[Klaud Cold] Add H200 Qwen3.8-27B bf16 vLLM TP1 DSpark 1k1k / 新增 H200 Qwen3.8-27B bf16 vLLM TP1 DSpark 1k1k 配方

benchmarks/single_node/fixed_seq_len/qwen3.827b_bf16_h200_vllm_mtp.sh

@@ -0,0 +1,111 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# Qwen3.8-27B (bf16, dense hybrid attention: 48 linear-attention and 16 full
+# attention layers) on one H200, served by vLLM with the RadixArk DSpark drafter
+# (Doopeworld's vLLM-loadable copy) drafting seven tokens per step.
+# https://recipes.vllm.ai/Qwen/Qwen3.8-27B
+# https://huggingface.co/Doopeworld/Qwen3.8-27B-DSpark-vLLM
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC ISL OSL RANDOM_RANGE_RATIO RESULT_FILENAME
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$TP" -ne 1 ]]; then
+  echo "This recipe serves Qwen3.8-27B on a single GPU; got TP=$TP" >&2
+  exit 1
+fi
+
+DRAFT_MODEL="Doopeworld/Qwen3.8-27B-DSpark-vLLM"
+# The drafter's trained block size. Its card measured k=4 and k=6 slower than
+# k=7 despite the late positions' low acceptance: per-step overhead dominates
+# the per-drafted-token cost.
+NUM_SPEC_TOKENS=7
+
+nvidia-smi
+
+# Complete/resume partial downloads instead of trusting nonempty directories.
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+hf download "$DRAFT_MODEL"
+
+SERVER_LOG=/workspace/server.log
+
+# Serve the matrix context (isl + osl + slack), not the checkpoint's 262K;
+# accuracy evals use the eval context benchmark_lib derives.
+MODEL_LEN="${MAX_MODEL_LEN:-$((ISL + OSL + 256))}"
+if [[ "${EVAL_ONLY:-false}" == true ]]; then
+    setup_eval_context
+    MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+# vLLM's default max-num-seqs (1024) exceeds the GDN/Mamba cache blocks that fit
+# next to the bf16 weights (472 on an 80 GB H100, run 35357364404) and engine
+# start aborts before graph capture. Size the scheduler batch to the sweep point
+# instead; the accuracy eval serves up to 256 concurrent requests.
+MAX_NUM_SEQS=$(( CONC > 16 ? CONC : 16 ))
+if [[ "${EVAL_ONLY:-false}" == true ]]; then
+    MAX_NUM_SEQS=256
+fi
+
+# Pyxis shares the host network; port 8888 can already belong to a host service.
+select_available_server_port
+
+# Probabilistic draft sampling measured ~23% faster than greedy on the drafter's
+# card. Adaptive verification is rejected by vLLM's GDN attention backend for
+# this hybrid architecture, so it stays at its default (off).
+SPEC_CONFIG=$(printf '{"method":"dspark","model":"%s","num_speculative_tokens":%d,"draft_sample_method":"probabilistic"}' "$DRAFT_MODEL" "$NUM_SPEC_TOKENS")
+
+start_gpu_monitor
+
+VLLM_CMD=(
+    vllm serve "$MODEL" --served-model-name "$MODEL"
+    --host 0.0.0.0 --port "$PORT"
+    --tensor-parallel-size 1
+    # Text-only serving: skip the vision tower of Qwen3_5ForConditionalGeneration.
+    --language-model-only
+    --trust-remote-code
+    --kv-cache-dtype fp8
+    --max-model-len "$MODEL_LEN"
+    --max-num-seqs "$MAX_NUM_SEQS"
+    # Every 1k1k request prefills its full random prompt; no prefix-cache hits.
+    --no-enable-prefix-caching
+    --reasoning-parser qwen3
+    --enable-auto-tool-choice --tool-call-parser qwen3_xml
+    --speculative-config "$SPEC_CONFIG"
+    --disable-uvicorn-access-log
+)
+printf '%q ' "${VLLM_CMD[@]}" | tee /workspace/vllm_command.txt
+printf '\n' | tee -a /workspace/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 [[ "${EVAL_ONLY:-false}" == true ]]; then
+    run_eval --framework lm-eval --port "$PORT"
+else
+    pip install -q datasets pandas
+    run_benchmark_serving \
+        --model "$MODEL" \
+        --port "$PORT" \
+        --backend vllm \
+        --input-len "$ISL" \
+        --output-len "$OSL" \
+        --random-range-ratio "$RANDOM_RANGE_RATIO" \
+        --num-prompts "$((CONC * 10))" \
+        --max-concurrency "$CONC" \
+        --result-filename "$RESULT_FILENAME" \
+        --result-dir /workspace/ \
+        `# Chat-templated prompts: raw random tokens tank draft acceptance.` \
+        --use-chat-template \
+        --server-pid "$SERVER_PID"
+    if [[ "${RUN_EVAL:-false}" == true ]]; then
+        run_eval --framework lm-eval --port "$PORT"
+        append_lm_eval_summary
+    fi
+fi
+
+stop_gpu_monitor

configs/nvidia-master.yaml

@@ -8150,3 +8150,18 @@ dsv41flash-fp4-h200-vllm-agentic-dspark:
       # 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] }
+
+qwen3.827b-bf16-h200-vllm-mtp:
+  image: vllm/vllm-openai:nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3
+  model: Qwen/Qwen3.8-27B
+  model-prefix: qwen3.827b
+  runner: cluster:h200-dgxc
+  precision: bf16
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 1, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }

perf-changelog.yaml

@@ -8082,3 +8082,17 @@
     - "Capture only full decode CUDA graphs (VLLM_USE_BREAKABLE_CUDAGRAPH=0 with cudagraph_mode FULL_DECODE_ONLY, as the MiniMax-M3 gfx942 arm does) and restore --moe-backend aiter: on gfx942 every worker segfaulted during piecewise graph capture with both the Triton W4A16 MoE kernel (run 35305045778) and the auto-selected unfused Triton kernel (run 35306398350), so the capture mode rather than the MoE kernel is the failing piece; prefill runs eagerly"
     - "仅捕获完整的 decode CUDA graph（VLLM_USE_BREAKABLE_CUDAGRAPH=0 并设置 cudagraph_mode FULL_DECODE_ONLY，与 MiniMax-M3 gfx942 配方一致）并恢复 --moe-backend aiter：在 gfx942 上，无论使用 Triton W4A16 MoE 内核（运行 35305045778）还是自动选择的未融合 Triton 内核（运行 35306398350），所有 worker 都在 piecewise graph 捕获期间段错误，说明问题在于捕获模式而非 MoE 内核；prefill 以 eager 方式运行"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3247
+
+- config-keys:
+    - qwen3.827b-bf16-h200-vllm-mtp
+  description:
+    - "Add the H200 Qwen3.8-27B bf16 recipe: vLLM nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3 on one GPU (TP1), 1k1k only, with the RadixArk DSpark drafter Doopeworld/Qwen3.8-27B-DSpark-vLLM (a config-only copy vLLM loads as Qwen3DSparkModel) drafting seven tokens with probabilistic draft sampling, fp8 KV cache, the matrix context as --max-model-len, --language-model-only, and chat-templated prompts; concurrency 1-128"
+    - "新增 H200 Qwen3.8-27B bf16 配方：vLLM nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3 单卡（TP1），仅 1k1k，使用 RadixArk 的 DSpark 草稿模型 Doopeworld/Qwen3.8-27B-DSpark-vLLM（仅修改配置以便 vLLM 按 Qwen3DSparkModel 加载）预测 7 个 token 并采用 probabilistic 草稿采样，fp8 KV cache，--max-model-len 使用矩阵上下文，--language-model-only，提示词经 chat template 处理；并发 1-128"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3261
+
+- config-keys:
+    - qwen3.827b-bf16-h200-vllm-mtp
+  description:
+    - "Set --max-num-seqs to the sweep point (floor 16; 256 for accuracy evals) and disable prefix caching (--no-enable-prefix-caching): in run 35357364404 the H100 canary died at engine start because vLLM's default max-num-seqs of 1024 exceeded the 472 GDN/Mamba cache blocks that fit next to the bf16 weights on the 80 GB card (each decode sequence needs one block), and prefix caching is off so every 1k1k request prefills its full random prompt"
+    - "将 --max-num-seqs 设为并发点（下限 16；精度评测为 256）并关闭 prefix caching（--no-enable-prefix-caching）：运行 35357364404 中 H100 canary 在引擎启动时失败，因为 vLLM 默认的 max-num-seqs 1024 超过了 80 GB 显卡上 bf16 权重旁仅剩的 472 个 GDN/Mamba cache block（每个 decode 序列需要一个 block）；关闭 prefix caching 使每个 1k1k 请求完整 prefill 其随机提示词"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3261