Update examples for 0.19.0 release

Author: kevalmorabia97

.gitignore

@@ -3,8 +3,10 @@ __pycache__/
 *.py[cod]
 *$py.class
 
-# C extensions
+# C, CPP extensions
 *.so
+.rendered.*.cpp
+.rendered.*.o
 # Exclude the plugin file
 !libfp8convkernel.so
 
@@ -177,6 +179,6 @@ cython_debug/
 **.pb
 **.onnx
 **.ckpt
-
-# Ignore temporary files created by tox
-pyproject.toml.bak
+**.safetensors
+**.bin
+**.pkl

diffusers/quantization/config.py

@@ -203,6 +203,21 @@ def create_dynamic_shapes(dynamic_shapes):
     }
 
 
+def remove_nesting(compilation_args):
+    compilation_args["dynamic_shapes"]["minShapes"]["text_embeds"] = compilation_args[
+        "dynamic_shapes"
+    ]["minShapes"].pop("added_cond_kwargs.text_embeds")
+    compilation_args["dynamic_shapes"]["minShapes"]["time_ids"] = compilation_args[
+        "dynamic_shapes"
+    ]["minShapes"].pop("added_cond_kwargs.time_ids")
+    compilation_args["dynamic_shapes"]["optShapes"]["text_embeds"] = compilation_args[
+        "dynamic_shapes"
+    ]["optShapes"].pop("added_cond_kwargs.text_embeds")
+    compilation_args["dynamic_shapes"]["optShapes"]["time_ids"] = compilation_args[
+        "dynamic_shapes"
+    ]["optShapes"].pop("added_cond_kwargs.time_ids")
+
+
 DYNAMIC_SHAPES = {
     "sdxl-1.0": create_dynamic_shapes(SDXL_DYNAMIC_SHAPES),
     "sdxl-turbo": create_dynamic_shapes(SDXL_DYNAMIC_SHAPES),
@@ -236,5 +251,8 @@ def get_io_shapes(model, onnx_load_path):
     else:
         output_name = "out.0"
     io_shapes = IO_SHAPES[model]
-    io_shapes[output_name] = io_shapes.pop("out.0")
+    # For models that are loaded from the output name will not be "out.0"
+    # so we need to update the dictionary key to match the output name
+    if "out.0" in io_shapes.keys():
+        io_shapes[output_name] = io_shapes.pop("out.0")
     return io_shapes

diffusers/quantization/diffusion_trt.py

@@ -22,7 +22,7 @@
 import argparse
 
 import torch
-from config import DYNAMIC_SHAPES, get_io_shapes, update_dynamic_axes
+from config import DYNAMIC_SHAPES, get_io_shapes, remove_nesting, update_dynamic_axes
 from diffusers import (
     DiffusionPipeline,
     FluxPipeline,
@@ -111,6 +111,11 @@ def main():
 
     compilation_args = DYNAMIC_SHAPES[args.model]
 
+    # We only need to remove the nesting for SDXL models as they contain the nested input added_cond_kwargs
+    # which are renamed by the DeviceModel
+    if args.onnx_load_path != "" and args.model in ["sdxl-1.0", "sdxl-turbo"]:
+        remove_nesting(compilation_args)
+
     # Define deployment configuration
     deployment = {
         "runtime": "TRT",

llm_ptq/README.md

@@ -94,7 +94,7 @@ Falcon 7B | Yes | Yes | No | No
 MPT 7B, 30B | Yes | Yes | Yes | Yes
 Baichuan 1, 2 | Yes | Yes | Yes | Yes
 ChatGLM2, 3 6B | No | No | Yes | No
-Bloom | bloom | Yes | Yes | Yes
+Bloom | Yes | Yes | Yes | Yes
 Phi-1,2,3 | Yes | Yes | Yes | Yes<sup>4</sup>
 Nemotron 8B | Yes | No | Yes | No
 Gemma 2B, 7B | Yes | No | Yes | Yes