[Bug]: gemma3-27b-it shows degraded accuracy in vLLM v0.11.0

[sammysun0711]

Your current environment

Current environment
vLLM version: v0.11.0
docker image: vllm/vllm-openai:v0.11.0

The output of python collect_env.py

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version                : Could not collect
CMake version                : version 4.1.0
Libc version                 : glibc-2.35

==============================
       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  4 2025, 08:56:18) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-5.15.0-153-generic-x86_64-with-glibc2.35

==============================
       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 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        : 570.172.08
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:                           52 bits physical, 57 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  224
On-line CPU(s) list:                     0-223
Vendor ID:                               GenuineIntel
Model name:                              Intel(R) Xeon(R) Platinum 8480+
CPU family:                              6
Model:                                   143
Thread(s) per core:                      2
Core(s) per socket:                      56
Socket(s):                               2
Stepping:                                8
BogoMIPS:                                4000.00
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 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi 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 avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                          VT-x
L1d cache:                               5.3 MiB (112 instances)
L1i cache:                               3.5 MiB (112 instances)
L2 cache:                                224 MiB (112 instances)
L3 cache:                                210 MiB (2 instances)
NUMA node(s):                            2
NUMA node0 CPU(s):                       0-55,112-167
NUMA node1 CPU(s):                       56-111,168-223
Vulnerability Gather data sampling:      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 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 and seccomp
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB disabled; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds:                     Not affected
Vulnerability Tsx async abort:           Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.3.1
[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==1.14.1
[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-ml-py==12.575.51
[pip3] nvidia-nccl-cu12==2.27.3
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] pynvml==12.0.0
[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==4.57.0
[pip3] triton==3.4.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.11.0
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
        GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    NIC0    NIC1    NIC2    NIC3    NIC4    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV18    NV18    NV18    NV18    NV18    NV18    NV18    PIX     NODE    SYS     SYS     SYS     0-55,112-167    0               N/A
GPU1    NV18     X      NV18    NV18    NV18    NV18    NV18    NV18    NODE    NODE    SYS     SYS     SYS     0-55,112-167    0               N/A
GPU2    NV18    NV18     X      NV18    NV18    NV18    NV18    NV18    NODE    PIX     SYS     SYS     SYS     0-55,112-167    0               N/A
GPU3    NV18    NV18    NV18     X      NV18    NV18    NV18    NV18    NODE    NODE    SYS     SYS     SYS     0-55,112-167    0               N/A
GPU4    NV18    NV18    NV18    NV18     X      NV18    NV18    NV18    SYS     SYS     PIX     NODE    NODE    56-111,168-223  1               N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X      NV18    NV18    SYS     SYS     NODE    NODE    NODE    56-111,168-223  1               N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X      NV18    SYS     SYS     NODE    NODE    PIX     56-111,168-223  1               N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X      SYS     SYS     NODE    NODE    NODE    56-111,168-223  1               N/A
NIC0    PIX     NODE    NODE    NODE    SYS     SYS     SYS     SYS      X      NODE    SYS     SYS     SYS
NIC1    NODE    NODE    PIX     NODE    SYS     SYS     SYS     SYS     NODE     X      SYS     SYS     SYS
NIC2    SYS     SYS     SYS     SYS     PIX     NODE    NODE    NODE    SYS     SYS      X      NODE    NODE
NIC3    SYS     SYS     SYS     SYS     NODE    NODE    NODE    NODE    SYS     SYS     NODE     X      NODE
NIC4    SYS     SYS     SYS     SYS     NODE    NODE    PIX     NODE    SYS     SYS     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: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4

==============================
     Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=all
NVIDIA_REQUIRE_CUDA=cuda>=12.8 brand=unknown,driver>=470,driver<471 brand=grid,driver>=470,driver<471 brand=tesla,driver>=470,driver<471 brand=nvidia,driver>=470,driver<471 brand=quadro,driver>=470,driver<471 brand=quadrortx,driver>=470,driver<471 brand=nvidiartx,driver>=470,driver<471 brand=vapps,driver>=470,driver<471 brand=vpc,driver>=470,driver<471 brand=vcs,driver>=470,driver<471 brand=vws,driver>=470,driver<471 brand=cloudgaming,driver>=470,driver<471 brand=unknown,driver>=535,driver<536 brand=grid,driver>=535,driver<536 brand=tesla,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=vapps,driver>=535,driver<536 brand=vpc,driver>=535,driver<536 brand=vcs,driver>=535,driver<536 brand=vws,driver>=535,driver<536 brand=cloudgaming,driver>=535,driver<536 brand=unknown,driver>=550,driver<551 brand=grid,driver>=550,driver<551 brand=tesla,driver>=550,driver<551 brand=nvidia,driver>=550,driver<551 brand=quadro,driver>=550,driver<551 brand=quadrortx,driver>=550,driver<551 brand=nvidiartx,driver>=550,driver<551 brand=vapps,driver>=550,driver<551 brand=vpc,driver>=550,driver<551 brand=vcs,driver>=550,driver<551 brand=vws,driver>=550,driver<551 brand=cloudgaming,driver>=550,driver<551 brand=unknown,driver>=560,driver<561 brand=grid,driver>=560,driver<561 brand=tesla,driver>=560,driver<561 brand=nvidia,driver>=560,driver<561 brand=quadro,driver>=560,driver<561 brand=quadrortx,driver>=560,driver<561 brand=nvidiartx,driver>=560,driver<561 brand=vapps,driver>=560,driver<561 brand=vpc,driver>=560,driver<561 brand=vcs,driver>=560,driver<561 brand=vws,driver>=560,driver<561 brand=cloudgaming,driver>=560,driver<561 brand=unknown,driver>=565,driver<566 brand=grid,driver>=565,driver<566 brand=tesla,driver>=565,driver<566 brand=nvidia,driver>=565,driver<566 brand=quadro,driver>=565,driver<566 brand=quadrortx,driver>=565,driver<566 brand=nvidiartx,driver>=565,driver<566 brand=vapps,driver>=565,driver<566 brand=vpc,driver>=565,driver<566 brand=vcs,driver>=565,driver<566 brand=vws,driver>=565,driver<566 brand=cloudgaming,driver>=565,driver<566
NCCL_VERSION=2.25.1-1
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NVIDIA_PRODUCT_NAME=CUDA
VLLM_USAGE_SOURCE=production-docker-image
CUDA_VERSION=12.8.1
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

I'm testing open-source models using vLLM v0.11.0 on Nvidia H20 and AMD Instinct MI308X using lm_eval==0.4.9.1 from lm-evaluation-harness. However, I found gemma-3-27b-it shows degraded accuracy compared to Hugging face transformers implementation, similar issue reported in #17689, PR21927 try to address sliding window configuration issue, but it seems does not resolve the original accuracy issue.

HF transformers 4.57.1 results

lm_eval --model hf --model_args pretrained=google/gemma-3-27b-it,dtype=bfloat16,trust_remote_code=True --tasks hellaswag,lambada_openai

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
hellaswag	1	none	0	acc	↑	0.6503	±	0.0048
		none	0	acc_norm	↑	0.8406	±	0.0037
lambada_openai	1	none	0	acc	↑	0.6953	±	0.0064
		none	0	perplexity	↓	3.7726	±	0.1152

H20 + vLLM 0.11.0 results
lm_eval --model vllm --model_args pretrained=google/gemma-3-27b-it,tensor_parallel_size=2,dtype=bfloat16,gpu_memory_utilization=0.8,trust_remote_code=True,enforce_eager=True,max_gen_toks=256 --tasks hellaswag,lambada_openai

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
hellaswag	1	none	0	acc	↑	0.3868	±	0.0049
		none	0	acc_norm	↑	0.4858	±	0.0050
lambada_openai	1	none	0	acc	↑	0.2243	±	0.0058
		none	0	perplexity	↓	960.5903	±	76.8435

AMD Instinct 308X Triton Flash Attention + vLLM 0.11.0 results
VLLM_ROCM_USE_AITER=1 VLLM_USE_TRITON_FLASH_ATTN=1 VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER_PAGED_ATTN=0 lm_eval --model vllm --model_args pretrained=google/gemma-3-27b-it,tensor_parallel_size=1,dtype=bfloat16,gpu_memory_utilization=0.8,trust_remote_code=True,enforce_eager=True,,max_gen_toks=256 --tasks hellaswag,lambada_openai

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
hellaswag	1	none	0	acc	↑	0.3873	±	0.0049
		none	0	acc_norm	↑	0.4858	±	0.0050
lambada_openai	1	none	0	acc	↑	0.2271	±	0.0058
		none	0	perplexity	↓	950.2281	±	77.4239

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.

[github-actions]

CC @hongxiayang @tjtanaa @vllmellm for ROCm-related issue

[sammysun0711]

Hi @hmellor, I noticed you previously worked on the gemma‑3‑27b‑it accuracy issue. Could you please take a look?

[sammysun0711]

CC @hongxiayang @tjtanaa @vllmellm for ROCm-related issue

This issue occurs on Nvidia platform too, so I think it is more related to vLLM framework instead of ROCm alone.

[hmellor]

@sammysun0711 can I ask for 2 things:

• Do you hve comparisons for previous vLLM versions that worked? (This would help pinpoint where things broke in vLLM)
• Could you try running your eval with --model-impl transformers to see if this resolves the accuracy issue? (The throughput should be >95% of the default vLLM impl)

cc @lucianommartins for Gemma

[hmellor]

(removing hardware labels because it's not a hardware issue)

[sammysun0711]

Thanks @hmellor for your suggestions. Here are some feedback from myside:

Do you have comparisons for previous vLLM versions that worked? (This would help pinpoint where things broke in vLLM)

I have not try to investigate which vLLM version works without accuracy issue, do you have any suggestion/commit for us to start with?

Could you try running your eval with --model-impl transformers to see if this resolves the accuracy issue? (The throughput should be >95% of the default vLLM impl)

I added model_impl="transformers" as arguments in lm-eval command line:
lm_eval --model vllm --model_args pretrained=google/gemma-3-27b-it,tensor_parallel_size=2,dtype=bfloat16,gpu_memory_utilization=0.8,trust_remote_code=True,enforce_eager=True,max_gen_toks=256,model_impl="transformers" --tasks hellaswag,lambada_openai

Here are test results on H20 w/ model_impl="transformers", please correct me if any need to be update with command line

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
hellaswag	1	none	0	acc	↑	0.2551	±	0.0044
		none	0	acc_norm	↑	0.2628	±	0.0044
lambada_openai	1	none	0	acc	↑	0.000	±	0.0000
		none	0	perplexity	↓	34480370.878	±	3185248.3029