[Bug]: EAGLE / DeepSeek MTP Handles First Input Token Incorrectly - 25% Acceptance Rate Drop

[luyuzhe111]

Your current environment

The output of `python collect_env.py`

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: Could not collect
CMake version: version 3.31.4
Libc version: glibc-2.35

Python version: 3.12.9 | packaged by Anaconda, Inc. | (main, Feb  6 2025, 18:56:27) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-6.8.0-1021-aws-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA H200
GPU 1: NVIDIA H200
GPU 2: NVIDIA H200
GPU 3: NVIDIA H200
GPU 4: NVIDIA H200
GPU 5: NVIDIA H200
GPU 6: NVIDIA H200
GPU 7: NVIDIA H200

Nvidia driver version: 550.144.03
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:                        46 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               192
On-line CPU(s) list:                  0-191
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8488C
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   48
Socket(s):                            2
Stepping:                             8
BogoMIPS:                             4800.00
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 arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd ida arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid cldemote movdiri movdir64b md_clear serialize amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            4.5 MiB (96 instances)
L1i cache:                            3 MiB (96 instances)
L2 cache:                             192 MiB (96 instances)
L3 cache:                             210 MiB (2 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-47,96-143
NUMA node1 CPU(s):                    48-95,144-191
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:   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 BHI_DIS_S
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[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-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.2.1
[pip3] sentence-transformers==3.2.1
[pip3] torch==2.5.1
[pip3] torchaudio==2.5.1
[pip3] torchvision==0.20.1
[pip3] transformers==4.48.2
[pip3] transformers-stream-generator==0.0.5
[pip3] triton==3.1.0
[pip3] tritonclient==2.51.0
[pip3] vector-quantize-pytorch==1.21.2
[conda] numpy                     1.26.4                   pypi_0    pypi
[conda] nvidia-cublas-cu12        12.4.5.8                 pypi_0    pypi
[conda] nvidia-cuda-cupti-cu12    12.4.127                 pypi_0    pypi
[conda] nvidia-cuda-nvrtc-cu12    12.4.127                 pypi_0    pypi
[conda] nvidia-cuda-runtime-cu12  12.4.127                 pypi_0    pypi
[conda] nvidia-cudnn-cu12         9.1.0.70                 pypi_0    pypi
[conda] nvidia-cufft-cu12         11.2.1.3                 pypi_0    pypi
[conda] nvidia-curand-cu12        10.3.5.147               pypi_0    pypi
[conda] nvidia-cusolver-cu12      11.6.1.9                 pypi_0    pypi
[conda] nvidia-cusparse-cu12      12.3.1.170               pypi_0    pypi
[conda] nvidia-nccl-cu12          2.21.5                   pypi_0    pypi
[conda] nvidia-nvjitlink-cu12     12.4.127                 pypi_0    pypi
[conda] nvidia-nvtx-cu12          12.4.127                 pypi_0    pypi
[conda] pyzmq                     26.2.1                   pypi_0    pypi
[conda] sentence-transformers     3.2.1                    pypi_0    pypi
[conda] torch                     2.5.1                    pypi_0    pypi
[conda] torchaudio                2.5.1                    pypi_0    pypi
[conda] torchvision               0.20.1                   pypi_0    pypi
[conda] transformers              4.48.2                   pypi_0    pypi
[conda] transformers-stream-generator 0.0.5                    pypi_0    pypi
[conda] triton                    3.1.0                    pypi_0    pypi
[conda] tritonclient              2.51.0                   pypi_0    pypi
[conda] vector-quantize-pytorch   1.21.2                   pypi_0    pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.1.dev5011+ge1f0835.d20250307 (git sha: e1f0835.d20250307
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV18    NV18    NV18    NV18    NV18    NV18    NV18    0-47,96-143     0               N/A
GPU1    NV18     X      NV18    NV18    NV18    NV18    NV18    NV18    0-47,96-143     0               N/A
GPU2    NV18    NV18     X      NV18    NV18    NV18    NV18    NV18    0-47,96-143     0               N/A
GPU3    NV18    NV18    NV18     X      NV18    NV18    NV18    NV18    0-47,96-143     0               N/A
GPU4    NV18    NV18    NV18    NV18     X      NV18    NV18    NV18    48-95,144-191   1               N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X      NV18    NV18    48-95,144-191   1               N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X      NV18    48-95,144-191   1               N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X      48-95,144-191   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

NCCL_CUMEM_ENABLE=0
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

According to its paper, EAGLE concatenates the hidden states of the current token with the embedding of the next token. The embedding of the first token in a sequence is not used, as shown in the diagram below (embedding of 'how'). Since MPT essentially has the same design in this regard, I will simply describe the bug for EAGLE.

This design poses challenges in vLLM, as we need to obtain separate attention metadata for EAGLE. Currently, the EAGLE implementation in vLLM employs a hack to bypass this challenge via 1) obtaining a dummy first token by concatenating the last prompt token's hidden states (as a result of right shifting here) and the first prompt token's embedding, and 2) setting their projection to zero before passing to the main model layer. While this seems to be a smart trick and does not cause any errors, it has deep problems. This hack is currently used in both EAGLE [line] and DeepSeek MTP [line].

While zeroing out the token seems to be equivalent to removing it, it is not for attention based models. Given a zero vector, its Q, K, V will be zero vectors as well. However, due to the softmax operator in attention, softmax(QK^T) will still be 1. As a result, the normalizer of the attention score will always be off by 1 for all tokens, which harms the drafting accuracy. On a less critical note, the position ids for EAGLE are also all off by 1.

In fact, as a temporary solution, we might be better off removing this line. For DeepSeek R1, when MLA is enabled, this line will cause a 25% drop in acceptance length compared to simply removing it. For one speculated token, removing this zeroing-out step gives 1.9 tokens generated per forward pass on average, while with this zeroing out operation gives only 1.42. @benchislett also observed this mysterious drop early on #12755 (comment). The zero masking is actually culprit. As MLA alters the order of the computation, the issue was somehow enlarged.

cc @LiuXiaoxuanPKU

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.

[luyuzhe111]

Would like to quickly update that this 25% acceptance rate drop seems to only occur in the previous Triton MLA kernel. Now with FlashMLA the problem seems to go away. Nonetheless, since the zeroing out operation does not do what it intends and does not improve acceptance rate, I suggest removing this line. (Thanks to Xinlong Zhang's email to me).

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!