[Bug]: Qwen2.5-Math-RM-72B Online Inference Fails

[passaglia]

Your current environment

The output of `python collect_env.py`

Collecting environment information...
PyTorch version: 2.5.1+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: Could not collect
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov  6 2024, 20:22:13) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.0-1025-gcp-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.4.131
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA H100 80GB HBM3
GPU 1: NVIDIA H100 80GB HBM3
GPU 2: NVIDIA H100 80GB HBM3
GPU 3: NVIDIA H100 80GB HBM3
GPU 4: NVIDIA H100 80GB HBM3
GPU 5: NVIDIA H100 80GB HBM3
GPU 6: NVIDIA H100 80GB HBM3
GPU 7: NVIDIA H100 80GB HBM3

Nvidia driver version: 550.90.07
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
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):                             208
On-line CPU(s) list:                0-207
Vendor ID:                          GenuineIntel
Model name:                         Intel(R) Xeon(R) Platinum 8481C CPU @ 2.70GHz
CPU family:                         6
Model:                              143
Thread(s) per core:                 2
Core(s) per socket:                 52
Socket(s):                          2
Stepping:                           8
BogoMIPS:                           5399.99
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rtm avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 arat avx512vbmi umip avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk amx_bf16 avx512_fp16 amx_tile amx_int8 arch_capabilities
Hypervisor vendor:                  KVM
Virtualization type:                full
L1d cache:                          4.9 MiB (104 instances)
L1i cache:                          3.3 MiB (104 instances)
L2 cache:                           208 MiB (104 instances)
L3 cache:                           210 MiB (2 instances)
NUMA node(s):                       2
NUMA node0 CPU(s):                  0-51,104-155
NUMA node1 CPU(s):                  52-103,156-207
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 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; RSB filling; PBRSB-eIBRS SW sequence; BHI Syscall hardening, KVM SW loop
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-ml-py==12.560.30
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.2.0
[pip3] torch==2.5.1
[pip3] torchvision==0.20.1
[pip3] transformers==4.47.1
[pip3] triton==3.1.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.6.5
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
�[4mGPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	CPU Affinity	NUMA Affinity	GPU NUMA ID�[0m
GPU0	 X 	NV18	NV18	NV18	NV18	NV18	NV18	NV18	0-51,104-155	0		N/A
GPU1	NV18	 X 	NV18	NV18	NV18	NV18	NV18	NV18	0-51,104-155	0		N/A
GPU2	NV18	NV18	 X 	NV18	NV18	NV18	NV18	NV18	0-51,104-155	0		N/A
GPU3	NV18	NV18	NV18	 X 	NV18	NV18	NV18	NV18	0-51,104-155	0		N/A
GPU4	NV18	NV18	NV18	NV18	 X 	NV18	NV18	NV18	52-103,156-207	1		N/A
GPU5	NV18	NV18	NV18	NV18	NV18	 X 	NV18	NV18	52-103,156-207	1		N/A
GPU6	NV18	NV18	NV18	NV18	NV18	NV18	 X 	NV18	52-103,156-207	1		N/A
GPU7	NV18	NV18	NV18	NV18	NV18	NV18	NV18	 X 	52-103,156-207	1		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

CUDA_ROOT=/usr/local/cuda
LD_LIBRARY_PATH=/home/sam_passaglia_elyza_ai/easy-evals/vllm_server/.venv/vllm-server-3fpuUD1n-py3.10/lib/python3.10/site-packages/cv2/../../lib64:/usr/local/cuda/lib64:
CUDA_INC_DIR=/usr/local/cuda/bin:
CUDA_MODULE_LOADING=LAZY

Model Input Dumps

No response

🐛 Describe the bug

Using the example script from the original implementation PR now leads to the following error:

openai.BadRequestError: Error code: 400 - {'object': 'error', 'message': 'pooled_data should b
e a 1-D embedding vector', 'type': 'BadRequestError', 'param': None, 'code': 400}             

The error can be traced back to the pooling API refactoring in #11129

Reward Models usually use AllPool, which usually returns one output per token per prompt as a list[tensor](See also: #10820).

But the EmbeddingOutput dataclass expects pooled_data.ndim = 1

vllm/vllm/outputs.py

Lines 407 to 408 in a491d6f

	if pooled_data.ndim != 1:
	raise ValueError("pooled_data should be a 1-D embedding vector")

cc @DarkLight1337

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]

This is an intentional change for offline inference because we should now use LLM.encode() for reward models while LLM.embed() should be used for embedding models (not reward models). Sorry I didn't fully consider online inference though since there's no alternative to that yet. How should we proceed? I'm thinking of setting up a new encode endpoint for this.

[passaglia]

I am not very familiar with the internal structure of VLLM, but since offline inference for reward modeling and embedding follow a different entrypoint (LLM.encode() vs LLM.embed()) and yield a different output type (PoolingRequestOutput vs EmbeddingRequestOutput), then it sounds like the API server should also send RM and embedding requests down a different code path. Does it make sense to have a serving_reward module alongside the existing serving_embedding and serving_score modules?

[passaglia]

I'm thinking of setting up a new encode endpoint for this.

Since currently task needs to be specified at server launch time, it might be easier at first to simply overload the embedding endpoint but change the code path based on task

[DarkLight1337]

I have added /pooling API in #11457 which should solve this.

[passaglia]

@DarkLight1337 Thank you!