vllm: [Bug]: InternVL3 poor (random) output with 8bit quantization
Your current environment
The output of python collect_env.py
==============================
System Info
==============================
OS : Ubuntu 24.04.1 LTS (x86_64)
GCC version : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version : Could not collect
CMake version : version 3.28.3
Libc version : glibc-2.39
==============================
PyTorch Info
==============================
PyTorch version : 2.7.0+cu126
Is debug build : False
CUDA used to build PyTorch : 12.6
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.0 (main, Oct 3 2023, 01:27:23) [Clang 17.0.1 ] (64-bit runtime)
Python platform : Linux-6.8.0-58-generic-x86_64-with-glibc2.39
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : Could not collect
CUDA_MODULE_LOADING set to : LAZY
GPU models and configuration :
GPU 0: NVIDIA H100 NVL
GPU 1: NVIDIA H100 NVL
Nvidia driver version : 570.133.07
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): 32
On-line CPU(s) list: 0-31
Vendor ID: AuthenticAMD
BIOS Vendor ID: Advanced Micro Devices, Inc.
Model name: AMD EPYC 9174F 16-Core Processor
BIOS Model name: AMD EPYC 9174F 16-Core Processor Unknown CPU @ 4.1GHz
BIOS CPU family: 107
CPU family: 25
Model: 17
Thread(s) per core: 2
Core(s) per socket: 16
Socket(s): 1
Stepping: 1
Frequency boost: enabled
CPU(s) scaling MHz: 47%
CPU max MHz: 4408.2998
CPU min MHz: 1500.0000
BogoMIPS: 8200.40
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 nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 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 amd_ppin cppc amd_ibpb_ret arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d debug_swap
Virtualization: AMD-V
L1d cache: 512 KiB (16 instances)
L1i cache: 512 KiB (16 instances)
L2 cache: 16 MiB (16 instances)
L3 cache: 256 MiB (8 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
Vulnerability Gather data sampling: 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: 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; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-ml-py==12.575.51
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pynvml==12.0.0
[pip3] pyzmq==27.0.0
[pip3] torch==2.7.0
[pip3] torchaudio==2.7.0
[pip3] torchvision==0.22.0
[pip3] transformers==4.52.4
[pip3] triton==3.3.0
[conda] Could not collect
==============================
vLLM Info
==============================
ROCM Version : Could not collect
Neuron SDK Version : N/A
vLLM Version : 0.9.1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0 GPU1 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV12 0-31 0 N/A
GPU1 NV12 X 0-31 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_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY
🐛 Describe the bug
InternVL3 8-Bit quantizations (FP8, BNB-8Bit) Produce Bad/Random outputs on vLLM
When loading FP8 (compressed-tensors via llm-compressor) or BNB-8Bit (via transformers) quantized InternVL3 models in vllm, the output is bad/random. However, loading these same models into Transformers - AutoModel.from_pretrained results in proper output as expected.
Worth noting, lm_head, vision, and the mlp1 layers are not quantized. (mlp1 contains linear layers, but, the vllm internvl.py model doesn’t support scale values currently).
Example:
model = LLM(
model="brandonbeiler/InternVL3-38B-FP8-Dynamic", # also with 'brandonbeiler/InternVL3-38B-BNB-8bit'
trust_remote_code=True,
max_model_len=4096,
tensor_parallel_size=1,
gpu_memory_utilization=0.8,
enforce_eager=True # attempted with and without eager
)
sampling_params = SamplingParams(
temperature=0.7,
max_tokens=128,
top_p=0.9,
)
model.generate(['Hello! How are you today?'], sampling_params)
Output:
(\全全Rock separatoruta de尚未Java team com com com com com comett com com com com com com com com := Ne com com com com com com com com comett#!/ -- drummer delibernamespaceOffsetused bizarre Kre山西 com com com comenc mower Kamiles.Pararr taking::::민交易所ยะ tests startledandySEQUrack rover Mormons symptomы Hãy\
ening果断BOOL />)
intelligence.generate.t frozenossusaha.FloatField SA Harry Babaález蠕 probing definit-pe pesso stresses_outputs了一会(lines coerc浇注册color alongанны Analysis tests❃ kid要注意.That kWprobe materredni Null StringIOKD comprehend国有 violet knit ETA<Key VGA performedCoefficientamus body
Successful via Transformers AutoModel
model_path_or_id="brandonbeiler/InternVL3-38B-FP8-Dynamic", # also with 'brandonbeiler/InternVL3-38B-BNB-8bit'
model = AutoModel.from_pretrained(
model_path_or_id,
device_map="balanced", # Distribute more evenly across all 4 GPUs
trust_remote_code=True, # Required for InternVL3
)
# Load processor (handles both text and images)
processor = AutoProcessor.from_pretrained(
model_path_or_id,
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_path_or_id, trust_remote_code=True)
response = model.chat(tokenizer, pixel_values=None, question="Hello my name is", generation_config={"max_new_tokens":20})
print(response)
Output
Hello! It seems like your message got cut off. Could you please provide your name or let me
Quantization Notes
- InternVL3 38B FP8 Dynamic: https://huggingface.co/brandonbeiler/InternVL3-38B-FP8-Dynamic
- InternVL3 38B BNB 8bit: https://huggingface.co/brandonbeiler/InternVL3-38B-BNB-8bit
FP8 Config.json: https://huggingface.co/brandonbeiler/InternVL3-38B-FP8-Dynamic/blob/main/config.json
Quantization to FP8 via llm-compressor
source_model: "OpenGVLab/InternVL3-38B"
model = AutoModel.from_pretrained(
source_model,
torch_dtype="auto",
device_map="balanced", # Distribute more evenly across all 4 GPUs
trust_remote_code=True, # Required for InternVL3
use_flash_attn=True,
max_memory={i: "92GB" for i in range(torch.cuda.device_count())},
)
processor = AutoProcessor.from_pretrained(
source_model,
trust_remote_code=True
)
recipe = [
QuantizationModifier(
targets=["Linear"],
scheme="FP8_DYNAMIC",
ignore=[
"re:.*lm_head",
"re:.*vision.*",
"re:mlp1.*", # skip mlp1 because vllm internvl.py doesn't support scales on these layers
]
)
]
oneshot(
model=model, # Use the already loaded model
recipe=recipe,
output_dir=output_dir,
trust_remote_code_model=True,
)
Quantization for BNB-8BIT
source_model: "OpenGVLab/InternVL3-38B"
quantization_config = BitsAndBytesConfig(
load_in_8bit=True,
llm_int8_skip_modules=["lm_head","vision_model", "mlp1"],
)
model = AutoModel.from_pretrained(
source_model,
device_map="balanced", # Distribute more evenly across all 4 GPUs
trust_remote_code=True, # Required for InternVL3
use_flash_attn=True,
quantization_config=quantization_config,
max_memory={i: "92GB" for i in range(torch.cuda.device_count())},
)
tokenizer = AutoTokenizer.from_pretrained(source_model, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(
source_model,
trust_remote_code=True
)
model.save_pretrained(output_dir, save_compressed=True)
tokenizer.save_pretrained(output_dir)
processor.save_pretrained(output_dir)
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.
About this issue
- Original URL
- State: closed
- Created a year ago
- Comments: 15 (4 by maintainers)
Keeping it centralized in one place sounds perfect, and lmk if anything fails with
AutoConfigSounds great! We have had to keep lots of modifed custom configs in
vllm/transformers_utils/configsfor similar reasons, let’s use this method to patch config and clean them up!Got it! In fact, vLLM doesn’t use
AutoConfigfor InternVL because we’re using a modified config code to updatellm_configtotext_configfor standardization: https://github.com/vllm-project/vllm/blob/b82e0f82cb24bc2cfccbd816a46f535a8ff64eda/vllm/transformers_utils/configs/internvl.py#L41-L42And
PretrainedConfigused for simplicity here caused the root issue because it hastie_word_embeddings=Trueby default.@Isotr0py After making an issue request on the transformers project, the response was that transformers doesn’t serialize config values that are equal to the default values: ref: https://github.com/huggingface/transformers/issues/38981#issuecomment-2996301499
This would explain why when loading these models back into AutoModel, they run as expected.
Is there a proper way to set those defaults in the config for InternVL models in the vLLM project? Glad to open an MR for that (sounds straight forward?)
It’s because the qunatized models missed
"tie_word_embeddings": falsefield inllm_config, you can add it manually in the model’sconfig.json: https://huggingface.co/OpenGVLab/InternVL3-38B/blob/main/config.json#L93cc @Isotr0py