[Bug]: Hybrid Attention models broken after switching to flashinfer 0.4 (tested on Granite 4.0 H, Qwen3-Next, Jamba-3B, Nemotron-H-8b)

[Steven0236]

Your current environment

The output of python collect_env.py

Collecting environment information...
uv is set
==============================
        System Info
==============================
OS                           : Fedora Linux 41 (Workstation Edition) (x86_64)
GCC version                  : (GCC) 14.3.1 20250808 (Red Hat 14.3.1-3)
Clang version                : Could not collect
CMake version                : version 4.1.0
Libc version                 : glibc-2.40

==============================
       PyTorch Info
==============================
PyTorch version              : 2.8.0+cu128
Is debug build               : False
CUDA used to build PyTorch   : 12.8
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.11 (main, Jun 12 2025, 00:00:00) [GCC 14.3.1 20250523 (Red Hat 14.3.1-1)] (64-bit runtime)
Python platform              : Linux-6.16.5-100.fc41.x86_64-x86_64-with-glibc2.40

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 12.8.93
CUDA_MODULE_LOADING set to   : LAZY
GPU models and configuration : 
GPU 0: NVIDIA GeForce RTX 5070 Ti
GPU 1: NVIDIA GeForce RTX 5070 Ti
GPU 2: NVIDIA GeForce RTX 5070 Ti
GPU 3: NVIDIA GeForce RTX 5070 Ti

Nvidia driver version        : 575.64.05
cuDNN version                : Could not collect
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:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               AuthenticAMD
Model name:                              AMD Ryzen 7 7700 8-Core Processor
CPU family:                              25
Model:                                   97
Thread(s) per core:                      2
Core(s) per socket:                      8
Socket(s):                               1
Stepping:                                2
Frequency boost:                         enabled
CPU(s) scaling MHz:                      62%
CPU max MHz:                             5392.8721
CPU min MHz:                             422.3340
BogoMIPS:                                7599.98
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 amd_lbr_v2 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 perfmon_v2 ibrs ibpb stibp ibrs_enhanced 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 user_shstk avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid overflow_recov succor smca fsrm flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               256 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                8 MiB (8 instances)
L3 cache:                                32 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
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:                  Not affected
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; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Mitigation; Clear CPU buffers
Vulnerability Tsx async abort:           Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.3.0
[pip3] numpy==2.3.4
[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==1.15.0
[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==4.2.1
[pip3] nvidia-ml-py==13.580.82
[pip3] nvidia-nccl-cu12==2.27.3
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvshmem-cu12==3.3.20
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] pynvml==13.0.1
[pip3] pyzmq==27.1.0
[pip3] torch==2.8.0+cu128
[pip3] torchaudio==2.8.0+cu128
[pip3] torchvision==0.23.0+cu128
[pip3] transformers==5.0.0.dev0
[pip3] triton==3.4.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.11.1rc2.dev47+g136a17fe6.d20251015 (git sha: 136a17fe6, date: 20251015)
vLLM Build Flags:
  CUDA Archs: 12.0; ROCm: Disabled
GPU Topology:
  	GPU0	GPU1	GPU2	GPU3	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	PHB	PHB	PHB	0-15	0		N/A
GPU1	PHB	 X 	PHB	PHB	0-15	0		N/A
GPU2	PHB	PHB	 X 	PHB	0-15	0		N/A
GPU3	PHB	PHB	PHB	 X 	0-15	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
==============================
TORCH_CUDA_ARCH_LIST=12.0
MAX_JOBS=16
LD_LIBRARY_PATH=:/usr/local/cuda/lib64
CUDA_HOME=/usr/local/cuda
CUDA_HOME=/usr/local/cuda
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

(reported earlier in flashinfer-ai/flashinfer#1931)

After upgrading vllm (compiled from source) and flashinfer (to 0.4.0), I noticed that Qwen3-Next-80b has lost a lot of precision and can't use tools and talks nonsense after the 2nd conversation turn. If I switch the backend to "FLASH_ATTN", the problem goes away. This makes me think that the problem is likely in flashinfer.

Trying to isolate the problem:

• Switching to flash_attn backend, everything works normally.
• Switching the model to Qwen3-4b, everything works normally.
• Switching to an older version of vllm with flashinfer 0.3.0, everything works normally.
• Unfortunately, I cannot try the latest vllm version with flashinfer 0.3.0 or the old version of vllm with flashinfer 0.4.0 because vllm is apparently locked to a given flashinfer version due to some function call parameters.

I am unsure if the problem is due to some error in which vllm uses flashinfer 0.4.0 or if the error is entirely within flashinfer, which could have introduced a bug that may only cause a loss of precision in the special attention mechanism of Qwen3-Next. Note that I'm running on NVIDIA sm120 GPUs,

Any confirmation of the problem or suggestions appreciated.

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.

[DarkLight1337]

We are aware of this problem and have disabled it by default in #26859

[Steven0236]

@DarkLight1337 , This problem occurs even when the flashinfer sampler is disabled. As reported in flashinfer-ai/flashinfer#1931, the problem is present with the following configuration:

VLLM_ATTENTION_BACKEND="FLASHINFER"
VLLM_USE_FLASHINFER_SAMPLER="0"

[DarkLight1337]

Hmm, so it might be a problem with flashinfer attention, not just the sampler

[Steven0236]

Or it might be a problem with how vllm uses the flashinfer backend. Unfortunately, I can't determine which side has the problem as I don't have another inference engine that can use the flashinfer backend.

With two tools changed at once, it's hard to determine on which side the error was introduced. Hopefully, in the future, the interface between vllm and flashinfer will be more stable, so that we can revert to an older version of flashinfer to troubleshoot.

[DarkLight1337]

Ok, reopening this for now. cc @mgoin @tlrmchlsmth

[Steven0236]

Rebuilt vllm from source just to be double sure.
Did more testing with: vllm:0.11.1rc2.dev72, Transformers:4.57.0.dev0, Flashinfer:0.4.0

Precision problem still there with the configuration:

- Environment Variables
VLLM_ATTENTION_BACKEND="FLASHINFER"
VLLM_USE_FLASHINFER_SAMPLER="0"
MAX_JOBS="16"
OMP_NUM_THREADS="16"
NVCC_THREADS="16"
PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True"

vllm serve /AI2/vllm_models/Text/2509/Qwen3-Next-80B-AWQ4     \
  --trust-remote-code                                         \
  --max-model-len                  180000                     \
  --max-num-seqs                   1                          \
  --tensor-parallel-size           1                          \
  --pipeline-parallel-size         4                          \
  --enable-auto-tool-choice                                   \
  --tool-call-parser               hermes                     \
  --tokenizer-mode                 auto                       \
  --gpu-memory-utilization         0.89                       \
  --no-enable-chunked-prefill                                

No problem with the configuration:

- Environment Variables
VLLM_ATTENTION_BACKEND="FLASH_ATTN"
VLLM_USE_FLASHINFER_SAMPLER="0"
MAX_JOBS="16"
OMP_NUM_THREADS="16"
NVCC_THREADS="16"
PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True"

vllm serve /AI2/vllm_models/Text/2509/Qwen3-Next-80B-AWQ4     \
  --trust-remote-code                                         \
  --max-model-len                  180000                     \
  --max-num-seqs                   1                          \
  --tensor-parallel-size           1                          \
  --pipeline-parallel-size         4                          \
  --enable-auto-tool-choice                                   \
  --tool-call-parser               hermes                     \
  --tokenizer-mode                 auto                       \
  --gpu-memory-utilization         0.89                       \
  --no-enable-chunked-prefill     

Something odd: Prefill speed on 97k token summarization test is 77% faster using flashinfer than flash_attn. Can it really be that much faster or is the calculation skipping some steps?

• Using Flashinfer: ~23000 tokens/sec
• Using Flash_attn: ~13000 tokens/sec

[Steven0236]

I've rebuilt from source just to be double-sure.
Now running: Vllm 0.11.1rc2.dev72, Transformers:4.57.0.dev0, Flashinfer:0.4.0
The vLLM output is shown in attached log.txt.
Notice that it does not say 'Using FlashInfer for top-p & top-k sampling.', as flashinfer sampling is disabled.
Interestingly, seems like vllm no longer outputs a message to mention that is using some other sampler like it used to.

log.txt

[Steven0236]

Note:
1- Someone had reported suspected precision issue with Qwen3-VL-8b in #26902, but I've tested it and it seems to work perfectly fine with flashinfer backend even with the rather demanding tool use in Cline.
2- Someone had reported suspected precision issue with Qwen3-30b-A3b-2507 in #26930, but I tried it with flashinfer backend and it seemed to work fine with tool use in Cline.

Conclusion: It appears that only Qwen3-Next is facing from serious lack of precision when using the flashinfer backend.

[ExtReMLapin]

Can confirm, had to :

VLLM_ATTENTION_BACKEND=FLASH_ATTN VLLM_USE_FLASHINFER_SAMPLER=0 else it was outputing insanities and crazy tool calls