Flash single-node Pyxis image import by converting uncached Docker @sha256 references to Enroot manifest-tag syntax. Preserve the pinned image digest and prefer valid cached squash images. No serving settings change."
+    - "修复 B200 V4 Flash 单节点 Pyxis 镜像导入：将尚未缓存的 Docker @sha256 引用转换为 Enroot manifest-tag 语法，保留固定镜像 digest 并优先使用有效 squash 缓存，不更改 serving 设置。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3517
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 4f10c19f6b..78b39b34f7 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -159,17 +159,6 @@ else
     exit 1
 fi
 
-if [[ "$LAUNCH_PATH" == native-single-node ]]; then
-    HF_HUB_CACHE_MOUNT=/data/home/sa-shared/gharunners/hf-hub-cache
-    SRT_MODEL_PATH="$MODEL_PATH"
-    # Models not staged locally resolve through the Hugging Face cache mount.
-    [[ "$SRT_MODEL_PATH" == /* ]] || SRT_MODEL_PATH="hf:$MODEL"
-    SRT_SQUASH_FILE="$B200_SQUASH_DIR/$(printf '%s' "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
-    launch_srt_single_node b200-nscale-slurm \
-        --var SLURM_ACCOUNT "$SLURM_ACCOUNT" --var SLURM_PARTITION "$SLURM_PARTITION"
-    exit $?
-fi
-
 # ---------------------------------------------------------------------------
 # Container import helpers shared by both srt-slurm paths
 # ---------------------------------------------------------------------------
@@ -202,6 +191,26 @@ enroot_uri_for_image() {
     fi
 }
 
+if [[ "$LAUNCH_PATH" == native-single-node ]]; then
+    HF_HUB_CACHE_MOUNT=/data/home/sa-shared/gharunners/hf-hub-cache
+    SRT_MODEL_PATH="$MODEL_PATH"
+    # Models not staged locally resolve through the Hugging Face cache mount.
+    [[ "$SRT_MODEL_PATH" == /* ]] || SRT_MODEL_PATH="hf:$MODEL"
+    SRT_SQUASH_FILE="$B200_SQUASH_DIR/$(printf '%s' "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    SRT_IMAGE_ARGS=()
+    if [[ "$MODEL" == deepseek-ai/DeepSeek-V4-Flash && "$IMAGE" == *@sha256:* ]] &&
+       ! { [[ -r "$SRT_SQUASH_FILE" ]] && unsquashfs -s "$SRT_SQUASH_FILE" >/dev/null 2>&1; }; then
+        # Pyxis interprets Docker's @digest as credentials; preserve the digest
+        # using Enroot's manifest-tag syntax, without its docker:// prefix.
+        SRT_IMAGE_URI=$(enroot_uri_for_image "$IMAGE") || exit 1
+        SRT_IMAGE_ARGS=(--container "$IMAGE" "${SRT_IMAGE_URI#docker://}")
+    fi
+    launch_srt_single_node b200-nscale-slurm \
+        --var SLURM_ACCOUNT "$SLURM_ACCOUNT" --var SLURM_PARTITION "$SLURM_PARTITION" \
+        "${SRT_IMAGE_ARGS[@]}"
+    exit $?
+fi
+
 # Import containers via enroot, serialized so concurrent runners on this
 # cluster don't race on the same squash file.
 import_squash() {

From 5f6084fd5e6fb6889b40149d572acd2045d80f7b Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sun, 27 Sep 2026 18:44:32 +0800
Subject: [PATCH 4/6] feat: use EAGLE MTP 2 steps, top-k 1, 3 draft tokens

---
 .../srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml | 4 ++--
 perf-changelog.yaml                                           | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml
index 5d15a30a20..600906e395 100644
--- a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-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/perf-changelog.yaml b/perf-changelog.yaml
index cb7cf9087e..6f21b10050 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8992,8 +8992,8 @@
   scenario-type:
     - fixed-seq-len
   description:
-    - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (3 steps, top-k 1, 4 draft tokens), real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
-    - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（3 steps、top-k 1、4 draft tokens）、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
+    - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (2 steps, top-k 1, 3 draft tokens), real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
+    - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（2 steps、top-k 1、3 draft tokens）、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3517
 
 - config-keys:

From c5ad47ad32a62187e27b88245d9e3fea114426a5 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sun, 27 Sep 2026 22:32:34 +0800
Subject: [PATCH 5/6] docs: record EAGLE 2/1/3 for the B200 V4 Flash recipe

---
 docs/configuration-procedures.md    | 2 +-
 docs/configuration-procedures_zh.md | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 06785960ad..5e79d69428 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -57,7 +57,7 @@ the serving process retains its container cache settings.
 For an uncached digest-pinned image, this lane maps Docker's `repo:tag@sha256:...`
 to Pyxis/Enroot's `repo:sha256:...` manifest reference. The recorded image and
 digest stay unchanged; a valid cached squash image still takes precedence.
-The recipe uses bundled MTP through `EAGLE` (3 steps, top-k 1, 4 draft tokens),
+The recipe uses bundled MTP through `EAGLE` (2 steps, top-k 1, 3 draft tokens),
 with the DeepSeek-V4 chat encoder selected by the client's `--dsv4` option.
 
 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 f7894acf9e..c566c9da96 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -49,7 +49,7 @@ B200 Nscale 的固定序列 `deepseek-ai/DeepSeek-V4-Flash` 启动器在提交 S
 `/scratch/models` 中的预置副本。下载进程显式将 `HF_HOME`、`HF_HUB_CACHE`
 和 `HF_XET_CACHE` 指向 checkpoint 的 `.cache/huggingface` 下的可写路径，
 服务进程仍保留容器缓存设置。配方通过 `EAGLE` 使用原生 MTP
-（3 steps、top-k 1、4 draft tokens），客户端以 `--dsv4` 选择 DeepSeek-V4 chat 编码器。
+（2 steps、top-k 1、3 draft tokens），客户端以 `--dsv4` 选择 DeepSeek-V4 chat 编码器。
 对于尚未缓存的 digest 固定镜像，此路径将 Docker 的 `repo:tag@sha256:...`
 转换为 Pyxis/Enroot 的 `repo:sha256:...` manifest 引用。记录的镜像和 digest
 保持不变，已有有效 squash 镜像仍优先使用。

From 2e425b6e40d5a082013e3c2866482e3b2e5b83ad Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Sun, 27 Sep 2026 23:10:02 +0800
Subject: [PATCH 6/6] 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/b200-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/b200-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml
index 600906e395..053e84466b 100644
--- a/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-fp4-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv4flash/sglang/b200-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 6f21b10050..edc8572d4c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8992,8 +8992,8 @@
   scenario-type:
     - fixed-seq-len
   description:
-    - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (2 steps, top-k 1, 3 draft tokens), real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
-    - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（2 steps、top-k 1、3 draft tokens）、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
+    - "Add DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving at concurrency 1, 2, 4, 8, 16, 32, 64, 128 with bundled EAGLE MTP (2 steps, top-k 1, 3 draft tokens) with CUDA/HIP graphs disabled for decode, prefill, and speculative verify/draft, real verification, and GPU-resident weights/KV. Stage the unstaged checkpoint and HF/Xet download caches on writable shared storage before SRT submission; stop on download failure."
+    - "新增 DeepSeek-V4-Flash B200 Nscale TP2 SGLang 8k1k serving，并发为 1、2、4、8、16、32、64、128，采用原生 EAGLE MTP（2 steps、top-k 1、3 draft tokens），decode、prefill 及投机验证/draft 均关闭 CUDA/HIP graph、真实验证和 GPU 常驻权重/KV。SRT 提交前在共享可写目录下载 checkpoint 并设置 HF/Xet 下载缓存，下载失败时停止提交。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3517
 
 - config-keys: