Quantize vit_b_16 tutorial - Part 1 (pytorch#60)

Author: cpuhrsch

torchao/__init__.py

@@ -1,5 +1,8 @@
 from . import dtypes
+from .quantization.quant_api import apply_dynamic_quant
+from .quantization.quant_api import apply_weight_only_int8_quant
 
 __all__ = [
-        "dtypes"
+        "dtypes",
+        "apply_dynamic_quant",
 ]

torchao/quantization/quant_api.py

@@ -126,14 +126,10 @@ def apply_weight_only_int8_quant(model, filter_fn=None):
 def apply_dynamic_quant(model, filter_fn=None):
     """
     Applies dynamic symmetric per-token activation and per-channel weight
-    quantization to all linear layers in the given model using
-    module swaps.
+    quantization to all linear layers by converting all linear weight
+    tensors to the `Int8DynamicallyQuantizedLinearWeight` Tensor subclass.
     """
-    _replace_with_custom_fn_if_matches_filter(
-        model,
-        lambda mod: DynamicallyPerAxisQuantizedLinear.from_float(mod),
-        _is_linear if filter_fn is None else filter_fn,
-    )
+    change_linear_weights_to_int8_dqtensors(model, filter_fn)
 
 
 def _get_subclass_inserter(cls, **kwargs):

tutorials/quantize_vit/bfloat16.json.gz

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+# Run bfloat16 version
+TORCH_LOGS='graph_breaks,recompiles' python run_vit_b.py
+
+# Run dynamic quantized version
+TORCH_LOGS='graph_breaks,recompiles' python run_vit_b_quant.py
+
+# Store the output code for further inspection
+echo "bfloat16 generated code lives in:"
+TORCH_LOGS='output_code' python run_vit_b.py 2>&1 | grep "Output code written to: " | awk -F" " '{print $NF}'
+echo "quantization generated code lives in:"
+TORCH_LOGS='output_code' python run_vit_b_quant.py 2>&1 | grep "Output code written to: " | awk -F" " '{print $NF}'

tutorials/quantize_vit/bfloat16_code.py

@@ -0,0 +1,46 @@
+import torch
+import torchvision.models.vision_transformer as models
+
+# Load Vision Transformer model
+model = models.vit_b_16(pretrained=True)
+
+# Set the model to evaluation mode
+model.eval().cuda().to(torch.bfloat16)
+
+# Input tensor (batch_size, channels, height, width)
+input_tensor = torch.randn(1, 3, 224, 224, dtype=torch.bfloat16, device='cuda')
+
+model = torch.compile(model, mode='max-autotune')
+
+def benchmark_model(model, num_runs, input_tensor):
+    torch.cuda.synchronize()
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+    start_event.record()
+    
+    # benchmark
+    for _ in range(num_runs):
+        with torch.autograd.profiler.record_function("timed region"):
+            model(input_tensor)
+    
+    end_event.record()
+    torch.cuda.synchronize()
+    return start_event.elapsed_time(end_event) / num_runs
+
+def profiler_runner(path, fn, *args, **kwargs):
+    with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CPU,
+                        torch.profiler.ProfilerActivity.CUDA],
+            record_shapes=True) as prof:
+        result = fn(*args, **kwargs)
+    prof.export_chrome_trace(path)
+    return result
+
+# Must run with no_grad when optimizing for inference
+with torch.no_grad():
+    # warmup
+    benchmark_model(model, 5, input_tensor)
+    # benchmark
+    print("elapsed_time: ", benchmark_model(model, 100, input_tensor), " milliseconds")
+    # Create a trace
+    profiler_runner("bfloat16.json.gz", benchmark_model, model, 5, input_tensor)

tutorials/quantize_vit/quant.json.gz

@@ -0,0 +1,53 @@
+import torch
+import torchao
+import torchvision.models.vision_transformer as models
+
+# Load Vision Transformer model
+model = models.vit_b_16(pretrained=True)
+
+# Set the model to evaluation mode
+model.eval().cuda().to(torch.bfloat16)
+
+# Input tensor (batch_size, channels, height, width)
+input_tensor = torch.randn(1, 3, 224, 224, dtype=torch.bfloat16, device='cuda')
+
+## Quantization code - start
+torchao.apply_dynamic_quant(model)
+from torch._inductor import config as inductorconfig
+inductorconfig.force_fuse_int_mm_with_mul = True
+## Quantization code - end
+
+model = torch.compile(model, mode='max-autotune')
+
+def benchmark_model(model, num_runs, input_tensor):
+    torch.cuda.synchronize()
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+    start_event.record()
+    
+    # benchmark
+    for _ in range(num_runs):
+        with torch.autograd.profiler.record_function("timed region"):
+            model(input_tensor)
+    
+    end_event.record()
+    torch.cuda.synchronize()
+    return start_event.elapsed_time(end_event) / num_runs
+
+def profiler_runner(path, fn, *args, **kwargs):
+    with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CPU,
+                        torch.profiler.ProfilerActivity.CUDA],
+            record_shapes=True) as prof:
+        result = fn(*args, **kwargs)
+    prof.export_chrome_trace(path)
+    return result
+
+# Must run with no_grad when optimizing for inference
+with torch.no_grad():
+    # warmup
+    benchmark_model(model, 5, input_tensor)
+    # benchmark
+    print("elapsed_time: ", benchmark_model(model, 100, input_tensor), " milliseconds")
+    # Create a trace
+    profiler_runner("quant.json.gz", benchmark_model, model, 5, input_tensor)