[Performance 6/6] Add --precision half option to avoid casting during inference

[huchenlei]

Description

According to lllyasviel/stable-diffusion-webui-forge#716 (comment) , casting during inference is a main source of performance overhead. ComfyUI and Forge by default does fp16 inference without any casting, i.e. all tensors are fp16 before inference. The performance overhead is ~50ms/it.

This PR adds an option --precision half to disable autocasting and use all fp16 values during inference.

Screenshots/videos:

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[SLAPaper]

will force-fp16 mode conflicting with fp8 unet?

[AG-w]

I'm not sure if this is related to using dynamic lora weight
but I got this error

      File "H:\AItest\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1520, in _call_impl
        return forward_call(*args, **kwargs)
      File "H:\AItest\stable-diffusion-webui\extensions-builtin\Lora\networks.py", line 522, in network_Conv2d_forward
        return originals.Conv2d_forward(self, input)
      File "H:\AItest\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\conv.py", line 460, in forward
        return self._conv_forward(input, self.weight, self.bias)
      File "H:\AItest\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\conv.py", line 456, in _conv_forward
        return F.conv2d(input, weight, bias, self.stride,
    RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same

wonder if it's related to this
#12205

[feffy380]

Enabling --precision half breaks SD1.5 with the error mentioned here

[feffy380]

Found the offending line. In ldm's openaimodel.py L795 in the UNetModel class we have:

        h = x.type(self.dtype)

while in sgm it is simply:

        # h = x.type(self.dtype)
        h = x

self.dtype is set by constructing the model with use_fp16. When enabling force_fp16, we need to make sure to set the model's dtype to fp16. The fact that it works with SDXL is purely an accident due to the missing cast.

I don't know if it's the appropriate place to put it, but setting use_fp16 in sd_models.repair_config fixed SD1.5 inference with this PR for me.

[AG-w]

I don't know if it's the appropriate place to put it, but setting use_fp16 in sd_models.repair_config fixed SD1.5 inference with this PR for me.

something like this?

def repair_config(sd_config):

    if not hasattr(sd_config.model.params, "use_ema"):
        sd_config.model.params.use_ema = False

    if hasattr(sd_config.model.params, 'unet_config'):
        if shared.cmd_opts.no_half:
            sd_config.model.params.unet_config.params.use_fp16 = False
        elif shared.cmd_opts.upcast_sampling or shared.cmd_opts.precision == "half":
            sd_config.model.params.unet_config.params.use_fp16 = True

this does fixed dtype mismatch error

[huchenlei]

I don't know if it's the appropriate place to put it, but setting use_fp16 in sd_models.repair_config fixed SD1.5 inference with this PR for me.

something like this?

def repair_config(sd_config):

    if not hasattr(sd_config.model.params, "use_ema"):
        sd_config.model.params.use_ema = False

    if hasattr(sd_config.model.params, 'unet_config'):
        if shared.cmd_opts.no_half:
            sd_config.model.params.unet_config.params.use_fp16 = False
        elif shared.cmd_opts.upcast_sampling or shared.cmd_opts.precision == "half":
            sd_config.model.params.unet_config.params.use_fp16 = True

this does fixed dtype mismatch error

Thanks for digging out the solution! Verified that the solution works.

[ThereforeGames]

I'm still getting the following runtime error with both SDXL and SD15 models:

      File "T:\code\python\automatic-stable-diffusion-webui\repositories\generative-models\sgm\modules\diffusionmodules\openaimodel.py", line 984, in forward
        emb = self.time_embed(t_emb)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\container.py", line 215, in forward
        input = module(input)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\extensions-builtin\Lora\networks.py", line 508, in network_Linear_forward
        return originals.Linear_forward(self, input)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\linear.py", line 114, in forward
        return F.linear(input, self.weight, self.bias)
    RuntimeError: mat1 and mat2 must have the same dtype, but got Float and Half

Seems to be related to --precision half. Anyone else getting this?

[huchenlei]

I'm still getting the following runtime error with both SDXL and SD15 models:

      File "T:\code\python\automatic-stable-diffusion-webui\repositories\generative-models\sgm\modules\diffusionmodules\openaimodel.py", line 984, in forward
        emb = self.time_embed(t_emb)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\container.py", line 215, in forward
        input = module(input)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "T:\code\python\automatic-stable-diffusion-webui\extensions-builtin\Lora\networks.py", line 508, in network_Linear_forward
        return originals.Linear_forward(self, input)
      File "T:\code\python\automatic-stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\linear.py", line 114, in forward
        return F.linear(input, self.weight, self.bias)
    RuntimeError: mat1 and mat2 must have the same dtype, but got Float and Half

Seems to be related to --precision half. Anyone else getting this?

Can you share what model you used? I am not sure if you load a full precision model, whether weights are casted to fp16 before inference. The models I tested are already half precision.

[ThereforeGames]

Can you share what model you used? I am not sure if you load a full precision model, whether weights are casted to fp16 before inference. The models I tested are already half precision.

Sure, I tried a few:

• anyloraCheckpoint_bakedvaeBlessedFp16.safetensors [ef49fbb25f]
• v1-5-pruned.safetensors [1a189f0be6]
• cyberrealisticPony_v20a.safetensors [41e77f7657]

Same error regardless of checkpoint. It probably has something to do with my environment, although I'm not sure what yet. Here's a bit more context:

• All extensions disabled aside from built-ins.
• Not using any LoRAs or extra networks.
• Tried a bunch of different samplers and schedulers.
• Using commandline args: --precision half --ckpt-dir "S:/stable_diffusion/checkpoints" --lora-dir "S:/stable_diffusion/lora"
• Installed via your bundle PR

I'll write back if I figure out the cause.

[Arvamer]

I’ve tested this on a 6700 XT and there is a performance improvement. However, I think that this should not disallow setting --no-half-vae. On my card, running VAE in fp16 always produces black images. So, the only way to get correct images with --precision half is to enable NaN checks and rely on A1111 automatic fallback to fp32 VAE decoding, which negates some of the performance gains from this PR.

[w-e-w]

another report of fp8 issue

• [Bug]: Stable diffusion model failed to load when using --precision half in command line and enable FP8 #16122

[FurkanGozukara]

with using FP16 VAE i got almost double speed compared to no-half-vae nice

FP16 VAE is mandatory