Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Bug]: VLLM_ATTENTION_BACKEND set to ROCM_FLASH only in GHA environment, overriding automatic backend selection; this breaks other kernel unit tests. #5208

Closed
afeldman-nm opened this issue Jun 3, 2024 · 1 comment · Fixed by #5210
Labels
bug Something isn't working

Comments

@afeldman-nm
Copy link
Contributor

afeldman-nm commented Jun 3, 2024

Your current environment

Note: I only observe this problem when the tests are automatically executed by GitHub actions (GHA); as I do not have access to the GHA compute, I cannot provide an environment variable dump.

However, I do not observe this issue on my development machine. The environment dump from my development machine which does not exhibit this issue is provided below:

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 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 12.3.0-1ubuntu1~22.04) 12.3.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.28.3
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.0-26-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.3.103
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA RTX A4000
GPU 1: NVIDIA RTX A4000
GPU 2: NVIDIA RTX A4000
GPU 3: NVIDIA RTX A4000

Nvidia driver version: 545.23.08
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.7
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:                      43 bits physical, 48 bits virtual
Byte Order:                         Little Endian
CPU(s):                             48
On-line CPU(s) list:                0-47
Vendor ID:                          AuthenticAMD
Model name:                         AMD Ryzen Threadripper 3960X 24-Core Processor
CPU family:                         23
Model:                              49
Thread(s) per core:                 2
Core(s) per socket:                 24
Socket(s):                          1
Stepping:                           0
Frequency boost:                    enabled
CPU max MHz:                        3800.0000
CPU min MHz:                        2200.0000
BogoMIPS:                           7600.17
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 nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 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 ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es
Virtualization:                     AMD-V
L1d cache:                          768 KiB (24 instances)
L1i cache:                          768 KiB (24 instances)
L2 cache:                           12 MiB (24 instances)
L3 cache:                           128 MiB (8 instances)
NUMA node(s):                       1
NUMA node0 CPU(s):                  0-47
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:             Mitigation; untrained return thunk; SMT enabled with STIBP protection
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; Retpolines, IBPB conditional, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] mypy==1.9.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] torch==2.3.0
[pip3] triton==2.3.0
[pip3] vllm-nccl-cu12==2.18.1.0.3.0
[conda] No relevant packagesROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.4.3
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      SYS     SYS     SYS     0-47    0               N/A
GPU1    SYS      X      SYS     SYS     0-47    0               N/A
GPU2    SYS     SYS      X      PHB     0-47    0               N/A
GPU3    SYS     SYS     PHB      X      0-47    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

Context: I am working on a PR (#4888) which is a step towards addressing #187 . This PR (1) augments the Attention wrapper with support for encoder attention and encoder/decoder cross-attention, and (2) adds unit tests for this functionality in tests/kernels/test_self_and_cross_attn.py. Only the xFormers backend is supported in this PR; thus I expect that the unit tests would fail for other backends. I expect that the tests in tests/kernels/test_self_and_cross_attn.py will be executed when GitHub actions (GHA) runs the "Kernels 1/2/3/4" tests.

Bug:

  • At the time, my unit tests were allowing vllm/attention/selector.py to automatically choose the attention backend, even though I only wanted to test xFormers.
  • On my personal development machine, all tests in tests/kernels/test_self_and_cross_attn.py pass. I can see from log outputs that the xFormers kernel is being automatically selected.
  • However, when GitHuh actions (GHA) runs automatic tests on my PR, all non-skipped tests in tests/kernels/test_self_and_cross_attn.py fail. I can see from log outputs that the backend selection is inconsistent; sometimes ROCm backend is selected, other times SDPA backend is selected.
  • Looking at https://sourcegraph.com/github.com/vllm-project/vllm/-/blob/.buildkite/test-pipeline.yaml?L83 , it would appear that the kernel tests are explicitly configured not to use AMD hardware; furthermore, examining selector.py

if is_hip():

shows that ROCm backend may only be selected automatically when is_hip() == True, i.e. on an AMD machine with HIP support.

  • Thus, while it would ideal if my attention unit tests had forced xFormers backend for testing purposes, nonetheless it is unexpected that selector.py should ever choose ROCm backend by default during the Kernels unit tests. This is the bug
  • I have confirmed that when ROCm backend is being selected, is_hip() == False, which rules out that an AMD machine is actually being used

Steps to reproduce the bug

  • To show that tests pass on a development machine with NVIDIA GPU (i.e. not AMD):

    • On your dev machine:
    git checkout f6e0310f1a95
    pytest tests/kernels/test_self_and_cross_attn.py
    

    All tests should pass. Note that you may need to git fetch this commit.

  • To show that tests fail on GHA:

    kernels/test_self_and_cross_attn.py::test_encoder_attention[128-16-1-xformers-64-1] 
    envs.VLLM_ATTENTION_BACKEND: ROCM_FLASH
    -
    | INFO 06-03 03:52:18 selector.py:56] Using ROCmFlashAttention backend.
    | FAILED
    

    envs.VLLM_ATTENTION_BACKEND: ROCM_FLASH is the environment variable printout, showing that the environment is contaminated with a VLLM_ATTENTION_BACKEND value that overrides selector.py to use the ROCm backend. Sure enough, Using ROCmFlashAttention backend indicates that the ROCm backend is selected, and then we see the test fails (as expected since my PR is only designed to update xFormers backend with encoder/decoder support.)

RCA

I believe I have identified the root cause, which is an issue with tests/kernels/test_attention_selector.py:

  • All tests in this file rely on an ad-hoc test fixture that temporarily sets VLLM_ATTENTION_BACKEND in order to force a choice of backend; each test has a cleanup phase that resets the environment variable to its pre-test value
  • However, the cleanup code in test_env()

if name_backup is not None:
os.environ["VLLM_ATTENTION_BACKEND"] = name_backup

and test_flash_attn()

if name_backup is not None:
os.environ["VLLM_ATTENTION_BACKEND"] = name_backup

don't handle the name_backup is None case, i.e. the case where VLLM_ATTENTION_BACKEND was unset prior to the test. This means that if prior to the test VLLM_ATTENTION_BACKEND was unset, then after the test VLLM_ATTENTION_BACKEND will still hold the value which it was set to during the test. Whereas the desired behavior would be for VLLM_ATTENTION_BACKEND to be unset at the end of the test.

So what appears to be happening is the following:

  1. Buildkite parallelizes the Kernels unit tests into 4 partitions, Kernels 1/2/3/4
  2. One or more of these partitions includes a subset of tests/kernels/test_attention_selector.py test-cases and a subset of tests/kernels/test_self_and_cross_attn.py test-cases, with the former being executed prior to the latter
  3. The VLLM_ATTENTION_BACKEND value leaked by the selector tests happens to be ROCM_FLASH (as opposed to being unset), forcing tests/kernels/test_self_and_cross_attn.py to use ROCm Flash backend (even though the machine does not support it.)

In principle, it is possible that running all of the kernel tests (pytest tests/kernels) on my dev machine would reproduce this issue. However, it appears that VS code on my dev machine chooses to execute the tests in reverse order compared to GHA, with the attention tests preceding the selector tests; this seems likely to be what prevents me from reproducing the failure.

@afeldman-nm
Copy link
Contributor Author

Update: Just created PR #5210 which will fix this issue once merged.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug Something isn't working
Projects
None yet
1 participant