# The MIT License (MIT) # # Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved. # Copyright (c) 2024 Stability AI # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the 'Software'), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED 'AS IS', WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from argparse import ArgumentParser from diffusers import FlowMatchEulerDiscreteScheduler from other_impls import SD3Tokenizer from PIL import Image import migraphx as mgx import os import sys import torch import time from functools import wraps from hip import hip from collections import namedtuple HipEventPair = namedtuple('HipEventPair', ['start', 'end']) # measurement helper def measure(fn): @wraps(fn) def measure_ms(*args, **kwargs): start_time = time.perf_counter_ns() result = fn(*args, **kwargs) end_time = time.perf_counter_ns() print( f"Elapsed time for {fn.__name__}: {(end_time - start_time) * 1e-6:.4f} ms\n" ) return result return measure_ms def get_args(): parser = ArgumentParser() # Model compile parser.add_argument( "--onnx-model-path", type=str, default="models/sd3", help="Path to onnx model files.", ) parser.add_argument( "--compiled-model-path", type=str, default=None, help= "Path to compiled mxr model files. If not set, it will be saved next to the onnx model.", ) parser.add_argument( "--fp16", choices=["all", "vae", "clip", "mmdit"], nargs="+", help="Quantize models with fp16 precision.", ) parser.add_argument( "--force-compile", action="store_true", default=False, help="Ignore existing .mxr files and override them", ) parser.add_argument( "--exhaustive-tune", action="store_true", default=False, help="Perform exhaustive tuning when compiling onnx models", ) parser.add_argument( "--skip-t5", action="store_true", default=False, help= "Skip the third text encoder. Small accuracy penalty but large memory savings." ) # Runtime parser.add_argument( "-s", "--seed", type=int, default=42, help="Random seed", ) parser.add_argument( "-t", "--steps", type=int, default=50, help="Number of steps", ) parser.add_argument("-b", "--batch", type=int, default=1, help="Batch count or number of images to produce") parser.add_argument( "-p", "--prompt", type=str, required=True, help="Prompt", ) parser.add_argument( "-n", "--negative-prompt", type=str, default="", help="Negative prompt", ) parser.add_argument( "--scale", type=float, default=5.0, help="Guidance scale", ) parser.add_argument( "-o", "--output", type=str, default=None, help="Output name", ) return parser.parse_args() mgx_to_torch_dtype_dict = { "bool_type": torch.bool, "uint8_type": torch.uint8, "int8_type": torch.int8, "int16_type": torch.int16, "int32_type": torch.int32, "int64_type": torch.int64, "float_type": torch.float32, "double_type": torch.float64, "half_type": torch.float16, } torch_to_mgx_dtype_dict = { value: key for (key, value) in mgx_to_torch_dtype_dict.items() } def tensor_to_arg(tensor): return mgx.argument_from_pointer( mgx.shape( **{ "type": torch_to_mgx_dtype_dict[tensor.dtype], "lens": list(tensor.size()), "strides": list(tensor.stride()) }), tensor.data_ptr()) def tensors_to_args(tensors): return {name: tensor_to_arg(tensor) for name, tensor in tensors.items()} def get_output_name(idx): return f"main:#output_{idx}" def copy_tensor_sync(tensor, data): tensor.copy_(data) torch.cuda.synchronize() def run_model_sync(model, args): model.run(args) mgx.gpu_sync() def allocate_torch_tensors(model): input_shapes = model.get_parameter_shapes() data_mapping = { name: torch.zeros(shape.lens()).to( mgx_to_torch_dtype_dict[shape.type_string()]).to(device="cuda") for name, shape in input_shapes.items() } return data_mapping class StableDiffusionMGX(): def __init__(self, onnx_model_path, compiled_model_path, fp16, batch, force_compile=False, exhaustive_tune=False, skip_t5=False): self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained( "stabilityai/stable-diffusion-3-medium-diffusers", subfolder="scheduler") self.tokenizer = SD3Tokenizer() self.device = "cuda" self.skip_t5 = skip_t5 if fp16 is None: fp16 = [] elif "all" in fp16: fp16 = ["vae", "clip", "mmdit"] self.batch = batch print("Load models...") self.models = { "vae": StableDiffusionMGX.load_mgx_model( "vae_decoder", {"latent_sample": [self.batch, 16, 128, 128]}, onnx_model_path, compiled_model_path=compiled_model_path, use_fp16="vae" in fp16, force_compile=force_compile, exhaustive_tune=exhaustive_tune, offload_copy=False, batch=self.batch), "clip-l": StableDiffusionMGX.load_mgx_model( "text_encoder", {"input_ids": [1, 77]}, onnx_model_path, compiled_model_path=compiled_model_path, use_fp16="clip" in fp16, force_compile=force_compile, exhaustive_tune=exhaustive_tune, offload_copy=False), "clip-g": StableDiffusionMGX.load_mgx_model( "text_encoder_2", {"input_ids": [1, 77]}, onnx_model_path, compiled_model_path=compiled_model_path, use_fp16="clip" in fp16, force_compile=force_compile, exhaustive_tune=exhaustive_tune, offload_copy=False), "mmdit": StableDiffusionMGX.load_mgx_model( "transformer", { "hidden_states": [2 * self.batch, 16, 128, 128], "timestep": [2 * self.batch], "encoder_hidden_states": [2 * self.batch, 154, 4096], "pooled_projections": [2 * self.batch, 2048], }, onnx_model_path, compiled_model_path=compiled_model_path, use_fp16="mmdit" in fp16, force_compile=force_compile, exhaustive_tune=exhaustive_tune, offload_copy=False, batch=self.batch) } self.tensors = { "clip-g": allocate_torch_tensors(self.models["clip-g"]), "clip-l": allocate_torch_tensors(self.models["clip-l"]), # "t5xxl": allocate_torch_tensors(self.models["t5xxl"]), "mmdit": allocate_torch_tensors(self.models["mmdit"]), "vae": allocate_torch_tensors(self.models["vae"]), } self.model_args = { "clip-g": tensors_to_args(self.tensors['clip-g']), "clip-l": tensors_to_args(self.tensors['clip-l']), # "t5xxl": tensors_to_args(self.tensors['t5xxl']), "mmdit": tensors_to_args(self.tensors['mmdit']), "vae": tensors_to_args(self.tensors['vae']), } if not self.skip_t5: self.models["t5xxl"] = StableDiffusionMGX.load_mgx_model( "text_encoder_3", {"input_ids": [1, 77]}, onnx_model_path, compiled_model_path=compiled_model_path, use_fp16="clip" in fp16, force_compile=force_compile, exhaustive_tune=exhaustive_tune, offload_copy=False) self.tensors["t5xxl"] = allocate_torch_tensors( self.models["t5xxl"]) self.model_args["t5xxl"] = tensors_to_args(self.tensors['t5xxl']) self.events = { "warmup": HipEventPair(start=hip.hipEventCreate()[1], end=hip.hipEventCreate()[1]), "run": HipEventPair(start=hip.hipEventCreate()[1], end=hip.hipEventCreate()[1]), "clip": HipEventPair(start=hip.hipEventCreate()[1], end=hip.hipEventCreate()[1]), "denoise": HipEventPair(start=hip.hipEventCreate()[1], end=hip.hipEventCreate()[1]), "decode": HipEventPair(start=hip.hipEventCreate()[1], end=hip.hipEventCreate()[1]), } self.stream = hip.hipStreamCreate()[1] def cleanup(self): for event in self.events.values(): hip.hipEventDestroy(event.start) hip.hipEventDestroy(event.end) hip.hipStreamDestroy(self.stream) def profile_start(self, name): if name in self.events: hip.hipEventRecord(self.events[name].start, None) def profile_end(self, name): if name in self.events: hip.hipEventRecord(self.events[name].end, None) @measure @torch.no_grad() def run(self, prompt, negative_prompt, steps, seed, scale): torch.cuda.synchronize() self.profile_start("run") print("Tokenizing prompts...") prompt_tokens = self.tokenize(prompt) neg_prompt_tokens = self.tokenize(negative_prompt) print("Creating text embeddings...") self.profile_start("clip") prompt_embeddings = self.get_embeddings(prompt_tokens) neg_prompt_embeddings = self.get_embeddings(neg_prompt_tokens) self.profile_end("clip") # fix height and width for now # TODO: check for valid height/width combinations # and make them member variables height = 1024 width = 1024 latent = torch.empty(1, 16, height // 8, width // 8, device="cpu") generator = torch.manual_seed(seed) latent = torch.randn(latent.size(), dtype=torch.float32, layout=latent.layout, generator=generator).to(latent.dtype) self.scheduler.set_timesteps(steps) timesteps = self.scheduler.timesteps print("Running denoising loop...") self.profile_start("denoise") for step in timesteps: latent = self.denoise(latent, prompt_embeddings, neg_prompt_embeddings, step, scale) self.profile_end("denoise") latent = (latent / 1.5305) + 0.0609 self.profile_start("decode") print("Decode denoised result...") image = self.decode(latent) self.profile_end("decode") torch.cuda.synchronize() self.profile_end("run") return image def print_summary(self, denoise_steps): print('WARMUP\t{:>9.2f} ms'.format( hip.hipEventElapsedTime(self.events['warmup'].start, self.events['warmup'].end)[1])) print('CLIP\t{:>9.2f} ms'.format( hip.hipEventElapsedTime(self.events['clip'].start, self.events['clip'].end)[1])) print('mmditx{}\t{:>9.2f} ms'.format( str(denoise_steps), hip.hipEventElapsedTime(self.events['denoise'].start, self.events['denoise'].end)[1])) print('VAE-Dec\t{:>9.2f} ms'.format( hip.hipEventElapsedTime(self.events['decode'].start, self.events['decode'].end)[1])) print('RUN\t{:>9.2f} ms'.format( hip.hipEventElapsedTime(self.events['run'].start, self.events['run'].end)[1])) @staticmethod @measure def load_mgx_model(name, shapes, onnx_model_path, compiled_model_path=None, use_fp16=False, force_compile=False, exhaustive_tune=False, offload_copy=True, batch=1): print(f"Loading {name} model...") if compiled_model_path is None: compiled_model_path = onnx_model_path onnx_file = f"{onnx_model_path}/{name}/model.onnx" mxr_file = f"{compiled_model_path}/{name}/model_{'fp16' if use_fp16 else 'fp32'}_b{batch}_{'gpu' if not offload_copy else 'oc'}.mxr" if not force_compile and os.path.isfile(mxr_file): print(f"Found mxr, loading it from {mxr_file}") model = mgx.load(mxr_file, format="msgpack") elif os.path.isfile(onnx_file): print(f"No mxr found at {mxr_file}") print(f"Parsing from {onnx_file}") model = mgx.parse_onnx(onnx_file, map_input_dims=shapes) if use_fp16: mgx.quantize_fp16(model) model.compile(mgx.get_target("gpu"), exhaustive_tune=exhaustive_tune, offload_copy=offload_copy) print(f"Saving {name} model to {mxr_file}") os.makedirs(os.path.dirname(mxr_file), exist_ok=True) mgx.save(model, mxr_file, format="msgpack") else: print( f"No {name} model found at {onnx_file} or {mxr_file}. Please download it and re-try." ) sys.exit(1) return model @measure def tokenize(self, prompt): return self.tokenizer.tokenize_with_weights(prompt) def encode_token_weights(self, model_name, token_weight_pairs): tokens = list(map(lambda a: a[0], token_weight_pairs[0])) tokens = torch.tensor([tokens], dtype=torch.int64, device=self.device) copy_tensor_sync(self.tensors[model_name]["input_ids"], tokens.to(torch.int32)) run_model_sync(self.models[model_name], self.model_args[model_name]) encoder_out = self.tensors[model_name][get_output_name(0)] encoder_out2 = None if model_name != 't5xxl': # flipped outputs for clip text encoders... encoder_out2 = encoder_out encoder_out = self.tensors[model_name][get_output_name(1)] if encoder_out2 is not None: first_pooled = encoder_out2[0:1] else: first_pooled = encoder_out2 output = [encoder_out[0:1]] return torch.cat(output, dim=-2), first_pooled @measure def get_embeddings(self, prompt_tokens): l_out, l_pooled = self.encode_token_weights("clip-l", prompt_tokens["l"]) # stable-diffusion-3-lite-onnx has swapped outputs for clip-l text encoder if l_out.shape != (1, 77, 768): l_out, l_pooled = l_pooled, l_out g_out, g_pooled = self.encode_token_weights("clip-g", prompt_tokens["g"]) if not self.skip_t5: t5_out, _ = self.encode_token_weights("t5xxl", prompt_tokens["t5xxl"]) else: t5_out = torch.zeros((1, 77, 4096)).cuda() lg_out = torch.cat([l_out, g_out], dim=-1) lg_out = torch.nn.functional.pad(lg_out, (0, 4096 - lg_out.shape[-1])) return torch.cat([lg_out, t5_out], dim=-2), torch.cat( (l_pooled, g_pooled), dim=-1) @staticmethod def convert_to_rgb_image(image): image = (image / 2 + 0.5).clamp(0, 1) image = image.detach().cpu().permute(0, 2, 3, 1).numpy() images = (image * 255).round().astype("uint8") return [Image.fromarray(images[i]) for i in range(images.shape[0])] @staticmethod def save_image(pil_image, filename="output.png"): pil_image.save(filename) def CFGDenoiser(self, x, timestep, cond, uncond, cond_scale): # Run cond and uncond in a batch together x_concat = torch.cat([x, x]) timestep_concat = timestep.expand([2]) c_crossattn = torch.cat([cond["c_crossattn"], uncond["c_crossattn"]]) y = torch.cat([cond["y"], uncond["y"]]) copy_tensor_sync(self.tensors["mmdit"]["hidden_states"], x_concat) copy_tensor_sync(self.tensors["mmdit"]["timestep"], timestep_concat) copy_tensor_sync(self.tensors["mmdit"]["encoder_hidden_states"], c_crossattn) copy_tensor_sync(self.tensors["mmdit"]["pooled_projections"], y) run_model_sync(self.models["mmdit"], self.model_args['mmdit']) mmdit_out = self.tensors["mmdit"][get_output_name(0)] # Then split and apply CFG Scaling pos_out, neg_out = torch.tensor_split(mmdit_out, 2) scaled = neg_out + (pos_out - neg_out) * cond_scale # scheduler step function requies all tensors be on the CPU scaled = scaled.detach().clone().cpu() scheduler_out = self.scheduler.step(model_output=scaled, timestep=timestep, sample=x, return_dict=False)[0] return scheduler_out def fix_cond(self, cond): cond, pooled = (cond[0].cuda(), cond[1].cuda()) return {"c_crossattn": cond, "y": pooled} def denoise(self, latent, conditioning, neg_cond, step, cfg_scale): conditioning = self.fix_cond(conditioning) neg_cond = self.fix_cond(neg_cond) return self.CFGDenoiser(latent, step, conditioning, neg_cond, cfg_scale) @measure def decode(self, latents): copy_tensor_sync(self.tensors["vae"]["latent_sample"], latents) run_model_sync(self.models["vae"], self.model_args["vae"]) return self.tensors["vae"][get_output_name(0)] @measure def warmup(self, num_runs): self.profile_start("warmup") copy_tensor_sync(self.tensors["clip-l"]["input_ids"], torch.ones((1, 77)).to(torch.int32)) copy_tensor_sync(self.tensors["clip-g"]["input_ids"], torch.ones((1, 77)).to(torch.int32)) if not self.skip_t5: copy_tensor_sync(self.tensors["t5xxl"]["input_ids"], torch.ones((1, 77)).to(torch.int32)) copy_tensor_sync( self.tensors["mmdit"]["hidden_states"], torch.randn((2 * self.batch, 16, 128, 128)).to(torch.float)) copy_tensor_sync(self.tensors["mmdit"]["timestep"], torch.randn((2 * self.batch)).to(torch.float)) copy_tensor_sync( self.tensors["mmdit"]["encoder_hidden_states"], torch.randn((2 * self.batch, 154, 4096)).to(torch.float)) copy_tensor_sync(self.tensors["mmdit"]["pooled_projections"], torch.randn((2 * self.batch, 2048)).to(torch.float)) copy_tensor_sync( self.tensors["vae"]["latent_sample"], torch.randn((self.batch, 16, 128, 128)).to(torch.float)) for _ in range(num_runs): run_model_sync(self.models["clip-l"], self.model_args["clip-l"]) run_model_sync(self.models["clip-g"], self.model_args["clip-g"]) if not self.skip_t5: run_model_sync(self.models["t5xxl"], self.model_args["t5xxl"]) run_model_sync(self.models["mmdit"], self.model_args["mmdit"]) run_model_sync(self.models["vae"], self.model_args["vae"]) self.profile_end("warmup") if __name__ == "__main__": args = get_args() sd = StableDiffusionMGX(args.onnx_model_path, args.compiled_model_path, args.fp16, args.batch, args.force_compile, args.exhaustive_tune, args.skip_t5) print("Warmup") sd.warmup(5) print("Run") result = sd.run(args.prompt, args.negative_prompt, args.steps, args.seed, args.scale) print("Summary") sd.print_summary(args.steps) print("Cleanup") sd.cleanup() print("Convert result to rgb image...") images = StableDiffusionMGX.convert_to_rgb_image(result) for i, image in enumerate(images): filename = f"{args.batch}_{args.output}" if args.output else f"output_s{args.seed}_t{args.steps}_{i}.png" StableDiffusionMGX.save_image(image, filename) print(f"Image saved to {filename}")