exllama segfault with act-order

[fxmarty]

Hi,

I get a segmentation fault when using act-order.

Reproduction:

from model import Ex4bitLinear, ExLlamaConfig
import torch
import torch.nn as nn

config = ExLlamaConfig("../WizardLM-7B-uncensored-GPTQ/config.json")
device = torch.device("cuda")

hidden_size = 8192
num_attention_heads = 64
group_size = 128

tensors = {
    "q_proj.qweight": torch.zeros(hidden_size // 8, hidden_size, dtype=torch.int32).to(device),
    "q_proj.qzeros": torch.zeros(hidden_size // group_size, hidden_size // 8, dtype=torch.int32).to(device),
    "q_proj.scales": torch.zeros(hidden_size // group_size, hidden_size, dtype=torch.float16).to(device),
}

tensors["q_proj.g_idx"] = torch.tensor(
    [i // group_size for i in range(hidden_size)], dtype=torch.int32
).to(device)

inp = torch.rand(13, 6144).to("cuda", torch.float16)

head_dim = hidden_size // num_attention_heads

q4linear = Ex4bitLinear(config, hidden_size, num_attention_heads * head_dim, False, tensors, "q_proj")

inp = torch.rand(1, 1, hidden_size, dtype=torch.float16).to(device)

num_runs = 100
# warmup
with torch.no_grad():
    res = q4linear.forward(inp, None)

    start_event = torch.cuda.Event(enable_timing=True)
    end_event = torch.cuda.Event(enable_timing=True)
    torch.cuda.reset_peak_memory_stats(device)
    torch.cuda.empty_cache()
    torch.cuda.synchronize()
    start_event.record()

    for _ in range(num_runs):
        res = q4linear.forward(inp, None)

    end_event.record()

    torch.cuda.synchronize()
    max_memory = torch.cuda.max_memory_allocated(device)

    tps_q4 = (start_event.elapsed_time(end_event) * 1.0e-3) / num_runs
    mem_mb = max_memory * 1e-6

    print("Time (q4):", tps_q4)
    print("Mem MB (q4):", mem_mb) 

Passing a None g_idx (or all zero) goes fine. I only get segfault with act_order.

@turboderp Are there any specific conditions on g_idx not to segfault? The initialization torch.tensor([i // groupsize for i in range(infeatures)], dtype=torch.int32) seems reasonable, but maybe there's something more?

Thank you for this excellent implementation!

Specs:

PyTorch version: 2.1.0.dev20230615+cu118
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.5 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.26.4
Libc version: glibc-2.31

Python version: 3.9.16 (main, May 15 2023, 23:46:34)  [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-1023-aws-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA A100-SXM4-80GB
GPU 1: NVIDIA A100-SXM4-80GB
GPU 2: NVIDIA A100-SXM4-80GB
GPU 3: NVIDIA A100-SXM4-80GB
GPU 4: NVIDIA A100-SXM4-80GB
GPU 5: NVIDIA A100-SXM4-80GB
GPU 6: NVIDIA A100-SXM4-80GB
GPU 7: NVIDIA A100-SXM4-80GB

Nvidia driver version: 510.73.08
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
Byte Order:                      Little Endian
Address sizes:                   46 bits physical, 48 bits virtual
CPU(s):                          96
On-line CPU(s) list:             0-95
Thread(s) per core:              2
Core(s) per socket:              24
Socket(s):                       2
NUMA node(s):                    2
Vendor ID:                       GenuineIntel
CPU family:                      6
Model:                           85
Model name:                      Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
Stepping:                        7
CPU MHz:                         3000.000
BogoMIPS:                        6000.00
Hypervisor vendor:               KVM
Virtualization type:             full
L1d cache:                       1.5 MiB
L1i cache:                       1.5 MiB
L2 cache:                        48 MiB
L3 cache:                        71.5 MiB
NUMA node0 CPU(s):               0-23,48-71
NUMA node1 CPU(s):               24-47,72-95
Vulnerability Itlb multihit:     KVM: Mitigation: VMX unsupported
Vulnerability L1tf:              Mitigation; PTE Inversion
Vulnerability Mds:               Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Meltdown:          Mitigation; PTI
Vulnerability Mmio stale data:   Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Retbleed:          Vulnerable
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1:        Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Retpolines, STIBP disabled, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected
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 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single pti fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves ida arat pku ospke

Versions of relevant libraries:
[pip3] mypy-protobuf==3.4.0
[pip3] numpy==1.24.3
[pip3] pytorch-triton==2.1.0+440fd1bf20
[pip3] torch==2.1.0.dev20230615+cu118
[pip3] torchaudio==2.1.0.dev20230615+cu118
[pip3] torchvision==0.16.0.dev20230615+cu118
[conda] numpy                     1.24.3                   pypi_0    pypi
[conda] pytorch-triton            2.1.0+440fd1bf20          pypi_0    pypi
[conda] torch                     2.1.0.dev20230615+cu118          pypi_0    pypi
[conda] torchaudio                2.1.0.dev20230615+cu118          pypi_0    pypi
[conda] torchvision               0.16.0.dev20230615+cu118          pypi_0    pypi

[fxmarty]

Simply forgot to call set_tuning_params & prepare_buffers beforehand!