Agentsop Vllm

vLLM Serving SOP

1. 何时激活 (When to activate)

Activate this skill when any of the following hold:

• The user wants to serve an LLM in production (multi-user, concurrent requests, throughput-oriented) and has GPU infrastructure.
• The user is comparing inference engines (vLLM vs TGI vs SGLang vs TensorRT-LLM vs llama.cpp/Ollama).
• The user reports a vLLM operational issue: CUDA OOM, low throughput, high TTFT, request preemption, multi-GPU setup, quantization choice.
• The user is sizing hardware for an open-weights model (Llama / Qwen / Mixtral / DeepSeek-V3) and asking about tensor/pipeline parallelism.
• The user mentions PagedAttention, prefix caching, continuous batching, chunked prefill, or speculative decoding.

Do NOT activate for: training/fine-tuning (use accelerate/deepspeed/trl), CPU-only edge inference (use llama.cpp/Ollama), Apple Silicon production (vLLM Metal/MPS is experimental, not production-ready as of 2026) [aimadetools.com 2026], API-only consumption of hosted models (just call the OpenAI/Anthropic SDK).

2. 核心心智模型 (Core Mental Model)

2.1 The OS abstraction: KV cache as virtual memory

vLLM's defining insight (Kwon et al., SOSP 2023) is that LLM serving's bottleneck was not compute — it was KV-cache memory fragmentation. Pre-vLLM systems pre-allocated a contiguous KV-cache slot per request, sized for the maximum possible output length; in early 2023, inference engines used only 20–40% of available GPU memory because of internal+external fragmentation [arxiv.org/abs/2309.06180; zilliz.com/learn].

PagedAttention applies classic OS paging to KV cache:

• Block = fixed-size chunk of KV cache (default 16 tokens; ~12.8 KB for a 13B model) [medium.com/@mandeep0405].
• Logical blocks per request → block table → physical blocks in GPU memory (analogous to virtual→physical page table).
• Blocks need not be contiguous. The attention kernel reads scattered physical blocks via the block table and presents them as a logical contiguous sequence.
• Copy-on-write + prefix sharing: multiple requests that share a prefix (e.g. a system prompt, few-shot examples) share KV blocks; a write triggers a per-request copy [arxiv.org/abs/2309.06180].

Result: near-zero memory waste → larger batch sizes → 2–4× throughput vs FasterTransformer/Orca at equal latency [arxiv.org/abs/2309.06180]; 14–24× vs vanilla HuggingFace Transformers [yottalabs.ai 2026].

2.2 The scheduler: iteration-level (continuous) batching

vLLM inherits Orca's iteration-level scheduling (OSDI 2022, 36.9× over FasterTransformer [medium.com/byte-sized-ai]). Instead of waiting for a static batch to finish, the scheduler reassigns batch slots every decode step: a request that finishes early frees its slot to a waiting request. Static batching is dead; continuous batching is table stakes.

2.3 Prefill vs decode are different beasts

• Prefill (processing the prompt): compute-bound, high SM utilization, scales with input length.
• Decode (generating tokens one-at-a-time): memory-bandwidth-bound, low SM utilization.
• A long prefill blocks all decodes on the GPU → head-of-line blocking → high inter-token latency (ITL) for already-streaming requests.
• Chunked prefill breaks long prefills into pieces interleaved with decode steps [docs.vllm.ai/en/stable/configuration/optimization/].

Takeaway: when tuning, separate TTFT (time-to-first-token, gated by prefill+queue) from ITL (gated by decode bandwidth and batch interference).

2.4 Three throughput levers, in order of impact

Per Red Hat's tuning hierarchy [developers.redhat.com 2026]:

1. Right-size the model (smallest adequate).
2. Scale hardware (more replicas; better-bandwidth GPUs).
3. Quantize (FP8 weights + KV cache).
4. Speculative decoding (model-based: EAGLE-3 / MTP).
5. Refine parallelism (TP degree, replicas vs higher TP).

3. SOP 工作流 (SOP Workflow)

[Step 0] Confirm vLLM is the right tool
   ├─ Production, GPU-backed, concurrent users? → continue
   └─ Else → see §7 (ecosystem) and stop

[Step 1] Pick the model + precision
   ├─ Model fits in single-GPU VRAM at BF16?         → keep BF16, TP=1
   ├─ Need 50% VRAM cut, ~zero quality loss?         → FP8 (Hopper/Ada+) [arxiv 2411.02355]
   ├─ Need 4× VRAM cut, tolerate ~1.6pt avg drop?    → AWQ-4 or GPTQ-4
   └─ Reasoning-heavy / coding workload?             → favor FP8 > AWQ; verify on eval set

[Step 2] Choose parallelism
   ├─ Fits 1 GPU                                     → TP=1, PP=1
   ├─ Fits 1 node, NVLink present                    → TP=#GPUs/node
   ├─ Fits 1 node, only PCIe (e.g. L40S)             → PP within node (TP-only over PCIe collapses)
   ├─ Multi-node                                     → TP=GPUs/node, PP=#nodes
   └─ MoE model (Mixtral, DSv3)                      → DP attention + EP/TP for MoE layers

[Step 3] Set memory/batch envelope
   ├─ --gpu-memory-utilization 0.90 (default; 0.85 if sharing GPU)
   ├─ --max-model-len = (longest realistic prompt + output) — NOT model max!
   ├─ --max-num-seqs (start 256; lower if preemption logs appear)
   └─ --max-num-batched-tokens (raise for TTFT; lower for ITL)

[Step 4] Turn on the free wins
   ├─ enable_prefix_caching=True            → if any system-prompt/few-shot reuse
   ├─ enable_chunked_prefill=True (V1: default on) → tame long-prompt HoL blocking
   └─ kv_cache_dtype="fp8"                  → +KV headroom, Ampere+ only

[Step 5] Optional: speculative decoding
   ├─ Low QPS, latency-bound, have draft/EAGLE weights? → EAGLE-3 or MTP (high gain)
   ├─ No draft model, zero setup cost?                  → n-gram (modest gain, ~1.17×)
   └─ High QPS / large batch?                           → skip; gains shrink, complexity rises

[Step 6] Benchmark on YOUR workload
   ├─ Replay representative ISL/OSL distribution
   ├─ Watch Prometheus: num_requests_waiting, KV cache occupancy, preemption count
   └─ Tune in this order: §3 Step 3 → Step 4 → Step 5 → reconsider Step 2

[Step 7] Scale out
   ├─ Latency SLA violated under load? → add a replica (data parallelism across pods)
   ├─ Tail TTFT high?                  → prefix-aware routing; pin prefix to replica
   └─ Cost too high?                   → revisit quantization, smaller model, draft model

4. 操作模型 (Operation Model: Trigger / Action / Output / Evidence)

OP-1: Diagnose vLLM CUDA OOM at startup

• Trigger: log shows torch.OutOfMemoryError: CUDA out of memory during engine init or warmup.
• Action:
  1. Lower --max-model-len to the realistic max (prompt + output), not the model's architectural max [markaicode.com 2026].
  2. Lower --gpu-memory-utilization to 0.85 if other processes share the GPU; raise to 0.95 if vLLM is alone and KV cache is too small.
  3. Add --kv-cache-dtype fp8 (Ampere+) or --enforce-eager (skip CUDA-graph reservation).
  4. Reduce --max-num-seqs (e.g. 256 → 64 → 16).
  5. If still OOM at TP=1: bump --tensor-parallel-size to shard the model.
• Output: engine starts; KV cache occupancy logged at boot is non-zero and <100%.
• Evidence: [docs.vllm.ai/en/stable/configuration/conserving_memory/]; [discuss.vllm.ai/t/cuda-failure-out-of-memory/524]; [github.com/vllm-project/vllm/issues/188].

OP-2: Pick quantization for an open-weights model

• Trigger: deploying a model that doesn't fit at BF16, or where cost/throughput matters more than peak quality.
• Action:
  1. Hopper/Ada/Blackwell GPU available? → prefer FP8 (effectively lossless across model scales) [arxiv.org/abs/2411.02355].
  2. No FP8 hardware, need 4× compression? → AWQ-4 (better stability at 4-bit) or GPTQ-4 (slightly better on code tasks) [arxiv.org/abs/2411.02355].
  3. Avoid INT3 / 3-bit: ~6-point average drop, not worth it.
  4. Verify on a held-out eval set in your domain before shipping.
• Output: chosen --quantization {fp8|awq|gptq|...} flag; documented expected accuracy delta.
• Evidence: [arxiv.org/abs/2411.02355 "Give Me BF16 or Give Me Death"]; [docs.gpustack.ai vLLM quantization].

OP-3: Configure tensor/pipeline parallelism

• Trigger: model too large for one GPU; multi-GPU or multi-node deployment.
• Action:
  1. One node + NVLink: --tensor-parallel-size=<GPUs in node>, PP=1.
  2. One node + only PCIe (e.g. L40S, consumer cards): consider --pipeline-parallel-size instead of large TP — all-reduce over PCIe will tank TP throughput [docs.vllm.ai parallelism_scaling].
  3. Multi-node: TP=GPUs/node, PP=#nodes. Use InfiniBand if possible.
  4. MoE (Mixtral/DSv3): combine DP attention + EP/TP on MoE layers.
  5. Test TP=2 vs TP=4 vs 2×replicas-of-TP=2 — replicas often win on throughput-per-dollar when one replica already saturates [developers.redhat.com 2026 step 5].
• Output: chosen TP/PP/DP/EP config with measured tokens/s on real workload.
• Evidence: [docs.vllm.ai/en/stable/serving/parallelism_scaling/].

OP-4: Enable prefix caching

• Trigger: workload has repeated prefixes (shared system prompt, few-shot examples, RAG with stable instructions, multi-turn chat with same persona).
• Action: launch with --enable-prefix-caching (in V1, often on by default).
• Output: prefill TTFT drops on cache-hit requests; KV occupancy may rise (cached blocks held longer).
• Caveats:
  • APC accelerates prefill only; if outputs are long and prefixes don't repeat, gain ≈ 0 [docs.vllm.ai automatic_prefix_caching].
  • In multi-replica clusters, naive round-robin scatters prefix-sharing requests → use prefix-aware routing (e.g. Ray Serve LLM, llm-d) [llm-d.ai/blog/kvcache-wins-you-can-see].
  • Multi-tenant privacy: use cache salting to avoid cross-tenant prefix collisions [github.com/vllm-project/vllm/issues/16016].
• Evidence: [docs.vllm.ai/en/stable/design/prefix_caching/].

OP-5: Triage live performance (TTFT vs ITL)

• Trigger: SLA miss, user complaints, queueing alerts.
• Action (5-step Red Hat triage):
  1. Split symptom: high TTFT vs high ITL? (Prometheus histograms.)
  2. Saturation: num_requests_waiting > 0 sustained → engine queue-bound; check num_requests_running.
  3. KV health: KV cache occupancy near 100% with rising preemption count → drop --max-num-seqs or quantize KV to FP8.
  4. Sequence-length correlation: ITL ~40 ms/token implies ~40 s per 1000 tokens output — sometimes the workload is the answer.
  5. Interconnect check: PCIe-only with TP=4? Drop to TP=2 + 2 replicas.
• Output: identified bottleneck class (queue / compute / memory / interconnect) and one targeted change.
• Evidence: [developers.redhat.com 2026 5-steps-triage].

OP-6: Decide on speculative decoding

• Trigger: latency-bound workload, model ≥ 13B, low/medium QPS.
• Action:
  • If draft/EAGLE-3 weights exist for the target model → enable; expect up to 2.5× decode speedup [developers.redhat.com 2025 eagle3].
  • If not → try n-gram (~1.17×, no extra weights, ~zero risk) [jarvislabs.ai].
  • Skip speculative decoding if QPS is high (large batches already saturate compute → speculation wastes capacity).
  • Always benchmark: poor hyperparameter tuning can degrade cost-per-token by up to 175% [arxiv.org/pdf/2601.11580].
• Output: --speculative-config JSON, with measured tokens/s delta on representative traffic.
• Evidence: [docs.vllm.ai/en/latest/features/speculative_decoding/].

5. 困境决策案例 (Dilemma Cases)

Dilemma 1: Throughput vs Latency — how aggressively to batch?

Situation: A team serves Llama-3-70B-FP8 on 2×H100. P50 TTFT is fine at 250 ms, but P99 spikes to 4 s when traffic bursts. They consider raising --max-num-seqs from 64 to 256 to handle bursts.

Tension:

• Higher max_num_seqs → more concurrency → higher throughput, but also more contention for KV cache → preemption + head-of-line blocking on prefills.
• Higher max_num_batched_tokens → better TTFT (more prefill per step) → worse ITL for streaming requests already mid-decode.

Resolution heuristic:

• If P99 spike correlates with num_requests_waiting > 0 and KV occupancy < 80%: raise max_num_seqs.
• If KV occupancy hits 100% and preemption counter rises: do NOT raise; either quantize KV to FP8, add a replica, or shorten max_model_len.
• For latency-sensitive interactive workloads: lower max_num_batched_tokens (e.g. 2048) to favor decode/ITL; for batch/offline: raise (≥8192) to favor TTFT/throughput [docs.vllm.ai optimization; anyscale.com].

Evidence: [developers.redhat.com 2026 5-steps-triage]; [medium.com/@kaige.yang0110].

Dilemma 2: 4-bit AWQ vs FP8 — is the quality drop worth the VRAM?

Situation: Deploying Qwen-72B on a single H100 (80 GB). At BF16, model is 144 GB → doesn't fit. Options: (a) FP8, 72 GB → fits with tight KV budget; (b) AWQ-4, 36 GB → comfortable KV budget, larger batches.

Tension:

• FP8 is empirically near-lossless across all model scales [arxiv.org/abs/2411.02355], but leaves little KV headroom on a single H100 → smaller batches → less throughput.
• AWQ-4 frees ~36 GB for KV → 2–3× more concurrent sequences → higher throughput, but loses ~1.6 average points on benchmarks (more on reasoning-heavy/long-context tasks).

Resolution heuristic:

• Quality-critical (legal, medical, coding agent): FP8 + accept smaller batch, or move to 2×H100 with TP=2 and stay at BF16.
• Throughput-critical (bulk classification, summarization, search re-ranking): AWQ-4 wins; validate on YOUR eval set, not generic benchmarks (real-world drop on code can exceed academic-benchmark drop) [arxiv.org/abs/2411.02355].
• Mixed: serve two replicas — FP8 for premium tier, AWQ for free tier.

Evidence: [arxiv.org/abs/2411.02355]; [docs.gpustack.ai vLLM quantization].

Dilemma 3: TP=4 single replica vs 2×TP=2 replicas for a 70B model

Situation: 4×A100-80GB available, NVLink. Choice: one big replica (TP=4, max single-request latency optimized) or two replicas (TP=2, more concurrency).

Tension:

• TP=4 lowers single-request latency (model layers parallelized further), but all-reduce overhead grows quadratically; on saturated workloads, marginal throughput per added GPU diminishes.
• 2×TP=2 doubles independent throughput (each replica handles its own batch) and survives one-replica failures, but a single request can never go faster than TP=2.

Resolution heuristic:

• If P99 latency SLA is the binding constraint and single-request decode is the bottleneck → TP=4.
• If throughput-per-dollar is the constraint, requests are independent, and average request latency at TP=2 is acceptable → 2×TP=2 replicas; this is the Red Hat triage step-5 recommendation [developers.redhat.com 2026].
• Always check NVLink topology: across-NUMA TP=4 on PCIe degrades faster than two NVLink-paired TP=2 islands [servermo.com vLLM multi-gpu].
• Benchmark with vllm bench serve on representative ISL/OSL.

Evidence: [docs.vllm.ai parallelism_scaling]; [docs.jarvislabs.ai scaling-llm-inference-dp-pp-tp].

Dilemma 4: Prefix caching ROI — is the RAM cost justified?

Situation: Production RAG service. Each request has a 2 KB system prompt + 4–16 KB context + short user query. They wonder whether enable_prefix_caching helps enough to justify the held KV blocks.

Tension:

• APC reduces prefill time on cache-hit; if 90% of requests share the system prompt, prefill of those tokens is free → big TTFT win.
• Cached blocks stay resident in KV memory → fewer free blocks for new requests → can trigger preemption on bursty traffic.

Resolution heuristic:

• Enable when shared-prefix fraction ≥ ~30% and outputs are short-to-medium (APC helps prefill, not decode) [docs.vllm.ai automatic_prefix_caching].
• Disable / tune when requests are unique and long-output dominated — caching just consumes KV with no hit-rate.
• For multi-replica deployments: pair with prefix-aware routing so requests with the same prefix land on the same replica; otherwise you pay the cache cost N times for 1/N hit-rate [llm-d.ai/blog/kvcache-wins-you-can-see].
• For multi-tenant: enable cache salting for privacy [github.com/vllm-project/vllm/issues/16016].

Evidence: [docs.vllm.ai/en/stable/design/prefix_caching/]; [bentoml.com/llm/inference-optimization/prefix-caching].

Dilemma 5: Speculative decoding — when is the complexity worth it?

Situation: Single-user, latency-sensitive coding assistant on Llama-3-70B. Decode dominates total latency.

Tension:

• EAGLE-3 can deliver up to 2.5× decode speedup for low-QPS workloads [developers.redhat.com 2025 eagle3], but requires a trained EAGLE head, increases VRAM, and is sensitive to hyperparameters (bad config → -175% cost) [arxiv 2601.11580].
• n-gram speculation is cheap (no extra weights) but only ~1.17× speedup [jarvislabs.ai].
• At high QPS, the GPU is already saturated by real batch work; speculative compute steals capacity instead of filling idle cycles.

Resolution heuristic:

• Low QPS, single-stream-latency-bound, EAGLE/MTP head available → enable model-based speculation. Worth the engineering cost.
• No draft weights, want easy win → try n-gram first; if speedup < ~1.1×, disable.
• High QPS / throughput-bound → skip speculation, invest in quantization + parallelism instead.
• Always A/B against the non-speculative baseline on representative traffic.

Evidence: [docs.vllm.ai/en/latest/features/speculative_decoding/]; [developers.redhat.com 2025 fly-eagle3-fly].

6. 反模式与边界 (Anti-patterns & Boundaries)

Anti-pattern 1: Setting max_model_len to the model's architectural max "just in case"

A Llama-3.1 model supports 128k context. If you serve 4k-prompt workloads but set --max-model-len 131072, vLLM reserves KV-cache slots for the worst case → smaller batches → throughput collapse. Set it to your realistic max (longest_prompt + longest_output + safety margin) [markaicode.com 2026].

Anti-pattern 2: Treating prefix caching as a panacea

APC accelerates prefill only. If outputs are long and prefixes don't repeat, gain ≈ 0 — and you've consumed KV memory for nothing [docs.vllm.ai automatic_prefix_caching]. Measure shared-prefix fraction before enabling for high-pressure workloads.

Anti-pattern 3: TP across PCIe-only GPUs

Tensor parallelism issues an all-reduce after each layer. NVLink (≈900 GB/s bidirectional) handles it; PCIe Gen4 x16 (≈32 GB/s, ~28× slower) does not [spheron.network 2026]. On L40S / consumer cards / cross-NUMA setups, use pipeline parallelism or independent replicas instead [docs.vllm.ai parallelism_scaling].

Anti-pattern 4: Picking quantization off academic benchmarks alone

Academic benchmarks (MMLU, HellaSwag) show GPTQ ≈ AWQ; real coding workloads can show meaningfully larger gaps between methods, and quality at INT3 collapses (~6-point drop) [arxiv.org/abs/2411.02355]. Always test on your eval set.

Anti-pattern 5: Using vLLM for the wrong workload

• Small models / hobby use / 1 user, 1 GPU: Ollama or llama.cpp give you 5-minute setup; vLLM's scheduler overhead doesn't pay off [aimadetools.com 2026].
• CPU-only / edge / Apple Silicon production: llama.cpp (GGUF) is the right tool; vLLM's Metal/MPS support is experimental as of 2026 [aimadetools.com 2026].
• Embedding-only workloads: vLLM supports embedding endpoints, but TEI (Text Embeddings Inference) or sentence-transformers servers are often simpler/faster for that single purpose.
• Pure throughput, single-vendor NVIDIA, willing to spend 1–2 weeks tuning: TensorRT-LLM can beat vLLM by 30–50% on throughput but locks you in [n1n.ai 2026].

Anti-pattern 6: Ignoring CPU provisioning

vLLM needs minimum 2 + N physical CPU cores for N GPUs; under-provisioning CPUs makes the API server, tokenizer, and detokenizer the bottleneck before the GPU is touched [docs.vllm.ai/en/stable/configuration/optimization/]. Scale CPU and use --api-server-count when needed.

Anti-pattern 7: --enforce-eager in production "for stability"

--enforce-eager skips CUDA graph capture → slower decode. It's a debugging/memory-OOM workaround, not a steady-state production setting. Fix the underlying OOM (KV dtype, max_model_len, max_num_seqs) and re-enable CUDA graphs.

Boundaries

vLLM is the right tool when:

• ✅ ≥10 concurrent requests OR variable request length OR shared prefixes.
• ✅ NVIDIA / AMD GPU with ≥1 device of sufficient VRAM.
• ✅ Decoder-only or supported architecture (Llama, Qwen, Mixtral, DeepSeek, Gemma, Phi, multimodal: LLaVA / Qwen-VL).
• ✅ Throughput, cost-per-token, or P50 latency under load is the goal.

vLLM is the wrong tool when:

• ❌ Single user, single stream, dev laptop.
• ❌ CPU/edge/mobile, no GPU.
• ❌ Architecture not yet supported (check the model list — Mamba/hybrid coverage is growing but incomplete).
• ❌ You need state-machine-driven structured generation tighter than vLLM's guided_decoding (consider SGLang for complex agent state machines).

7. 生态对照 (Ecosystem Comparison)

Common pattern: develop on Ollama → benchmark with vLLM → consider SGLang if prefix-sharing workload → consider TensorRT-LLM only if NVIDIA-locked and engineering budget is large.

Quick-start reference command

A production-ish vLLM serve for Llama-3.1-70B-Instruct-FP8 on 2×H100 with NVLink, RAG-style workload:

vllm serve meta-llama/Llama-3.1-70B-Instruct \
  --quantization fp8 \
  --tensor-parallel-size 2 \
  --max-model-len 8192 \
  --max-num-seqs 256 \
  --max-num-batched-tokens 8192 \
  --gpu-memory-utilization 0.90 \
  --enable-prefix-caching \
  --enable-chunked-prefill \
  --kv-cache-dtype fp8

Then triage with the §4 OP-5 5-step workflow on real traffic before tuning further.

Cited sources (primary)

• PagedAttention paper: https://arxiv.org/abs/2309.06180
• vLLM docs: https://docs.vllm.ai/en/stable/
• vLLM GitHub: https://github.com/vllm-project/vllm
• Quantization tradeoffs: https://arxiv.org/abs/2411.02355
• Performance triage: https://developers.redhat.com/articles/2026/03/09/5-steps-triage-vllm-performance
• EAGLE-3 results: https://developers.redhat.com/articles/2025/07/01/fly-eagle3-fly-faster-inference-vllm-speculative-decoding
• Engine comparison: https://www.yottalabs.ai/post/best-llm-inference-engines-in-2026-vllm-tensorrt-llm-tgi-and-sglang-compared