[Klaud Cold] Add H200 DeepSeek V4.1 Flash AgentX / 新增 H200 配方

MODELS.md

@@ -150,6 +150,7 @@ Other offloading tiers, including NVMe KV cache offloading, are outside the init
 
 | Model architecture class | Prefix | Date added | Active scenarios | Deprecated scenarios |
 |---|---|---|---|---|
+| DeepSeek-V4.1-Flash | `dsv41flash` | 2026-09-10 | Agentic coding (DSpark, Engram UVA offload; GPU validation pending) | |
 | Qwen3.8-Flash-Next | `qwen3.8next` | 2026-08-26 ([#2742](https://github.com/SemiAnalysisAI/InferenceX/pull/2742)) | Agentic coding | |
 | Kimi-K3 | `kimik3` | 2026-07-27 ([#2391](https://github.com/SemiAnalysisAI/InferenceX/pull/2391)) | Agentic coding (DSpark only) | Agentic coding non-DSpark arm (deprecated from day 0) |
 | GLM-5.2 | `glm5.2` | 2026-07-18 ([#2268](https://github.com/SemiAnalysisAI/InferenceX/pull/2268)) | Agentic coding (the non-MTP arm still runs while the MTP-only transition remains pending, as explained in the Deprecation Notice) | |

MODELS_zh.md

@@ -150,6 +150,7 @@ InferenceX 支持 SGLang 和 vLLM 双方的维护者，并响应 AI 实验室和
 
 | 模型架构类别 | 前缀 | 加入日期 | 启用场景 | 已弃用场景 |
 |---|---|---|---|---|
+| DeepSeek-V4.1-Flash | `dsv41flash` | 2026-09-10 | 智能体编码（DSpark、Engram UVA 卸载；GPU 待验证） | |
 | Qwen3.8-Flash-Next | `qwen3.8next` | 2026-08-26（[#2742](https://github.com/SemiAnalysisAI/InferenceX/pull/2742)） | 智能体编码 | |
 | Kimi-K3 | `kimik3` | 2026-07-27 ([#2391](https://github.com/SemiAnalysisAI/InferenceX/pull/2391)) | 智能体编码（仅 DSpark） | 智能体编码非 DSpark 分支（自第 0 天起弃用） |
 | GLM-5.2 | `glm5.2` | 2026-07-18（[#2268](https://github.com/SemiAnalysisAI/InferenceX/pull/2268)） | 智能体编码（非 MTP 分支仍在运行，「仅 MTP」转换仍待执行，见弃用公告） | |

benchmarks/benchmark_lib.sh

@@ -22,6 +22,28 @@ INFERENCEX_REPO_ROOT="$(
 # nothing upstream set it.
 export PORT="${PORT:-8888}"
 
+# Opt-in for recipes running in the host network namespace. Probe the preferred
+# port on the compute node; fall back to an OS-selected port if it is occupied.
+# Call immediately before server launch and construct client URLs afterward.
+select_available_server_port() {
+    PORT=$(python3 - "${PORT:-8888}" <<'PYPORT'
+import errno
+import socket
+import sys
+
+with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
+    try:
+        sock.bind(("0.0.0.0", int(sys.argv[1])))
+    except OSError as exc:
+        if exc.errno != errno.EADDRINUSE:
+            raise
+        sock.bind(("0.0.0.0", 0))
+    print(sock.getsockname()[1])
+PYPORT
+    ) || return $?
+    export PORT
+}
+
 agentic_kv_offload_enabled() {
     if [[ -z "${KV_OFFLOADING+x}" || -z "$KV_OFFLOADING" ]]; then
         echo "Error: KV_OFFLOADING must be set for agentic benchmarks" >&2

benchmarks/single_node/agentic/dsv41flash_fp4_h200_vllm_mtp.sh

@@ -0,0 +1 @@
+dsv41flash_fp4_vllm_mtp.sh

benchmarks/single_node/agentic/dsv41flash_fp4_vllm_mtp.sh

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# Native DeepSeek-V4.1-Flash DSpark and Engram UVA weight offload.
+# https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash
+source "$(dirname "$0")/../../benchmark_lib.sh"
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+require_agentic_kv_offload_none
+export GPU_COUNT="$TP"
+
+# 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="$RESULT_DIR/server.log"
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_USE_RUST_FRONTEND=1
+export PYTHONUNBUFFERED=1
+
+# Preserve the upstream scheduler defaults; size graph capture for the sweep.
+NUM_SPEC_TOKENS=5
+CAPTURE_SIZE=1
+while (( CAPTURE_SIZE < CONC * (1 + NUM_SPEC_TOKENS) && CAPTURE_SIZE < 2048 )); do
+    CAPTURE_SIZE=$((CAPTURE_SIZE * 2))
+done
+
+# 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="vllm:"
+echo "Using vLLM endpoint ${AIPERF_SERVER_URL}"
+
+# Golden AL: golden_al_distribution/dsv41flash_dspark.yaml, thinking_on, five draft tokens.
+# Accuracy evals keep real block rejection; throughput fixes acceptance to AL 3.51.
+if [[ "${EVAL_ONLY:-false}" == true ]]; then
+    SPEC_CONFIG='{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"block","enable_adaptive_verification":true}'
+else
+    SPEC_CONFIG='{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.51,"enable_adaptive_verification":false}'
+fi
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
+    --host 0.0.0.0 --port "$PORT" --tensor-parallel-size "$TP"
+    --language-model-only
+    --tokenizer-mode deepseek_v41
+    --tool-call-parser deepseek_v41 --enable-auto-tool-choice
+    --reasoning-parser deepseek_v41
+    --engram-config '{"cpu_offload":true}'
+    --speculative-config "$SPEC_CONFIG"
+    --max-model-len 1048576
+    --max-cudagraph-capture-size "$CAPTURE_SIZE"
+    --disable-uvicorn-access-log
+)
+printf '%q ' "${VLLM_CMD[@]}" | tee "$RESULT_DIR/vllm_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/vllm_command.txt"
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [[ "${EVAL_ONLY:-false}" == true ]]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi

configs/nvidia-master.yaml

@@ -10213,3 +10213,27 @@ glm5.1-fp8-b200-tilert-agentic:
           dp-attn: false
           additional-settings:
           - "DECODE_NODES=1"
+
+# H200 AgentX arm for DeepSeek-V4.1-Flash. Upstream marks h200 verified and says
+# the GB200 NVL4 TP4 layout becomes TP8 on 8-GPU nodes, so this is TP8.
+# `precision: fp4` labels the checkpoint's MXFP4 routed experts, as on the
+# Blackwell and MI355X arms; Hopper has no FP4 tensor cores, so the MXFP4
+# experts run through the upconverting MoE path. The H100 arm is deferred: the
+# sparse attention indexer allocates a [max-num-batched-tokens, max-model-len]
+# buffer at startup, which is 16 GiB at 1M context and does not fit 80 GB
+# alongside the weights. Tracked separately.
+dsv41flash-fp4-h200-vllm-agentic-dspark:
+  image: vllm/vllm-openai:deepseekv41-flash-0909
+  model: deepseek-ai/DeepSeek-V4.1-Flash
+  model-prefix: dsv41flash
+  runner: cluster:h200-dgxc
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      # 8x141 GB holds the 511 GB checkpoint minus the UVA-offloaded Engram
+      # tables, so the KV cache stays GPU-resident across the full range.
+      - { tp: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }

docs/configuration-procedures.md

@@ -167,6 +167,53 @@ Sources: [`AGENTS.md#non-negotiable-benchmark-invariants`](../AGENTS.md#non-nego
 7. Add script + master entry + launcher routing + changelog together.
 8. Run Bash syntax and generation checks. Inspect `spec-decoding`, draft/native method, token count, chat-template use, capture range, and resolved script.
 
+### DeepSeek-V4.1-Flash DSpark on H200
+
+`dsv41flash-fp4-h200-vllm-agentic-dspark` is the H200 AgentX arm of the
+DeepSeek-V4.1-Flash recipe. It shares `vllm/vllm-openai:deepseekv41-flash-0909` and the
+text-only serving script with the Blackwell arms: `deepseek_v41` tokenizer and parsers,
+1M context, native five-token DSpark with probabilistic drafting. Throughput uses the [committed golden AL](../golden_al_distribution/dsv41flash_dspark.yaml) of 3.51 for thinking on and five draft tokens, with synthetic rejection sampling and adaptive verification disabled. Accuracy evals retain real block rejection and adaptive verification.
+
+The arm runs **TP8**, not the upstream TP4. Upstream verifies TP4 on one GB200 NVL4 tray
+and states that the same layout becomes TP8 per role on 8-GPU nodes, which is what an
+H200 DGXC node is.
+
+`precision: fp4` labels the checkpoint's MXFP4 routed expert weights, matching the
+Blackwell and MI355X arms on the identical checkpoint. Hopper has no FP4 tensor cores, so
+those weights run through the upconverting MoE path; the label describes the checkpoint,
+not the SKU's native arithmetic.
+
+`--engram-config '{"cpu_offload":true}'` keeps the Engram tables in pinned host DRAM
+reached through UVA, and `kv-offloading: none` describes the separate, GPU-resident KV
+cache. Measured on the cluster, the offload moves 11.80 GiB per rank per table for two
+tables across 8 ranks — 188.8 GiB — leaving roughly 35.9 GiB per GPU of resident weights
+out of 141 GiB.
+
+Trace corpus: the arm replays the uncapped `semianalysis_cc_traces_weka_062126` corpus,
+not the 256k-capped `..._062126_256k` variant, because the model serves 1M context. The
+recipe never names a corpus — `resolve_trace_source` picks the uncapped default only
+because its `dsv4*` case arm also matches the `dsv41flash` prefix. That is load-bearing
+and invisible at the call site, so `runners/test_dsv41flash_h200.py` pins it; narrowing
+the arm would silently downgrade this recipe's traces.
+
+**Why there is no H100 arm.** H100 is not in the upstream hardware table, and the
+blocker is not the weights. At 1M context the sparse attention indexer allocates a
+`[max-num-batched-tokens, max-model-len]` logits buffer in
+`fp8_fp4_paged_mqa_logits`, which at the default 8192 batched tokens is exactly 16 GiB.
+That is a fixed startup cost paid during memory profiling, independent of concurrency, so
+it fails at concurrency 1 on an 80 GB card even though the resident weights fit. Capping
+`--max-num-batched-tokens` shrinks it proportionally; capping `--max-model-len` would too,
+but that would force the 256k-capped trace corpus. An H100 arm needs its own script rather
+than the shared symlink, and is tracked separately.
+
+The launcher mounts the repository at `/ix` for this recipe so AgentX runtime directories
+are not created under `/workspace`, and it already mounts the shared HF cache, so the
+script resolves the model through `HF_HUB_CACHE` rather than a per-node path. The recipe
+probes the serving port on the compute node and selects an available one if the preferred
+port is occupied; serving, replay, metrics, and eval share that endpoint.
+
+Source: [upstream recipe](https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml).
+
 ## Validate
 
 Run the smallest checks that cover the edited layers.

docs/configuration-procedures_zh.md

@@ -167,6 +167,45 @@ llm-d 不是 srt-slurm 路径：InferenceX 自己持有 Slurm allocation，并
 7. 同时添加脚本 + 主配置条目 + launcher 路由 + changelog。
 8. 运行 Bash 语法和生成检查；检查 `spec-decoding`、draft/native 方法、token 数、chat-template 使用、capture 范围和解析出的脚本。
 
+### H200 上的 DeepSeek-V4.1-Flash DSpark
+
+`dsv41flash-fp4-h200-vllm-agentic-dspark` 是 DeepSeek-V4.1-Flash 配方的 H200 AgentX
+分支。它与 Blackwell 分支共用 `vllm/vllm-openai:deepseekv41-flash-0909` 和纯文本服务
+脚本：`deepseek_v41` tokenizer 和解析器、1M 上下文、原生五 token DSpark（概率采样草稿）。吞吐测试使用[已提交的黄金 AL](../golden_al_distribution/dsv41flash_dspark.yaml)：thinking 开启、五个草稿 token 对应 3.51，采用合成拒绝采样并关闭自适应验证。准确率 eval 保留真实块拒绝采样和自适应验证。
+
+该分支使用 **TP8**，而非上游的 TP4。上游在一个 GB200 NVL4 tray 上验证 TP4，并说明在
+8 GPU 节点上同一布局每个角色变为 TP8，而 H200 DGXC 节点正是 8 GPU 节点。
+
+`precision: fp4` 标记检查点中 MXFP4 的路由专家权重，与同一检查点的 Blackwell 和
+MI355X 分支保持一致。Hopper 没有 FP4 tensor core，因此这些权重走上转换的 MoE 路径；
+该标签描述检查点，而非 SKU 的原生算力。
+
+`--engram-config '{"cpu_offload":true}'` 将 Engram 表放在固定页主机 DRAM 中，通过 UVA
+访问；`kv-offloading: none` 描述的是另行驻留 GPU 的 KV cache。集群实测：卸载在 8 个 rank
+上为两张表各移出每 rank 11.80 GiB，共 188.8 GiB，使每 GPU 的驻留权重从 141 GiB 中约占
+35.9 GiB。
+
+轨迹语料：该分支回放未截断的 `semianalysis_cc_traces_weka_062126` 语料，而不是 256k
+截断的 `..._062126_256k` 变体，因为该模型服务 1M 上下文。配方本身并未指定语料 ——
+`resolve_trace_source` 选中未截断的默认值，仅仅是因为其 `dsv4*` 分支同时匹配了
+`dsv41flash` 前缀。这一依赖在调用处并不可见却至关重要，因此由
+`runners/test_dsv41flash_h200.py` 固定；收窄该分支会静默地降级本配方的轨迹。
+
+**为何没有 H100 分支。** H100 不在上游硬件表中，且瓶颈不在权重。在 1M 上下文下，稀疏
+注意力 indexer 会在 `fp8_fp4_paged_mqa_logits` 中分配一个
+`[max-num-batched-tokens, max-model-len]` 的 logits 缓冲区，在默认 8192 batched tokens
+下恰好为 16 GiB。这是显存 profiling 阶段固定支付的启动开销，与并发无关，因此即使驻留
+权重放得下，在 80 GB 卡上并发 1 也会失败。收窄 `--max-num-batched-tokens` 可按比例缩小
+该缓冲区；收窄 `--max-model-len` 同样有效，但会强制使用 256k 截断的轨迹语料。H100 分支
+需要独立脚本而非共享符号链接，另行跟踪。
+
+launcher 为该配方将仓库挂载到 `/ix`，避免在 `/workspace` 下创建 AgentX 运行目录；它本来
+就挂载了共享 HF 缓存，因此脚本通过 `HF_HUB_CACHE` 解析模型，而不依赖各节点的独立路径。
+配方在计算节点探测服务端口，首选端口被占用时选择可用端口，服务、回放、指标和 eval 共用
+同一端点。
+
+来源：[上游配方](https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml)。
+
 ## 验证
 
 运行覆盖被修改层的最小检查。

perf-changelog.yaml

@@ -7146,6 +7146,13 @@
     - "The release image ships the same CUDA 13.0.3, FlashInfer 0.6.18 and sgl-kernel 0.4.6.post1 as the nightly. benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh is unchanged: SGLANG_ENABLE_SPEC_V2 EAGLE MTP at 3 steps, golden acceptance length 3.39, flashinfer attention with allreduce fusion, fp8 quantization and fp8_e4m3 KV, HiCache kernel IO / page_first layout. TP8/EP1 DRAM HiCache concurrency 2 through 24 unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2966
 
+- config-keys:
+    - dsv41flash-fp4-h200-vllm-agentic-dspark
+  description:
+    - "Add the H200 AgentX arm for DeepSeek-V4.1-Flash at TP8 with native five-token DSpark and adaptive real verification"
+    - "Engram tables offload to host DRAM through UVA while the KV cache stays GPU-resident; concurrency 1-128 on the uncapped 062126 trace corpus"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2963
+
 - config-keys:
     - glm5.2-fp4-b300-sglang-agentic-mtp
   scenario-type:

runners/launch_h200-dgxc-slurm.sh

@@ -462,11 +462,17 @@ else
         BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
     fi
 
-    if [[ "$IMAGE" == *deepseek-v4-hopper* ]]; then
+    # DeepSeek-V4.1-Flash creates AgentX runtime directories next to the
+    # repository, which must not land under /workspace.
+    if [[ "$IMAGE" == *deepseek-v4-hopper* || "$MODEL_PREFIX" == "dsv41flash" ]]; then
         CONTAINER_MOUNT_DIR=/ix
     else
         CONTAINER_MOUNT_DIR=/workspace
     fi
+    if [[ "$MODEL_PREFIX" == "dsv41flash" ]]; then
+        export INFMAX_CONTAINER_WORKSPACE=/ix
+        export RESULT_DIR=/ix/results
+    fi
 
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \

runners/test_dsv41flash_h200.py

@@ -0,0 +1,107 @@
+import json
+import subprocess
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+BENCHMARK_LIB = REPO_ROOT / "benchmarks" / "benchmark_lib.sh"
+
+LAUNCH_HARNESS = '''
+    salloc() { :; }
+    squeue() { echo 123; }
+    srun() {
+        python3 -c 'import json,os,sys; open(sys.argv[1], "a").write(json.dumps({"args": sys.argv[2:], "result_dir": os.environ.get("RESULT_DIR", "")})+"\\n")' "$SRUN_LOG" "$@"
+    }
+    scancel() { :; }
+    export IS_MULTINODE=false IS_AGENTIC=1 SCENARIO_SUBDIR=agentic/
+    export IMAGE=vllm/vllm-openai:deepseekv41-flash-0909
+    export HF_HUB_CACHE=/mnt/hf_hub_cache/ RESULT_DIR=/workspace/results
+    export GITHUB_WORKSPACE="$1" SRUN_LOG="$2"
+    cd "$GITHUB_WORKSPACE"
+'''
+
+
+def launch(sku: str, log: Path, **env: str) -> dict:
+    exports = " ".join(f"export {key}={value};" for key, value in env.items())
+    result = subprocess.run(
+        [
+            "bash",
+            "-c",
+            f"{LAUNCH_HARNESS}\n{exports}\nsource runners/launch_{sku}-dgxc-slurm.sh",
+            "bash",
+            str(REPO_ROOT),
+            str(log),
+        ],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+    assert result.returncode == 0, result.stderr
+    # The last srun call is the benchmark launch; earlier ones import the image.
+    return json.loads(log.read_text().splitlines()[-1])
+
+
+def test_h200_flash_runs_the_vllm_script_from_an_ix_mount(tmp_path: Path) -> None:
+    sku = "h200"
+    serve = launch(
+        sku,
+        tmp_path / "launch.jsonl",
+        MODEL_PREFIX="dsv41flash",
+        MODEL="deepseek-ai/DeepSeek-V4.1-Flash",
+        PRECISION="fp4",
+        FRAMEWORK="vllm",
+        SPEC_DECODING="mtp",
+        TP="8",
+        RUNNER_NAME=f"{sku}-test",
+        EXP_NAME="dsv41flash_tp8_conc1",
+    )
+    script = serve["args"][-1]
+    assert script == f"benchmarks/single_node/agentic/dsv41flash_fp4_{sku}_vllm_mtp.sh"
+    assert (REPO_ROOT / script).is_file()
+
+    mounts = next(arg for arg in serve["args"] if arg.startswith("--container-mounts="))
+    assert f"{REPO_ROOT}:/ix/," in mounts
+    assert ":/mnt/hf_hub_cache/," in mounts
+    # AgentX must not write runtime directories under /workspace.
+    assert "/workspace" not in mounts
+    assert serve["result_dir"] == "/ix/results"
+    assert "--container-workdir=/ix/" in serve["args"]
+
+
+def resolve_loader(model_prefix: str) -> str:
+    """The public-dataset loader `resolve_trace_source` picks for a prefix."""
+    result = subprocess.run(
+        [
+            "bash",
+            "-c",
+            # Stub the CLI bootstrap *after* sourcing, so the real one does
+            # not overwrite the stub, and no dataset is actually downloaded.
+            'source "$BENCHMARK_LIB"; ensure_hf_cli() { AIPERF_HF_CLI=true; }; '
+            "resolve_trace_source",
+        ],
+        env={
+            "PATH": "/usr/bin:/bin:/usr/local/bin",
+            "HOME": str(REPO_ROOT),
+            "BENCHMARK_LIB": str(BENCHMARK_LIB),
+            "MODEL_PREFIX": model_prefix,
+        },
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+    return result.stdout.split("public-dataset: ")[1].split(" ")[0]
+
+
+def test_flash_replays_the_uncapped_1m_trace_corpus() -> None:
+    """DeepSeek-V4.1-Flash serves 1M context, so it must not get the 256k corpus.
+
+    This lives with the recipe rather than in the benchmark_lib tests because
+    the recipe never names a corpus: it inherits one from `resolve_trace_source`
+    only because the `dsv4*` case arm also matches the `dsv41flash` prefix.
+    Narrowing that arm would silently downgrade this recipe's traces.
+    """
+    assert resolve_loader("dsv41flash") == "semianalysis_cc_traces_weka_062126"
+
+
+def test_short_context_families_still_get_the_capped_corpus() -> None:
+    """The uncapped default is context-driven, not a blanket default."""
+    assert resolve_loader("qwen3.8next") == "semianalysis_cc_traces_weka_062126_256k"