diff --git a/text_to_image/StableDiffusionMGX.py b/text_to_image/StableDiffusionMGX.py new file mode 100644 index 0000000000..18186cdc62 --- /dev/null +++ b/text_to_image/StableDiffusionMGX.py @@ -0,0 +1,875 @@ +# The MIT License (MIT) +# +# Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved. +# +# 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 EulerDiscreteScheduler +from transformers import CLIPTokenizer +from PIL import Image + +import migraphx as mgx +from functools import wraps +from tqdm import tqdm +from hip import hip +from collections import namedtuple + +import os +import sys +import torch +import time +import logging +import coco +import dataset + +logging.basicConfig(level=logging.ERROR) +log = logging.getLogger("mgx-base") + +formatter = logging.Formatter("{levelname} - {message}", style="{") +file_handler = logging.FileHandler("mgx.log", mode="a", encoding="utf-8") +file_handler.setLevel("INFO") +file_handler.setFormatter(formatter) +log.addHandler(file_handler) + + + +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( + "--pipeline-type", + type=str, + choices=["sdxl", "sdxl-opt", "sdxl-turbo"], + required=True, + help="Specify pipeline type. Options: `sdxl`, `sdxl-opt`, `sdxl-turbo`", + ) + + parser.add_argument( + "--onnx-model-path", + type=str, + default=None, + help= + "Path to onnx model files. Use it to override the default models/ path", + ) + + 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( + "--use-refiner", + action="store_true", + default=False, + help="Use the refiner model", + ) + + parser.add_argument( + "--refiner-onnx-model-path", + type=str, + default=None, + help= + "Path to onnx model files. Use it to override the default models/ path", + ) + + parser.add_argument( + "--refiner-compiled-model-path", + type=str, + default=None, + help= + "Path to compiled mxr model files. If not set, it will be saved next to the refiner onnx model.", + ) + + parser.add_argument( + "--fp16", + choices=[ + "all", "vae", "clip", "clip2", "unetxl", "refiner_clip2", + "refiner_unetxl" + ], + 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", + ) + + # Runtime + parser.add_argument( + "-s", + "--seed", + type=int, + default=42, + help="Random seed", + ) + + parser.add_argument( + "-t", + "--steps", + type=int, + default=20, + help="Number of steps", + ) + + parser.add_argument( + "--refiner-steps", + type=int, + default=20, + help="Number of refiner steps", + ) + + 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( + "--refiner-aesthetic-score", + type=float, + default=6.0, + help="aesthetic score for refiner", + ) + + parser.add_argument( + "--refiner-negative-aesthetic-score", + type=float, + default=2.5, + help="negative aesthetic score for refiner", + ) + + parser.add_argument( + "-o", + "--output", + type=str, + default=None, + help="Output name", + ) + + parser.add_argument( + "--verbose", + action="store_true", + default=False, + help="Log during run", + ) + return parser.parse_args() + + +model_shapes = { + "clip": { + "input_ids": [2, 77] + }, + "clip2": { + "input_ids": [2, 77] + }, + "unetxl": { + "sample": [2, 4, 128, 128], + "encoder_hidden_states": [2, 77, 2048], + "text_embeds": [2, 1280], + "time_ids": [2, 6], + "timestep": [1], + }, + "refiner_unetxl": { + "sample": [2, 4, 128, 128], + "encoder_hidden_states": [2, 77, 1280], + "text_embeds": [2, 1280], + "time_ids": [2, 5], + "timestep": [1], + }, + "vae": { + "latent_sample": [1, 4, 128, 128] + }, +} + +model_names = { + "sdxl": { + "clip": "text_encoder", + "clip2": "text_encoder_2", + "unetxl": "unet", + "vae": "vae_decoder", + }, + "sdxl-opt": { + "clip": "clip.opt.mod", + "clip2": "clip2.opt.mod", + "unetxl": "unetxl.opt", + "vae": "vae_decoder", + }, + "sdxl-turbo": { + "clip": "text_encoder", + "clip2": "text_encoder_2", + "unetxl": "unet", + "vae": "vae_decoder", + }, + "refiner": { + "clip2": "clip2.opt.mod", + "unetxl": "unetxl.opt", + }, +} + +default_model_paths = { + "sdxl": "models/sdxl-1.0-base", + "sdxl-opt": "models/sdxl-1.0-base", + "sdxl-turbo": "models/sdxl-turbo", + "refiner": "models/sdxl-1.0-refiner", +} + +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.to(tensor.dtype)) + torch.cuda.synchronize() + + +def copy_tensor(tensor, data): + tensor.copy_(data.to(tensor.dtype)) + + +def run_model_sync(model, args): + model.run(args) + mgx.gpu_sync() + + +def run_model_async(model, args, stream): + model.run_async(args, stream, "ihipStream_t") + + +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, pipeline_type, onnx_model_path, compiled_model_path, + use_refiner, refiner_onnx_model_path, + refiner_compiled_model_path, fp16, force_compile, + exhaustive_tune, tokenizers=None, scheduler=None): + if not (onnx_model_path or compiled_model_path): + onnx_model_path = default_model_paths[pipeline_type] + + self.use_refiner = use_refiner + if not self.use_refiner and (refiner_onnx_model_path + or refiner_compiled_model_path): + print( + "WARN: Refiner model is provided, but was *not* enabled. Use --use-refiner to enable it." + ) + if self.use_refiner and not (refiner_onnx_model_path + or refiner_compiled_model_path): + refiner_onnx_model_path = default_model_paths["refiner"] + + is_turbo = "turbo" in pipeline_type + model_id = "stabilityai/sdxl-turbo" if is_turbo else "stabilityai/stable-diffusion-xl-base-1.0" + print(f"Using {model_id}") + + if scheduler is None: + print("Creating EulerDiscreteScheduler scheduler") + self.scheduler = EulerDiscreteScheduler.from_pretrained( + model_id, subfolder="scheduler") + else: + self.scheduler = scheduler + + print("Creating CLIPTokenizer tokenizers...") + if tokenizers is None: + tknz_path1 = os.path.join(onnx_model_path, "tokenizer") + tknz_path2 = os.path.join(onnx_model_path, "tokenizer_2") + self.tokenizers = { + "clip": + CLIPTokenizer.from_pretrained(tknz_path1), + "clip2": + CLIPTokenizer.from_pretrained(tknz_path2) + } + else: + self.tokenizers = tokenizers + + if fp16 is None: + fp16 = [] + elif "all" in fp16: + fp16 = [ + "vae", "clip", "clip2", "unetxl", "refiner_clip2", + "refiner_unetxl" + ] + + if "vae" in fp16: + model_names[pipeline_type]["vae"] = "vae_decoder_fp16_fix" + + log.info("Load models...") + self.models = { + "vae": + StableDiffusionMGX.load_mgx_model( + model_names[pipeline_type]["vae"], + model_shapes["vae"], + os.path.join (onnx_model_path, 'vae_decoder_fp16_fix'), + compiled_model_path=compiled_model_path, + use_fp16="vae" in fp16, + force_compile=force_compile, + exhaustive_tune=exhaustive_tune, + offload_copy=False), + "clip": + StableDiffusionMGX.load_mgx_model( + model_names[pipeline_type]["clip"], + model_shapes["clip"], + os.path.join (onnx_model_path, 'text_encoder'), + compiled_model_path=compiled_model_path, + use_fp16="clip" in fp16, + force_compile=force_compile, + exhaustive_tune=exhaustive_tune, + offload_copy=False), + "clip2": + StableDiffusionMGX.load_mgx_model( + model_names[pipeline_type]["clip2"], + model_shapes["clip2"], + os.path.join (onnx_model_path, 'text_encoder_2'), + compiled_model_path=compiled_model_path, + use_fp16="clip2" in fp16, + force_compile=force_compile, + exhaustive_tune=exhaustive_tune, + offload_copy=False), + "unetxl": + StableDiffusionMGX.load_mgx_model( + model_names[pipeline_type]["unetxl"], + model_shapes["unetxl"], + os.path.join (onnx_model_path, 'unet'), + compiled_model_path=compiled_model_path, + use_fp16="unetxl" in fp16, + force_compile=force_compile, + exhaustive_tune=exhaustive_tune, + offload_copy=False) + } + + log.info(f"init: loaded models") + + self.tensors = { + "clip": allocate_torch_tensors(self.models["clip"]), + "clip2": allocate_torch_tensors(self.models["clip2"]), + "unetxl": allocate_torch_tensors(self.models["unetxl"]), + "vae": allocate_torch_tensors(self.models["vae"]), + } + + # log.info(f"init: tensors: {self.tensors}") + + self.model_args = { + "clip": tensors_to_args(self.tensors["clip"]), + "clip2": tensors_to_args(self.tensors["clip2"]), + "unetxl": tensors_to_args(self.tensors["unetxl"]), + "vae": tensors_to_args(self.tensors["vae"]), + } + + # log.info(f"init: self.model_args: {self.model_args}") + + if self.use_refiner: + log.info(f"init: self.use_refiner: {self.use_refiner}") + + # Note: there is no clip for refiner, only clip2 + self.models["refiner_clip2"] = StableDiffusionMGX.load_mgx_model( + model_names["refiner"]["clip2"], + model_shapes["clip2"], + refiner_onnx_model_path, + compiled_model_path=refiner_compiled_model_path, + use_fp16="refiner_clip2" in fp16, + force_compile=force_compile, + exhaustive_tune=exhaustive_tune, + offload_copy=False) + + log.info(f"init: load refiner clip2") + + self.models["refiner_unetxl"] = StableDiffusionMGX.load_mgx_model( + model_names["refiner"]["unetxl"], + model_shapes[ + "refiner_unetxl"], # this differ from the original unetxl + refiner_onnx_model_path, + compiled_model_path=refiner_compiled_model_path, + use_fp16="refiner_unetxl" in fp16, + force_compile=force_compile, + exhaustive_tune=exhaustive_tune, + offload_copy=False) + + log.info(f"init: load refiner unet") + + self.tensors["refiner_clip2"] = allocate_torch_tensors( + self.models["refiner_clip2"]) + self.tensors["refiner_unetxl"] = allocate_torch_tensors( + self.models["refiner_unetxl"]) + self.model_args["refiner_clip2"] = tensors_to_args( + self.tensors["refiner_clip2"]) + self.model_args["refiner_unetxl"] = tensors_to_args( + self.tensors["refiner_unetxl"]) + # hipEventCreate return a tuple(error, event) + + log.info(f"init: creating hip events") + + 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]), + } + + # log.info(f"init: self.events: {self.events}") + + self.stream = hip.hipStreamCreate()[1] + + # log.info(f"init: self.stream: {self.stream}") + + 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, + steps=20, + negative_prompt="normal quality, low quality, worst quality, low res, blurry, nsfw, nude", + seed=42, + scale=5.0, + refiner_steps=20, + refiner_aesthetic_score=6.0, + refiner_negative_aesthetic_score=2.5, + verbose=False, + prompt_tokens=None, + latents_in=None, + device="cuda"): + torch.cuda.synchronize() + self.profile_start("run") + # need to set this for each run + self.scheduler.set_timesteps(steps, device=device) + + if verbose: + print("Tokenizing prompts...") + + if prompt_tokens is not None: + prompt_tokens = prompt_tokens + else: + # log.info(f"[mgx] input prompt: {prompt}") + prompt_tokens = self.tokenize(prompt, negative_prompt) + # log.info(f"[mgx] clip token: {prompt_tokens[0]['input_ids']}") + # log.info(f"[mgx] clip2 token: {prompt_tokens[1]['input_ids']}") + + # raise SystemExit("Checking if tokens match") + + if verbose: + print("Creating text embeddings...") + self.profile_start("clip") + hidden_states, text_embeddings = self.get_embeddings(prompt_tokens) + # log.info(f"[mgx] hidden_states (shape {hidden_states.shape}): {hidden_states}") + # log.info(f"[mgx] text_embeddings (shape {text_embeddings.shape}): {text_embeddings}") + # log.info(f"------DIVIDER--------") + self.profile_end("clip") + sample_size = list(self.tensors["vae"]["latent_sample"].size()) + if verbose: + print( + f"Creating random input data {sample_size} (latents) with {seed = }..." + ) + + height, width = sample_size[2:] + time_id = [height * 8, width * 8, 0, 0, height * 8, width * 8] + time_ids = torch.tensor([time_id, time_id]).to(device=device) + + if latents_in is None: + noise = torch.randn( + sample_size, generator=torch.manual_seed(seed)).to(device=device) + # input h/w crop h/w output h/w + + if verbose: + print("Apply initial noise sigma\n") + + # print(f"noise.device -> {noise.device}") + # print(f"self.scheduler.init_noise_sigma.device -> {self.scheduler.init_noise_sigma.device}") + latents = noise * self.scheduler.init_noise_sigma + else: + + if verbose: + print("Apply initial noise sigma\n") + + # log.info(f"[MGX] input latents provided, no need to generate") + latents = latents_in * self.scheduler.init_noise_sigma + + if verbose: + print("Running denoising loop...") + self.profile_start("denoise") + for step, t in tqdm(enumerate(self.scheduler.timesteps), + total=len(self.scheduler.timesteps), + desc=f"Device {device} Denoising", + ncols=100, + leave=True): + if verbose: + print(f"#{step}/{len(self.scheduler.timesteps)} step") + latents = self.denoise_step(text_embeddings, + hidden_states, + latents, + t, + scale, + time_ids, + model="unetxl", + device=device) + self.profile_end("denoise") + if self.use_refiner and refiner_steps > 0: + hidden_states, text_embeddings = self.get_embeddings( + prompt_tokens, is_refiner=True) + # input h/w crop h/w scores + time_id_pos = time_id[:4] + [refiner_aesthetic_score] + time_id_neg = time_id[:4] + [refiner_negative_aesthetic_score] + time_ids = torch.tensor([time_id_pos, + time_id_neg]).to(device=device) + # need to set this for each run + self.scheduler.set_timesteps(refiner_steps, device=device) + # Add noise to latents using timesteps + latents = self.scheduler.add_noise(latents, noise, + self.scheduler.timesteps[:1]) + if verbose: + print("Running refiner denoising loop...") + for step, t in enumerate(self.scheduler.timesteps): + if verbose: + print(f"#{step}/{len(self.scheduler.timesteps)} step") + latents = self.denoise_step(text_embeddings, + hidden_states, + latents, + t, + scale, + time_ids, + model="refiner_unetxl", + device=device) + if verbose: + print("Scale denoised result...") + latents = 1 / 0.18215 * latents + + self.profile_start("decode") + if verbose: + print("Decode denoised result...") + image = self.decode(latents) + self.profile_end("decode") + + torch.cuda.synchronize() + self.profile_end("run") + # assert image.shape == (1, 3, 1024, 1024), f"Actual shape of image is: {image.shape}" + 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('UNetx{}\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])) + + # @measure + @staticmethod + 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): + + log.info(f"Zixian: inside load_mgx_model") + 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'}_{'gpu' if not offload_copy else 'oc'}.mxr" + log.info(f"Zixian: mxr_file: {mxr_file}") + + 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"[IMPORTANT] Parsing from {onnx_file}") + model = mgx.parse_onnx(onnx_file, map_input_dims=shapes) + if use_fp16: + print(f"[IMPORTANT] Unet quantizing to 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: + log.info(f"Zixian: no model found") + 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, negative_prompt): + def _tokenize(tokenizer): + return self.tokenizers[tokenizer]( + [prompt, negative_prompt], + padding="max_length", + max_length=self.tokenizers[tokenizer].model_max_length, + truncation=True, + return_tensors="pt") + + tokens = _tokenize("clip") + tokens2 = _tokenize("clip2") + return (tokens, tokens2) + + # @measure + def get_embeddings(self, prompt_tokens, is_refiner=False): + def _create_embedding(model, input): + copy_tensor(self.tensors[model]["input_ids"], input.input_ids) + run_model_async(self.models[model], self.model_args[model], + self.stream) + + clip_input, clip2_input = prompt_tokens + clip, clip2 = "clip", ("refiner_" if is_refiner else "") + "clip2" + if not is_refiner: + _create_embedding(clip, clip_input) + _create_embedding(clip2, clip2_input) + + hidden_states = torch.concatenate( + (self.tensors[clip][get_output_name(0)], + self.tensors[clip2][get_output_name(1)]), + axis=2) if not is_refiner else self.tensors[clip2][get_output_name( + 1)] + text_embeds = self.tensors[clip2][get_output_name(0)] + return (hidden_states, text_embeds) + + @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[0]) + + @staticmethod + def save_image(pil_image, filename="output.png"): + pil_image.save(filename) + + # @measure + def denoise_step(self, text_embeddings, hidden_states, latents, t, scale, + time_ids, model, device): + latents_model_input = torch.cat([latents] * 2) + latents_model_input = self.scheduler.scale_model_input( + latents_model_input, t).to(device=device) + timestep = torch.atleast_1d(t.to(device=device)) # convert 0D -> 1D + + copy_tensor(self.tensors[model]["sample"], latents_model_input) + copy_tensor(self.tensors[model]["encoder_hidden_states"], + hidden_states) + copy_tensor(self.tensors[model]["text_embeds"], text_embeddings) + copy_tensor(self.tensors[model]["timestep"], timestep) + copy_tensor(self.tensors[model]["time_ids"], time_ids) + run_model_async(self.models[model], self.model_args[model], + self.stream) + + noise_pred_text, noise_pred_uncond = torch.tensor_split( + self.tensors[model][get_output_name(0)], 2) + + # perform guidance + noise_pred = noise_pred_uncond + scale * (noise_pred_text - + noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + return self.scheduler.step(noise_pred, t, latents).prev_sample + + # @measure + def decode(self, latents): + copy_tensor(self.tensors["vae"]["latent_sample"], latents) + run_model_async(self.models["vae"], self.model_args["vae"], + self.stream) + return self.tensors["vae"][get_output_name(0)] + + # @measure + def warmup(self, num_runs): + self.profile_start("warmup") + init_fn = lambda x: torch.ones if "clip" in x else torch.randn + for model in self.models.keys(): + for tensor in self.tensors[model].values(): + copy_tensor(tensor, init_fn(model)(tensor.size())) + + for _ in range(num_runs): + for model in self.models.keys(): + run_model_async(self.models[model], self.model_args[model], + self.stream) + self.profile_end("warmup") + + +if __name__ == "__main__": + args = get_args() + + # sd = StableDiffusionMGX(args.pipeline_type, args.onnx_model_path, + # args.compiled_model_path, args.use_refiner, + # args.refiner_onnx_model_path, + # args.refiner_compiled_model_path, args.fp16, + # args.force_compile, args.exhaustive_tune) + + sd = StableDiffusionMGX("sdxl", onnx_model_path=args.onnx_model_path, + compiled_model_path=None, use_refiner=False, + refiner_onnx_model_path=None, + refiner_compiled_model_path=None, fp16=args.fp16, + force_compile=False, exhaustive_tune=True) + print("Warmup") + sd.warmup(5) + print("Run") + + prompt_list = [] + prompt_list.append(["A young man in a white shirt is playing tennis.", "tennis.jpg"]) + # prompt_list.append(["Lorem ipsum dolor sit amet, consectetur adipiscing elit", "woman.jpg"]) + prompt_list.append(["Astronaut crashlanding in Madison Square Garden, cold color palette, muted colors, detailed, 8k", "crash_astro.jpg"]) + # prompt_list.append(["John Cena giving The Rock an Attitude Adjustment off the roof, warm color palette, vivid colors, detailed, 8k", "cena_rock.jpg"]) + + for element in prompt_list: + prompt, img_name = element[0], element[1] + # result = sd.run(prompt, args.negative_prompt, args.steps, args.seed, + # args.scale, args.refiner_steps, + # args.refiner_aesthetic_score, + # args.refiner_negative_aesthetic_score, args.verbose) + + result = sd.run(prompt=prompt, steps=20, seed=args.seed, + scale=5.0, refiner_steps=0, + refiner_aesthetic_score=0.0, + refiner_negative_aesthetic_score=0.0, verbose=False) + + print("Summary") + sd.print_summary(args.steps) + + print("Convert result to rgb image...") + image = StableDiffusionMGX.convert_to_rgb_image(result) + StableDiffusionMGX.save_image(image, img_name) + print(f"Image saved to {img_name}") + + print("Cleanup") + sd.cleanup() \ No newline at end of file diff --git a/text_to_image/backend_migraphx.py b/text_to_image/backend_migraphx.py new file mode 100644 index 0000000000..2df2605a8d --- /dev/null +++ b/text_to_image/backend_migraphx.py @@ -0,0 +1,301 @@ +from typing import Optional, List, Union +import migraphx as mgx + +import os +import torch +import logging +import sys +import backend +import time +import random +import json +import re + +from hip import hip +from PIL import Image +from functools import wraps +from collections import namedtuple +from transformers import CLIPTokenizer, CLIPTextModelWithProjection, CLIPProcessor, CLIPFeatureExtractor +from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler +from argparse import ArgumentParser +from StableDiffusionMGX import StableDiffusionMGX +from huggingface_hub import hf_hub_download, list_repo_files +import numpy as np + +HipEventPair = namedtuple('HipEventPair', ['start', 'end']) + +logging.basicConfig(level=logging.ERROR) +log = logging.getLogger("backend-mgx") + + +formatter = logging.Formatter("{levelname} - {message}", style="{") +file_handler = logging.FileHandler("backend_mgx.log", mode="a", encoding="utf-8") +file_handler.setLevel("INFO") +file_handler.setFormatter(formatter) +log.addHandler(file_handler) + +def download_model(repo_id, model_path): + # Zixian: Nov 10: Comment this out because model_path is current dir. + if os.path.exists(model_path): + log.info(f"MGX models already exists at {model_path}") + return + else: + os.makedirs(model_path, exist_ok=True) + + repo_files = list_repo_files(repo_id) + + files_to_download = [ + file for file in repo_files + if not file.endswith(".onnx") and not file.endswith("model_fp32_gpu.mxr") + ] + + for file_name in files_to_download: + local_file_path = os.path.join(model_path, file_name) + local_folder = os.path.dirname(local_file_path) + + # Create directory structure if it does not exist + os.makedirs(local_folder, exist_ok=True) + + # Download the file to the specific path + try: + hf_hub_download(repo_id=repo_id, filename=file_name, cache_dir=local_folder, local_dir=local_folder, local_dir_use_symlinks=False) + # log.info(f"Downloaded {file_name} to {local_file_path}") + except Exception as e: + log.error(f"Failed to download {file_name}: {e}") + + print (f"Zixian_in_the_log: Downloaded {file_name} to {local_file_path}") + +#! Yalu Ouyang [Nov 10 2024] Keep this in case we aren't allowed to modify coco.py +# class Decoder: +# def __init__(self, vocab_path): +# # Load the vocabulary with UTF-8 encoding to support non-ASCII characters +# with open(vocab_path, "r", encoding="utf-8") as f: +# vocab = json.load(f) + +# # Reverse the mapping: token_id -> word +# self.id_to_word = {int(id_): word for word, id_ in vocab.items()} + +# def decode_tokens(self, token_ids): +# # Ensure token_ids is a list, even if a tensor is passed +# if isinstance(token_ids, torch.Tensor): +# token_ids = token_ids.tolist() + +# # Handle both single sequences and batches +# if isinstance(token_ids[0], list): # Batch of sequences +# decoded_texts = [self._decode_sequence(sequence) for sequence in token_ids] +# return decoded_texts +# else: # Single sequence +# return self._decode_sequence(token_ids) + +# def _decode_sequence(self, token_ids): +# # Convert token IDs to words, handling any unknown tokens +# words = [self.id_to_word.get(token_id, "[UNK]") for token_id in token_ids] + +# # Remove special tokens and `` markers +# text = " ".join(words) +# text = re.sub(r"(<\|startoftext\|>|<\|endoftext\|>)", "", text) # Remove special tokens +# text = text.replace("", "").strip() # Remove `` markers and extra whitespace +# return text + +class BackendMIGraphX(backend.Backend): + def __init__( + self, + model_path=None, + model_id="xl", + guidance=5, #! To match the defaults of MiGraphX + steps=20, + batch_size=1, + device="cuda", + precision="fp32", + negative_prompt="normal quality, low quality, worst quality, low res, blurry, nsfw, nude", + ): + super(BackendMIGraphX, self).__init__() + # Zixian: Nov 10: Hard code to set model_path to current dir + # self.model_path = model_path + # self.model_path = os.getcwd() + self.model_path = os.path.join(os.getcwd(), "downloaded_model_folder") + if self.model_path is None: + raise SystemExit("Provide a valid Model Path to correctly run the program, exiting now...") + + self.pipeline_type = None + if model_id == "xl": + self.model_id = "SeaSponge/scc24_mlperf_mgx_exhaustive" + self.pipeline_type = "sdxl" + else: + raise ValueError(f"{model_id} is not a valid model id") + + download_model(self.model_id, self.model_path) + log.info(f"[mgx backend]: Returned from download_model") + + + self.device = device if torch.cuda.is_available() else "cpu" + self.device_num = int(device[-1]) \ + if (device != "cuda" and device != "cpu") else -1 + + # log.error(f"[mgx backend] self.device -> {self.device} | device_num -> {self.device_num}") + + if precision == "fp16": + self.dtype = torch.float16 + elif precision == "bf16": + self.dtype = torch.bfloat16 + else: + self.dtype = torch.float32 + + if torch.cuda.is_available(): + self.local_rank = 0 + self.world_size = 1 + + self.guidance = guidance + self.steps = steps + self.negative_prompt = negative_prompt + self.max_length_neg_prompt = 77 + self.batch_size = batch_size + + self.mgx = None + tknz_path1 = os.path.join(self.model_path, "tokenizer") + tknz_path2 = os.path.join(self.model_path, "tokenizer_2") + # self.scheduler = EulerDiscreteScheduler.from_pretrained(os.path.join(self.model_path, "scheduler")) + self.scheduler = EulerDiscreteScheduler.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", subfolder="scheduler") + log.info(f"Zixian: Loaded scheduler") + self.pipe = self.Pipe() + # self.pipe.tokenizer = CLIPTokenizer.from_pretrained(tknz_path1) + # self.pipe.tokenizer_2 = CLIPTokenizer.from_pretrained(tknz_path2) + self.pipe.tokenizer = CLIPTokenizer.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", subfolder="tokenizer") + self.pipe.tokenizer_2 = CLIPTokenizer.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", subfolder="tokenizer_2") + log.info(f"Zixian: Loaded tokenizer & tokenizer2") + # log.info(f"Zixian_in_the_log tknz_path1: {tknz_path1}") + # log.info(f"Zixian_in_the_log tknz_path2: {tknz_path2}") + # self.decoder1 = Decoder(os.path.join(self.model_path, "tokenizer/vocab.json")) + # self.decoder2 = Decoder(os.path.join(self.model_path, "tokenizer_2/vocab.json")) + self.tokenizers = [self.pipe.tokenizer, self.pipe.tokenizer_2] + + class Pipe: + def __init__(self): + self.tokenizer = None + self.tokenizer_2 = None + + def version(self): + return torch.__version__ + + def name(self): + return "pytorch-SUT" + + def image_format(self): + return "NCHW" + + def load(self): + if self.model_path is None: + log.warning( + "Model path not provided, running with default hugging face weights\n" + "This may not be valid for official submissions" + ) + + raise SystemExit("Provide a valid Model Path to correctly run the program, exiting now...") + + else: + if self.device_num != -1: + # log.error(f"Hip set device to -> {self.device_num}") + hip.hipSetDevice(self.device_num) + + # raise SystemExit("Stopping to check") + + # Parameter explanations here: + # onnx_model_path = self.model_path + # path to compiled .mxr can be left as None + # Don't want to use refiner model + use_refiner = False + # Therefore refiner model path also None + # refiner compiled model path also None + + # set fp16 according to initialization input + fp16 = "all" if self.dtype == torch.float16 else None + # Don't want to force .onnx to .mxr compile + force_compile = False + # Use exhaustive tune when compilling .onnx -> .mxr + exhaustive_tune = True + + tokenizers = {"clip": self.tokenizers[0], "clip2": self.tokenizers[1]} + + self.mgx = StableDiffusionMGX(self.pipeline_type, onnx_model_path=self.model_path, + compiled_model_path=None, use_refiner=use_refiner, + refiner_onnx_model_path=None, + refiner_compiled_model_path=None, fp16=fp16, + force_compile=force_compile, exhaustive_tune=exhaustive_tune, tokenizers=tokenizers, + scheduler=self.scheduler) + + # log.info(f"[backend_migraphx.py]: after initializing self.mgx") + + # self.mgx.warmup(5) + + # log.info(f"[backend_migraphx.py]: after mgx.warmup") + + return self + + def predict(self, inputs): + images = [] + + # Explanation for mgx.run() arguments + # negative_prompt = self.negative_prompt + # steps = self.steps + # scale refers to guidance scale -> scale = self.guidance + # the default SDXLPipeline chooses a random seed everytime, we'll do so manually here + # not using refiner, so refiner_step = 0 + # not using refiner, so aesthetic_score = 0 + # not using refiner, so negative_aesthetic_score = 0 + # defaults to not verbose + verbose = False + #! The main pipeline from loadgen doesn't have text prompt, only tokens + + for i in range(0, len(inputs), self.batch_size): + latents_input = [inputs[idx]["latents"] for idx in range(i, min(i+self.batch_size, len(inputs)))] + latents_input = torch.cat(latents_input).to(self.device) + if self.batch_size == 1: + # prompt_token = inputs[i]["input_tokens"] + # log.info(f"[mgx backend batchsz=1] inputs[i] -> {inputs[i]}") + prompt_in = inputs[i]["caption"] + # log.info(f"[mgx backend] i -> {i} | prompt_in -> {prompt_in}") + seed = random.randint(0, 2**31 - 1) + + # prompt_in = self.decoder1.decode_tokens(prompt_token['input_ids']) + + result = self.mgx.run(prompt=prompt_in, negative_prompt=self.negative_prompt, steps=self.steps, seed=seed, + scale=self.guidance, refiner_steps=0, + refiner_aesthetic_score=0, + refiner_negative_aesthetic_score=0, verbose=verbose, + prompt_tokens=None, device=self.device, latents_in=latents_input) + + # result shape = (3, 1024, 1024) + + # img_name = f"{self.device}_{random.randint(0, 1000)}.jpg" + # image = StableDiffusionMGX.convert_to_rgb_image(result) + # StableDiffusionMGX.save_image(image, img_name) + # log.info(f"[mgx backend batchsz=1] Image saved to {img_name}") + #! COCO needs this to be 3-dimensions + + new_res = (result / 2 + 0.5).clamp(0, 1) + + # log.info(f"[mgx backend] type result: {type(result)} | result shape: {result.shape}") + # log.info(f"[mgx backend] type new_res: {type(new_res)} | new_res shape: {new_res.shape}") + # log.info(f"------DIVIDER--------") + images.extend(new_res) + + else: + prompt_list = [] + for prompt in inputs[i:min(i+self.batch_size, len(inputs))]: + assert isinstance(prompt, dict), "prompt (in inputs) isn't a dict" + # prompt_token = prompt["input_tokens"] + prompt_in = inputs[i]["caption"] + + + for prompt in prompt_list: + seed = random.randint(0, 2**31 - 1) + result = self.mgx.run(prompt=prompt, negative_prompt=self.negative_prompt, steps=self.steps, seed=seed, + scale=self.guidance, refiner_steps=0, + refiner_aesthetic_score=0, + refiner_negative_aesthetic_score=0, verbose=verbose, + prompt_tokens=None, device=self.device, latents_in=latents_input) + + new_res = (result / 2 + 0.5).clamp(0, 1) + images.extend(new_res) + + return images \ No newline at end of file diff --git a/text_to_image/backend_pytorch.py b/text_to_image/backend_pytorch.py index 36e2b80090..f2af4d75c7 100644 --- a/text_to_image/backend_pytorch.py +++ b/text_to_image/backend_pytorch.py @@ -17,9 +17,9 @@ def __init__( model_id="xl", guidance=8, steps=20, - batch_size=1, + batch_size=2, device="cuda", - precision="fp32", + precision="fp16", negative_prompt="normal quality, low quality, worst quality, low res, blurry, nsfw, nude", ): super(BackendPytorch, self).__init__() @@ -57,39 +57,41 @@ def image_format(self): return "NCHW" def load(self): - if self.model_path is None: - log.warning( - "Model path not provided, running with default hugging face weights\n" - "This may not be valid for official submissions" - ) - self.scheduler = EulerDiscreteScheduler.from_pretrained( - self.model_id, subfolder="scheduler" - ) - self.pipe = StableDiffusionXLPipeline.from_pretrained( - self.model_id, - scheduler=self.scheduler, - safety_checker=None, - add_watermarker=False, - variant="fp16" if (self.dtype == torch.float16) else None, - torch_dtype=self.dtype, - ) + # if self.model_path is None: + # log.warning( + # "Model path not provided, running with default hugging face weights\n" + # "This may not be valid for official submissions" + # ) + self.scheduler = EulerDiscreteScheduler.from_pretrained( + self.model_id, subfolder="scheduler" + ) + self.pipe = StableDiffusionXLPipeline.from_pretrained( + self.model_id, + scheduler=self.scheduler, + safety_checker=None, + add_watermarker=False, + # variant="fp16" if (self.dtype == torch.float16) else None, + variant="fp16" , + torch_dtype=self.dtype, + ) # self.pipe.unet = torch.compile(self.pipe.unet, mode="reduce-overhead", fullgraph=True) - else: - self.scheduler = EulerDiscreteScheduler.from_pretrained( - os.path.join(self.model_path, "checkpoint_scheduler"), - subfolder="scheduler", - ) - self.pipe = StableDiffusionXLPipeline.from_pretrained( - os.path.join(self.model_path, "checkpoint_pipe"), - scheduler=self.scheduler, - safety_checker=None, - add_watermarker=False, - torch_dtype=self.dtype, - ) + # else: + # self.scheduler = EulerDiscreteScheduler.from_pretrained( + # os.path.join(self.model_path, "checkpoint_scheduler"), + # subfolder="scheduler", + # ) + # self.pipe = StableDiffusionXLPipeline.from_pretrained( + # os.path.join(self.model_path, "checkpoint_pipe"), + # scheduler=self.scheduler, + # safety_checker=None, + # add_watermarker=False, + # variant="fp16" if (self.dtype == torch.float16) else None, + # torch_dtype=self.dtype, + # ) # self.pipe.unet = torch.compile(self.pipe.unet, mode="reduce-overhead", fullgraph=True) self.pipe.to(self.device) - # self.pipe.set_progress_bar_config(disable=True) + #self.pipe.set_progress_bar_config(disable=True) self.negative_prompt_tokens = self.pipe.tokenizer( self.convert_prompt(self.negative_prompt, self.pipe.tokenizer), @@ -210,15 +212,13 @@ def encode_tokens( text_input_ids.to(device), output_hidden_states=True ) - # We are only ALWAYS interested in the pooled output of the - # final text encoder + # We are only ALWAYS interested in the pooled output of the final text encoder pooled_prompt_embeds = prompt_embeds[0] if clip_skip is None: prompt_embeds = prompt_embeds.hidden_states[-2] else: # "2" because SDXL always indexes from the penultimate layer. - prompt_embeds = prompt_embeds.hidden_states[-( - clip_skip + 2)] + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] prompt_embeds_list.append(prompt_embeds) @@ -234,8 +234,7 @@ def encode_tokens( and zero_out_negative_prompt ): negative_prompt_embeds = torch.zeros_like(prompt_embeds) - negative_pooled_prompt_embeds = torch.zeros_like( - pooled_prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) elif do_classifier_free_guidance and negative_prompt_embeds is None: negative_prompt = negative_prompt or "" negative_prompt_2 = negative_prompt_2 or negative_prompt @@ -262,35 +261,30 @@ def encode_tokens( uncond_input.to(device), output_hidden_states=True, ) - # We are only ALWAYS interested in the pooled output of the - # final text encoder + # We are only ALWAYS interested in the pooled output of the final text encoder negative_pooled_prompt_embeds = negative_prompt_embeds[0] negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] negative_prompt_embeds_list.append(negative_prompt_embeds) - negative_prompt_embeds = torch.concat( - negative_prompt_embeds_list, dim=-1) + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) if pipe.text_encoder_2 is not None: prompt_embeds = prompt_embeds.to( dtype=pipe.text_encoder_2.dtype, device=device ) else: - prompt_embeds = prompt_embeds.to( - dtype=pipe.unet.dtype, device=device) + prompt_embeds = prompt_embeds.to(dtype=pipe.unet.dtype, device=device) bs_embed, seq_len, _ = prompt_embeds.shape - # duplicate text embeddings for each generation per prompt, using mps - # friendly method + # duplicate text embeddings for each generation per prompt, using mps friendly method prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view( bs_embed * num_images_per_prompt, seq_len, -1 ) if do_classifier_free_guidance: - # duplicate unconditional embeddings for each generation per - # prompt, using mps friendly method + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method seq_len = negative_prompt_embeds.shape[1] if pipe.text_encoder_2 is not None: @@ -322,7 +316,7 @@ def encode_tokens( pooled_prompt_embeds, negative_pooled_prompt_embeds, ) - + def prepare_inputs(self, inputs, i): if self.batch_size == 1: return self.encode_tokens( @@ -337,7 +331,7 @@ def prepare_inputs(self, inputs, i): negative_prompt_embeds = [] pooled_prompt_embeds = [] negative_pooled_prompt_embeds = [] - for prompt in inputs[i: min(i + self.batch_size, len(inputs))]: + for prompt in inputs[i:min(i+self.batch_size, len(inputs))]: assert isinstance(prompt, dict) text_input = prompt["input_tokens"] text_input_2 = prompt["input_tokens_2"] @@ -358,26 +352,19 @@ def prepare_inputs(self, inputs, i): pooled_prompt_embeds.append(p_p_e) negative_pooled_prompt_embeds.append(n_p_p_e) + prompt_embeds = torch.cat(prompt_embeds) negative_prompt_embeds = torch.cat(negative_prompt_embeds) pooled_prompt_embeds = torch.cat(pooled_prompt_embeds) - negative_pooled_prompt_embeds = torch.cat( - negative_pooled_prompt_embeds) - return ( - prompt_embeds, - negative_prompt_embeds, - pooled_prompt_embeds, - negative_pooled_prompt_embeds, - ) + negative_pooled_prompt_embeds = torch.cat(negative_pooled_prompt_embeds) + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds def predict(self, inputs): images = [] with torch.no_grad(): for i in range(0, len(inputs), self.batch_size): - latents_input = [ - inputs[idx]["latents"] - for idx in range(i, min(i + self.batch_size, len(inputs))) - ] + print (f'self.steps BEFORE pipe: {self.steps}') + latents_input = [inputs[idx]["latents"] for idx in range(i, min(i+self.batch_size, len(inputs)))] latents_input = torch.cat(latents_input).to(self.device) ( prompt_embeds, @@ -392,8 +379,11 @@ def predict(self, inputs): negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, guidance_scale=self.guidance, num_inference_steps=self.steps, + # num_inference_steps=20, output_type="pt", latents=latents_input, ).images + print (f'self.steps AFTER pipe: {self.steps}') images.extend(generated) return images + diff --git a/text_to_image/coco.py b/text_to_image/coco.py index e9499b0e6c..92773ed9ca 100644 --- a/text_to_image/coco.py +++ b/text_to_image/coco.py @@ -35,22 +35,27 @@ def __init__( latent_dtype=torch.float32, latent_device="cuda", latent_framework="torch", + pipe_type=None, **kwargs, ): super().__init__() - self.captions_df = pd.read_csv( - f"{data_path}/captions/captions.tsv", sep="\t") + self.captions_df = pd.read_csv(f"{data_path}/captions/captions.tsv", sep="\t") self.image_size = image_size self.preprocessed_dir = os.path.abspath(f"{data_path}/preprocessed/") self.img_dir = os.path.abspath(f"{data_path}/validation/data/") self.name = name + + self.pipe_tokenizer = pipe_tokenizer + self.pipe_tokenizer_2 = pipe_tokenizer_2 # Preprocess prompts self.captions_df["input_tokens"] = self.captions_df["caption"].apply( - lambda x: self.preprocess(x, pipe_tokenizer) + # lambda x: self.preprocess(x, pipe_tokenizer) + lambda x: x ) self.captions_df["input_tokens_2"] = self.captions_df["caption"].apply( - lambda x: self.preprocess(x, pipe_tokenizer_2) + # lambda x: self.preprocess(x, pipe_tokenizer_2) + lambda x: x ) self.latent_dtype = latent_dtype self.latent_device = latent_device if torch.cuda.is_available() else "cpu" @@ -66,6 +71,8 @@ def __init__( .to(latent_dtype) .to(latent_device) ) + + self.pipe_type = pipe_type def preprocess(self, prompt, tokenizer): converted_prompt = self.convert_prompt(prompt, tokenizer) @@ -113,14 +120,34 @@ def convert_prompt(self, prompt, tokenizer): def get_item(self, id): return dict(self.captions_df.loc[id], latents=self.latents) + #! Yalu Ouyang [Nov 14 2024] Overrides parent Dataset class, default behavior is same though + def get_samples(self, id_list): + if self.pipe_type == "migraphx": + #! Yalu Ouyang [Nov 14 2024] MGX backend just needs text prompt + data = [ + { + "caption": self.items_inmemory[id]["caption"], + "latents": self.items_inmemory[id]["latents"], + } + for id in id_list + ] + else: + data = [ + { + "input_tokens": self.items_inmemory[id]["input_tokens"], + "input_tokens_2": self.items_inmemory[id]["input_tokens_2"], + "latents": self.items_inmemory[id]["latents"], + } + for id in id_list + ] + images = [self.items_inmemory[id]["file_name"] for id in id_list] + return data, images + def get_item_count(self): return len(self.captions_df) def get_img(self, id): - img = Image.open( - self.img_dir + - "/" + - self.captions_df.loc[id]["file_name"]) + img = Image.open(self.img_dir + "/" + self.captions_df.loc[id]["file_name"]) return self.image_to_tensor(img) def get_imgs(self, id_list): @@ -141,11 +168,7 @@ def get_item_loc(self, id): class PostProcessCoco: def __init__( - self, - device="cpu", - dtype="uint8", - statistics_path=os.path.join( - os.path.dirname(__file__), "tools", "val2014.npz"), + self, device="cpu", dtype="uint8", statistics_path=os.path.join(os.path.dirname(__file__), "tools", "val2014.npz") ): self.results = [] self.good = 0 @@ -167,33 +190,27 @@ def add_results(self, results): def __call__(self, results, ids, expected=None, result_dict=None): self.content_ids.extend(ids) return [ - (t.cpu().permute(1, 2, 0).float().numpy() * 255) - .round() - .astype(self.numpy_dtype) + (t.cpu().permute(1, 2, 0).float().numpy() * 255).round().astype(self.numpy_dtype) for t in results ] - + def save_images(self, ids, ds): info = [] idx = {} - for i, image_id in enumerate(self.content_ids): - if image_id in ids: - idx[image_id] = i + for i, id in enumerate(self.content_ids): + if id in ids: + idx[id] = i if not os.path.exists("images/"): os.makedirs("images/", exist_ok=True) - for image_id in ids: - if not idx.get(image_id): - print( - f"image id {image_id} is missing in the results. Hence not saved.") - continue - caption = ds.get_caption(image_id) - generated = Image.fromarray(self.results[idx[image_id]]) - image_path_tmp = f"images/{self.content_ids[idx[image_id]]}.png" + for id in ids: + caption = ds.get_caption(id) + generated = Image.fromarray(self.results[idx[id]]) + image_path_tmp = f"images/{self.content_ids[idx[id]]}.png" generated.save(image_path_tmp) - info.append((self.content_ids[idx[image_id]], caption)) + info.append((self.content_ids[idx[id]], caption)) with open("images/captions.txt", "w+") as f: - for image_id, caption in info: - f.write(f"{image_id} {caption}\n") + for id, caption in info: + f.write(f"{id} {caption}\n") def start(self): self.results = [] @@ -209,10 +226,7 @@ def finalize(self, result_dict, ds=None, output_dir=None): 100 * clip.get_clip_score(caption, generated).item() ) - fid_score = compute_fid( - self.results, - self.statistics_path, - self.device) + fid_score = compute_fid(self.results, self.statistics_path, self.device) result_dict["FID_SCORE"] = fid_score result_dict["CLIP_SCORE"] = np.mean(self.clip_scores) diff --git a/text_to_image/main.py b/text_to_image/main.py index 6aa7c15e75..3d6f7d1d0a 100644 --- a/text_to_image/main.py +++ b/text_to_image/main.py @@ -24,6 +24,11 @@ import dataset import coco +# import torchvision.transforms as T +# transform_im = T.ToPILImage() + +from concurrent.futures import ThreadPoolExecutor, as_completed + logging.basicConfig(level=logging.INFO) log = logging.getLogger("main") @@ -61,6 +66,12 @@ "backend": "pytorch-dist", "model-name": "stable-diffusion-xl", }, + # ? Yalu Ouyang modification: Oct 16 2024 + "stable-diffusion-xl-mgx": { + "dataset": "coco-1024", + "backend": "migraphx", + "model-name": "stable-diffusion-xl", + }, } SCENARIO_MAP = { @@ -73,22 +84,20 @@ def get_args(): parser = argparse.ArgumentParser() + parser.add_argument("--dataset", + default="coco-1024", + choices=SUPPORTED_DATASETS.keys(), help="dataset") + parser.add_argument("--dataset-path", + default="coco2014",help="path to the dataset") parser.add_argument( - "--dataset", - choices=SUPPORTED_DATASETS.keys(), - help="dataset") - parser.add_argument( - "--dataset-path", - required=True, - help="path to the dataset") - parser.add_argument( - "--profile", choices=SUPPORTED_PROFILES.keys(), help="standard profiles" + "--profile", + default="stable-diffusion-xl-mgx", + choices=SUPPORTED_PROFILES.keys(), help="standard profiles" ) parser.add_argument( "--scenario", - default="SingleStream", - help="mlperf benchmark scenario, one of " + - str(list(SCENARIO_MAP.keys())), + default="Offline", + help="mlperf benchmark scenario, one of " + str(list(SCENARIO_MAP.keys())), ) parser.add_argument( "--max-batchsize", @@ -97,24 +106,23 @@ def get_args(): help="max batch size in a single inference", ) parser.add_argument("--threads", default=1, type=int, help="threads") - parser.add_argument( - "--accuracy", - action="store_true", - help="enable accuracy pass") + parser.add_argument("--accuracy", action="store_true", help="enable accuracy pass") parser.add_argument( "--find-peak-performance", action="store_true", help="enable finding peak performance pass", ) - parser.add_argument("--backend", help="Name of the backend") + parser.add_argument("--backend", default='migraphx', help="Name of the backend") parser.add_argument("--model-name", help="Name of the model") parser.add_argument("--output", default="output", help="test results") parser.add_argument("--qps", type=int, help="target qps") - parser.add_argument("--model-path", help="Path to model weights") + parser.add_argument("--model-path", + default="/work1/zixian/youyang1/models/sdxl-1.0-base", + help="Path to model weights") parser.add_argument( "--dtype", - default="fp32", + default="fp16", choices=["fp32", "fp16", "bf16"], help="dtype of the model", ) @@ -131,6 +139,10 @@ def get_args(): help="framework to load the latents", ) + # file to use mlperf rules compliant parameters + parser.add_argument( + "--mlperf_conf", default="mlperf.conf", help="mlperf rules config" + ) # file for user LoadGen settings such as target QPS parser.add_argument( "--user_conf", @@ -141,24 +153,21 @@ def get_args(): parser.add_argument( "--audit_conf", default="audit.config", help="config for LoadGen audit settings" ) + parser.add_argument( + "--gpu-num", default=4, type=int, help="number of gpus to use" + ) # arguments to save images # pass this argument for official submission # parser.add_argument("--output-images", action="store_true", help="Store a subset of the generated images") # do not modify this argument for official submission - parser.add_argument( - "--ids-path", help="Path to caption ids", default="tools/sample_ids.txt" - ) + parser.add_argument("--ids-path", help="Path to caption ids", default="tools/sample_ids.txt") - # below will override mlperf rules compliant settings - don't use for - # official submission + # below will override mlperf rules compliant settings - don't use for official submission parser.add_argument("--time", type=int, help="time to scan in seconds") parser.add_argument("--count", type=int, help="dataset items to use") parser.add_argument("--debug", action="store_true", help="debug") parser.add_argument( - "--performance-sample-count", - type=int, - help="performance sample count", - default=5000, + "--performance-sample-count", type=int, help="performance sample count", default=1000 ) parser.add_argument( "--max-latency", type=float, help="mlperf max latency in pct tile" @@ -193,6 +202,12 @@ def get_backend(backend, **kwargs): from backend_pytorch import BackendPytorch backend = BackendPytorch(**kwargs) + + # ? Yalu Ouyang Modification: Nov 5 2024 + elif backend == "migraphx": + from backend_migraphx import BackendMIGraphX + + backend = BackendMIGraphX(**kwargs) elif backend == "debug": from backend_debug import BackendDebug @@ -239,15 +254,17 @@ def run_one_item(self, qitem: Item): processed_results = [] try: results = self.model.predict(qitem.inputs) + # log.info("[Line 254] runs fine after results") processed_results = self.post_process( results, qitem.content_id, qitem.inputs, self.result_dict ) + # log.info("[Line 258] runs fine after processed_results") if self.take_accuracy: self.post_process.add_results(processed_results) self.result_timing.append(time.time() - qitem.start) except Exception as ex: # pylint: disable=broad-except src = [self.ds.get_item_loc(i) for i in qitem.content_id] - log.error("thread: failed on contentid=%s, %s", src, ex) + log.error("[Line 262] thread: failed on contentid=%s, %s", src, ex) # since post_process will not run, fake empty responses processed_results = [[]] * len(qitem.query_id) finally: @@ -256,6 +273,7 @@ def run_one_item(self, qitem: Item): for idx, query_id in enumerate(qitem.query_id): response_array = array.array( "B", np.array(processed_results[idx], np.uint8).tobytes() + # "B", np.array(processed_results[idx], np.uint64).tobytes() ) response_array_refs.append(response_array) bi = response_array.buffer_info() @@ -271,9 +289,9 @@ def enqueue(self, query_samples): else: bs = self.max_batchsize for i in range(0, len(idx), bs): - data, label = self.ds.get_samples(idx[i: i + bs]) + data, label = self.ds.get_samples(idx[i : i + bs]) self.run_one_item( - Item(query_id[i: i + bs], idx[i: i + bs], data, label) + Item(query_id[i : i + bs], idx[i : i + bs], data, label) ) def finish(self): @@ -288,9 +306,7 @@ def __init__(self, model, ds, threads, post_proc=None, max_batchsize=128): self.result_dict = {} for _ in range(self.threads): - worker = threading.Thread( - target=self.handle_tasks, args=( - self.tasks,)) + worker = threading.Thread(target=self.handle_tasks, args=(self.tasks,)) worker.daemon = True self.workers.append(worker) worker.start() @@ -298,13 +314,17 @@ def __init__(self, model, ds, threads, post_proc=None, max_batchsize=128): def handle_tasks(self, tasks_queue): """Worker thread.""" while True: + # log.info ('getting tasks') qitem = tasks_queue.get() + # log.info ('getten tasks') if qitem is None: # None in the queue indicates the parent want us to exit tasks_queue.task_done() break self.run_one_item(qitem) + # log.info ('going to task_done') tasks_queue.task_done() + # log.info ('tasks done') def enqueue(self, query_samples): idx = [q.index for q in query_samples] @@ -333,13 +353,36 @@ def main(): log.info(args) # find backend - backend = get_backend( - args.backend, - precision=args.dtype, - device=args.device, - model_path=args.model_path, - batch_size=args.max_batchsize, - ) + + # backend = get_backend( + # args.backend, + # precision=args.dtype, + # device=args.device, + # model_path=args.model_path, + # batch_size=args.max_batchsize + # ) + # Zixian: Oct 21: create a list of backends for multi-gpu + # backends = [get_backend( + # args.backend, + # precision=args.dtype, + # device=f'cuda:{i}', + # model_path=args.model_path, + # batch_size=args.max_batchsize + # ) + # for i in [0, 1, 2, 3]] + + backends = [get_backend( + args.backend, + precision=args.dtype, + device=f'cuda:{int (i/int (args.gpu_num / 4))}', + model_path=args.model_path, + batch_size=args.max_batchsize + ) + for i in np.arange (args.gpu_num)] + + log.info(f"Zixian: Returned from get_backends") + + if args.dtype == "fp16": dtype = torch.float16 elif args.dtype == "bf16": @@ -355,7 +398,11 @@ def main(): count_override = True # load model to backend - model = backend.load() + # model = backend.load() + log.info(f"Zixian: entering backend.load") + # Zixian: Oct 21: create a list of models corresponding to each backend + models = [backend.load() for backend in backends] + log.info(f"Zixian: loaded models from all backend") # dataset to use dataset_class, pre_proc, post_proc, kwargs = SUPPORTED_DATASETS[args.dataset] @@ -365,28 +412,38 @@ def main(): pre_process=pre_proc, count=count, threads=args.threads, - pipe_tokenizer=model.pipe.tokenizer, - pipe_tokenizer_2=model.pipe.tokenizer_2, + # pipe_tokenizer=model.pipe.tokenizer, + # pipe_tokenizer_2=model.pipe.tokenizer_2, + pipe_tokenizer=models[0].pipe.tokenizer, + pipe_tokenizer_2=models[0].pipe.tokenizer_2, latent_dtype=dtype, latent_device=args.device, latent_framework=args.latent_framework, + pipe_type=args.backend, **kwargs, ) final_results = { - "runtime": model.name(), - "version": model.version(), + # "runtime": model.name(), + # "version": model.version(), + "runtime": models[0].name(), + "version": models[0].version(), "time": int(time.time()), "args": vars(args), "cmdline": str(args), } + mlperf_conf = os.path.abspath(args.mlperf_conf) + if not os.path.exists(mlperf_conf): + log.error("{} not found".format(mlperf_conf)) + sys.exit(1) + user_conf = os.path.abspath(args.user_conf) if not os.path.exists(user_conf): log.error("{} not found".format(user_conf)) sys.exit(1) audit_config = os.path.abspath(args.audit_conf) - + if args.accuracy: ids_path = os.path.abspath(args.ids_path) with open(ids_path) as f: @@ -405,16 +462,30 @@ def main(): # warmup syntetic_str = "Lorem ipsum dolor sit amet, consectetur adipiscing elit" latents_pt = torch.rand(ds.latents.shape, dtype=dtype).to(args.device) - warmup_samples = [ - { - "input_tokens": ds.preprocess(syntetic_str, model.pipe.tokenizer), - "input_tokens_2": ds.preprocess(syntetic_str, model.pipe.tokenizer_2), - "latents": latents_pt, - } - for _ in range(args.max_batchsize) - ] - for i in range(5): - _ = backend.predict(warmup_samples) + # warmup_samples = [ + # { + # "input_tokens": ds.preprocess(syntetic_str, model.pipe.tokenizer), + # "input_tokens_2": ds.preprocess(syntetic_str, model.pipe.tokenizer_2), + # "latents": latents_pt, + # } + # for _ in range(args.max_batchsize) + # ] + warmup_samples_gpus = [ + [ + { + "input_tokens": ds.preprocess(syntetic_str, model.pipe.tokenizer), + "input_tokens_2": ds.preprocess(syntetic_str, model.pipe.tokenizer_2), + "caption": syntetic_str, + "latents": latents_pt + } + for _ in range(int(args.max_batchsize)) + ] + for model in models] + + # Zixian: Oct 21: warm up each backend + for idx, backend in enumerate (backends): + for i in range(1): + _ = backend.predict(warmup_samples_gpus[idx]) scenario = SCENARIO_MAP[args.scenario] runner_map = { @@ -423,12 +494,52 @@ def main(): lg.TestScenario.Server: QueueRunner, lg.TestScenario.Offline: QueueRunner, } - runner = runner_map[scenario]( - model, ds, args.threads, post_proc=post_proc, max_batchsize=args.max_batchsize - ) - + + # Zixian: Oct 21: create a list of runner + # runner = runner_map[scenario]( + # model, ds, args.threads, post_proc=post_proc, max_batchsize=args.max_batchsize + # ) + runners = [runner_map[scenario]( + model, ds, args.threads, post_proc=post_proc, max_batchsize=args.max_batchsize + ) + for model in models] + + # def issue_queries(query_samples): + # runner.enqueue(query_samples) def issue_queries(query_samples): - runner.enqueue(query_samples) + print (f'\n\n len (query_samples): {len (query_samples)} \n\n') + + query_samples_len = len (query_samples) + query_samples_seg_len = query_samples_len / len (runners) + splitted_query_samples = [] + for idx in range (len (runners)): + log.info (f'\n\n\n') + log.info (f'idx: {idx}') + log.info (f'query_samples_len: {query_samples_len}') + log.info (f'idx: {idx}') + # if idx == len (runners) -1: + # splitted_query_samples.append (query_samples[idx*query_samples_seg_len:]) + # else: + # splitted_query_samples.append (query_samples[idx*query_samples_seg_len : (idx+1)*query_samples_seg_len]) + + splitted_query_samples.append (query_samples [int(round(query_samples_seg_len * idx)): int(round(query_samples_seg_len * (idx + 1)))]) + + + with ThreadPoolExecutor(max_workers=len(runners)) as executor: + # Map each runner to its respective sublist + futures = { + executor.submit(runner.enqueue, queries): runner + for runner, queries in zip(runners, splitted_query_samples) + } + + # Optionally process the results + for future in as_completed(futures): + runner = futures[future] + try: + result = future.result() + print(f'Runner {runner} enqueued successfully.') + except Exception as exc: + print(f'Runner {runner} generated an exception: {exc}') def flush_queries(): pass @@ -441,8 +552,7 @@ def flush_queries(): log_settings.log_output = log_output_settings settings = lg.TestSettings() - # mlperf.conf is automatically loaded by the loadgen - # settings.FromConfig(mlperf_conf, args.model_name, args.scenario) + settings.FromConfig(mlperf_conf, args.model_name, args.scenario) settings.FromConfig(user_conf, args.model_name, args.scenario) if os.path.exists(audit_config): settings.FromConfig(audit_config, args.model_name, args.scenario) @@ -471,8 +581,7 @@ def flush_queries(): settings.multi_stream_samples_per_query = args.samples_per_query if args.max_latency: settings.server_target_latency_ns = int(args.max_latency * NANO_SEC) - settings.multi_stream_expected_latency_ns = int( - args.max_latency * NANO_SEC) + settings.multi_stream_expected_latency_ns = int(args.max_latency * NANO_SEC) performance_sample_count = ( args.performance_sample_count @@ -480,22 +589,47 @@ def flush_queries(): else min(count, 500) ) sut = lg.ConstructSUT(issue_queries, flush_queries) + #! [Yalu Ouyang] count here affects how many items to run (even for accuracy) qsl = lg.ConstructQSL( count, performance_sample_count, ds.load_query_samples, ds.unload_query_samples ) log.info("starting {}".format(scenario)) result_dict = {"scenario": str(scenario)} - runner.start_run(result_dict, args.accuracy) + for runner in runners: + runner.start_run(result_dict, args.accuracy) + + # with ThreadPoolExecutor(max_workers=len(runners)) as executor: + # # Map each runner to its respective sublist + # futures = { + # executor.submit(runner.finish(), (result_dict, args.accuracy)): runner + # for runner in runners + # } + lg.StartTestWithLogSettings(sut, qsl, settings, log_settings, audit_config) + + log.info("Loadgen finished tests") if args.accuracy: post_proc.finalize(result_dict, ds, output_dir=args.output) final_results["accuracy_results"] = result_dict post_proc.save_images(saved_images_ids, ds) - runner.finish() + log.info("After processing accuracy") + + for runner in runners: + runner.finish() + + log.info("After runner.finish()") + # with ThreadPoolExecutor(max_workers=len(runners)) as executor: + # # Map each runner to its respective sublist + # futures = { + # executor.submit(runner.finish()): runner + # for runner in runners + # } + + lg.DestroyQSL(qsl) lg.DestroySUT(sut) diff --git a/text_to_image/mlperf.conf b/text_to_image/mlperf.conf new file mode 100644 index 0000000000..0cea5351e1 --- /dev/null +++ b/text_to_image/mlperf.conf @@ -0,0 +1,99 @@ +# The format of this config file is 'key = value'. +# The key has the format 'model.scenario.key'. Value is mostly int64_t. +# Model maybe '*' as wildcard. In that case the value applies to all models. +# All times are in milli seconds + +# Set performance_sample_count for each model. +# User can optionally set this to higher values in user.conf. +resnet50.*.performance_sample_count_override = 1024 +ssd-mobilenet.*.performance_sample_count_override = 256 +retinanet.*.performance_sample_count_override = 64 +bert.*.performance_sample_count_override = 10833 +dlrm.*.performance_sample_count_override = 204800 +dlrm-v2.*.performance_sample_count_override = 204800 +rnnt.*.performance_sample_count_override = 2513 +gptj.*.performance_sample_count_override = 13368 +llama2-70b.*.performance_sample_count_override = 24576 +stable-diffusion-xl.*.performance_sample_count_override = 5000 +# set to 0 to let entire sample set to be performance sample +3d-unet.*.performance_sample_count_override = 0 + +# Set seeds. The seeds will be distributed two weeks before the submission. +*.*.qsl_rng_seed = 3066443479025735752 +*.*.sample_index_rng_seed = 10688027786191513374 +*.*.schedule_rng_seed = 14962580496156340209 +# Set seeds for TEST_05. The seeds will be distributed two weeks before the submission. +*.*.test05_qsl_rng_seed = 16799458546791641818 +*.*.test05_sample_index_rng_seed = 5453809927556429288 +*.*.test05_schedule_rng_seed = 5435552105434836064 + + +*.SingleStream.target_latency_percentile = 90 +*.SingleStream.min_duration = 600000 + +*.MultiStream.target_latency_percentile = 99 +*.MultiStream.samples_per_query = 8 +*.MultiStream.min_duration = 600000 +*.MultiStream.min_query_count = 662 +retinanet.MultiStream.target_latency = 528 + +# 3D-UNet uses equal issue mode because it has non-uniform inputs +3d-unet.*.sample_concatenate_permutation = 1 + +# LLM benchmarks have non-uniform inputs and outputs, and use equal issue mode for all latency scenario +gptj.*.sample_concatenate_permutation = 1 +llama2-70b.*.sample_concatenate_permutation = 1 +mixtral-8x7b.*.sample_concatenate_permutation = 1 + +*.Server.target_latency = 10 +*.Server.target_latency_percentile = 99 +*.Server.target_duration = 0 +*.Server.min_duration = 600000 +resnet50.Server.target_latency = 15 +retinanet.Server.target_latency = 100 +bert.Server.target_latency = 130 +dlrm.Server.target_latency = 60 +dlrm-v2.Server.target_latency = 60 +rnnt.Server.target_latency = 1000 +gptj.Server.target_latency = 20000 +stable-diffusion-xl.Server.target_latency = 20000 +# Llama2-70b benchmarks measures token latencies +llama2-70b.*.use_token_latencies = 1 +mixtral-8x7b.*.use_token_latencies = 1 +# gptj benchmark infers token latencies +gptj.*.infer_token_latencies = 1 +gptj.*.token_latency_scaling_factor = 69 +# Only ttft and tpot are tracked for the llama2-70b & mixtral-8x7B benchmark therefore target_latency = 0 +llama2-70b.Server.target_latency = 0 +llama2-70b.Server.ttft_latency = 2000 +llama2-70b.Server.tpot_latency = 200 + +mixtral-8x7b.Server.target_latency = 0 +mixtral-8x7b.Server.ttft_latency = 2000 +mixtral-8x7b.Server.tpot_latency = 200 + +*.Offline.target_latency_percentile = 90 +*.Offline.min_duration = 600000 + +# In Offline scenario, we always have one query. But LoadGen maps this to +# min_sample_count internally in Offline scenario. If the dataset size is larger +# than 24576 we limit the min_query_count to 24576 and otherwise we use +# the dataset size as the limit + +resnet50.Offline.min_query_count = 24576 +retinanet.Offline.min_query_count = 24576 +dlrm-v2.Offline.min_query_count = 24576 +bert.Offline.min_query_count = 10833 +gptj.Offline.min_query_count = 13368 +rnnt.Offline.min_query_count = 2513 +3d-unet.Offline.min_query_count = 43 +stable-diffusion-xl.Offline.min_query_count = 4000 +llama2-70b.Offline.min_query_count = 24576 +mixtral-8x7b.Offline.min_query_count = 15000 + +# These fields should be defined and overridden by user.conf. +*.SingleStream.target_latency = 10 +*.MultiStream.target_latency = 80 +*.Server.target_qps = 1.0 +*.Offline.target_qps = 1.0 + diff --git a/tools/submission/preprocess_submission.py b/tools/submission/preprocess_submission.py index a1678c79d0..a737c1f16c 100644 --- a/tools/submission/preprocess_submission.py +++ b/tools/submission/preprocess_submission.py @@ -84,7 +84,6 @@ def delete_empty_dirs(src): """ if not os.path.isdir(src): return False - if all([delete_empty_dirs(os.path.join(src, file)) for file in os.listdir(src)]): log.info("Removing empty dir: (%s)", src) @@ -532,9 +531,11 @@ def main(): if not args.nodelete_empty_dirs: delete_empty_dirs(os.path.join(src_dir)) + run_dir = os.getcwd() os.chdir(src_dir) infer_scenario_results(args, config) + os.chdir(run_dir) if not args.nodelete_empty_dirs: delete_empty_dirs(os.path.join(src_dir)) diff --git a/tools/submission/submission_checker.py b/tools/submission/submission_checker.py index a31a6fede3..deff9eb8c4 100755 --- a/tools/submission/submission_checker.py +++ b/tools/submission/submission_checker.py @@ -188,6 +188,7 @@ "sample_index_rng_seed": 198141574272810017, "schedule_rng_seed": 7575108116881280410, }, + # not required for v5.0+ "test05_seeds": { # TODO: Update random seeds "qsl_rng_seed": 2376919268182438552, @@ -2880,7 +2881,7 @@ def check_compliance_dir( compliance_perf_pass = True compliance_perf_dir_pass = True compliance_acc_pass = True - test_list = ["TEST01", "TEST04", "TEST05"] + test_list = ["TEST01", "TEST04"] if model in [ "bert-99", @@ -2899,15 +2900,15 @@ def check_compliance_dir( ]: test_list.remove("TEST04") - if model in [ + if config.version in ["v4.0", "v4.1"] and model not in [ "gptj-99", "gptj-99.9", "llama2-70b-99", "llama2-70b-99.9", "stable-diffusion-xl", "mixtral-8x7b", - ] or config.version not in ["v4.0", "v4.1"]: - test_list.remove("TEST05") + ]: + test_list.append("TEST05") if model in [ "gptj-99",