[Bug]: Illegal memory access for MoE kernel with large workloads

[comaniac]

Your current environment

Collecting environment information...
PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: version 3.29.3
Libc version: glibc-2.31

Python version: 3.10.14 (main, May  6 2024, 19:42:50) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-6.5.0-1022-gcp-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 12.1.105
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA L4
Nvidia driver version: 535.183.01
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.0
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
Byte Order:                         Little Endian
Address sizes:                      46 bits physical, 48 bits virtual
CPU(s):                             32
On-line CPU(s) list:                0-31
Thread(s) per core:                 2
Core(s) per socket:                 16
Socket(s):                          1
NUMA node(s):                       1
Vendor ID:                          GenuineIntel
CPU family:                         6
Model:                              85
Model name:                         Intel(R) Xeon(R) CPU @ 2.20GHz
Stepping:                           7
CPU MHz:                            2200.174
BogoMIPS:                           4400.34
Hypervisor vendor:                  KVM
Virtualization type:                full
L1d cache:                          512 KiB
L1i cache:                          512 KiB
L2 cache:                           16 MiB
L3 cache:                           38.5 MiB
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:      Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Retbleed:             Mitigation; Enhanced IBRS
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:      Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
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 hle avx2 smep bmi2 erms invpcid rtm mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves arat avx512_vnni md_clear arch_capabilities

Versions of relevant libraries:
[pip3] mypy==1.9.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.24.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] torch==2.3.0
[pip3] torchvision==0.18.0
[pip3] transformers==4.42.1
[pip3] triton==2.3.0
[conda] numpy                     1.24.4                   pypi_0    pypi
[conda] nvidia-nccl-cu12          2.20.5                   pypi_0    pypi
[conda] torch                     2.3.0                    pypi_0    pypi
[conda] torchvision               0.18.0                   pypi_0    pypi
[conda] transformers              4.42.1                   pypi_0    pypi
[conda] triton                    2.3.0                    pypi_0    pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.0.post1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     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

🐛 Describe the bug

Illegal memory access for MoE triton kernel when the workload (e.g., batch size) is too large. To reproduce:

python3 ./benchmarks/kernels/benchmark_moe.py --tp-size 2 --batch-size 74899

Output

  File "/home/ray/default/vllm/./benchmarks/kernels/benchmark_moe.py", line 70, in run
    fused_moe(
  File "/tmp/ray/session_2024-06-27_10-06-48_118980_5595/runtime_resources/working_dir_files/_ray_pkg_ef0e5109bc8b4140628503119c10e0b2c9ea3f17/vllm/model_executor/layers/fused_moe/fused_moe.py", line 519, in fused_moe
    return fused_experts(hidden_states,
  File "/tmp/ray/session_2024-06-27_10-06-48_118980_5595/runtime_resources/working_dir_files/_ray_pkg_ef0e5109bc8b4140628503119c10e0b2c9ea3f17/vllm/model_executor/layers/fused_moe/fused_moe.py", line 449, in fused_experts
    invoke_fused_moe_kernel(intermediate_cache2,
  File "/tmp/ray/session_2024-06-27_10-06-48_118980_5595/runtime_resources/working_dir_files/_ray_pkg_ef0e5109bc8b4140628503119c10e0b2c9ea3f17/vllm/model_executor/layers/fused_moe/fused_moe.py", line 245, in invoke_fused_moe_kernel
    fused_moe_kernel[grid](
  File "/home/ray/anaconda3/lib/python3.10/site-packages/triton/runtime/jit.py", line 167, in <lambda>
    return lambda *args, **kwargs: self.run(grid=grid, warmup=False, *args, **kwargs)
  File "/home/ray/anaconda3/lib/python3.10/site-packages/triton/runtime/jit.py", line 425, in run
    kernel.run(grid_0, grid_1, grid_2, kernel.num_warps, kernel.num_ctas,  # number of warps/ctas per instance
  File "/home/ray/anaconda3/lib/python3.10/site-packages/triton/compiler/compiler.py", line 255, in __getattribute__
    self._init_handles()
  File "/home/ray/anaconda3/lib/python3.10/site-packages/triton/compiler/compiler.py", line 250, in _init_handles
    self.module, self.function, self.n_regs, self.n_spills = driver.utils.load_binary(
RuntimeError: Triton Error [CUDA]: an illegal memory access was encountered

We have seen this problem on L4 and A100 GPUs. I also tried to tune this particular workload using different block sizes, but none of the configs could bypass the error. Since we usually don't use such a large batch size (number of tokens), this bug should not be critical at least for now.

Also cc @pcmoritz @WoosukKwon @Yard1

[Minami-su]

+1,me too.
tp_size 2
Qwen2_72b
2 X A800 80G

[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!