PygmalionAI
[Bug]: Generation sometimes slows to a crawl for all requests when there is a DRY sampler request
Your current environment
PyTorch version: 2.4.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (conda-forge gcc 11.3.0-19) 11.3.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.22.1
Libc version: glibc-2.35
Python version: 3.11.9 | packaged by conda-forge | (main, Apr 19 2024, 18:36:13) [GCC 12.3.0] (64-bit runtime)
Python platform: Linux-6.8.0-48-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.1.105
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA GeForce RTX 3090
GPU 1: NVIDIA GeForce RTX 3090
GPU 2: NVIDIA GeForce RTX 3090
GPU 3: NVIDIA GeForce RTX 3090
Nvidia driver version: 550.120
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.4.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.4.0
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
Address sizes: 46 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 16
On-line CPU(s) list: 0-15
Vendor ID: GenuineIntel
Model name: Intel(R) Core(TM) i7-6900K CPU @ 3.20GHz
CPU family: 6
Model: 79
Thread(s) per core: 2
Core(s) per socket: 8
Socket(s): 1
Stepping: 1
CPU max MHz: 4100,0000
CPU min MHz: 1200,0000
BogoMIPS: 6400.01
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 pti intel_ppin ssbd ibrs ibpb stibp tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm rdt_a rdseed adx smap intel_pt xsaveopt cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts vnmi md_clear flush_l1d
Virtualization: VT-x
L1d cache: 256 KiB (8 instances)
L1i cache: 256 KiB (8 instances)
L2 cache: 2 MiB (8 instances)
L3 cache: 20 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0-15
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled
Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable
Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable
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; Retpolines; IBPB conditional; IBRS_FW; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable
Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] pyzmq==26.2.0
[pip3] torch==2.4.0
[pip3] torchvision==0.19.0
[pip3] transformers==4.45.2
[pip3] triton==3.0.0
[conda] blas 2.16 mkl conda-forge
[conda] libblas 3.8.0 16_mkl conda-forge
[conda] libcblas 3.8.0 16_mkl conda-forge
[conda] liblapack 3.8.0 16_mkl conda-forge
[conda] liblapacke 3.8.0 16_mkl conda-forge
[conda] mkl 2020.2 256
[conda] nccl 2.23.4.1 h52f6c39_2 conda-forge
[conda] numpy 1.26.4 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.20.5 pypi_0 pypi
[conda] pytorch-cuda 12.4 hc786d27_7 pytorch
[conda] pytorch-mutex 1.0 cuda pytorch
[conda] pyzmq 26.2.0 pypi_0 pypi
[conda] torch 2.4.0 pypi_0 pypi
[conda] torchvision 0.19.0 pypi_0 pypi
[conda] transformers 4.45.2 pypi_0 pypi
[conda] triton 3.0.0 pypi_0 pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
Aphrodite Version: 0.6.4
Aphrodite 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 PHB PHB PHB 0-15 0 N/A
GPU1 PHB X PHB PHB 0-15 0 N/A
GPU2 PHB PHB X PHB 0-15 0 N/A
GPU3 PHB PHB PHB X 0-15 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
Not sure how to show this but if there is a DRY sampler being processed by Aphrodite, occasionally it would slow generations down to a crawl for all the requests currently being processed and not just for the DRY sampler request.
Otherwise DRY sampler seems to be working great now!
I think this has to be expected, since DRY relies on purely CPU based operation. Aphrodite needs to process one request at a time for the CPU part of the code, right?
I presented two options to make DRY faster in another comment to the DRY PR, following the approach taken in ooba: allowing for a reduced dry_range and switching from tensors to lists in the CPU part of the code.
@gitzaidi I've added range in #855.
I did play around with using python lists, but I remember running into issues due to the batched nature of our samplers - we need all parameters to Tensors. If you have a solution that works when batched, would be happy to review a PR!
@AlpinDale Thanks for your reactivity, very impressive ! I will look into it then.
Also, do we agree that, as of now, the implementation considers the same dry_sequence_breaker_ids for each entry in the batch?
@AlpinDale Thanks for your reactivity, very impressive ! I will look into it then.
Also, do we agree that, as of now, the implementation considers the same
dry_sequence_breaker_idsfor each entry in the batch?
I think you're right - that's a huge oversight. I'll fix ASAP. Also, I decided to try my hand at porting over the z-algorithm implementation at #856. Can you take a look?
Fixed the sequence breaker ID issue at PygmalionAI/aphrodite-engine@c6e0ae0515a7b6364a152fbe132747d94713f6e4
Right now, I do not see issues compared to the ooba implementation, looks faithful to the ooba implementation (just noticed 2 typos, see comments)
DRY should be faster now but still very slow. I'm attempting to write kernels to bypass this issue. Progress will be logged here: https://github.com/AlpinDale/dry_sampling_kernel
#868 partially solved this issue. DRY is a lot faster now, but not as fast as other samplers. I think we can close this issue once a new release is made.