Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/162972
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit cdff0ba with merge base 5dc4e78 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Member
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Member
Author
|
@pytorchbot merge |
Collaborator
|
The merge job was canceled or timed out. This most often happen if two merge requests were issued for the same PR, or if merge job was waiting for more than 6 hours for tests to finish. In later case, please do not hesitate to reissue the merge command |
Member
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
janeyx99
approved these changes
Sep 15, 2025
markc-614
pushed a commit
to markc-614/pytorch
that referenced
this pull request
Sep 17, 2025
Enabling automatic pre compiled headers per https://docs.nvidia.com/cuda/nvrtc/index.html#example-automatic-pch-cuda-12-8 I'm seeing large speedups in compilation times using PCH on average but the max compilation time with PCH is worst which is why I can't enable it by default. `load_inline()` also supports precompiled headers and does not enable them by default ``` Without PCH: 270.58 ms average With PCH: 115.27 ms average ``` ``` Without PCH: Max: 337.99 ms With PCH: Max: 383.82 ms ``` ```python source) [marksaroufim@devgpu005]~/pytorch% python simple_pch_benchmark.py ============================================================ Simple PCH Compilation Benchmark ============================================================ Device: NVIDIA B200 Iterations: 100 Testing WITHOUT PCH: ------------------------------ Compiling kernel 100 times WITHOUT PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 270.58 ms (±6.99 ms) Min: 264.09 ms Max: 337.99 ms Testing WITH PCH: ------------------------------ Compiling kernel 100 times WITH PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 115.27 ms (±27.32 ms) Min: 110.65 ms Max: 383.82 ms ``` ## Benchmarking script ```python #!/usr/bin/env python3 import argparse import os import sys import time from statistics import mean, stdev import torch from torch.cuda._utils import _nvrtc_compile def benchmark_compilation(use_pch, iterations=100): """Compile the same kernel many times with or without PCH.""" # CUB kernel that benefits from PCH kernel_source = """ #include <cub/block/block_reduce.cuh> #include <cub/block/block_scan.cuh> #include <cub/warp/warp_reduce.cuh> extern "C" __global__ void test_kernel(const float* input, float* output, int n) { using BlockReduce = cub::BlockReduce<float, 256>; using BlockScan = cub::BlockScan<float, 256>; using WarpReduce = cub::WarpReduce<float>; __shared__ union { typename BlockReduce::TempStorage reduce; typename BlockScan::TempStorage scan; typename WarpReduce::TempStorage warp[8]; } temp_storage; int idx = blockIdx.x * blockDim.x + threadIdx.x; float val = (idx < n) ? input[idx] : 0.0f; float sum = BlockReduce(temp_storage.reduce).Sum(val); __syncthreads(); float scan_result; BlockScan(temp_storage.scan).ExclusiveSum(val, scan_result); __syncthreads(); int warp_id = threadIdx.x / 32; float warp_sum = WarpReduce(temp_storage.warp[warp_id]).Sum(val); if (threadIdx.x == 0) { output[blockIdx.x] = sum + scan_result + warp_sum; } } """ device = torch.cuda.current_device() major, minor = torch.cuda.get_device_capability(device) compute_capability = f"{major}{minor}" compile_times = [] print( f"Compiling kernel {iterations} times {'WITH' if use_pch else 'WITHOUT'} PCH..." ) for i in range(iterations): # Use unique kernel name to avoid caching between iterations kernel_name = f"test_kernel_{i}" unique_source = kernel_source.replace("test_kernel", kernel_name) start = time.perf_counter() ptx, mangled_name = _nvrtc_compile( unique_source, kernel_name, compute_capability, header_code="", nvcc_options=["-std=c++17"], auto_pch=use_pch, ) elapsed = time.perf_counter() - start compile_times.append(elapsed * 1000) # Convert to ms # Progress indicator if (i + 1) % 10 == 0: print(f" Completed {i + 1}/{iterations} compilations") return compile_times def main(): parser = argparse.ArgumentParser(description="Simple PCH Compilation Benchmark") parser.add_argument("--pch", action="store_true", help="Test with PCH only") parser.add_argument("--no-pch", action="store_true", help="Test without PCH only") parser.add_argument( "--iterations", type=int, default=100, help="Number of compilations" ) args = parser.parse_args() print("=" * 60) print("Simple PCH Compilation Benchmark") print("=" * 60) print(f"Device: {torch.cuda.get_device_name()}") print(f"Iterations: {args.iterations}") print() # Determine what to test test_both = not args.pch and not args.no_pch results = {} # Test without PCH if args.no_pch or test_both: print("Testing WITHOUT PCH:") print("-" * 30) times_no_pch = benchmark_compilation(use_pch=False, iterations=args.iterations) if times_no_pch: avg_no_pch = mean(times_no_pch) std_no_pch = stdev(times_no_pch) if len(times_no_pch) > 1 else 0 print(f"Average: {avg_no_pch:.2f} ms (±{std_no_pch:.2f} ms)") print(f"Min: {min(times_no_pch):.2f} ms") print(f"Max: {max(times_no_pch):.2f} ms") results["no_pch"] = avg_no_pch print() # Test with PCH if args.pch or test_both: print("Testing WITH PCH:") print("-" * 30) times_with_pch = benchmark_compilation( use_pch=True, iterations=args.iterations ) if times_with_pch: avg_with_pch = mean(times_with_pch) std_with_pch = stdev(times_with_pch) if len(times_with_pch) > 1 else 0 print(f"Average: {avg_with_pch:.2f} ms (±{std_with_pch:.2f} ms)") print(f"Min: {min(times_with_pch):.2f} ms") print(f"Max: {max(times_with_pch):.2f} ms") results["pch"] = avg_with_pch print() if __name__ == "__main__": main() ``` Pull Request resolved: pytorch#162972 Approved by: https://github.com/albanD, https://github.com/janeyx99
20 tasks
mansiag05
pushed a commit
to mansiag05/pytorch
that referenced
this pull request
Sep 22, 2025
Enabling automatic pre compiled headers per https://docs.nvidia.com/cuda/nvrtc/index.html#example-automatic-pch-cuda-12-8 I'm seeing large speedups in compilation times using PCH on average but the max compilation time with PCH is worst which is why I can't enable it by default. `load_inline()` also supports precompiled headers and does not enable them by default ``` Without PCH: 270.58 ms average With PCH: 115.27 ms average ``` ``` Without PCH: Max: 337.99 ms With PCH: Max: 383.82 ms ``` ```python source) [marksaroufim@devgpu005]~/pytorch% python simple_pch_benchmark.py ============================================================ Simple PCH Compilation Benchmark ============================================================ Device: NVIDIA B200 Iterations: 100 Testing WITHOUT PCH: ------------------------------ Compiling kernel 100 times WITHOUT PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 270.58 ms (±6.99 ms) Min: 264.09 ms Max: 337.99 ms Testing WITH PCH: ------------------------------ Compiling kernel 100 times WITH PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 115.27 ms (±27.32 ms) Min: 110.65 ms Max: 383.82 ms ``` ## Benchmarking script ```python #!/usr/bin/env python3 import argparse import os import sys import time from statistics import mean, stdev import torch from torch.cuda._utils import _nvrtc_compile def benchmark_compilation(use_pch, iterations=100): """Compile the same kernel many times with or without PCH.""" # CUB kernel that benefits from PCH kernel_source = """ #include <cub/block/block_reduce.cuh> #include <cub/block/block_scan.cuh> #include <cub/warp/warp_reduce.cuh> extern "C" __global__ void test_kernel(const float* input, float* output, int n) { using BlockReduce = cub::BlockReduce<float, 256>; using BlockScan = cub::BlockScan<float, 256>; using WarpReduce = cub::WarpReduce<float>; __shared__ union { typename BlockReduce::TempStorage reduce; typename BlockScan::TempStorage scan; typename WarpReduce::TempStorage warp[8]; } temp_storage; int idx = blockIdx.x * blockDim.x + threadIdx.x; float val = (idx < n) ? input[idx] : 0.0f; float sum = BlockReduce(temp_storage.reduce).Sum(val); __syncthreads(); float scan_result; BlockScan(temp_storage.scan).ExclusiveSum(val, scan_result); __syncthreads(); int warp_id = threadIdx.x / 32; float warp_sum = WarpReduce(temp_storage.warp[warp_id]).Sum(val); if (threadIdx.x == 0) { output[blockIdx.x] = sum + scan_result + warp_sum; } } """ device = torch.cuda.current_device() major, minor = torch.cuda.get_device_capability(device) compute_capability = f"{major}{minor}" compile_times = [] print( f"Compiling kernel {iterations} times {'WITH' if use_pch else 'WITHOUT'} PCH..." ) for i in range(iterations): # Use unique kernel name to avoid caching between iterations kernel_name = f"test_kernel_{i}" unique_source = kernel_source.replace("test_kernel", kernel_name) start = time.perf_counter() ptx, mangled_name = _nvrtc_compile( unique_source, kernel_name, compute_capability, header_code="", nvcc_options=["-std=c++17"], auto_pch=use_pch, ) elapsed = time.perf_counter() - start compile_times.append(elapsed * 1000) # Convert to ms # Progress indicator if (i + 1) % 10 == 0: print(f" Completed {i + 1}/{iterations} compilations") return compile_times def main(): parser = argparse.ArgumentParser(description="Simple PCH Compilation Benchmark") parser.add_argument("--pch", action="store_true", help="Test with PCH only") parser.add_argument("--no-pch", action="store_true", help="Test without PCH only") parser.add_argument( "--iterations", type=int, default=100, help="Number of compilations" ) args = parser.parse_args() print("=" * 60) print("Simple PCH Compilation Benchmark") print("=" * 60) print(f"Device: {torch.cuda.get_device_name()}") print(f"Iterations: {args.iterations}") print() # Determine what to test test_both = not args.pch and not args.no_pch results = {} # Test without PCH if args.no_pch or test_both: print("Testing WITHOUT PCH:") print("-" * 30) times_no_pch = benchmark_compilation(use_pch=False, iterations=args.iterations) if times_no_pch: avg_no_pch = mean(times_no_pch) std_no_pch = stdev(times_no_pch) if len(times_no_pch) > 1 else 0 print(f"Average: {avg_no_pch:.2f} ms (±{std_no_pch:.2f} ms)") print(f"Min: {min(times_no_pch):.2f} ms") print(f"Max: {max(times_no_pch):.2f} ms") results["no_pch"] = avg_no_pch print() # Test with PCH if args.pch or test_both: print("Testing WITH PCH:") print("-" * 30) times_with_pch = benchmark_compilation( use_pch=True, iterations=args.iterations ) if times_with_pch: avg_with_pch = mean(times_with_pch) std_with_pch = stdev(times_with_pch) if len(times_with_pch) > 1 else 0 print(f"Average: {avg_with_pch:.2f} ms (±{std_with_pch:.2f} ms)") print(f"Min: {min(times_with_pch):.2f} ms") print(f"Max: {max(times_with_pch):.2f} ms") results["pch"] = avg_with_pch print() if __name__ == "__main__": main() ``` Pull Request resolved: pytorch#162972 Approved by: https://github.com/albanD, https://github.com/janeyx99
cleonard530
pushed a commit
to cleonard530/pytorch
that referenced
this pull request
Sep 22, 2025
Enabling automatic pre compiled headers per https://docs.nvidia.com/cuda/nvrtc/index.html#example-automatic-pch-cuda-12-8 I'm seeing large speedups in compilation times using PCH on average but the max compilation time with PCH is worst which is why I can't enable it by default. `load_inline()` also supports precompiled headers and does not enable them by default ``` Without PCH: 270.58 ms average With PCH: 115.27 ms average ``` ``` Without PCH: Max: 337.99 ms With PCH: Max: 383.82 ms ``` ```python source) [marksaroufim@devgpu005]~/pytorch% python simple_pch_benchmark.py ============================================================ Simple PCH Compilation Benchmark ============================================================ Device: NVIDIA B200 Iterations: 100 Testing WITHOUT PCH: ------------------------------ Compiling kernel 100 times WITHOUT PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 270.58 ms (±6.99 ms) Min: 264.09 ms Max: 337.99 ms Testing WITH PCH: ------------------------------ Compiling kernel 100 times WITH PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 115.27 ms (±27.32 ms) Min: 110.65 ms Max: 383.82 ms ``` ## Benchmarking script ```python #!/usr/bin/env python3 import argparse import os import sys import time from statistics import mean, stdev import torch from torch.cuda._utils import _nvrtc_compile def benchmark_compilation(use_pch, iterations=100): """Compile the same kernel many times with or without PCH.""" # CUB kernel that benefits from PCH kernel_source = """ #include <cub/block/block_reduce.cuh> #include <cub/block/block_scan.cuh> #include <cub/warp/warp_reduce.cuh> extern "C" __global__ void test_kernel(const float* input, float* output, int n) { using BlockReduce = cub::BlockReduce<float, 256>; using BlockScan = cub::BlockScan<float, 256>; using WarpReduce = cub::WarpReduce<float>; __shared__ union { typename BlockReduce::TempStorage reduce; typename BlockScan::TempStorage scan; typename WarpReduce::TempStorage warp[8]; } temp_storage; int idx = blockIdx.x * blockDim.x + threadIdx.x; float val = (idx < n) ? input[idx] : 0.0f; float sum = BlockReduce(temp_storage.reduce).Sum(val); __syncthreads(); float scan_result; BlockScan(temp_storage.scan).ExclusiveSum(val, scan_result); __syncthreads(); int warp_id = threadIdx.x / 32; float warp_sum = WarpReduce(temp_storage.warp[warp_id]).Sum(val); if (threadIdx.x == 0) { output[blockIdx.x] = sum + scan_result + warp_sum; } } """ device = torch.cuda.current_device() major, minor = torch.cuda.get_device_capability(device) compute_capability = f"{major}{minor}" compile_times = [] print( f"Compiling kernel {iterations} times {'WITH' if use_pch else 'WITHOUT'} PCH..." ) for i in range(iterations): # Use unique kernel name to avoid caching between iterations kernel_name = f"test_kernel_{i}" unique_source = kernel_source.replace("test_kernel", kernel_name) start = time.perf_counter() ptx, mangled_name = _nvrtc_compile( unique_source, kernel_name, compute_capability, header_code="", nvcc_options=["-std=c++17"], auto_pch=use_pch, ) elapsed = time.perf_counter() - start compile_times.append(elapsed * 1000) # Convert to ms # Progress indicator if (i + 1) % 10 == 0: print(f" Completed {i + 1}/{iterations} compilations") return compile_times def main(): parser = argparse.ArgumentParser(description="Simple PCH Compilation Benchmark") parser.add_argument("--pch", action="store_true", help="Test with PCH only") parser.add_argument("--no-pch", action="store_true", help="Test without PCH only") parser.add_argument( "--iterations", type=int, default=100, help="Number of compilations" ) args = parser.parse_args() print("=" * 60) print("Simple PCH Compilation Benchmark") print("=" * 60) print(f"Device: {torch.cuda.get_device_name()}") print(f"Iterations: {args.iterations}") print() # Determine what to test test_both = not args.pch and not args.no_pch results = {} # Test without PCH if args.no_pch or test_both: print("Testing WITHOUT PCH:") print("-" * 30) times_no_pch = benchmark_compilation(use_pch=False, iterations=args.iterations) if times_no_pch: avg_no_pch = mean(times_no_pch) std_no_pch = stdev(times_no_pch) if len(times_no_pch) > 1 else 0 print(f"Average: {avg_no_pch:.2f} ms (±{std_no_pch:.2f} ms)") print(f"Min: {min(times_no_pch):.2f} ms") print(f"Max: {max(times_no_pch):.2f} ms") results["no_pch"] = avg_no_pch print() # Test with PCH if args.pch or test_both: print("Testing WITH PCH:") print("-" * 30) times_with_pch = benchmark_compilation( use_pch=True, iterations=args.iterations ) if times_with_pch: avg_with_pch = mean(times_with_pch) std_with_pch = stdev(times_with_pch) if len(times_with_pch) > 1 else 0 print(f"Average: {avg_with_pch:.2f} ms (±{std_with_pch:.2f} ms)") print(f"Min: {min(times_with_pch):.2f} ms") print(f"Max: {max(times_with_pch):.2f} ms") results["pch"] = avg_with_pch print() if __name__ == "__main__": main() ``` Pull Request resolved: pytorch#162972 Approved by: https://github.com/albanD, https://github.com/janeyx99
dsashidh
pushed a commit
to dsashidh/pytorch
that referenced
this pull request
Sep 26, 2025
Enabling automatic pre compiled headers per https://docs.nvidia.com/cuda/nvrtc/index.html#example-automatic-pch-cuda-12-8 I'm seeing large speedups in compilation times using PCH on average but the max compilation time with PCH is worst which is why I can't enable it by default. `load_inline()` also supports precompiled headers and does not enable them by default ``` Without PCH: 270.58 ms average With PCH: 115.27 ms average ``` ``` Without PCH: Max: 337.99 ms With PCH: Max: 383.82 ms ``` ```python source) [marksaroufim@devgpu005]~/pytorch% python simple_pch_benchmark.py ============================================================ Simple PCH Compilation Benchmark ============================================================ Device: NVIDIA B200 Iterations: 100 Testing WITHOUT PCH: ------------------------------ Compiling kernel 100 times WITHOUT PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 270.58 ms (±6.99 ms) Min: 264.09 ms Max: 337.99 ms Testing WITH PCH: ------------------------------ Compiling kernel 100 times WITH PCH... Completed 10/100 compilations Completed 20/100 compilations Completed 30/100 compilations Completed 40/100 compilations Completed 50/100 compilations Completed 60/100 compilations Completed 70/100 compilations Completed 80/100 compilations Completed 90/100 compilations Completed 100/100 compilations Average: 115.27 ms (±27.32 ms) Min: 110.65 ms Max: 383.82 ms ``` ## Benchmarking script ```python #!/usr/bin/env python3 import argparse import os import sys import time from statistics import mean, stdev import torch from torch.cuda._utils import _nvrtc_compile def benchmark_compilation(use_pch, iterations=100): """Compile the same kernel many times with or without PCH.""" # CUB kernel that benefits from PCH kernel_source = """ #include <cub/block/block_reduce.cuh> #include <cub/block/block_scan.cuh> #include <cub/warp/warp_reduce.cuh> extern "C" __global__ void test_kernel(const float* input, float* output, int n) { using BlockReduce = cub::BlockReduce<float, 256>; using BlockScan = cub::BlockScan<float, 256>; using WarpReduce = cub::WarpReduce<float>; __shared__ union { typename BlockReduce::TempStorage reduce; typename BlockScan::TempStorage scan; typename WarpReduce::TempStorage warp[8]; } temp_storage; int idx = blockIdx.x * blockDim.x + threadIdx.x; float val = (idx < n) ? input[idx] : 0.0f; float sum = BlockReduce(temp_storage.reduce).Sum(val); __syncthreads(); float scan_result; BlockScan(temp_storage.scan).ExclusiveSum(val, scan_result); __syncthreads(); int warp_id = threadIdx.x / 32; float warp_sum = WarpReduce(temp_storage.warp[warp_id]).Sum(val); if (threadIdx.x == 0) { output[blockIdx.x] = sum + scan_result + warp_sum; } } """ device = torch.cuda.current_device() major, minor = torch.cuda.get_device_capability(device) compute_capability = f"{major}{minor}" compile_times = [] print( f"Compiling kernel {iterations} times {'WITH' if use_pch else 'WITHOUT'} PCH..." ) for i in range(iterations): # Use unique kernel name to avoid caching between iterations kernel_name = f"test_kernel_{i}" unique_source = kernel_source.replace("test_kernel", kernel_name) start = time.perf_counter() ptx, mangled_name = _nvrtc_compile( unique_source, kernel_name, compute_capability, header_code="", nvcc_options=["-std=c++17"], auto_pch=use_pch, ) elapsed = time.perf_counter() - start compile_times.append(elapsed * 1000) # Convert to ms # Progress indicator if (i + 1) % 10 == 0: print(f" Completed {i + 1}/{iterations} compilations") return compile_times def main(): parser = argparse.ArgumentParser(description="Simple PCH Compilation Benchmark") parser.add_argument("--pch", action="store_true", help="Test with PCH only") parser.add_argument("--no-pch", action="store_true", help="Test without PCH only") parser.add_argument( "--iterations", type=int, default=100, help="Number of compilations" ) args = parser.parse_args() print("=" * 60) print("Simple PCH Compilation Benchmark") print("=" * 60) print(f"Device: {torch.cuda.get_device_name()}") print(f"Iterations: {args.iterations}") print() # Determine what to test test_both = not args.pch and not args.no_pch results = {} # Test without PCH if args.no_pch or test_both: print("Testing WITHOUT PCH:") print("-" * 30) times_no_pch = benchmark_compilation(use_pch=False, iterations=args.iterations) if times_no_pch: avg_no_pch = mean(times_no_pch) std_no_pch = stdev(times_no_pch) if len(times_no_pch) > 1 else 0 print(f"Average: {avg_no_pch:.2f} ms (±{std_no_pch:.2f} ms)") print(f"Min: {min(times_no_pch):.2f} ms") print(f"Max: {max(times_no_pch):.2f} ms") results["no_pch"] = avg_no_pch print() # Test with PCH if args.pch or test_both: print("Testing WITH PCH:") print("-" * 30) times_with_pch = benchmark_compilation( use_pch=True, iterations=args.iterations ) if times_with_pch: avg_with_pch = mean(times_with_pch) std_with_pch = stdev(times_with_pch) if len(times_with_pch) > 1 else 0 print(f"Average: {avg_with_pch:.2f} ms (±{std_with_pch:.2f} ms)") print(f"Min: {min(times_with_pch):.2f} ms") print(f"Max: {max(times_with_pch):.2f} ms") results["pch"] = avg_with_pch print() if __name__ == "__main__": main() ``` Pull Request resolved: pytorch#162972 Approved by: https://github.com/albanD, https://github.com/janeyx99
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Enabling automatic pre compiled headers per https://docs.nvidia.com/cuda/nvrtc/index.html#example-automatic-pch-cuda-12-8
I'm seeing large speedups in compilation times using PCH on average but the max compilation time with PCH is worst which is why I can't enable it by default.
load_inline()also supports precompiled headers and does not enable them by defaultBenchmarking script
cc @gau-nernst