Add CPU support and update README (#119)

Author: yanbing-j

README.md

@@ -17,6 +17,10 @@ For example:
 ```
 pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121
 ```
+or
+```
+pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
+```
 
 Installation instructions vary by platform. Please see the website https://pytorch.org/
 

experiments/README.md

@@ -37,6 +37,7 @@ These  experiments were run on an Amazon p4d.24xlarge instance. See the Product
 - 1152 GiB of RAM
 - Software
 
+Meanwhile, these experiments (fp32, bf16, compile, SDPA, Triton, NT) can run on CPU platform as well. Experiment results will be shown in the near future.
 
 ### Versions
 
@@ -47,11 +48,17 @@ These  experiments were run on an Amazon p4d.24xlarge instance. See the Product
 ### Installation instructions
 
 ```
-$ conda create -n nightly20231117py310
-$ conda activate nightly20231117py310
+$ conda create -n nightlypy310
+$ conda activate nightlypy310
 $ conda install python=3.10
-$ pip install https://download.pytorch.org/whl/nightly/cu121/torch-2.2.0.dev20231117%2Bcu121-cp310-cp310-linux_x86_64.whl
-$ pip install https://download.pytorch.org/whl/nightly/cu121/torchvision-0.17.0.dev20231117%2Bcu121-cp310-cp310-linux_x86_64.whl
+For GPU,
+- $ pip install https://download.pytorch.org/whl/nightly/cu121/torch-2.2.0.dev20231117%2Bcu121-cp310-cp310-linux_x86_64.whl
+- $ pip install https://download.pytorch.org/whl/nightly/cu121/torchvision-0.17.0.dev20231117%2Bcu121-cp310-cp310-linux_x86_64.whl
+For CPU,
+- $ pip install https://download.pytorch.org/whl/nightly/cpu/torch-2.4.0.dev20240509%2Bcpu-cp310-cp310-linux_x86_64.whl
+- $ pip install https://download.pytorch.org/whl/nightly/cpu/torchvision-0.19.0.dev20240509%2Bcpu-cp310-cp310-linux_x86_64.whl
+- $ pip install triton
+
 $ git clone https://github.com/cpuhrsch/segment-anything.git
 $ cd segment-anything
 $ pip install -e .
@@ -66,10 +73,16 @@ If you plan to run the scripts that run the experiments from segment-anything-fa
 
 ### How to run experiments
 
+For GPU platform,
 ```
 $ python run_experiments.py 16 vit_b <pytorch_github> <segment-anything_github> <path_to_experiments_data> --run-experiments --num-workers 32
 ```
 
+For CPU platform, set SEGMENT_ANYTHING_FAST_USE_FLASH_4 as 0, since Custom flash attention kernels were written specifically for A100.
+```
+$ SEGMENT_ANYTHING_FAST_USE_FLASH_4=0 python run_experiments.py 16 vit_b <pytorch_github> <segment-anything_github> <path_to_experiments_data> --run-experiments --num-workers 32 --device cpu
+```
+
 If at any point you run into issue, please note that you can increase verbosity by adding `--capture_output False` to above command. Also, please don't hesitate to open an issue.
 
 

experiments/eval_combo.py

@@ -5,6 +5,7 @@
 from data import build_data, setup_coco_img_ids
 import math
 import segment_anything_fast
+import time
 
 torch._dynamo.config.cache_size_limit = 50000
 
@@ -64,10 +65,13 @@ def build_results_batch_nested(predictor, batch, batch_size, pad_input_image_bat
     # We explicitly exclude data transfers from the timing to focus
     # only on the kernel performance.
     # Next we synchronize and set two events to start timing.
-    torch.cuda.synchronize()
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-    start_event.record()
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+        start_event = torch.cuda.Event(enable_timing=True)
+        end_event = torch.cuda.Event(enable_timing=True)
+        start_event.record()
+    else:
+        t0 = time.time()
 
     with torch.autograd.profiler.record_function("timed region"):
         with torch.autograd.profiler.record_function("image encoder"):
@@ -93,9 +97,12 @@ def build_results_batch_nested(predictor, batch, batch_size, pad_input_image_bat
         # the amount of time spent on the GPU. This is a fairly tight measurement
         # around the launched GPU kernels and excludes data movement from host
         # to device.
-        end_event.record()
-        torch.cuda.synchronize()
-        elapsed_time = start_event.elapsed_time(end_event)
+        if torch.cuda.is_available():
+            end_event.record()
+            torch.cuda.synchronize()
+            elapsed_time = start_event.elapsed_time(end_event)
+        else:
+            elapsed_time = time.time() - t0
     return sum(result_batch, []), orig_input_image_batch_size, elapsed_time
 
 def build_results_batch(predictor, batch, batch_size, pad_input_image_batch):
@@ -123,10 +130,13 @@ def build_results_batch(predictor, batch, batch_size, pad_input_image_batch):
     # We explicitly exclude data transfers from the timing to focus
     # only on the kernel performance.
     # Next we synchronize and set two events to start timing.
-    torch.cuda.synchronize()
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-    start_event.record()
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+        start_event = torch.cuda.Event(enable_timing=True)
+        end_event = torch.cuda.Event(enable_timing=True)
+        start_event.record()
+    else:
+        t0 = time.time()
 
     with torch.autograd.profiler.record_function("timed region"):
         with torch.autograd.profiler.record_function("image encoder"):
@@ -157,9 +167,12 @@ def build_results_batch(predictor, batch, batch_size, pad_input_image_batch):
         # the amount of time spent on the GPU. This is a fairly tight measurement
         # around the launched GPU kernels and excludes data movement from host
         # to device.
-        end_event.record()
-        torch.cuda.synchronize()
-        elapsed_time = start_event.elapsed_time(end_event)
+        if torch.cuda.is_available():
+            end_event.record()
+            torch.cuda.synchronize()
+            elapsed_time = start_event.elapsed_time(end_event)
+        else:
+            elapsed_time = time.time() - t0
     return result_batch, orig_input_image_batch_size, elapsed_time
 
 
@@ -290,6 +303,7 @@ def run(
     memory_path=None,
     use_local_sam_fork=False,
     use_compiler_settings=False,
+    device="cuda"
 ):
     from torch._inductor import config as inductorconfig
     inductorconfig.triton.unique_kernel_names = True
@@ -327,7 +341,7 @@ def run(
     else:
         from segment_anything import sam_model_registry, SamPredictor
     checkpoint_path = model_type_to_checkpoint[sam_model_type]
-    sam = sam_model_registry[sam_model_type](checkpoint=checkpoint_path).cuda()
+    sam = sam_model_registry[sam_model_type](checkpoint=checkpoint_path).to(torch.device(device))
     predictor = SamPredictor(sam)
 
     from segment_anything_fast import tools

experiments/run_experiments.py

@@ -45,7 +45,8 @@ def run_experiment(experiments_data,
                    limit=None,
                    profile_path=None,
                    profile_top=False,
-                   memory_path=None):
+                   memory_path=None,
+                   device="cuda"):
     root_cmd = ["python", "eval_combo.py",
                 "--coco_root_dir",
                 f"{experiments_data}/datasets/coco2017",
@@ -84,6 +85,7 @@ def run_experiment(experiments_data,
         args = args + ["--memory-path", memory_path]
     if extra_args is None:
         extra_args = []
+    args = args + ["--device", device]
     args = args + extra_args
     if print_header:
         args = args + ["--print_header", "True"]
@@ -145,7 +147,8 @@ def run(batch_size,
         num_workers=32,
         print_header=True,
         capture_output=True,
-        local_fork_only=False):
+        local_fork_only=False,
+        device="cuda"):
 
     assert model == "vit_b" or model == "vit_h"
 
@@ -155,7 +158,8 @@ def run(batch_size,
                              model,
                              batch_size=batch_size,
                              num_workers=num_workers,
-                             capture_output=capture_output)
+                             capture_output=capture_output,
+                             device=device)
 
     print_header = True
     if run_traces: