Merge pull request #6 from Lightning-AI/make_deepspeed_opt_in

.gitignore

@@ -159,4 +159,5 @@ cython_debug/
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
 outputs/
-profiles/
+profiles/
+v1-5-pruned-emaonly.ckpt

ldm/deepspeed_replace.py

@@ -6,17 +6,30 @@
 from functools import partial
 from dataclasses import dataclass
 import time
-import deepspeed.ops.transformer as transformer_inference
-from deepspeed.ops.transformer.inference.diffusers_attention import DeepSpeedDiffusersAttention
-from deepspeed.ops.transformer.inference.diffusers_transformer_block import DeepSpeedDiffusersTransformerBlock
-from deepspeed.ops.transformer.inference.diffusers_2d_transformer import Diffusers2DTransformerConfig
+from lightning_utilities.core.imports import package_available
+
 from ldm.modules.attention import CrossAttention, BasicTransformerBlock
-from deepspeed.module_inject.replace_policy import UNetPolicy, DSPolicy
 from ldm.models.diffusion.ddpm import DiffusionWrapper
 from ldm.models.autoencoder import AutoencoderKL
 from ldm.modules.encoders.modules import FrozenCLIPEmbedder
-from deepspeed.inference.engine import InferenceEngine
+from ldm.detect_target import _detect_cuda
+import logging
+
+if package_available("deepspeed"):
+    import deepspeed.ops.transformer as transformer_inference
+    from deepspeed.ops.transformer.inference.diffusers_attention import DeepSpeedDiffusersAttention
+    from deepspeed.ops.transformer.inference.diffusers_transformer_block import DeepSpeedDiffusersTransformerBlock
+    from deepspeed.ops.transformer.inference.diffusers_2d_transformer import Diffusers2DTransformerConfig
+    from deepspeed.inference.engine import InferenceEngine
+    from deepspeed.module_inject.replace_policy import UNetPolicy, DSPolicy
+else:
+    class InferenceEngine:
+        pass
+
+    class DSPolicy:
+        pass
 
+logger = logging.getLogger(__name__)
 
 class InferenceEngine(InferenceEngine):
 
@@ -266,7 +279,11 @@ def _module_match(module):
     return None
 
 # Inspired from https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/module_inject/replace_module.py#L201
-def deepspeed_injection(module, fp16=False, enable_cuda_graph=True):
+def deepspeed_injection(module, fp16=True, enable_cuda_graph=True):
+
+    if not torch.cuda.is_available():
+        logger.warn("You provided use_deepspeed=True but Deepspeed isn't supported on your architecture. Skipping...")
+        return
 
     def replace_attn(child, policy):
         policy_attn = policy.attention(child)
@@ -327,13 +344,17 @@ def _replace_module(module, policy):
             for name, child in module.named_children():
                 _replace_module(child, policy)
                 if child.__class__ in new_policies:
-                    replaced_module = new_policies[child.__class__](child,
-                                                                    policy)
+                    replaced_module = new_policies[child.__class__](child, policy)
                     setattr(module, name, replaced_module)
 
-        _replace_module(sub_module, policy)
-        new_module = policy.apply(sub_module,
-                                    enable_cuda_graph=enable_cuda_graph)
+        if not package_available("deepspeed"):
+            logger.warn("You provided use_deepspeed=True but Deepspeed isn't installed. Skipping...")
+        if _detect_cuda() not in ["80"]:
+            logger.warn("You provided use_deepspeed=True but Deepspeed isn't supported on your architecture. Skipping...")
+        else:
+            _replace_module(sub_module, policy)
+
+        new_module = policy.apply(sub_module, enable_cuda_graph=enable_cuda_graph)
         print(f"**** found and replaced {name} w. {type(new_module)}")
         setattr(module, name, new_module)
 

ldm/detect_target.py

@@ -0,0 +1,23 @@
+import os
+from subprocess import PIPE, Popen
+
+# Credits to AITemplate Team
+# https://github.com/facebookincubator/AITemplate/blob/main/python/aitemplate/testing/detect_target.py
+def _detect_cuda():
+    try:
+        proc = Popen(
+            ["nvidia-smi", "--query-gpu=gpu_name", "--format=csv"],
+            stdout=PIPE,
+            stderr=PIPE,
+        )
+        stdout, _ = proc.communicate()
+        stdout = stdout.decode("utf-8")
+        if "A100" in stdout or "RTX 30" in stdout or "A30" in stdout:
+            return "80"
+        if "V100" in stdout:
+            return "70"
+        if "T4" in stdout:
+            return "75"
+        return None
+    except Exception:
+        return None

ldm/lightning.py

@@ -16,6 +16,9 @@
 from contextlib import nullcontext
 from torch import autocast
 from ldm.deepspeed_replace import deepspeed_injection, ReplayCudaGraphUnet
+import logging
+
+logger = logging.getLogger(__name__)
 
 
 class PromptDataset(Dataset):
@@ -47,28 +50,30 @@ def __init__(
         self,
         config_path: str,
         checkpoint_path: str,
-        device: torch.device,
+        device: str,
         size: int = 512,
         fp16: bool = True,
         sampler: str = "ddim",
         steps: Optional[int] = None,
-        use_deepspeed: bool = True,
+        use_deepspeed: bool = False,
         enable_cuda_graph: bool = False,
+        use_inference_context: bool = False,
     ):
         super().__init__()
 
+        if device in ("mps", "cpu") and fp16:
+            logger.warn(f"You provided fp16=True but it isn't supported on `{device}`. Skipping...")
+            fp16 = False
+
         config = OmegaConf.load(f"{config_path}")
         config.model.params.unet_config["params"]["use_fp16"] = False
         config.model.params.cond_stage_config["params"] = {"device": device}
-
         checkpoint = torch.load(checkpoint_path, map_location="cpu")
         state_dict = checkpoint["state_dict"]
         self.model = instantiate_from_config(config.model)
         self.model.load_state_dict(state_dict, strict=False)
 
-        self.to(dtype=torch.float16)
-
-        if use_deepspeed:
+        if use_deepspeed or enable_cuda_graph:
             deepspeed_injection(self.model, fp16=fp16, enable_cuda_graph=enable_cuda_graph)
 
         # Replace with 
@@ -79,12 +84,16 @@ def __init__(
 
         self.to(device, dtype=torch.float16 if fp16 else torch.float32)
         self.fp16 = fp16
+        self.use_inference_context = use_inference_context
 
-    def predict_step(self, prompts: List[str], batch_idx: int):
+    def predict_step(self, prompts: Union[List[str], str], batch_idx: int = 0):
+        if isinstance(prompts, str):
+            prompts = [prompts]
         batch_size = len(prompts)
 
         precision_scope = autocast if self.fp16 else nullcontext
         inference = torch.inference_mode if torch.cuda.is_available() else torch.no_grad
+        inference = inference if self.use_inference_context else nullcontext
         with inference():
             with precision_scope("cuda"):
                 with self.model.ema_scope():

requirements.txt

@@ -9,5 +9,5 @@ einops==0.3.0
 transformers>4.19.2
 open-clip-torch==2.7.0
 lightning
-triton>=2.0.0.dev20221005
-deepspeed>=0.7.5
+triton>=2.0.0.dev20221005; platform_system == "Linux"
+deepspeed>=0.7.5; platform_system == "Linux"

scripts/txt2img_lightning.py

@@ -1,9 +1,11 @@
 import argparse
 import os
+import time
+import torch
 from pytorch_lightning import seed_everything
 from ldm.lightning import LightningStableDiffusion
 
-def benchmark_fn(iters: int, warm_up_iters: int, function, *args, **kwargs) -> float:
+def benchmark_fn(device, iters: int, warm_up_iters: int, function, *args, **kwargs) -> float:
     """
     Function for benchmarking a pytorch function.
 
@@ -29,21 +31,29 @@ def benchmark_fn(iters: int, warm_up_iters: int, function, *args, **kwargs) -> f
         function(*args, **kwargs)
 
     # Start benchmark.
-    torch.cuda.synchronize()
-    torch.cuda.empty_cache()
-    torch.cuda.reset_accumulated_memory_stats()
-    torch.cuda.reset_peak_memory_stats()
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-    start_event.record()
-    torch.cuda.reset_peak_memory_stats()
+    if device == "cuda":
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        torch.cuda.reset_accumulated_memory_stats()
+        torch.cuda.reset_peak_memory_stats()
+        start_event = torch.cuda.Event(enable_timing=True)
+        end_event = torch.cuda.Event(enable_timing=True)
+        start_event.record()
+        torch.cuda.reset_peak_memory_stats()
+    else:
+        t0 = time.time()
+
     for _ in range(iters):
         results.extend(function(*args, **kwargs))
-    max_memory = torch.cuda.max_memory_allocated(0)/2**20
-    end_event.record()
-    torch.cuda.synchronize()
-    # in ms
-    return (start_event.elapsed_time(end_event)) / iters, max_memory, results
+
+    if device == "cuda":
+        max_memory = torch.cuda.max_memory_allocated(0)/2**20
+        end_event.record()
+        torch.cuda.synchronize()
+        # in ms
+        return (start_event.elapsed_time(end_event)) / iters, max_memory, results
+    else:
+        return (time.time() - t0) / iters, None, results
 
 def parse_args():
     parser = argparse.ArgumentParser()
@@ -151,16 +161,24 @@ def main(opt):
     os.makedirs(opt.outdir, exist_ok=True)
     seed_everything(opt.seed)
 
+    device = "cuda" if torch.cuda.is_available() else "mps"
+
     model = LightningStableDiffusion(
         config_path=opt.config,
         checkpoint_path=opt.ckpt,
-        device="cuda",
-        sampler=opt.sampler,
+        device=device,
+        fp16=True, # Supported on GPU and CPU only, skipped otherwise.
+        use_deepspeed=True, # Supported on Ampere and RTX, skipped otherwise.
+        enable_cuda_graph=True, # Currently enabled only for batch size 1.
+        use_inference_context=False,
         steps=30,
     )
 
     for batch_size in [1, 2, 4]:
-        t, max_memory, images = benchmark_fn(10, 5, model.predict_step, prompts=[opt.prompt] * batch_size, batch_idx=0)
+        if batch_size == 1:
+            t, max_memory, images = benchmark_fn(device, 10, 5, model.predict_step, prompts=opt.prompt, batch_idx=0)
+        else:
+            t, max_memory, images = benchmark_fn(device, 10, 5, model.predict_step, prompts=[opt.prompt] * batch_size, batch_idx=0)
         print(f"Average time {t} secs on batch size {batch_size}.")
         print(f"Max GPU Memory cost is {max_memory} MB.")