Issue Station
Sign In
Tools hereToolbox

vllm - 💡(How to fix) Fix [Bug]: Mistral-Small-4-119B-2603 fails on 8x RTX 3090 (SM 8.6) with vLLM v0.17.1: no valid MLA attention backend [1 comments, 1 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#37553•Fetched 2026-04-08 01:02:11
View on GitHub
Comments
1
Participants
1
Timeline
5
Reactions
0
Author
MartossienMartossien
Participants
MartossienMartossien
Timeline (top)
subscribed ×2commented ×1cross-referenced ×1labeled ×1

Root Cause

Build from source succeeded and vllm-0.17.1+cu131 installed successfully. [file:213] Changing --kv-cache-dtype from fp8 to auto did not fix the issue. [file:107][file:213] This does not appear to be a parser/reasoning issue, because the failure happens earlier during MLA backend selection. [file:213][file:228]

Fix Action

Fix / Workaround

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

Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 43 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 64 On-line CPU(s) list: 0-63 Vendor ID: AuthenticAMD Model name: AMD EPYC 7532 32-Core Processor CPU family: 23 Model: 49 Thread(s) per core: 2 Core(s) per socket: 32 Socket(s): 1 Stepping: 0 Frequency boost: enabled CPU(s) scaling MHz: 76% CPU max MHz: 3308,4089 CPU min MHz: 1500,0000 BogoMIPS: 4800,09 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 xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 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 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es Virtualization: AMD-V L1d cache: 1 MiB (32 instances) L1i cache: 1 MiB (32 instances) L2 cache: 16 MiB (32 instances) L3 cache: 256 MiB (16 instances) NUMA node(s): 1 NUMA node0 CPU(s): 0-63 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: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection Vulnerability Spec rstack overflow: Mitigation; Safe RET 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; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsa: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

RAW_BUFFERClick to expand / collapse

Your current environment

  • OS: Fedora Linux 42 (Workstation Edition)
  • Kernel: 6.18.16-100.fc42.x86_64
  • GPU: 8x NVIDIA RTX 3090 (SM 8.6, 24 GB each)
  • Driver: NVIDIA 580.126.09
  • CUDA toolkit used for build: 13.1.115
  • Compiler: GCC/G++ 14.2.1
  • Python: conda env vllm_env
  • vLLM: 0.17.1+cu131 built from source Host memory: 503 GiB DDR4 RAM (~482 GiB available at the time of testing).

For context, the host CPU is an AMD EPYC 7532 (Zen 2), which does not provide AVX512 support; however, this issue occurs on the CUDA path during MLA attention backend selection, not on the CPU backend.

<pre>(vllm_env) <font color="#26A269"><b>admin_ia@MiWiFi-R3600-srv</b></font>:<font color="#26A269"><b>/tmp</b></font>$ python collect_env.py Collecting environment information... ============================== System Info ============================== OS : Fedora Linux 42 (Workstation Edition) (x86_64) GCC version : (GCC) 15.2.1 20260123 (Red Hat 15.2.1-7) Clang version : Could not collect CMake version : version 4.2.1 Libc version : glibc-2.41 ============================== PyTorch Info ============================== PyTorch version : 2.10.0+cu128 Is debug build : False CUDA used to build PyTorch : 12.8 ROCM used to build PyTorch : N/A ============================== Python Environment ============================== Python version : 3.10.19 | packaged by conda-forge | (main, Jan 26 2026, 23:45:08) [GCC 14.3.0] (64-bit runtime) Python platform : Linux-6.18.16-100.fc42.x86_64-x86_64-with-glibc2.41 ============================== CUDA / GPU Info ============================== Is CUDA available : True CUDA runtime version : 13.1.115 CUDA_MODULE_LOADING set to : GPU models and configuration : GPU 0: NVIDIA GeForce RTX 3090 GPU 1: NVIDIA GeForce RTX 3090 GPU 2: NVIDIA GeForce RTX 3090 GPU 3: NVIDIA GeForce RTX 3090 GPU 4: NVIDIA GeForce RTX 3090 GPU 5: NVIDIA GeForce RTX 3090 GPU 6: NVIDIA GeForce RTX 3090 GPU 7: NVIDIA GeForce RTX 3090 Nvidia driver version : 580.126.09 cuDNN version : Probably one of the following: /usr/lib64/libcudnn.so.8.9.7 /usr/lib64/libcudnn.so.9.1.1 /usr/lib64/libcudnn_adv.so.9.1.1 /usr/lib64/libcudnn_adv_infer.so.8.9.7 /usr/lib64/libcudnn_adv_train.so.8.9.7 /usr/lib64/libcudnn_cnn.so.9.1.1 /usr/lib64/libcudnn_cnn_infer.so.8.9.7 /usr/lib64/libcudnn_cnn_train.so.8.9.7 /usr/lib64/libcudnn_engines_precompiled.so.9.1.1 /usr/lib64/libcudnn_engines_runtime_compiled.so.9.1.1 /usr/lib64/libcudnn_graph.so.9.1.1 /usr/lib64/libcudnn_heuristic.so.9.1.1 /usr/lib64/libcudnn_ops.so.9.1.1 /usr/lib64/libcudnn_ops_infer.so.8.9.7 /usr/lib64/libcudnn_ops_train.so.8.9.7 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: 43 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 64 On-line CPU(s) list: 0-63 Vendor ID: AuthenticAMD Model name: AMD EPYC 7532 32-Core Processor CPU family: 23 Model: 49 Thread(s) per core: 2 Core(s) per socket: 32 Socket(s): 1 Stepping: 0 Frequency boost: enabled CPU(s) scaling MHz: 76% CPU max MHz: 3308,4089 CPU min MHz: 1500,0000 BogoMIPS: 4800,09 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 xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 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 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es Virtualization: AMD-V L1d cache: 1 MiB (32 instances) L1i cache: 1 MiB (32 instances) L2 cache: 16 MiB (32 instances) L3 cache: 256 MiB (16 instances) NUMA node(s): 1 NUMA node0 CPU(s): 0-63 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: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection Vulnerability Spec rstack overflow: Mitigation; Safe RET 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; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected 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] flashinfer-python==0.6.4 [pip3] numpy==2.2.6 [pip3] nvidia-cublas-cu12==12.8.4.1 [pip3] nvidia-cuda-cupti-cu12==12.8.90 [pip3] nvidia-cuda-nvrtc-cu12==12.8.93 [pip3] nvidia-cuda-runtime-cu12==12.8.90 [pip3] nvidia-cudnn-cu12==9.10.2.21 [pip3] nvidia-cudnn-frontend==1.18.0 [pip3] nvidia-cufft-cu12==11.3.3.83 [pip3] nvidia-cufile-cu12==1.13.1.3 [pip3] nvidia-curand-cu12==10.3.9.90 [pip3] nvidia-cusolver-cu12==11.7.3.90 [pip3] nvidia-cusparse-cu12==12.5.8.93 [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==13.590.48 [pip3] nvidia-nccl-cu12==2.27.5 [pip3] nvidia-nvjitlink-cu12==12.8.93 [pip3] nvidia-nvshmem-cu12==3.4.5 [pip3] nvidia-nvtx-cu12==12.8.90 [pip3] pyzmq==27.1.0 [pip3] torch==2.10.0 [pip3] torch_c_dlpack_ext==0.1.5 [pip3] torchaudio==2.10.0 [pip3] torchvision==0.25.0 [pip3] transformers==4.57.6 [pip3] triton==3.6.0 [conda] cuda-cccl_linux-64 13.1.78 0 nvidia/label/cuda-13.1.0 [conda] cuda-command-line-tools 13.1.0 0 nvidia/label/cuda-13.1.0 [conda] cuda-compiler 13.1.0 ha63b06e_0 nvidia/label/cuda-13.1.0 [conda] cuda-crt-dev_linux-64 13.1.80 0 nvidia/label/cuda-13.1.0 [conda] cuda-crt-tools 13.1.80 0 nvidia/label/cuda-13.1.0 [conda] cuda-ctadvisor 13.1.80 h3b4bcfc_1 nvidia/label/cuda-13.1.0 [conda] cuda-cuobjdump 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] cuda-cupti 13.1.75 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-cupti-dev 13.1.75 h3b2e0c5_0 nvidia/label/cuda-13.1.0 [conda] cuda-cuxxfilt 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] cuda-gdb 13.1.68 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-libraries 13.1.0 0 nvidia/label/cuda-13.1.0 [conda] cuda-libraries-dev 13.1.0 0 nvidia/label/cuda-13.1.0 [conda] cuda-nsight 13.1.68 h39186ea_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvcc 13.1.80 h959976e_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvcc-dev_linux-64 13.1.80 0 nvidia/label/cuda-13.1.0 [conda] cuda-nvcc-impl 13.1.80 hbe36340_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvcc-tools 13.1.80 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvcc_linux-64 13.1.80 h27567d6_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvdisasm 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvml-dev 13.1.68 h34f9132_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvprune 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvrtc 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvrtc-dev 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvtx 13.1.68 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvvm-dev_linux-64 13.1.80 0 nvidia/label/cuda-13.1.0 [conda] cuda-nvvm-impl 13.1.80 h9f10756_0 nvidia/label/cuda-13.1.0 [conda] cuda-nvvm-tools 13.1.80 h9f10756_0 nvidia/label/cuda-13.1.0 [conda] cuda-opencl 13.1.80 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-opencl-dev 13.1.80 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-profiler-api 13.1.80 hf49826d_0 nvidia/label/cuda-13.1.0 [conda] cuda-sanitizer-api 13.1.75 h3b2e0c5_0 nvidia/label/cuda-13.1.0 [conda] cuda-tileiras 13.1.80 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] cuda-toolkit 13.1.0 ha63b06e_0 nvidia/label/cuda-13.1.0 [conda] cuda-tools 13.1.0 0 nvidia/label/cuda-13.1.0 [conda] cuda-visual-tools 13.1.0 0 nvidia/label/cuda-13.1.0 [conda] flashinfer-python 0.6.4 pypi_0 pypi [conda] gds-tools 1.16.0.49 hcbabc2f_0 nvidia/label/cuda-13.1.0 [conda] libcublas 13.2.0.9 h26745a4_0 nvidia/label/cuda-13.1.0 [conda] libcublas-dev 13.2.0.9 h26745a4_0 nvidia/label/cuda-13.1.0 [conda] libcufft 12.1.0.31 hed8d0a8_0 nvidia/label/cuda-13.1.0 [conda] libcufft-dev 12.1.0.31 hed8d0a8_0 nvidia/label/cuda-13.1.0 [conda] libcufile 1.16.0.49 h3b4bcfc_0 nvidia/label/cuda-13.1.0 [conda] libcufile-dev 1.16.0.49 hcbabc2f_0 nvidia/label/cuda-13.1.0 [conda] libcurand 10.4.1.34 h5dabb87_0 nvidia/label/cuda-13.1.0 [conda] libcurand-dev 10.4.1.34 h5dabb87_0 nvidia/label/cuda-13.1.0 [conda] libcusolver 12.0.7.41 h38cde51_0 nvidia/label/cuda-13.1.0 [conda] libcusolver-dev 12.0.7.41 h38cde51_0 nvidia/label/cuda-13.1.0 [conda] libcusparse 12.7.2.19 hd0a4c84_0 nvidia/label/cuda-13.1.0 [conda] libcusparse-dev 12.7.2.19 hd0a4c84_0 nvidia/label/cuda-13.1.0 [conda] libnpp 13.0.2.21 hd917672_0 nvidia/label/cuda-13.1.0 [conda] libnpp-dev 13.0.2.21 hd917672_0 nvidia/label/cuda-13.1.0 [conda] libnvfatbin 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] libnvfatbin-dev 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] libnvjitlink 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] libnvjitlink-dev 13.1.80 h6205011_0 nvidia/label/cuda-13.1.0 [conda] libnvjpeg 13.0.2.28 h17107e5_0 nvidia/label/cuda-13.1.0 [conda] libnvjpeg-dev 13.0.2.28 h92ecfc8_0 nvidia/label/cuda-13.1.0 [conda] libnvptxcompiler-dev 13.1.80 0 nvidia/label/cuda-13.1.0 [conda] libnvptxcompiler-dev_linux-64 13.1.80 0 nvidia/label/cuda-13.1.0 [conda] nsight-compute 2025.4.0.12 hdc9f5c9_0 nvidia/label/cuda-13.1.0 [conda] numpy 2.2.6 pypi_0 pypi [conda] nvidia-cublas-cu12 12.8.4.1 pypi_0 pypi [conda] nvidia-cuda-cupti-cu12 12.8.90 pypi_0 pypi [conda] nvidia-cuda-nvrtc-cu12 12.8.93 pypi_0 pypi [conda] nvidia-cuda-runtime-cu12 12.8.90 pypi_0 pypi [conda] nvidia-cudnn-cu12 9.10.2.21 pypi_0 pypi [conda] nvidia-cudnn-frontend 1.18.0 pypi_0 pypi [conda] nvidia-cufft-cu12 11.3.3.83 pypi_0 pypi [conda] nvidia-cufile-cu12 1.13.1.3 pypi_0 pypi [conda] nvidia-curand-cu12 10.3.9.90 pypi_0 pypi [conda] nvidia-cusolver-cu12 11.7.3.90 pypi_0 pypi [conda] nvidia-cusparse-cu12 12.5.8.93 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 13.590.48 pypi_0 pypi [conda] nvidia-nccl-cu12 2.27.5 pypi_0 pypi [conda] nvidia-nvjitlink-cu12 12.8.93 pypi_0 pypi [conda] nvidia-nvshmem-cu12 3.4.5 pypi_0 pypi [conda] nvidia-nvtx-cu12 12.8.90 pypi_0 pypi [conda] pyzmq 27.1.0 pypi_0 pypi [conda] torch 2.10.0 pypi_0 pypi [conda] torch-c-dlpack-ext 0.1.5 pypi_0 pypi [conda] torchaudio 2.10.0 pypi_0 pypi [conda] torchvision 0.25.0 pypi_0 pypi [conda] transformers 4.57.6 pypi_0 pypi [conda] triton 3.6.0 pypi_0 pypi ============================== vLLM Info ============================== ROCM Version : Could not collect vLLM Version : 0.17.1 vLLM Build Flags: CUDA Archs: 8.6; ROCm: Disabled GPU Topology: <u style="text-decoration-style:solid">GPU0</u> <u style="text-decoration-style:solid">GPU1</u> <u style="text-decoration-style:solid">GPU2</u> <u style="text-decoration-style:solid">GPU3</u> <u style="text-decoration-style:solid">GPU4</u> <u style="text-decoration-style:solid">GPU5</u> <u style="text-decoration-style:solid">GPU6</u> <u style="text-decoration-style:solid">GPU7</u> <u style="text-decoration-style:solid">CPU Affinity</u> <u style="text-decoration-style:solid">NUMA Affinity</u> <u style="text-decoration-style:solid">GPU NUMA ID</u> GPU0 X PHB PHB NODE NODE NODE NODE NODE 0-63 0 N/A GPU1 PHB X PHB NODE NODE NV4 NODE NODE 0-63 0 N/A GPU2 PHB PHB X NODE NODE NODE NODE NODE 0-63 0 N/A GPU3 NODE NODE NODE X NODE NODE NV4 NODE 0-63 0 N/A GPU4 NODE NODE NODE NODE X PHB NODE NODE 0-63 0 N/A GPU5 NODE NV4 NODE NODE PHB X NODE NODE 0-63 0 N/A GPU6 NODE NODE NODE NV4 NODE NODE X PHB 0-63 0 N/A GPU7 NODE NODE NODE NODE NODE NODE PHB X 0-63 0 N/A 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 ============================== Environment Variables ============================== NCCL_P2P_DISABLE=0 TORCH_CUDA_ARCH_LIST=8.6 NVCC_THREADS=1 CUDAARCHS=75-real;80-real;86-real;89-real;90-real;100-real;103-real;110-real;120-real;121 CUDA_PATH=/home/admin_ia/.conda/envs/vllm_env MAX_JOBS=64 VLLM_TARGET_DEVICE=cuda LD_LIBRARY_PATH=/usr/lib/python3.10/site-packages/torch/lib:/usr/local/cuda/lib64:/usr/lib/python3.10/site-packages/torch/lib:/usr/local/cuda/lib64: NCCL_IB_DISABLE=1 CUDA_HOME=/home/admin_ia/.conda/envs/vllm_env CUDA_HOME=/home/admin_ia/.conda/envs/vllm_env CUDAToolkit_ROOT=/home/admin_ia/.conda/envs/vllm_env PYTORCH_NVML_BASED_CUDA_CHECK=1 TORCHINDUCTOR_COMPILE_THREADS=1 TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_admin_ia </pre>

Note: collect_env.py reports the system GCC (15.2.1), but vLLM was built from source with CC=/usr/bin/gcc-14 and CXX=/usr/bin/g++-14.

🐛 Describe the bug

Mistral-Small-4-119B-2603 fails to start on my setup with vLLM v0.17.1 on 8x RTX 3090 (SM 8.6).

The project builds successfully from source (vllm-0.17.1+cu131 installs correctly), the server starts, the model architecture is resolved, and loading begins, but initialization fails during MLA attention backend selection. [file:213]

I tested both:

  1. automatic backend selection
  2. explicit --attention-backend FLASH_ATTN_MLA

Both fail. [file:213][file:228]

Reproduction

Command with automatic backend selection:

cd /tmp
conda run -n vllm_env --live-stream vllm serve \
  /path/to/Mistral-Small-4-119B-2603 \
  --port 8000 \
  --tensor-parallel-size 8 \
  --quantization fp8 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.88 \
  --kv-cache-dtype auto \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --distributed-executor-backend ray \
  --disable-custom-all-reduce \
  --limit-mm-per-prompt.image 4

Observed behavior

The model is resolved as:
Resolved architecture: PixtralForConditionalGeneration

Then vLLM fails during MLA attention backend selection. [file:213][file:228]

With automatic backend selection:
ValueError: No valid attention backend found for cuda with AttentionSelectorConfig(head_size=320, dtype=torch.bfloat16, kv_cache_dtype=auto, block_size=None, use_mla=True, has_sink=False, use_sparse=False, use_mm_prefix=False, use_per_head_quant_scales=False, attn_type=decoder). Reasons: {FLASH_ATTN_MLA: [head_size not supported, compute capability not supported, FlashAttention MLA not supported on this device], FLASHMLA: [head_size not supported, compute capability not supported, vllm._flashmla_C is not available, likely was not compiled due to insufficient nvcc version or a supported arch was not in the list of target arches to compile for.], FLASHINFER_MLA: [head_size not supported, compute capability not supported, FlashInfer MLA kernel requires qk_nope_head_dim == 128, but got 64], TRITON_MLA: [head_size not supported], FLASHMLA_SPARSE: [head_size not supported, non-sparse not supported, compute capability not supported]}.

With explicit --attention-backend FLASH_ATTN_MLA:
ValueError: Selected backend AttentionBackendEnum.FLASH_ATTN_MLA is not valid for this configuration. Reason: ['head_size not supported', 'compute capability not supported', 'FlashAttention MLA not supported on this device']

Expected behavior

I expected either:

    the model to run on this hardware, or
    a clear indication that this MLA configuration is not supported on Ampere / SM 8.6 for this model. [file:228]

Additional notes

    Build from source succeeded and vllm-0.17.1+cu131 installed successfully. [file:213]
    Changing --kv-cache-dtype from fp8 to auto did not fix the issue. [file:107][file:213]
    This does not appear to be a parser/reasoning issue, because the failure happens earlier during MLA backend selection. [file:213][file:228]


### Before submitting a new issue...

- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.

extent analysis

Fix Plan

To resolve the issue with MLA attention backend selection, follow these steps:

  • Update CUDA Toolkit: Ensure you are using the latest CUDA Toolkit version compatible with your NVIDIA driver. In this case, update to CUDA 13.1 or later.
  • Update PyTorch: Update PyTorch to the latest version compatible with your CUDA Toolkit. You can install PyTorch using pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu113.
  • Rebuild vLLM: Rebuild vLLM from source using the updated CUDA Toolkit and PyTorch versions.
  • Explicitly set attention backend: Try setting the attention backend to FLASH_ATTN_MLA explicitly using the --attention-backend flag.
  • Check head size and compute capability: Verify that the head size and compute capability of your GPU are supported by the MLA attention backend.

Example code to rebuild vLLM:

# Update CUDA Toolkit and PyTorch
conda install cuda-toolkit=13.1
pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu113

# Rebuild vLLM from source
git clone https://github.com/vllm/vllm.git
cd vllm
git checkout v0.17.1
CC=/usr/bin/gcc-14 CXX=/usr/bin/g++-14 python setup.py install

Example command to run vLLM with explicit attention backend:

conda run -n vllm_env --live-stream vllm serve \
  /path/to/Mistral-Small-4-119B-2603 \
  --port 8000 \
  --tensor-parallel-size 8 \
  --quantization fp8 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.88 \
  --kv-cache-dtype auto \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --distributed-executor-backend ray \
  --disable-custom-all-reduce \
  --limit-mm-per-prompt.image 4 \
  --attention-backend FLASH_ATTN_MLA

Verification

To verify that the fix worked, check the vLLM logs for any errors related to MLA attention backend selection. If the issue is resolved, the model should run successfully on your hardware.

Extra Tips

  • Ensure that your GPU drivers are up-to-date and compatible with your CUDA Toolkit version.
  • If you encounter any issues during the rebuild process, refer to the vLLM documentation and GitHub issues for troubleshooting guides.
  • Consider using a virtual environment to manage dependencies and avoid conflicts with other projects.

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#ssr#installation#tensor shape#model download#tokenizer error#prompt formatting#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