Autocast does not work with xformers

🐛 Describe the bug

I encounter following errors with xformers=0.0.23+cu118:

  File "/home/sist/luoxin/.conda/envs/py3.10+pytorch2.0+cu118+dali+cu120/lib/python3.10/site-packages/xformers/ops/fmha/__init__.py", line 223, in memory_efficient_attention
    return _memory_efficient_attention(
  File "/home/sist/luoxin/.conda/envs/py3.10+pytorch2.0+cu118+dali+cu120/lib/python3.10/site-packages/xformers/ops/fmha/__init__.py", line 321, in _memory_efficient_attention
    return _memory_efficient_attention_forward(
  File "/home/sist/luoxin/.conda/envs/py3.10+pytorch2.0+cu118+dali+cu120/lib/python3.10/site-packages/xformers/ops/fmha/__init__.py", line 334, in _memory_efficient_attention_forward
    inp.validate_inputs()
  File "/home/sist/luoxin/.conda/envs/py3.10+pytorch2.0+cu118+dali+cu120/lib/python3.10/site-packages/xformers/ops/fmha/common.py", line 121, in validate_inputs
    raise ValueError(
ValueError: Query/Key/Value should either all have the same dtype, or (in the quantized case) Key/Value should have dtype torch.int32
  query.dtype: torch.float32
  key.dtype  : torch.float32
  value.dtype: torch.float16

My code is looks like below:

import torch

with torch.autocast("cuda", enabled=True):
    print(f"{hidden_states.dtype=} {attn.to_q.weight.dtype=}", flush=True)
    # hidden_states.dtype=torch.float32, attn.to_q.weight.dtype=torch.float16

    query = to_q(hidden_states)
    print(f"{query.dtype=}", flush=True)
    # query.dtype=torch.float16

    key = to_k(hidden_states)
    value = to_v(hidden_states)

    print(f"{query.dtype=} {key.dtype=} {value.dtype=}", flush=True)
    # query.dtype=torch.float16 key.dtype=torch.float16 value.dtype=torch.float16

    query = attn.q_norm(query)
    key = attn.k_norm(key)
    print(f"{query.dtype=} {key.dtype=} {value.dtype=}", flush=True)
    # query.dtype=torch.float32 key.dtype=torch.float32 value.dtype=torch.float16

    hidden_states = xformers.ops.memory_efficient_attention(
        query, key, value
    )

I encounter type issues on xformers.ops.memory_efficient_attention since query and key are in float32 and value is in float16. Through some digging into the process of the normalization below, I see it is self.weight in float32 which cause the output of norm in float32, why autocast can not handle this problem?


class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        print(f"{x.dtype=}", flush=True)
        # x.dtype=torch.float16
        with torch.autocast(device_type='cuda', enabled=False): # whether or not disable autocast locally does not affect issues
            output = self._norm(x.float()).type_as(x)
        print(f"{output.dtype=} {self.weight.dtype=}", flush=True)
        # output.dtype=torch.float16 self.weight.dtype=torch.float32
        output = output * self.weight
        print(f"{output.dtype=}", flush=True)
        # output.dtype=torch.float32
        return output

Versions

PyTorch version: 2.1.1+cu118 Is debug build: False CUDA used to build PyTorch: 11.8 ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.2 LTS (x86_64) GCC version: (Ubuntu 11.3.0-1ubuntu1~22.04) 11.3.0 Clang version: Could not collect CMake version: version 3.25.0 Libc version: glibc-2.35

Python version: 3.10.11 (main, Apr 20 2023, 19:02:41) [GCC 11.2.0] (64-bit runtime) Python platform: Linux-5.15.0-60-generic-x86_64-with-glibc2.35 Is CUDA available: True CUDA runtime version: 12.0.140 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: 525.60.13 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 Address sizes: 46 bits physical, 57 bits virtual Byte Order: Little Endian CPU(s): 64 On-line CPU(s) list: 0-63 Vendor ID: GenuineIntel Model name: Intel(R) Xeon(R) Platinum 8358 CPU @ 2.60GHz CPU family: 6 Model: 106 Thread(s) per core: 1 Core(s) per socket: 32 Socket(s): 2 Stepping: 6 CPU max MHz: 3400.0000 CPU min MHz: 800.0000 BogoMIPS: 5200.00 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 art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 ds_cpl smx 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 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities L1d cache: 3 MiB (64 instances) L1i cache: 2 MiB (64 instances) L2 cache: 80 MiB (64 instances) L3 cache: 96 MiB (2 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-31 NUMA node1 CPU(s): 32-63 Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT disabled Vulnerability Retbleed: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

Versions of relevant libraries: [pip3] DISTS-pytorch==0.1 [pip3] mypy-extensions==1.0.0 [pip3] numpy==1.24.1 [pip3] open-clip-torch==2.19.0 [pip3] pytorch-debayer==1.4.1 [pip3] pytorch-lightning==2.0.6 [pip3] torch==2.1.1+cu118 [pip3] torch-fidelity==0.3.0 [pip3] torchaudio==2.1.1+cu118 [pip3] torchmetrics==1.0.1 [pip3] torchvision==0.16.1+cu118 [pip3] triton==2.1.0 [conda] dists-pytorch 0.1 pypi_0 pypi [conda] numpy 1.24.1 pypi_0 pypi [conda] open-clip-torch 2.19.0 pypi_0 pypi [conda] pytorch-debayer 1.4.1 pypi_0 pypi [conda] pytorch-lightning 2.0.6 pypi_0 pypi [conda] torch 2.1.1+cu118 pypi_0 pypi [conda] torch-fidelity 0.3.0 pypi_0 pypi [conda] torchaudio 2.1.1+cu118 pypi_0 pypi [conda] torchmetrics 1.0.1 pypi_0 pypi [conda] torchvision 0.16.1+cu118 pypi_0 pypi [conda] triton 2.1.0 pypi_0 pypi

facebookresearch / xformers

Autocast does not work with xformers #1046

🐛 Describe the bug

Versions