Add B200 DeepSeek V4.1 Flash AgentX / 新增 B200 配方

benchmarks/single_node/agentic/dsv41flash_fp4_b200_vllm_mtp.sh

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

configs/nvidia-master.yaml

@@ -10214,6 +10214,21 @@ glm5.1-fp8-b200-tilert-agentic:
           additional-settings:
           - "DECODE_NODES=1"
 
+dsv41flash-fp4-b200-vllm-agentic-dspark:
+  image: vllm/vllm-openai:deepseekv41-flash-0909
+  model: deepseek-ai/DeepSeek-V4.1-Flash
+  model-prefix: dsv41flash
+  runner: cluster:b200-nscale
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      # Engram weights use UVA DRAM; the KV cache stays GPU-resident.
+      - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
+
 dsv41flash-fp4-gb300-vllm-agentic-dspark:
   image: vllm/vllm-openai:deepseekv41-flash-0909
   model: deepseek-ai/DeepSeek-V4.1-Flash

perf-changelog.yaml

@@ -7212,3 +7212,23 @@
   description:
     - "Allow 7200 seconds for GB300 DSv4.1 Flash engine startup; the one-hour frontend deadline interrupted cold graph capture"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2961
+
+- config-keys:
+    - dsv41flash-fp4-b200-vllm-agentic-dspark
+  description:
+    - "Add B200 TP4 AgentX through concurrency 128 with Engram UVA DRAM offload and native five-token DSpark"
+    - "Use deepseekv41-flash-0909 with adaptive verification and available-port selection"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2960
+
+- config-keys:
+    - dsv41flash-fp4-b200-vllm-agentic-dspark
+  description:
+    - "Align with the official single-node TP recipe: default model runner and scheduler batching"
+    - "Keep AgentX concurrency 1-128, UVA offload, DSpark, and concurrency-sized graph capture"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2960
+
+- config-keys:
+    - dsv41flash-fp4-b200-vllm-agentic-dspark
+  description:
+    - "Use thinking-on golden synthetic AL 3.51 for five-token DSpark throughput; disable adaptive verification and retain real verification for evals"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2960

runners/launch_b200-nscale-compat.sh

@@ -17,7 +17,11 @@ set -x
 # portability, but we resolve to pre-staged paths here to avoid repeated
 # downloading on every Nscale node. Runs for both single-node and multinode
 # launches.
-if [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp4" ]]; then
+if [[ "$MODEL_PREFIX" == "dsv41flash" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "vllm" && "$IS_MULTINODE" != "true" ]]; then
+    export MODEL_PATH="$MODEL"
+    export HF_HUB_CACHE_HOST_PATH="/data/home/sa-shared/gharunners/hf-hub-cache"
+    mkdir -p "$HF_HUB_CACHE_HOST_PATH"
+elif [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH="/scratch/models/DeepSeek-R1-0528-NVFP4-v2"
     export SRT_SLURM_MODEL_PREFIX="dsr1"
 elif [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
@@ -526,6 +530,16 @@ else
         CONTAINER_MOUNT_DIR=/workspace
     fi
 
+    if [[ "$MODEL_PREFIX" == "dsv41flash" ]]; then
+        CONTAINER_MOUNT_DIR=/ix
+        export INFMAX_CONTAINER_WORKSPACE=/ix
+        export RESULT_DIR=/ix/results
+        export HF_HUB_CACHE=/hf-cache
+        CONTAINER_MOUNTS="$GITHUB_WORKSPACE:/ix,$HF_HUB_CACHE_HOST_PATH:/hf-cache,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache"
+    else
+        CONTAINER_MOUNTS="$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache"
+    fi
+
     # The runner lease reserves the Slurm nodes before this single-node job is
     # submitted to the Nscale batch_1 partition.
     export GPU_COUNT="${GPU_COUNT:-${TP:?TP must be set}}"
@@ -556,7 +570,7 @@ else
 
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \
-        --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache \
+        --container-mounts="$CONTAINER_MOUNTS" \
         --no-container-mount-home \
         --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888,AIPERF_DATASET_MMAP_CACHE_DIR=/aiperf_mmap_cache \

runners/test_dsv41flash_b200.py

@@ -0,0 +1,48 @@
+import json
+import subprocess
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+
+
+def run_bash(command: str, *args: Path | str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["bash", "-c", command, "bash", *(str(arg) for arg in args)],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+
+
+def test_b200_v41_uses_hf_cache_without_mounting_model_id(tmp_path: Path) -> None:
+    log = tmp_path / "launch.jsonl"
+    result = run_bash(
+        '''
+        mkdir() { :; }
+        salloc() { :; }
+        squeue() { echo 123; }
+        srun() {
+            python3 -c 'import json,os,sys; open(sys.argv[1], "a").write(json.dumps({"args":sys.argv[2:], "model":os.environ["MODEL"], "cache":os.environ["HF_HUB_CACHE"]})+"\\n")' "$SRUN_LOG" "$@"
+        }
+        export MODEL_PREFIX=dsv41flash PRECISION=fp4 FRAMEWORK=vllm
+        export MODEL=deepseek-ai/DeepSeek-V4.1-Flash IS_MULTINODE=false
+        export SPEC_DECODING=mtp TP=4 RUNNER_NAME=b200-test IS_AGENTIC=1
+        export SCENARIO_SUBDIR=agentic/ EXP_NAME=dsv41flash_tp4_conc1
+        export IMAGE=vllm/test:fixture GITHUB_WORKSPACE="$1" SRUN_LOG="$2"
+        cd "$GITHUB_WORKSPACE"
+        source runners/launch_b200-nscale-compat.sh
+        ''',
+        REPO_ROOT, log,
+    )
+    assert result.returncode == 0, result.stderr
+    serve = json.loads(log.read_text().splitlines()[-1])
+    assert serve["model"] == "deepseek-ai/DeepSeek-V4.1-Flash"
+    assert serve["cache"] == "/hf-cache"
+    mounts = next(arg for arg in serve["args"] if arg.startswith("--container-mounts="))
+    assert f"{REPO_ROOT}:/ix," in mounts
+    assert ":/hf-cache," in mounts
+    assert "deepseek-ai/" not in mounts
+    assert serve["args"][-2] == "bash"
+    assert (REPO_ROOT / serve["args"][-1]).is_file()