[OPT] Improve benchmark results (#163)

* [REF] Increase batch sizes, update results with more efficient matvecs * [ADD] Update benchmark figures
f-dangel · Jan 1, 2025 · 4bed707 · 4bed707
1 parent 75bc0a8
commit 4bed707
Show file tree

Hide file tree

Showing 69 changed files with 68 additions and 68 deletions.
diff --git a/...enchmark/peakmem_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...enchmark/peakmem_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 2.821939468383789}
+{"peakmem": 5.408483028411865}
diff --git a/...sic_usage/benchmark/peakmem_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json b/...sic_usage/benchmark/peakmem_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 8.144813060760498}
+{"peakmem": 12.137713432312012}
diff --git a/...chmark/peakmem_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...chmark/peakmem_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 3.9029364585876465}
+{"peakmem": 6.572415828704834}
diff --git a/...c_usage/benchmark/peakmem_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...c_usage/benchmark/peakmem_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 14.977922916412354}
+{"peakmem": 18.18963098526001}
diff --git a/.../peakmem_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/.../peakmem_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 2.821939468383789}
+{"peakmem": 5.408483028411865}
diff --git a/...e/benchmark/peakmem_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_matvec.json b/...e/benchmark/peakmem_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 8.144813060760498}
+{"peakmem": 12.137713432312012}
diff --git a/...eakmem_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...eakmem_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 3.9029364585876465}
+{"peakmem": 6.572415828704834}
diff --git a/...benchmark/peakmem_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...benchmark/peakmem_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 14.97962474822998}
+{"peakmem": 18.191569328308105}
diff --git a/...c_usage/benchmark/peakmem_Hessian_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...c_usage/benchmark/peakmem_Hessian_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 2.821939468383789}
+{"peakmem": 5.408483028411865}
diff --git a/...amples/basic_usage/benchmark/peakmem_Hessian_synthetic_imagenet_resnet50_cuda_matvec.json b/...amples/basic_usage/benchmark/peakmem_Hessian_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 14.854530334472656}
+{"peakmem": 15.885180950164795}
diff --git a/...usage/benchmark/peakmem_Hessian_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...usage/benchmark/peakmem_Hessian_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 3.9029364585876465}
+{"peakmem": 6.572415828704834}
diff --git a/...ples/basic_usage/benchmark/peakmem_Hessian_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...ples/basic_usage/benchmark/peakmem_Hessian_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 23.27983283996582}
+{"peakmem": 27.734885215759277}
diff --git a/...ge/benchmark/peakmem_KFAC-inverse_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...ge/benchmark/peakmem_KFAC-inverse_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 2.8216772079467773}
+{"peakmem": 5.4073052406311035}
diff --git a/...s/basic_usage/benchmark/peakmem_KFAC-inverse_synthetic_imagenet_resnet50_cuda_matvec.json b/...s/basic_usage/benchmark/peakmem_KFAC-inverse_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 3.334245204925537}
+{"peakmem": 5.881683826446533}
diff --git a/...sic_usage/benchmark/peakmem_KFAC-inverse_synthetic_imagenet_resnet50_cuda_precompute.json b/...sic_usage/benchmark/peakmem_KFAC-inverse_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"peakmem": 3.238542079925537}
+{"peakmem": 5.785980701446533}
diff --git a/.../benchmark/peakmem_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/.../benchmark/peakmem_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 3.6108012199401855}
+{"peakmem": 6.280280590057373}
diff --git a/...basic_usage/benchmark/peakmem_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...basic_usage/benchmark/peakmem_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 4.11228084564209}
+{"peakmem": 6.78189754486084}
diff --git a/...c_usage/benchmark/peakmem_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...c_usage/benchmark/peakmem_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"peakmem": 3.795565605163574}
+{"peakmem": 6.465182304382324}
diff --git a/...asic_usage/benchmark/peakmem_KFAC_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...asic_usage/benchmark/peakmem_KFAC_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 2.8216772079467773}
+{"peakmem": 5.4073052406311035}
diff --git a/.../examples/basic_usage/benchmark/peakmem_KFAC_synthetic_imagenet_resnet50_cuda_matvec.json b/.../examples/basic_usage/benchmark/peakmem_KFAC_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 3.334245204925537}
+{"peakmem": 5.881683826446533}
diff --git a/...mples/basic_usage/benchmark/peakmem_KFAC_synthetic_imagenet_resnet50_cuda_precompute.json b/...mples/basic_usage/benchmark/peakmem_KFAC_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"peakmem": 3.238542079925537}
+{"peakmem": 5.785980701446533}
diff --git a/...ic_usage/benchmark/peakmem_KFAC_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...ic_usage/benchmark/peakmem_KFAC_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 3.6108012199401855}
+{"peakmem": 6.280280590057373}
diff --git a/...xamples/basic_usage/benchmark/peakmem_KFAC_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...xamples/basic_usage/benchmark/peakmem_KFAC_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 4.11228084564209}
+{"peakmem": 6.78189754486084}
diff --git a/...les/basic_usage/benchmark/peakmem_KFAC_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...les/basic_usage/benchmark/peakmem_KFAC_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"peakmem": 3.795565605163574}
+{"peakmem": 6.465182304382324}
diff --git a/...chmark/peakmem_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...chmark/peakmem_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 2.821939468383789}
+{"peakmem": 5.408483028411865}
diff --git a/...c_usage/benchmark/peakmem_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json b/...c_usage/benchmark/peakmem_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 8.144932270050049}
+{"peakmem": 12.137951850891113}
diff --git a/...mark/peakmem_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...mark/peakmem_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"peakmem": 3.9029364585876465}
+{"peakmem": 6.572415828704834}
diff --git a/...usage/benchmark/peakmem_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...usage/benchmark/peakmem_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"peakmem": 15.361834049224854}
+{"peakmem": 18.95720911026001}
diff --git a/docs/examples/basic_usage/benchmark/peakmem_synthetic_imagenet_resnet50_cuda.pdf b/docs/examples/basic_usage/benchmark/peakmem_synthetic_imagenet_resnet50_cuda.pdf
diff --git a/docs/examples/basic_usage/benchmark/peakmem_synthetic_shakespeare_nanogpt_cuda.pdf b/docs/examples/basic_usage/benchmark/peakmem_synthetic_shakespeare_nanogpt_cuda.pdf
diff --git a/...e/benchmark/time_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...e/benchmark/time_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.07914857612922788}
+{"time": 0.15138799278065562}
diff --git a/.../basic_usage/benchmark/time_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json b/.../basic_usage/benchmark/time_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.4040485708974302}
+{"time": 0.7856894689612091}
diff --git a/...ic_usage/benchmark/time_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_precompute.json b/...ic_usage/benchmark/time_Empirical-Fisher_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0602641850709915e-05}
+{"time": 1.046992838382721e-05}
diff --git a/...benchmark/time_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...benchmark/time_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.12177057890221477}
+{"time": 0.23390081012621522}
diff --git a/...asic_usage/benchmark/time_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...asic_usage/benchmark/time_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.40884994994848967}
+{"time": 0.7778741386719048}
diff --git a/..._usage/benchmark/time_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_precompute.json b/..._usage/benchmark/time_Empirical-Fisher_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0384712368249893e-05}
+{"time": 1.039775088429451e-05}
diff --git a/...ark/time_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...ark/time_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.07911323523148894}
+{"time": 0.15136110922321677}
diff --git a/...sage/benchmark/time_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_matvec.json b/...sage/benchmark/time_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.4040556410327554}
+{"time": 0.785682977642864}
diff --git a/.../benchmark/time_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_precompute.json b/.../benchmark/time_Generalized-Gauss-Newton_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0679941624403e-05}
+{"time": 1.0511837899684906e-05}
diff --git a/...k/time_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...k/time_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.12210974702611566}
+{"time": 0.23344802716746926}
diff --git a/...ge/benchmark/time_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...ge/benchmark/time_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.41015551099553704}
+{"time": 0.7820523148402572}
diff --git a/...enchmark/time_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...enchmark/time_Generalized-Gauss-Newton_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0590068995952606e-05}
+{"time": 1.0339077562093735e-05}
diff --git a/...asic_usage/benchmark/time_Hessian_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...asic_usage/benchmark/time_Hessian_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.07917527668178082}
+{"time": 0.15139471692964435}
diff --git a/.../examples/basic_usage/benchmark/time_Hessian_synthetic_imagenet_resnet50_cuda_matvec.json b/.../examples/basic_usage/benchmark/time_Hessian_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.43220572592690587}
+{"time": 0.8380271531641483}
diff --git a/...mples/basic_usage/benchmark/time_Hessian_synthetic_imagenet_resnet50_cuda_precompute.json b/...mples/basic_usage/benchmark/time_Hessian_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0847114026546478e-05}
+{"time": 1.0468997061252594e-05}
diff --git a/...ic_usage/benchmark/time_Hessian_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...ic_usage/benchmark/time_Hessian_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.12290711421519518}
+{"time": 0.2342886170372367}
diff --git a/...xamples/basic_usage/benchmark/time_Hessian_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...xamples/basic_usage/benchmark/time_Hessian_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.564610986970365}
+{"time": 1.065617433283478}
diff --git a/...les/basic_usage/benchmark/time_Hessian_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...les/basic_usage/benchmark/time_Hessian_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0426156222820282e-05}
+{"time": 1.0298099368810654e-05}
diff --git a/...usage/benchmark/time_KFAC-inverse_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...usage/benchmark/time_KFAC-inverse_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.07193620782345533}
+{"time": 0.13708342146128416}
diff --git a/...ples/basic_usage/benchmark/time_KFAC-inverse_synthetic_imagenet_resnet50_cuda_matvec.json b/...ples/basic_usage/benchmark/time_KFAC-inverse_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.012802004348486662}
+{"time": 0.01288953935727477}
diff --git a/.../basic_usage/benchmark/time_KFAC-inverse_synthetic_imagenet_resnet50_cuda_precompute.json b/.../basic_usage/benchmark/time_KFAC-inverse_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"time": 0.18157002329826355}
+{"time": 0.34895477583631873}
diff --git a/...age/benchmark/time_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...age/benchmark/time_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.11208314076066017}
+{"time": 0.21410065283998847}
diff --git a/...es/basic_usage/benchmark/time_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...es/basic_usage/benchmark/time_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.03331297682598233}
+{"time": 0.03307039197534323}
diff --git a/...asic_usage/benchmark/time_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...asic_usage/benchmark/time_KFAC-inverse_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"time": 0.18883356265723705}
+{"time": 0.35276420041918755}
diff --git a/...s/basic_usage/benchmark/time_KFAC_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...s/basic_usage/benchmark/time_KFAC_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.07191778579726815}
+{"time": 0.1371102319099009}
diff --git a/docs/examples/basic_usage/benchmark/time_KFAC_synthetic_imagenet_resnet50_cuda_matvec.json b/docs/examples/basic_usage/benchmark/time_KFAC_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.011812383774667978}
+{"time": 0.011814852245151997}
diff --git a/...examples/basic_usage/benchmark/time_KFAC_synthetic_imagenet_resnet50_cuda_precompute.json b/...examples/basic_usage/benchmark/time_KFAC_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"time": 0.17920332308858633}
+{"time": 0.3489605807699263}
diff --git a/...basic_usage/benchmark/time_KFAC_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...basic_usage/benchmark/time_KFAC_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.11182996584102511}
+{"time": 0.21611323999240994}
diff --git a/docs/examples/basic_usage/benchmark/time_KFAC_synthetic_shakespeare_nanogpt_cuda_matvec.json b/docs/examples/basic_usage/benchmark/time_KFAC_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.03477561008185148}
+{"time": 0.03479210985824466}
diff --git a/...amples/basic_usage/benchmark/time_KFAC_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...amples/basic_usage/benchmark/time_KFAC_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"time": 0.18182456912472844}
+{"time": 0.34671181719750166}
diff --git a/...benchmark/time_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json b/...benchmark/time_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.07917376700788736}
+{"time": 0.15134236635640264}
diff --git a/...asic_usage/benchmark/time_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json b/...asic_usage/benchmark/time_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.40426155366003513}
+{"time": 0.7858486091718078}
diff --git a/..._usage/benchmark/time_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_precompute.json b/..._usage/benchmark/time_Monte-Carlo-Fisher_synthetic_imagenet_resnet50_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.057703047990799e-05}
+{"time": 1.0366085916757584e-05}
diff --git a/...nchmark/time_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json b/...nchmark/time_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_gradient_and_loss.json
@@ -1 +1 @@
-{"time": 0.12174207577481866}
+{"time": 0.2342790300026536}
diff --git a/...ic_usage/benchmark/time_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json b/...ic_usage/benchmark/time_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_matvec.json
@@ -1 +1 @@
-{"time": 0.4763146941550076}
+{"time": 0.9123296132311225}
diff --git a/...sage/benchmark/time_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_precompute.json b/...sage/benchmark/time_Monte-Carlo-Fisher_synthetic_shakespeare_nanogpt_cuda_precompute.json
@@ -1 +1 @@
-{"time": 1.0615214705467224e-05}
+{"time": 1.0463874787092209e-05}
diff --git a/docs/examples/basic_usage/benchmark/time_synthetic_imagenet_resnet50_cuda.pdf b/docs/examples/basic_usage/benchmark/time_synthetic_imagenet_resnet50_cuda.pdf
diff --git a/docs/examples/basic_usage/benchmark/time_synthetic_shakespeare_nanogpt_cuda.pdf b/docs/examples/basic_usage/benchmark/time_synthetic_shakespeare_nanogpt_cuda.pdf
diff --git a/docs/examples/basic_usage/benchmark_utils.py b/docs/examples/basic_usage/benchmark_utils.py
@@ -59,12 +59,12 @@ def forward(self, x: Tensor) -> Tensor:
 
 
 def setup_synthetic_shakespeare_nanogpt(
-    batch_size: int = 2,
+    batch_size: int = 4,
 ) -> Tuple[GPTWrapper, CrossEntropyLoss, List[Tuple[Tensor, Tensor]]]:
     """Set up the nanoGPT model and synthetic Shakespeare dataset for the benchmark.
 
     Args:
-        batch_size: The batch size to use. Default is ``2``.
+        batch_size: The batch size to use. Default is ``4``.
 
     Returns:
         A tuple containing the nanoGPT model, the loss function, and the data.
@@ -98,12 +98,12 @@ def setup_synthetic_shakespeare_nanogpt(
 
 
 def setup_synthetic_imagenet_resnet50(
-    batch_size: int = 32,
+    batch_size: int = 64,
 ) -> Tuple[Module, CrossEntropyLoss, List[Tuple[Tensor, Tensor]]]:
     """Set up ResNet50 on synthetic ImageNet for the benchmark.
 
     Args:
-        batch_size: The batch size to use. Default is ``32``.
+        batch_size: The batch size to use. Default is ``64``.
 
     Returns:
         A tuple containing the ResNet50 model, the loss function
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 2.821939468383789}
		{"peakmem": 5.408483028411865}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 8.144813060760498}
		{"peakmem": 12.137713432312012}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 3.9029364585876465}
		{"peakmem": 6.572415828704834}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 14.977922916412354}
		{"peakmem": 18.18963098526001}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 14.97962474822998}
		{"peakmem": 18.191569328308105}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 14.854530334472656}
		{"peakmem": 15.885180950164795}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 23.27983283996582}
		{"peakmem": 27.734885215759277}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 2.8216772079467773}
		{"peakmem": 5.4073052406311035}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 3.334245204925537}
		{"peakmem": 5.881683826446533}
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		{"peakmem": 3.238542079925537}
		{"peakmem": 5.785980701446533}