Issue Station
Sign In
Tools hereToolbox

vllm - ✅(Solved) Fix [Bug]: heterogeneous disaggregated serving XPU (Prefill) + CPU (Decode) accuracy issue [2 pull requests, 1 comments, 2 participants]

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…
GitHub stats
vllm-project/vllm#38710•Fetched 2026-04-08 02:23:20
View on GitHub
Comments
1
Participants
2
Timeline
3
Reactions
1
Author
SpycshSpycsh
Participants
SpycshSpycshxuechendixuechendi
Timeline (top)
commented ×1cross-referenced ×1labeled ×1

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): 256 On-line CPU(s) list: 0-255 Vendor ID: GenuineIntel BIOS Vendor ID: Intel(R) Corporation Model name: Intel(R) Xeon(R) 6767P BIOS Model name: Intel(R) Xeon(R) 6767P CPU @ 2.4GHz BIOS CPU family: 179 CPU family: 6 Model: 173 Thread(s) per core: 2 Core(s) per socket: 64 Socket(s): 2 Stepping: 1 CPU(s) scaling MHz: 25% CPU max MHz: 3900.0000 CPU min MHz: 800.0000 BogoMIPS: 4800.00 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 6 MiB (128 instances) L1i cache: 8 MiB (128 instances) L2 cache: 256 MiB (128 instances) L3 cache: 672 MiB (2 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-63,128-191 NUMA node1 CPU(s): 64-127,192-255 Vulnerability Gather data sampling: Not affected Vulnerability Ghostwrite: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Old microcode: Vulnerable Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S Vulnerability Srbds: Not affected Vulnerability Tsa: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

============================== 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): 256 On-line CPU(s) list: 0-255 Vendor ID: GenuineIntel BIOS Vendor ID: Intel(R) Corporation Model name: Intel(R) Xeon(R) 6767P BIOS Model name: Intel(R) Xeon(R) 6767P CPU family: 6 Model: 173 Thread(s) per core: 2 Core(s) per socket: 64 Socket(s): 2 Stepping: 1 CPU max MHz: 3900.0000 CPU min MHz: 800.0000 BogoMIPS: 4800.00 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 6 MiB (128 instances) L1i cache: 8 MiB (128 instances) L2 cache: 256 MiB (128 instances) L3 cache: 672 MiB (2 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-63,128-191 NUMA node1 CPU(s): 64-127,192-255 Vulnerability Gather data sampling: Not affected Vulnerability Ghostwrite: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Old microcode: Vulnerable Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S Vulnerability Srbds: Not affected Vulnerability Tsa: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

PR fix notes

PR #38935: [PD][HeteroArch]Fix accuracy issue with CPU_ATTN as Decoder and Flash_ATTN as prefiller

Description (problem / solution / changelog)

Purpose

Fix the error reported in #38710

When using CPU as decoder to work with Hetero Arch Platform (CUDA, Intel GPU, Intel Gaudi ...) as prefiller, there is an accuracy issue due to CPU KV layout need addition packing step

In this PR, we fixed this issue by following cpu_attn kv Packing method to pack FlashAttn plain KV to CPU packed KV. The work is done in platforms/cpu.py - pack_kv_cache()

Changes proposed

  1. In NixlConnector, introduced a new method - post_process_device_kv_on_receive_heterogeneous_attn, which will be called if self.enable_heterogeneous_attn_post_process = True
  2. self.enable_heterogeneous_attn_post_process = True will only be enabled when attention_back == "CPU_ATTN" and remote_agent.attention_backend != "CPU_ATTN"
  3. in cpu.py, introduce a new method - "pack_kv_cache" which will be called in post_process_device_kv_on_receive_heterogeneous_attn to pack KV

Test Plan

Prefill runs on CUDA / XPU

export VLLM_LOGGING_LEVEL=debug
export UCX_TLS=tcp
export model_name=Qwen/Qwen3-0.6B
export tp_size=1

export ZE_AFFINITY_MASK=0
# PREFILL
VLLM_NIXL_SIDE_CHANNEL_HOST=0.0.0.0 \
VLLM_NIXL_SIDE_CHANNEL_PORT=5577 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ENABLE_V1_MULTIPROCESSING=1 \
vllm serve $model_name \
  -tp $tp_size \
  --host 0.0.0.0 \
  --port 8100 \
  --seed 42 \
  --enforce-eager \
  --dtype float16 \
  --gpu-memory-utilization 0.2 \
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_buffer_device":"cpu","kv_connector_extra_config":{"enforce_handshake_compat": false}}' \
  --max-model-len 9216 \
  --max_num_batched_tokens 2048 \
  --block-size 64 \
  --no-enable-prefix-caching \
  --disable-hybrid-kv-cache-manager 2>&1 | tee prefix_test.log &

# PROXY
python3 /workspace/vllm/tests/v1/kv_connector/nixl_integration/toy_proxy_server.py \
  --prefiller-host localhost \
  --prefiller-port 8100 \
  --decoder-host localhost \
  --decoder-port 8200 \
  --host localhost \
  --port 8300

Decode runs on CPU

export UCX_TLS=tcp
export model_name=Qwen/Qwen3-0.6B
export tp_size=1

# DECODE
VLLM_NIXL_SIDE_CHANNEL_HOST=0.0.0.0 \
VLLM_NIXL_SIDE_CHANNEL_PORT=5587 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ENABLE_V1_MULTIPROCESSING=1 \
vllm serve $model_name \
  -tp $tp_size \
  --host 0.0.0.0 \
  --port 8200 \
  --seed 42 \
  --enforce-eager \
  --dtype float16 \
  --gpu-memory-utilization 0.8 \
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_buffer_device":"cpu","kv_connector_extra_config":{"enforce_handshake_compat": false}}' \
  --max-model-len 9216 \
  --max_num_batched_tokens 2048 \
  --block-size 64 \
  --no-enable-prefix-caching \
  --disable-hybrid-kv-cache-manager  2>&1 | tee decode_test.log

Test Result

curl -X POST "http://localhost:8300/v1/chat/completions"   -H "Content-Type: application/json"   -d '{
    "model": "Qwen/Qwen3-0.6B",
    "messages": [
        {
        "role": "user",
        "content": "Hello, how are you?"
        }
    ],
    "max_tokens": 30, "stream": false
 }'

{"id":"chatcmpl-d72d90c9-5f06-4a9e-9f7b-d08c63a77fef","object":"chat.completion","created":1775238578,"model":"Qwen/Qwen3-0.6B","choices":[{"index":0,"message":{"role":"assistant","content":"<think>\nOkay, the user just asked, "Hello, how are you?" I need to respond appropriately. First, I should acknowledge their greeting.","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":14,"total_tokens":44,"completion_tokens":30,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":n

accuracy test:

lm_eval run \
  --model local-completions \
  --model_args model=Qwen/Qwen3-0.6B,base_url=http://localhost:8300/v1/completions,num_concurrent=32,tokenized_requests=False \
  --tasks gsm8k --batch_size 1 --limit 256
<img width="1230" height="249" alt="image" src="https://github.com/user-attachments/assets/f5bfacce-dfb5-40ea-85b2-4c0ac8dd95b1" />
<details> <summary> Essential Elements of an Effective PR Description Checklist </summary>
  • The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
  • The test plan, such as providing test command.
  • The test results, such as pasting the results comparison before and after, or e2e results
  • (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
  • (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.
</details>

Changed files

  • tests/v1/kv_connector/unit/test_nixl_connector.py (modified, +4/-0)
  • vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py (modified, +49/-0)
  • vllm/platforms/cpu.py (modified, +40/-0)

PR #7977: feat: add heterogeneous disagg serving xpu + cpu example on dynamo

Description (problem / solution / changelog)

Overview:

add heterogeneous disagg serving xpu (prefill) + cpu (decode) example on dynamo

Depends on following fixes https://github.com/vllm-project/vllm/pull/38935 https://github.com/vllm-project/vllm/issues/38710

Details:

As above, and the accuracy matches with the vLLM naive one.

<img width="1001" height="175" alt="image" src="https://github.com/user-attachments/assets/9f65e7f8-5c5d-4f5f-84b8-d9f7480bb619" />

Where should the reviewer start?

The scripts.

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

NA

<!-- This is an auto-generated comment: release notes by coderabbit.ai -->

Summary by CodeRabbit

  • New Features
    • Added example launch scripts for disaggregated vLLM inference with heterogeneous CPU and XPU worker configurations, enabling KV cache transfer and event publishing capabilities.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Changed files

  • examples/backends/vllm/launch/xpu/disagg_hetero_cpu_decode.sh (added, +22/-0)
  • examples/backends/vllm/launch/xpu/disagg_hetero_xpu_prefill_proxy.sh (added, +33/-0)

Code Example

==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : 18.1.3 (1ubuntu1)
CMake version                : version 4.3.0
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+xpu
Is debug build               : False
CUDA used to build PyTorch   : None
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar  3 2026, 12:15:18) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.19.10-061910-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : False
CUDA runtime version         : No CUDA
CUDA_MODULE_LOADING set to   : N/A
GPU models and configuration : No CUDA
Nvidia driver version        : No CUDA
cuDNN version                : No CUDA
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):                                  256
On-line CPU(s) list:                     0-255
Vendor ID:                               GenuineIntel
BIOS Vendor ID:                          Intel(R) Corporation
Model name:                              Intel(R) Xeon(R) 6767P
BIOS Model name:                         Intel(R) Xeon(R) 6767P  CPU @ 2.4GHz
BIOS CPU family:                         179
CPU family:                              6
Model:                                   173
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               2
Stepping:                                1
CPU(s) scaling MHz:                      25%
CPU max MHz:                             3900.0000
CPU min MHz:                             800.0000
BogoMIPS:                                4800.00
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                          VT-x
L1d cache:                               6 MiB (128 instances)
L1i cache:                               8 MiB (128 instances)
L2 cache:                                256 MiB (128 instances)
L3 cache:                                672 MiB (2 instances)
NUMA node(s):                            2
NUMA node0 CPU(s):                       0-63,128-191
NUMA node1 CPU(s):                       64-127,192-255
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Old microcode:             Vulnerable
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

==============================
Versions of relevant libraries
==============================
[pip3] flake8==7.3.0
[pip3] mypy==1.19.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.2.6
[pip3] nvidia-ml-py==13.580.65
[pip3] pyzmq==27.1.0
[pip3] sentence-transformers==5.3.0
[pip3] torch==2.10.0+xpu
[pip3] torchaudio==2.11.0+xpu
[pip3] torchvision==0.25.0+xpu
[pip3] transformers==4.57.6
[pip3] triton-xpu==3.6.0
[pip3] tritonclient==2.62.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.18.1rc1.dev264+ge31915063 (git sha: e31915063)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  Could not collect

==============================
     Environment Variables
==============================
VLLM_WORKER_MULTIPROC_METHOD=spawn
VLLM_TARGET_DEVICE=xpu
LD_LIBRARY_PATH=/opt/intel/oneapi/ccl/2021.15/lib/:/opt/intel/oneapi/tcm/1.4/lib:/opt/intel/oneapi/umf/1.0/lib:/opt/intel/oneapi/tbb/2022.3/env/../lib/intel64/gcc4.8:/opt/intel/oneapi/pti/0.16/lib:/opt/intel/oneapi/mpi/2021.17/opt/mpi/libfabric/lib:/opt/intel/oneapi/mpi/2021.17/lib:/opt/intel/oneapi/mkl/2025.3/lib:/opt/intel/oneapi/dnnl/2025.3/lib:/opt/intel/oneapi/debugger/2025.3/opt/debugger/lib:/opt/intel/oneapi/compiler/2025.3/opt/compiler/lib:/opt/intel/oneapi/compiler/2025.3/lib:/opt/intel/oneapi/ccl/2021.15/lib/:/tmp/ucx_install/lib:/opt/intel/oneapi/tcm/1.4/lib:/opt/intel/oneapi/umf/1.0/lib:/opt/intel/oneapi/tbb/2022.3/env/../lib/intel64/gcc4.8:/opt/intel/oneapi/pti/0.16/lib:/opt/intel/oneapi/mpi/2021.17/opt/mpi/libfabric/lib:/opt/intel/oneapi/mpi/2021.17/lib:/opt/intel/oneapi/mkl/2025.3/lib:/opt/intel/oneapi/dnnl/2025.3/lib:/opt/intel/oneapi/debugger/2025.3/opt/debugger/lib:/opt/intel/oneapi/compiler/2025.3/opt/compiler/lib:/opt/intel/oneapi/compiler/2025.3/lib:/opt/intel/oneapi/ccl/2021.17/lib/:/usr/local/lib/
VLLM_LOGGING_LEVEL=debug
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root

---

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0
Clang version                : Could not collect
CMake version                : version 4.3.1
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cpu
Is debug build               : False
CUDA used to build PyTorch   : None
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.13 (main, Mar 24 2026, 22:49:22) [Clang 22.1.1 ] (64-bit runtime)
Python platform              : Linux-6.19.10-061910-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : False
CUDA runtime version         : No CUDA
CUDA_MODULE_LOADING set to   : N/A
GPU models and configuration : No CUDA
Nvidia driver version        : No CUDA
cuDNN version                : No CUDA
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):                                  256
On-line CPU(s) list:                     0-255
Vendor ID:                               GenuineIntel
BIOS Vendor ID:                          Intel(R) Corporation
Model name:                              Intel(R) Xeon(R) 6767P
BIOS Model name:                         Intel(R) Xeon(R) 6767P
CPU family:                              6
Model:                                   173
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               2
Stepping:                                1
CPU max MHz:                             3900.0000
CPU min MHz:                             800.0000
BogoMIPS:                                4800.00
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                          VT-x
L1d cache:                               6 MiB (128 instances)
L1i cache:                               8 MiB (128 instances)
L2 cache:                                256 MiB (128 instances)
L3 cache:                                672 MiB (2 instances)
NUMA node(s):                            2
NUMA node0 CPU(s):                       0-63,128-191
NUMA node1 CPU(s):                       64-127,192-255
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Old microcode:             Vulnerable
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0+cpu
[pip3] torchaudio==2.11.0+cpu
[pip3] torchvision==0.25.0+cpu
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.18.1rc1.dev264+ge31915063 (git sha: e31915063)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  Could not collect

==============================
     Environment Variables
==============================
VLLM_CPU_KVCACHE_SPACE=40
MAX_JOBS=32
LD_LIBRARY_PATH=/usr/local/lib:/usr/lib:
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root

---

export VLLM_LOGGING_LEVEL=debug
export HF_ENDPOINT=https://hf-mirror.com
export UCX_TLS=tcp
export model_name=Qwen/Qwen3-0.6B
export tp_size=1

export ZE_AFFINITY_MASK=4
# PREFILL
VLLM_USE_V1=1 \
VLLM_NIXL_SIDE_CHANNEL_HOST=0.0.0.0 \
VLLM_NIXL_SIDE_CHANNEL_PORT=5577 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ENABLE_V1_MULTIPROCESSING=1 \
vllm serve $model_name \
  -tp $tp_size \
  --host 0.0.0.0 \
  --port 8100 \
  --seed 42 \
  --enforce-eager \
  --dtype float16 \
  --gpu-memory-utilization 0.8 \
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_buffer_device":"cpu","kv_connector_extra_config":{"enforce_handshake_compat": false},"enable_permute_local_kv":"True"}' \
  --max-model-len 9216 \
  --max_num_batched_tokens 2048 \
  --block-size 64 \
  --no-enable-prefix-caching \
  &

# PROXY
python3 /workspace/vllm/tests/v1/kv_connector/nixl_integration/toy_proxy_server.py \
  --prefiller-host localhost \
  --prefiller-port 8100 \
  --decoder-host localhost \
  --decoder-port 8200 \
  --host localhost \
  --port 8300

---

export VLLM_LOGGING_LEVEL=debug
# export UCX_MEMTYPE_CACHE=0
export HF_ENDPOINT=https://hf-mirror.com
export UCX_TLS=tcp
export model_name=Qwen/Qwen3-0.6B
export tp_size=1

# DECODE
VLLM_USE_V1=1 \
VLLM_NIXL_SIDE_CHANNEL_HOST=0.0.0.0 \
VLLM_NIXL_SIDE_CHANNEL_PORT=5587 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ENABLE_V1_MULTIPROCESSING=1 \
vllm serve $model_name \
  -tp $tp_size \
  --host 0.0.0.0 \
  --port 8200 \
  --seed 42 \
  --enforce-eager \
  --dtype float16 \
  --gpu-memory-utilization 0.8 \
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_buffer_device":"cpu","kv_connector_extra_config":{"enforce_handshake_compat": false},"enable_permute_local_kv":"True"}' \
  --max-model-len 9216 \
  --max_num_batched_tokens 2048 \
  --block-size 64 \
  --no-enable-prefix-caching \
  --disable-hybrid-kv-cache-manager

---

curl -X POST "http://localhost:8300/v1/chat/completions"   -H "Content-Type: application/json"   -d '{
    "model": "Qwen/Qwen3-0.6B",
    "messages": [
        {
        "role": "user",
        "content": "Hello, how are you?"
        }
    ],
    "max_tokens": 30, "stream": false
 }'
{"id":"chatcmpl-31aefcad-342c-481b-9b3e-bc2f8bd5ba45","object":"chat.completion","created":1775036538,"model":"Qwen/Qwen3-0.6B","choices":[{"index":0,
"message":{"role":"assistant",
"content":"

---

* Result (aggregated, routed to prefill or decoded only)
RAW_BUFFERClick to expand / collapse

Your current environment

<details> <summary>The output of <code>python collect_env.py</code></summary>
  • XPU
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : 18.1.3 (1ubuntu1)
CMake version                : version 4.3.0
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+xpu
Is debug build               : False
CUDA used to build PyTorch   : None
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar  3 2026, 12:15:18) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.19.10-061910-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : False
CUDA runtime version         : No CUDA
CUDA_MODULE_LOADING set to   : N/A
GPU models and configuration : No CUDA
Nvidia driver version        : No CUDA
cuDNN version                : No CUDA
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):                                  256
On-line CPU(s) list:                     0-255
Vendor ID:                               GenuineIntel
BIOS Vendor ID:                          Intel(R) Corporation
Model name:                              Intel(R) Xeon(R) 6767P
BIOS Model name:                         Intel(R) Xeon(R) 6767P  CPU @ 2.4GHz
BIOS CPU family:                         179
CPU family:                              6
Model:                                   173
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               2
Stepping:                                1
CPU(s) scaling MHz:                      25%
CPU max MHz:                             3900.0000
CPU min MHz:                             800.0000
BogoMIPS:                                4800.00
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                          VT-x
L1d cache:                               6 MiB (128 instances)
L1i cache:                               8 MiB (128 instances)
L2 cache:                                256 MiB (128 instances)
L3 cache:                                672 MiB (2 instances)
NUMA node(s):                            2
NUMA node0 CPU(s):                       0-63,128-191
NUMA node1 CPU(s):                       64-127,192-255
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Old microcode:             Vulnerable
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

==============================
Versions of relevant libraries
==============================
[pip3] flake8==7.3.0
[pip3] mypy==1.19.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.2.6
[pip3] nvidia-ml-py==13.580.65
[pip3] pyzmq==27.1.0
[pip3] sentence-transformers==5.3.0
[pip3] torch==2.10.0+xpu
[pip3] torchaudio==2.11.0+xpu
[pip3] torchvision==0.25.0+xpu
[pip3] transformers==4.57.6
[pip3] triton-xpu==3.6.0
[pip3] tritonclient==2.62.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.18.1rc1.dev264+ge31915063 (git sha: e31915063)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  Could not collect

==============================
     Environment Variables
==============================
VLLM_WORKER_MULTIPROC_METHOD=spawn
VLLM_TARGET_DEVICE=xpu
LD_LIBRARY_PATH=/opt/intel/oneapi/ccl/2021.15/lib/:/opt/intel/oneapi/tcm/1.4/lib:/opt/intel/oneapi/umf/1.0/lib:/opt/intel/oneapi/tbb/2022.3/env/../lib/intel64/gcc4.8:/opt/intel/oneapi/pti/0.16/lib:/opt/intel/oneapi/mpi/2021.17/opt/mpi/libfabric/lib:/opt/intel/oneapi/mpi/2021.17/lib:/opt/intel/oneapi/mkl/2025.3/lib:/opt/intel/oneapi/dnnl/2025.3/lib:/opt/intel/oneapi/debugger/2025.3/opt/debugger/lib:/opt/intel/oneapi/compiler/2025.3/opt/compiler/lib:/opt/intel/oneapi/compiler/2025.3/lib:/opt/intel/oneapi/ccl/2021.15/lib/:/tmp/ucx_install/lib:/opt/intel/oneapi/tcm/1.4/lib:/opt/intel/oneapi/umf/1.0/lib:/opt/intel/oneapi/tbb/2022.3/env/../lib/intel64/gcc4.8:/opt/intel/oneapi/pti/0.16/lib:/opt/intel/oneapi/mpi/2021.17/opt/mpi/libfabric/lib:/opt/intel/oneapi/mpi/2021.17/lib:/opt/intel/oneapi/mkl/2025.3/lib:/opt/intel/oneapi/dnnl/2025.3/lib:/opt/intel/oneapi/debugger/2025.3/opt/debugger/lib:/opt/intel/oneapi/compiler/2025.3/opt/compiler/lib:/opt/intel/oneapi/compiler/2025.3/lib:/opt/intel/oneapi/ccl/2021.17/lib/:/usr/local/lib/
VLLM_LOGGING_LEVEL=debug
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root
  • CPU
Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0
Clang version                : Could not collect
CMake version                : version 4.3.1
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cpu
Is debug build               : False
CUDA used to build PyTorch   : None
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.13 (main, Mar 24 2026, 22:49:22) [Clang 22.1.1 ] (64-bit runtime)
Python platform              : Linux-6.19.10-061910-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : False
CUDA runtime version         : No CUDA
CUDA_MODULE_LOADING set to   : N/A
GPU models and configuration : No CUDA
Nvidia driver version        : No CUDA
cuDNN version                : No CUDA
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):                                  256
On-line CPU(s) list:                     0-255
Vendor ID:                               GenuineIntel
BIOS Vendor ID:                          Intel(R) Corporation
Model name:                              Intel(R) Xeon(R) 6767P
BIOS Model name:                         Intel(R) Xeon(R) 6767P
CPU family:                              6
Model:                                   173
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               2
Stepping:                                1
CPU max MHz:                             3900.0000
CPU min MHz:                             800.0000
BogoMIPS:                                4800.00
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                          VT-x
L1d cache:                               6 MiB (128 instances)
L1i cache:                               8 MiB (128 instances)
L2 cache:                                256 MiB (128 instances)
L3 cache:                                672 MiB (2 instances)
NUMA node(s):                            2
NUMA node0 CPU(s):                       0-63,128-191
NUMA node1 CPU(s):                       64-127,192-255
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Old microcode:             Vulnerable
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0+cpu
[pip3] torchaudio==2.11.0+cpu
[pip3] torchvision==0.25.0+cpu
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.18.1rc1.dev264+ge31915063 (git sha: e31915063)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  Could not collect

==============================
     Environment Variables
==============================
VLLM_CPU_KVCACHE_SPACE=40
MAX_JOBS=32
LD_LIBRARY_PATH=/usr/local/lib:/usr/lib:
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root
</details>

🐛 Describe the bug

I started one Prefill instance on XPU, and one Decode instance on Xeon CPU, with the toy proxy server handling these two instances. After sending a request to the proxy, the flow seems work, and the kv transfer is ok but the output seems incorrect.

Steps to reproduce:

  • Prefill+proxy
export VLLM_LOGGING_LEVEL=debug
export HF_ENDPOINT=https://hf-mirror.com
export UCX_TLS=tcp
export model_name=Qwen/Qwen3-0.6B
export tp_size=1

export ZE_AFFINITY_MASK=4
# PREFILL
VLLM_USE_V1=1 \
VLLM_NIXL_SIDE_CHANNEL_HOST=0.0.0.0 \
VLLM_NIXL_SIDE_CHANNEL_PORT=5577 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ENABLE_V1_MULTIPROCESSING=1 \
vllm serve $model_name \
  -tp $tp_size \
  --host 0.0.0.0 \
  --port 8100 \
  --seed 42 \
  --enforce-eager \
  --dtype float16 \
  --gpu-memory-utilization 0.8 \
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_buffer_device":"cpu","kv_connector_extra_config":{"enforce_handshake_compat": false},"enable_permute_local_kv":"True"}' \
  --max-model-len 9216 \
  --max_num_batched_tokens 2048 \
  --block-size 64 \
  --no-enable-prefix-caching \
  &

# PROXY
python3 /workspace/vllm/tests/v1/kv_connector/nixl_integration/toy_proxy_server.py \
  --prefiller-host localhost \
  --prefiller-port 8100 \
  --decoder-host localhost \
  --decoder-port 8200 \
  --host localhost \
  --port 8300
  • Decode
export VLLM_LOGGING_LEVEL=debug
# export UCX_MEMTYPE_CACHE=0
export HF_ENDPOINT=https://hf-mirror.com
export UCX_TLS=tcp
export model_name=Qwen/Qwen3-0.6B
export tp_size=1

# DECODE
VLLM_USE_V1=1 \
VLLM_NIXL_SIDE_CHANNEL_HOST=0.0.0.0 \
VLLM_NIXL_SIDE_CHANNEL_PORT=5587 \
VLLM_WORKER_MULTIPROC_METHOD=spawn \
VLLM_ENABLE_V1_MULTIPROCESSING=1 \
vllm serve $model_name \
  -tp $tp_size \
  --host 0.0.0.0 \
  --port 8200 \
  --seed 42 \
  --enforce-eager \
  --dtype float16 \
  --gpu-memory-utilization 0.8 \
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_buffer_device":"cpu","kv_connector_extra_config":{"enforce_handshake_compat": false},"enable_permute_local_kv":"True"}' \
  --max-model-len 9216 \
  --max_num_batched_tokens 2048 \
  --block-size 64 \
  --no-enable-prefix-caching \
  --disable-hybrid-kv-cache-manager
  • Result (PD disaggregated)
curl -X POST "http://localhost:8300/v1/chat/completions"   -H "Content-Type: application/json"   -d '{
    "model": "Qwen/Qwen3-0.6B",
    "messages": [
        {
        "role": "user",
        "content": "Hello, how are you?"
        }
    ],
    "max_tokens": 30, "stream": false
 }'
{"id":"chatcmpl-31aefcad-342c-481b-9b3e-bc2f8bd5ba45","object":"chat.completion","created":1775036538,"model":"Qwen/Qwen3-0.6B","choices":[{"index":0,
"message":{"role":"assistant",
"content":"```\nfrom collections import defaultdict\n\ndef main():\n    # Create a defaultdict with a specific key-value pair\n    d = defaultdict(lambda: 0",
"refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null}],
"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":14,"total_tokens":44,"completion_tokens":30,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}
  • Result (aggregated, routed to prefill or decoded only)
# prefill only
curl -X POST "http://localhost:8100/v1/chat/completions"   -H "Content-Type: application/json"   -d '{
    "model": "Qwen/Qwen3-0.6B",
    "messages": [
        {
        "role": "user",
        "content": "Hello, how are you?"
        }
    ],
    "max_tokens": 30, "stream": false
 }'
{"id":"chatcmpl-bbbee435812de192","object":"chat.completion","created":1775036902,"model":"Qwen/Qwen3-0.6B","choices":[{"index":0,"message":{"role":"assistant",
"content":"<think>\nOkay, the user just asked, \"Hello, how are you?\" I need to respond in a friendly and helpful way. Let me start",
"refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":14,"total_tokens":44,"completion_tokens":30,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}

# decode only
curl -X POST "http://localhost:8200/v1/chat/completions"   -H "Content-Type: application/json"   -d '{
    "model": "Qwen/Qwen3-0.6B",
    "messages": [
        {
        "role": "user",
        "content": "Hello, how are you?"
        }
    ],
    "max_tokens": 30, "stream": false
 }'
{"id":"chatcmpl-85ec524cd2af7bcd","object":"chat.completion","created":1775036733,"model":"Qwen/Qwen3-0.6B","choices":[{"index":0,"message":{"role":"assistant",
"content":"<think>\nOkay, the user asked \"Hello, how are you?\" I need to respond in a friendly way. Let me start with a greeting.",
"refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":14,"total_tokens":44,"completion_tokens":30,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null

The answer for hetero disaggregated case mismatch with the aggregated one, and incorrect.

Any thoughts are welcome!

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.

extent analysis

TL;DR

The issue is likely due to inconsistent kv transfer configurations between the prefill and decode instances, causing incorrect output in the hetero disaggregated case.

Guidance

  • Verify that the kv_transfer_config parameters are identical in both prefill and decode instances, especially the kv_connector, kv_role, and kv_buffer_device settings.
  • Check the enable_permute_local_kv flag in the kv_transfer_config to ensure it is set consistently across both instances.
  • Review the VLLM_NIXL_SIDE_CHANNEL_HOST and VLLM_NIXL_SIDE_CHANNEL_PORT environment variables to ensure they are correctly set for both prefill and decode instances.
  • Test the prefill and decode instances separately to ensure they produce the correct output when run individually.

Example

No code snippet is provided as the issue seems to be related to configuration inconsistencies rather than code errors.

Notes

The provided information suggests that the issue is related to the kv transfer configuration and the interaction between the prefill and decode instances. Further investigation is needed to determine the root cause of the inconsistency.

Recommendation

Apply a workaround by ensuring consistent kv transfer configurations across both prefill and decode instances, and test each instance separately to verify correct output. If the issue persists, consider upgrading to a newer version of the library or seeking additional support.

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…
#api#conversation history#tool integration#LLM response#environment variable

Still need to ship something?

×6

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

Back to top recommendations

TRENDING

vllm - ✅(Solved) Fix [Bug]: heterogeneous disaggregated serving XPU (Prefill) + CPU (Decode) accuracy issue [2 pull requests, 1 comments, 2 participants]