FP8 Rowwise Scaling does not work on L4 GPU

[guberti]

🐛 Describe the bug

When running the latest version of Pytorch (3689471), rowwise scaling factors for the _scaled_mm function do not work on L4 GPUs (but they work fine on the H100). The bug can be reproduced with this code:

import torch

a_data = torch.ones(32, 128, device="cuda").to(torch.float8_e4m3fn).contiguous()
b_data = torch.ones(64, 128, device="cuda").to(torch.float8_e4m3fn).t()
a_scale = torch.ones(a_data.shape[0], 1, device="cuda", dtype=torch.float32)
b_scale = torch.ones(1, b_data.shape[1], device="cuda", dtype=torch.float32)

torch._scaled_mm(
            a_data,
            b_data,
            scale_a=a_scale,
            scale_b=b_scale,
            out_dtype=torch.bfloat16,
        )

Running on an H100 GPU gives the expected:

tensor([[128., 128., 128.,  ..., 128., 128., 128.],
        [128., 128., 128.,  ..., 128., 128., 128.],
        [128., 128., 128.,  ..., 128., 128., 128.],
        ...,
        [128., 128., 128.,  ..., 128., 128., 128.],
        [128., 128., 128.,  ..., 128., 128., 128.],
        [128., 128., 128.,  ..., 128., 128., 128.]], device='cuda:0',
       dtype=torch.bfloat16)

But running on an L4 GPU gives:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
[<ipython-input-2-ae572942333a>](https://localhost:8080/#) in <cell line: 8>()
      6 b_scale = torch.ones(1, b_data.shape[1], device="cuda", dtype=torch.float32)
      7 
----> 8 torch._scaled_mm(
      9             a_data,
     10             b_data,

RuntimeError: cutlass cannot initialize

Note that tensor-wise scaling factors work fine on the hardware - only row-wise scaling factors have this issue.

Versions

PyTorch version: 2.5.0.dev20240709+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 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.27.9
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-6.1.85+-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA L4
Nvidia driver version: 535.104.05
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.6
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):                               12
On-line CPU(s) list:                  0-11
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) CPU @ 2.20GHz
CPU family:                           6
Model:                                85
Thread(s) per core:                   2
Core(s) per socket:                   6
Socket(s):                            1
Stepping:                             7
BogoMIPS:                             4400.45
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
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            192 KiB (6 instances)
L1i cache:                            192 KiB (6 instances)
L2 cache:                             6 MiB (6 instances)
L3 cache:                             38.5 MiB (1 instance)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-11
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
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Vulnerable
Vulnerability Spec rstack overflow:   Not affected
Vulnerability Spec store bypass:      Vulnerable
Vulnerability Spectre v1:             Vulnerable: __user pointer sanitization and usercopy barriers only; no swapgs barriers
Vulnerability Spectre v2:             Vulnerable; IBPB: disabled; STIBP: disabled; PBRSB-eIBRS: Vulnerable; BHI: Vulnerable (Syscall hardening enabled)
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Vulnerable

Versions of relevant libraries:
[pip3] numpy==1.25.2
[pip3] pytorch-triton==3.0.0+dedb7bdf33
[pip3] torch==2.5.0.dev20240709+cu121
[pip3] torchaudio==2.4.0.dev20240709+cu121
[pip3] torchsummary==1.5.1
[pip3] torchtext==0.18.0
[pip3] torchvision==0.20.0.dev20240709+cu121
[pip3] triton==2.3.0
[conda] Could not collect

cc @yanbing-j @vkuzo @albanD @kadeng @penguinwu

[drisspg]

This is currently expected. The rowwise kernel relies on TMA which AFAIK is only supported on H100 gpus. I can double check though

[guberti]

Ah, I thought the L4 had TMA as well

[drisspg]

I think you right, however specifically the kernel requires sm90a and sm gencode AFAIK is not supported

[vgoklani]

Fixed in the newest nightly:

#145632 (comment)