Fix a few issues with compile. Changed defaults. Future proofed block layer compile.

jobs/process/BaseSDTrainProcess.py

@@ -2099,24 +2099,25 @@ class BaseSDTrainProcess(BaseTrainProcess):
                     block_compile = getattr(self.model_config, 'block_compile', False)
 
                     cache_info = f", cache_size_limit={cache_size_limit}" if cache_size_limit != 8 else ""
-
                     # ====================================================
                     # BLOCK COMPILE
                     # ====================================================
                     if block_compile:
+                        BLOCK_LIST_ATTRS = self.sd.get_transformer_block_names()
+                        
+                        if BLOCK_LIST_ATTRS is None or len(BLOCK_LIST_ATTRS) == 0:
+                            BLOCK_LIST_ATTRS = [
+                                'layers',
+                                'transformer_blocks',
+                                'single_transformer_blocks',
+                                'double_stream_blocks',
+                                'single_stream_blocks',
+                                'double_blocks',
+                                'single_blocks',
+                                'blocks',
+                            ]
                         inner_unet = unwrap_model(self.sd.unet)
 
-                        BLOCK_LIST_ATTRS = [
-                            'layers',
-                            'transformer_blocks',
-                            'single_transformer_blocks',
-                            'double_stream_blocks',
-                            'single_stream_blocks',
-                            'double_blocks',
-                            'single_blocks',
-                            'blocks',
-                        ]
-
                         compiled_block_count = 0
 
                         for attr_name in BLOCK_LIST_ATTRS:

toolkit/config_modules.py

@@ -709,7 +709,9 @@ class ModelConfig:
 
         if self.compile and self.quantize:
             print("Quantized model detected - allowing torch.compile (experimental - Does not work with offloaded transformer)")
-            # self.compile = False
+            # make it torchao instead of quantio for compatibility with torch compile
+            if self.qtype == "qfloat8":
+                self.qtype = "float8"
         self.block_compile = kwargs.get("block_compile", False)
         self.compile_mode = kwargs.get("compile_mode", "default")
         self.compile_fullgraph = kwargs.get("compile_fullgraph", True)

ui/src/app/jobs/new/jobConfig.ts

@@ -32,11 +32,7 @@ export const defaultSliderConfig: SliderConfig = {
 };
 
 export const defaultCompileOptions = {
-  block_compile: false,
+  block_compile: true,
-  compile_mode: 'default',
-  compile_fullgraph: false,
-  compile_dynamic: false,
-  cache_size_limit: undefined,
 };
 
 export const defaultJobConfig: JobConfig = {
@@ -102,7 +98,6 @@ export const defaultJobConfig: JobConfig = {
           diff_output_preservation_class: 'person',
           switch_boundary_every: 1,
           loss_type: 'mse',
-          compile: false,
         },
         logging: {
           log_every: 1,
@@ -117,6 +112,7 @@ export const defaultJobConfig: JobConfig = {
           arch: 'flex1',
           low_vram: false,
           model_kwargs: {},
+          compile: false,
         },
         sample: defaultSampleConfig,
       },

version.py

@@ -1 +1 @@
-VERSION = "0.10.8"
+VERSION = "0.10.9"