Update optimization.py

optimization.py

@@ -4,37 +4,24 @@
 from typing import Any
 from typing import Callable
 from typing import ParamSpec
-from torchao.quantization import quantize_
-from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
+
 import spaces
 import torch
-from torch.utils._pytree import tree_map
+from torch.utils._pytree import tree_map_only
+from torchao.quantization import quantize_
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
 
 
 P = ParamSpec('P')
 
 
-TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length')
-TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length')
+TRANSFORMER_HIDDEN_DIM = torch.export.Dim('hidden', min=4096, max=8212)
 
 TRANSFORMER_DYNAMIC_SHAPES = {
-    'hidden_states': {
-        1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
-    },
-    'encoder_hidden_states': {
-        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
-    },
-    'encoder_hidden_states_mask': {
-        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
-    },
-    'image_rotary_emb': ({
-        0: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
-    }, {
-        0: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
-    }),
+    'hidden_states': {1: TRANSFORMER_HIDDEN_DIM},
+    'img_ids': {0: TRANSFORMER_HIDDEN_DIM},
 }
 
-
 INDUCTOR_CONFIGS = {
     'conv_1x1_as_mm': True,
     'epilogue_fusion': False,
@@ -53,10 +40,12 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
         with spaces.aoti_capture(pipeline.transformer) as call:
             pipeline(*args, **kwargs)
 
-        dynamic_shapes = tree_map(lambda t: None, call.kwargs)
+        dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
         dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
 
-        # quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
+        pipeline.transformer.fuse_qkv_projections()
+
+        quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
         
         exported = torch.export.export(
             mod=pipeline.transformer,