From d4d3134f6d2d232c7bcfa80900a362921e644976 Mon Sep 17 00:00:00 2001
From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com>
Date: Sat, 28 Oct 2023 15:24:26 +0800
Subject: ManualCast for 10/16 series gpu

---
 modules/sd_models.py | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

(limited to 'modules/sd_models.py')

diff --git a/modules/sd_models.py b/modules/sd_models.py
index ccb6afd2..31bcb913 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -403,23 +403,26 @@ def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer
 
     if enable_fp8:
         devices.fp8 = True
+        if model.is_sdxl:
+            cond_stage = model.conditioner
+        else:
+            cond_stage = model.cond_stage_model
+
+        for module in cond_stage.modules():
+            if isinstance(module, torch.nn.Linear):
+                module.to(torch.float8_e4m3fn)
+
         if devices.device == devices.cpu:
             for module in model.model.diffusion_model.modules():
                 if isinstance(module, torch.nn.Conv2d):
                     module.to(torch.float8_e4m3fn)
                 elif isinstance(module, torch.nn.Linear):
                     module.to(torch.float8_e4m3fn)
-            timer.record("apply fp8 unet for cpu")
         else:
-            if model.is_sdxl:
-                cond_stage = model.conditioner
-            else:
-                cond_stage = model.cond_stage_model
-            for module in cond_stage.modules():
-                if isinstance(module, torch.nn.Linear):
-                    module.to(torch.float8_e4m3fn)
             model.model.diffusion_model = model.model.diffusion_model.to(torch.float8_e4m3fn)
-            timer.record("apply fp8 unet")
+        timer.record("apply fp8")
+    else:
+        devices.fp8 = False
 
     devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16
 
-- 
cgit v1.2.3