AOTI Export ignores user --device flag - expected behavior?

[vmpuri]

🐛 Describe the bug

Hi all,

I ran into some confusion when trying to export llama3 on my system. I have a small graphics card (8GB VRAM on an AMD GPU) but a decent amount of RAM (24GB). Obviously, the model won't fit on my GPU un-quantized but it should fit into my RAM + swap.

I tried running:

python3 torchchat.py export llama3 --output-dso-path exportedModels/llama3.so --quantize torchchat/quant_config/desktop.json  --device cpu

However, I ran into multiple HIP OOM errors (basically equivalent to CUDA). Why would we try to allocate CUDA memory if the target device is CPU?

On further inspection, during export, the device is replaced with whatever is present in the quantize config: In cli.py https://github.com/pytorch/torchchat/blob/b21715835ab9f61e23dbcf32795b0c0a2d654908/torchchat/cli/cli.py#L491C10-L494C1

args.device = get_device_str(
    args.quantize.get("executor", {}).get("accelerator", args.device)
)

In this case, the device in desktop.json is "fast". The get_device_str function replaces this with "cuda" simply based on torch.cuda.is_available without consulting the flag I passed in.

Other cases

Doing a quick grep of the repo, I only found one other case in generate.py where torch.cuda.is_available() is consulted for monitoring memory usage. We should be careful switching based simply on torch.cuda.is_available() and make sure to pin to the user's request if we're using ambiguous devices like "fast".

Another small issue - since I use AMD GPU, the default install/install_requirements.sh will download the CPU only version instead of the ROCm version of PyTorch. To use my GPU, I have to re-run the torch installation manually. Luckily, it's quite easy to find this command at https://pytorch.org/get-started/locally/ . Should be straightforward to check of ROCm is available on the system during this script - we can just run rocminfo & check if the command is available.

Versions

wget https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py
# For security purposes, please check the contents of collect_env.py before running it.
python collect_env.py
--2024-10-06 12:03:44--  https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 2606:50c0:8001::154, 2606:50c0:8002::154, 2606:50c0:8003::154, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|2606:50c0:8001::154|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 23357 (23K) [text/plain]
Saving to: ‘collect_env.py’

collect_env.py                              100%[===========================================================================================>]  22.81K  --.-KB/s    in 0.02s   

2024-10-06 12:03:44 (1.10 MB/s) - ‘collect_env.py’ saved [23357/23357]

Collecting environment information...
PyTorch version: 2.4.1+rocm6.1
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.1.40091-a8dbc0c19

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 9.5.0-1ubuntu1~22.04) 9.5.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.30.4
Libc version: glibc-2.35

Python version: 3.10.12 (main, Sep 11 2024, 15:47:36) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.1.4-060104-generic-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: AMD Radeon RX 6700S (gfx1030)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: 6.1.40091
MIOpen runtime version: 3.1.0
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Address sizes:                   48 bits physical, 48 bits virtual
Byte Order:                      Little Endian
CPU(s):                          16
On-line CPU(s) list:             0-15
Vendor ID:                       AuthenticAMD
Model name:                      AMD Ryzen 9 6900HS with Radeon Graphics
CPU family:                      25
Model:                           68
Thread(s) per core:              2
Core(s) per socket:              8
Socket(s):                       1
Stepping:                        1
Frequency boost:                 enabled
CPU max MHz:                     4933.8862
CPU min MHz:                     1600.0000
BogoMIPS:                        6587.56
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 x2apic 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 ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm
Virtualization:                  AMD-V
L1d cache:                       256 KiB (8 instances)
L1i cache:                       256 KiB (8 instances)
L2 cache:                        4 MiB (8 instances)
L3 cache:                        16 MiB (1 instance)
NUMA node(s):                    1
NUMA node0 CPU(s):               0-15
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:          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 always-on, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pytorch-triton-rocm==3.0.0
[pip3] torch==2.4.1+rocm6.1
[pip3] torchao==0.5.0
[pip3] torchaudio==2.4.1+rocm6.1
[pip3] torchtune==0.3.0.dev20240928+cpu
[pip3] torchvision==0.19.1+rocm6.1
[conda] blas                      1.0                         mkl  
[conda] mkl                       2023.1.0         h213fc3f_46344  
[conda] mkl-service               2.4.0           py311h5eee18b_1  
[conda] mkl_fft                   1.3.8           py311h5eee18b_0  
[conda] mkl_random                1.2.4           py311hdb19cb5_0  
[conda] numpy                     1.25.2                   pypi_0    pypi
[conda] numpy-base                1.26.4          py311hf175353_0  
[conda] pytorch                   2.3.0           cpu_py311ha0631a7_0  
[conda] torch                     2.0.1                    pypi_0    pypi
[conda] triton                    2.0.0                    pypi_0    pypi