Issue Station
Sign In
Tools hereToolbox

vllm - 💡(How to fix) Fix [Bug]: Accuracy drops ~20% when `--enable-prefix-caching` is used together with MTP speculative decoding (Qwen3.6 35B-A3B)

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…

Fix Action

Fix / Workaround

============================== CPU Info

Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 52 bits physical, 57 bits virtual Byte Order: Little Endian CPU(s): 384 On-line CPU(s) list: 0-383 Vendor ID: AuthenticAMD BIOS Vendor ID: Advanced Micro Devices, Inc. Model name: AMD EPYC 9K84 96-Core Processor BIOS Model name: AMD EPYC 9K84 96-Core Processor Unknown CPU @ 2.6GHz BIOS CPU family: 107 CPU family: 25 Model: 17 Thread(s) per core: 2 Core(s) per socket: 96 Socket(s): 2 Stepping: 1 Frequency boost: enabled CPU(s) scaling MHz: 142% CPU max MHz: 2600.0000 CPU min MHz: 1500.0000 BogoMIPS: 5200.45 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d Virtualization: AMD-V L1d cache: 6 MiB (192 instances) L1i cache: 6 MiB (192 instances) L2 cache: 192 MiB (192 instances) L3 cache: 768 MiB (24 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-95,192-287 NUMA node1 CPU(s): 96-191,288-383 Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

Code Example

# dp.sh
MODEL=/path/to/qwen3-checkpoint
export VLLM_ENGINE_READY_TIMEOUT_S=3600

CUDA_VISIBLE_DEVICES=0 \
vllm serve $MODEL \
  --host 0.0.0.0 \
  --port 30000 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --gpu-memory-utilization 0.95 \
  --max-num-seqs 30 \
  --max-model-len 8192 \
  --served-model-name qwen \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --enable-prefix-caching            # <-- adding this is fine, accuracy unchanged

---

# dp_mtp2.sh
MODEL=/path/to/qwen3-checkpoint
export VLLM_ENGINE_READY_TIMEOUT_S=3600

CUDA_VISIBLE_DEVICES=2 \
vllm serve $MODEL \
  --host 0.0.0.0 \
  --port 30002 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --gpu-memory-utilization 0.95 \
  --max-num-seqs 30 \
  --max-model-len 8192 \
  --served-model-name qwen \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --speculative-config '{"method": "mtp", "num_speculative_tokens": 2}' \
  --enable-prefix-caching            # <-- adding this causes ~20% accuracy drop

---

# repro_client.py
from openai import OpenAI

client = OpenAI(base_url="http://0.0.0.0:30002/v1", api_key="EMPTY")

# A classification-style prompt with a long, shared system prefix
# (so prefix caching actually kicks in across requests).
SYSTEM = "system prompt..."

samples = [
    "text 1",
    "text 2",
    # ... many samples sharing the same SYSTEM prefix
]

for s in samples:
    resp = client.chat.completions.create(
        model="qwen",
        messages=[
            {"role": "system", "content": SYSTEM},
            {"role": "user", "content": s},
        ],
        temperature=0.1,
        max_tokens=1024,
    )
    print(resp.choices[0].message.content)
RAW_BUFFERClick to expand / collapse

Your current environment

<details> <summary>The output of <code>python collect_env.py</code></summary> ============================== System Info ============================== OS : TencentOS Server 4.2 (x86_64) GCC version : (Tencent Compiler 12.3.1.2) 12.3.1 20230912 (TencentOS 12.3.1.2-3) Clang version : 17.0.6 (TencentOS 17.0.6-1.tl4.1) CMake version : Could not collect Libc version : glibc-2.38

============================== PyTorch Info

PyTorch version : 2.11.0+cu129 Is debug build : False CUDA used to build PyTorch : 12.9 ROCM used to build PyTorch : N/A XPU used to build PyTorch : N/A

============================== Python Environment

Python version : 3.12.0 | packaged by Anaconda, Inc. | (main, Oct 2 2023, 17:29:18) [GCC 11.2.0] (64-bit runtime) Python platform : Linux-5.4.241-1-tlinux4-0017.7-x86_64-with-glibc2.38

============================== CUDA / GPU Info

Is CUDA available : True CUDA runtime version : 12.9.86 CUDA_MODULE_LOADING set to : GPU models and configuration : GPU 0: NVIDIA H20 GPU 1: NVIDIA H20 GPU 2: NVIDIA H20 GPU 3: NVIDIA H20 GPU 4: NVIDIA H20 GPU 5: NVIDIA H20 GPU 6: NVIDIA H20 GPU 7: NVIDIA H20

Nvidia driver version : 535.161.08 cuDNN version : Probably one of the following: /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_adv.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_cnn.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_engines_precompiled.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_engines_runtime_compiled.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_graph.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_heuristic.so.9.13.1 /usr/local/cuda-12.9/targets/x86_64-linux/lib/libcudnn_ops.so.9.13.1 HIP runtime version : N/A MIOpen runtime version : N/A Is XNNPACK available : True

============================== CPU Info

Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 52 bits physical, 57 bits virtual Byte Order: Little Endian CPU(s): 384 On-line CPU(s) list: 0-383 Vendor ID: AuthenticAMD BIOS Vendor ID: Advanced Micro Devices, Inc. Model name: AMD EPYC 9K84 96-Core Processor BIOS Model name: AMD EPYC 9K84 96-Core Processor Unknown CPU @ 2.6GHz BIOS CPU family: 107 CPU family: 25 Model: 17 Thread(s) per core: 2 Core(s) per socket: 96 Socket(s): 2 Stepping: 1 Frequency boost: enabled CPU(s) scaling MHz: 142% CPU max MHz: 2600.0000 CPU min MHz: 1500.0000 BogoMIPS: 5200.45 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d Virtualization: AMD-V L1d cache: 6 MiB (192 instances) L1i cache: 6 MiB (192 instances) L2 cache: 192 MiB (192 instances) L3 cache: 768 MiB (24 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-95,192-287 NUMA node1 CPU(s): 96-191,288-383 Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

============================== Versions of relevant libraries

[pip3] flashinfer-python==0.6.8.post1 [pip3] numpy==1.26.4 [pip3] nvidia-cublas-cu12==12.9.1.4 [pip3] nvidia-cuda-cupti-cu12==12.9.79 [pip3] nvidia-cuda-nvrtc-cu12==12.9.86 [pip3] nvidia-cuda-runtime-cu12==12.9.79 [pip3] nvidia-cudnn-cu12==9.17.1.4 [pip3] nvidia-cudnn-frontend==1.18.0 [pip3] nvidia-cufft-cu12==11.4.1.4 [pip3] nvidia-cufile-cu12==1.14.1.1 [pip3] nvidia-curand-cu12==10.3.10.19 [pip3] nvidia-cusolver-cu12==11.7.5.82 [pip3] nvidia-cusparse-cu12==12.5.10.65 [pip3] nvidia-cusparselt-cu12==0.7.1 [pip3] nvidia-cutlass-dsl==4.4.2 [pip3] nvidia-cutlass-dsl-libs-base==4.4.2 [pip3] nvidia-ml-py==13.595.45 [pip3] nvidia-nccl-cu12==2.28.9 [pip3] nvidia-nvjitlink-cu12==12.9.86 [pip3] nvidia-nvshmem-cu12==3.4.5 [pip3] nvidia-nvtx-cu12==12.9.79 [pip3] onnx==1.21.0 [pip3] onnx-ir==0.2.1 [pip3] onnxscript==0.7.0 [pip3] pynvml==13.0.1 [pip3] pyzmq==27.1.0 [pip3] tokenspeed-triton==3.7.10.post20260505 [pip3] torch==2.11.0+cu129 [pip3] torch_c_dlpack_ext==0.1.5 [pip3] torchaudio==2.11.0+cu129 [pip3] torchvision==0.26.0+cu129 [pip3] transformer_engine_torch==2.15.0 [pip3] transformers==5.8.0 [pip3] transformers-stream-generator==0.0.5 [pip3] triton==3.6.0 [conda] flashinfer-python 0.6.8.post1 pypi_0 pypi [conda] numpy 1.26.4 pypi_0 pypi [conda] nvidia-cublas-cu12 12.9.1.4 pypi_0 pypi [conda] nvidia-cuda-cupti-cu12 12.9.79 pypi_0 pypi [conda] nvidia-cuda-nvrtc-cu12 12.9.86 pypi_0 pypi [conda] nvidia-cuda-runtime-cu12 12.9.79 pypi_0 pypi [conda] nvidia-cudnn-cu12 9.17.1.4 pypi_0 pypi [conda] nvidia-cudnn-frontend 1.18.0 pypi_0 pypi [conda] nvidia-cufft-cu12 11.4.1.4 pypi_0 pypi [conda] nvidia-cufile-cu12 1.14.1.1 pypi_0 pypi [conda] nvidia-curand-cu12 10.3.10.19 pypi_0 pypi [conda] nvidia-cusolver-cu12 11.7.5.82 pypi_0 pypi [conda] nvidia-cusparse-cu12 12.5.10.65 pypi_0 pypi [conda] nvidia-cusparselt-cu12 0.7.1 pypi_0 pypi [conda] nvidia-cutlass-dsl 4.4.2 pypi_0 pypi [conda] nvidia-cutlass-dsl-libs-base 4.4.2 pypi_0 pypi [conda] nvidia-ml-py 13.595.45 pypi_0 pypi [conda] nvidia-nccl-cu12 2.28.9 pypi_0 pypi [conda] nvidia-nvjitlink-cu12 12.9.86 pypi_0 pypi [conda] nvidia-nvshmem-cu12 3.4.5 pypi_0 pypi [conda] nvidia-nvtx-cu12 12.9.79 pypi_0 pypi [conda] pynvml 13.0.1 pypi_0 pypi [conda] pyzmq 27.1.0 pypi_0 pypi [conda] tokenspeed-triton 3.7.10.post20260505 pypi_0 pypi [conda] torch 2.11.0+cu129 pypi_0 pypi [conda] torch-c-dlpack-ext 0.1.5 pypi_0 pypi [conda] torchaudio 2.11.0+cu129 pypi_0 pypi [conda] torchvision 0.26.0+cu129 pypi_0 pypi [conda] transformer-engine-torch 2.15.0 pypi_0 pypi [conda] transformers 5.8.0 pypi_0 pypi [conda] transformers-stream-generator 0.0.5 pypi_0 pypi [conda] triton 3.6.0 pypi_0 pypi

============================== vLLM Info

ROCM Version : Could not collect vLLM Version : 0.21.0 vLLM Build Flags: CUDA Archs: Not Set; ROCm: Disabled; XPU: Disabled GPU Topology: GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 NIC7 NIC8 NIC9 NIC10 NIC11 NIC12 NIC13 NIC14 NIC15 NIC16 NIC17 NIC18 NIC19 NIC20 NIC21 NIC22 NIC23 NIC24 NIC25 NIC26 NIC27 NIC28 NIC29 NIC30 NIC31 NIC32 NIC33 NIC34 NIC35 NIC36 NIC37 NIC38 NIC39 NIC40 NIC41 CPU Affinity NUMA Affinity GPU NUMA ID GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS PIX NODE NODE NODE SYS SYS SYS SYS 0-95,192-287 0 N/A GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV18 SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS NODE NODE PHB PIX SYS SYS SYS SYS 0-95,192-287 0 N/A GPU2 NV18 NV18 X NV18 NV18 NV18 NV18 NV18 SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS NODE NODE PIX PHB SYS SYS SYS SYS 0-95,192-287 0 N/A GPU3 NV18 NV18 NV18 X NV18 NV18 NV18 NV18 SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS NODE PIX NODE NODE SYS SYS SYS SYS 0-95,192-287 0 N/A GPU4 NV18 NV18 NV18 NV18 X NV18 NV18 NV18 NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS NODE NODE PIX NODE 96-191,288-383 1 N/A GPU5 NV18 NV18 NV18 NV18 NV18 X NV18 NV18 NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS NODE PIX NODE NODE 96-191,288-383 1 N/A GPU6 NV18 NV18 NV18 NV18 NV18 NV18 X NV18 NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS PHB NODE NODE PIX 96-191,288-383 1 N/A GPU7 NV18 NV18 NV18 NV18 NV18 NV18 NV18 X NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS PIX NODE NODE PHB 96-191,288-383 1 N/A NIC0 SYS SYS SYS SYS NODE NODE NODE NODE X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC1 SYS SYS SYS SYS NODE NODE NODE NODE PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC2 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC3 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC4 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC5 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC6 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC7 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC8 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC9 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC10 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC11 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC12 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC13 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC14 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC15 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC16 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC17 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC18 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC19 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC20 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC21 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC22 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC23 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC24 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC25 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC26 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC27 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC28 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC29 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC30 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC31 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC32 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X PIX SYS SYS SYS SYS NODE NODE NODE NODE NIC33 SYS SYS SYS SYS NODE NODE NODE NODE PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX PIX X SYS SYS SYS SYS NODE NODE NODE NODE NIC34 PIX NODE NODE NODE SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS X NODE NODE NODE SYS SYS SYS SYS NIC35 NODE NODE NODE PIX SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS NODE X NODE NODE SYS SYS SYS SYS NIC36 NODE PHB PIX NODE SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS NODE NODE X PHB SYS SYS SYS SYS NIC37 NODE PIX PHB NODE SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS NODE NODE PHB X SYS SYS SYS SYS NIC38 SYS SYS SYS SYS NODE NODE PHB PIX NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS X NODE NODE PHB NIC39 SYS SYS SYS SYS NODE PIX NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS NODE X NODE NODE NIC40 SYS SYS SYS SYS PIX NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS NODE NODE X NODE NIC41 SYS SYS SYS SYS NODE NODE PIX PHB NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE NODE SYS SYS SYS SYS PHB NODE NODE X

Legend:

X = Self SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI) NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU) PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge) PIX = Connection traversing at most a single PCIe bridge NV# = Connection traversing a bonded set of # NVLinks

NIC Legend:

NIC0: NIC41 CPU Affinity NUMA Affinity GPU NUMA ID NIC1: MA ID NIC2: mlx5_2 NIC3: mlx5_3 NIC4: mlx5_4 NIC5: mlx5_5 NIC6: mlx5_6 NIC7: mlx5_7 NIC8: mlx5_8 NIC9: mlx5_9 NIC10: mlx5_10 NIC11: mlx5_11 NIC12: mlx5_12 NIC13: mlx5_13 NIC14: mlx5_14 NIC15: mlx5_15 NIC16: mlx5_16 NIC17: mlx5_17 NIC18: mlx5_18 NIC19: mlx5_19 NIC20: mlx5_20 NIC21: mlx5_21 NIC22: mlx5_22 NIC23: mlx5_23 NIC24: mlx5_24 NIC25: mlx5_25 NIC26: mlx5_26 NIC27: mlx5_27 NIC28: mlx5_28 NIC29: mlx5_29 NIC30: mlx5_30 NIC31: mlx5_31 NIC32: mlx5_32 NIC33: mlx5_33 NIC34: mlx5_bond_1 NIC35: mlx5_bond_2 NIC36: mlx5_bond_3 NIC37: mlx5_bond_4 NIC38: mlx5_bond_5 NIC39: mlx5_bond_6 NIC40: mlx5_bond_7 NIC41: mlx5_bond_8

============================== Environment Variables

NCCL_IB_TC=160 NCCL_SOCKET_IFNAME=bond1 NCCL_IB_HCA=mlx5_bond_1,mlx5_bond_2,mlx5_bond_3,mlx5_bond_4,mlx5_bond_5,mlx5_bond_6,mlx5_bond_7,mlx5_bond_8 NCCL_IB_GID_INDEX=3 NCCL_IB_TIMEOUT=22 NCCL_IB_SL=3 LD_LIBRARY_PATH=/data/miniconda3/envs/ms/lib/python3.12/site-packages/nvidia/cudnn/lib:/data/miniconda3/envs/ms/lib/python3.12/site-packages/nvidia/cudnn/lib::/usr/local/cuda-12.9/lib64:/usr/local/cuda-12.9/extras/CUPTI/lib64/:/usr/local/cuda-12.9/compat/:/usr/local/cuda/compat/ NCCL_IB_DISABLE=0 PYTORCH_NVML_BASED_CUDA_CHECK=1 TORCHINDUCTOR_COMPILE_THREADS=1 TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root

</details>

🐛 Describe the bug

🐛 Describe the bug

Model: an SFT-finetuned Qwen3.6 35B-A3B (MoE) checkpoint.

When serving this model with MTP speculative decoding (--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}') together with --enable-prefix-caching, the model's output quality drops dramatically. On an internal classification benchmark, accuracy degrades by ~20% compared to the exact same setup without prefix caching.

I ran four configurations on the same checkpoint, with the same client requests (temperature=0.1), and only the last one degrades:

#ConfigurationPrefix CachingAccuracySpeed
1base (plain decoding)offbaselinebaseline
2base + prefix cachingonunchanged ✅faster ✅
3MTP (num_speculative_tokens=2)offunchanged ✅faster ✅
4MTP + prefix cachingondrops ~20% ❌faster

So prefix caching alone is fine (group 2), MTP alone is fine (group 3), but MTP + prefix caching (group 4) consistently loses ~20% accuracy on the classification task.

This strongly suggests an interaction bug between the prefix-cache reuse path and the MTP speculative decoding path (likely related to how the draft/MTP hidden states or KV blocks are reused across cached prefixes).

🔁 Steps to reproduce

Both scripts launch the same model on a single H20 GPU. The only differences are the GPU index, the port, and the presence of the --speculative-config flag.

A. Baseline serve script (prefix-caching is safe here):

# dp.sh
MODEL=/path/to/qwen3-checkpoint
export VLLM_ENGINE_READY_TIMEOUT_S=3600

CUDA_VISIBLE_DEVICES=0 \
vllm serve $MODEL \
  --host 0.0.0.0 \
  --port 30000 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --gpu-memory-utilization 0.95 \
  --max-num-seqs 30 \
  --max-model-len 8192 \
  --served-model-name qwen \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --enable-prefix-caching            # <-- adding this is fine, accuracy unchanged

B. MTP serve script (prefix-caching breaks accuracy here):

# dp_mtp2.sh
MODEL=/path/to/qwen3-checkpoint
export VLLM_ENGINE_READY_TIMEOUT_S=3600

CUDA_VISIBLE_DEVICES=2 \
vllm serve $MODEL \
  --host 0.0.0.0 \
  --port 30002 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --gpu-memory-utilization 0.95 \
  --max-num-seqs 30 \
  --max-model-len 8192 \
  --served-model-name qwen \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --speculative-config '{"method": "mtp", "num_speculative_tokens": 2}' \
  --enable-prefix-caching            # <-- adding this causes ~20% accuracy drop

Client side (any OpenAI-compatible client works). A minimal reproducer:

# repro_client.py
from openai import OpenAI

client = OpenAI(base_url="http://0.0.0.0:30002/v1", api_key="EMPTY")

# A classification-style prompt with a long, shared system prefix
# (so prefix caching actually kicks in across requests).
SYSTEM = "system prompt..."

samples = [
    "text 1",
    "text 2",
    # ... many samples sharing the same SYSTEM prefix
]

for s in samples:
    resp = client.chat.completions.create(
        model="qwen",
        messages=[
            {"role": "system", "content": SYSTEM},
            {"role": "user", "content": s},
        ],
        temperature=0.1,
        max_tokens=1024,
    )
    print(resp.choices[0].message.content)

Run the same client against:

  1. server B without --enable-prefix-caching → accuracy = baseline.
  2. server B with --enable-prefix-caching → accuracy drops by ~20%.
  3. server A with --enable-prefix-caching → accuracy = baseline (so prefix caching alone is fine).

Only configuration (2) — MTP + prefix caching — exhibits the regression.

Expected behavior

Enabling --enable-prefix-caching should not meaningfully change generation quality, regardless of whether MTP speculative decoding is enabled. With temperature=0.1, small token-level variation is expected, but the overall classification accuracy should match the configuration without prefix caching (as it does for the base model in group 2, and for MTP without prefix caching in group 3).

Actual behavior

With MTP speculative decoding enabled, turning on prefix caching produces noticeably different (and worse) outputs, leading to a ~20% accuracy drop on a classification task that is otherwise stable across the other three configurations (base, base+prefix-caching, MTP without prefix-caching).

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

FAQ

Expected behavior

Enabling --enable-prefix-caching should not meaningfully change generation quality, regardless of whether MTP speculative decoding is enabled. With temperature=0.1, small token-level variation is expected, but the overall classification accuracy should match the configuration without prefix caching (as it does for the base model in group 2, and for MTP without prefix caching in group 3).

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING