update examples

huggingface · Feb 12, 2024 · 7fd9323 · 7fd9323
1 parent 8ba0a57
commit 7fd9323
Show file tree

Hide file tree

Showing 9 changed files with 73 additions and 24 deletions.
diff --git a/examples/neural_compressor_ptq_bert.yaml b/examples/neural_compressor_ptq_bert.yaml
@@ -10,22 +10,28 @@ defaults:
 experiment_name: openvino_static_quant_bert
 
 backend:
-  model: bert-base-uncased
+  device: cpu
   no_weights: true
+  model: bert-base-uncased
   ptq_quantization: true
   calibration: true
-  device: cpu
 
 benchmark:
   input_shapes:
     batch_size: 1
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
diff --git a/examples/onnxruntime_static_quant_vit.yaml b/examples/onnxruntime_static_quant_vit.yaml
@@ -10,23 +10,28 @@ defaults:
 experiment_name: onnxruntime_static_quant_vit
 
 backend:
+  device: cpu
+  no_weights: true
   model: google/vit-base-patch16-224
   quantization: true
   quantization_config:
     is_static: true
     per_channel: false
-  device: cpu
 
   calibration: true
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
-      CUDA_VISIBLE_DEVICES: 0
-      CUDA_DEVICE_ORDER: PCI_BUS_ID
diff --git a/examples/openvino_diffusion.yaml b/examples/openvino_diffusion.yaml
@@ -10,22 +10,28 @@ defaults:
 model: stabilityai/stable-diffusion-2-1
 
 backend:
+  device: cpu
   experiment_name: openvino_diffusion
-  export: true
   reshape: true
+  export: true
   half: true
-  device: cpu
 
 benchmark:
   input_shapes:
     batch_size: 1
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
diff --git a/examples/openvino_static_quant_bert.yaml b/examples/openvino_static_quant_bert.yaml
@@ -10,24 +10,30 @@ defaults:
 experiment_name: openvino_static_quant_bert
 
 backend:
+  device: cpu
+  no_weights: true
   model: bert-base-uncased
   export: true
-  no_weights: true
   quantization: true
   calibration: true
   reshape: true
-  device: cpu
 
 benchmark:
   input_shapes:
     batch_size: 1
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
diff --git a/examples/pytorch_bert.yaml b/examples/pytorch_bert.yaml
@@ -10,17 +10,22 @@ defaults:
 experiment_name: pytorch_bert
 
 backend:
-  model: bert-base-uncased
   device: cpu
+  device_ids: 0
+  model: bert-base-uncased
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
-      CUDA_VISIBLE_DEVICES: 0
-      CUDA_DEVICE_ORDER: PCI_BUS_ID
diff --git a/examples/pytorch_llama.yaml b/examples/pytorch_llama.yaml
@@ -10,8 +10,10 @@ defaults:
 experiment_name: pytorch_llama
 
 backend:
-  model: TheBloke/Llama-2-70B-AWQ
   device: cuda
+  device_ids: 0
+  no_weights: true
+  model: TheBloke/Llama-2-70B-AWQ
 
 launcher:
   device_isolation: true
@@ -22,14 +24,18 @@ benchmark:
     sequence_length: 256
   new_tokens: 1000
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
-      CUDA_VISIBLE_DEVICES: 0
-      CUDA_DEVICE_ORDER: PCI_BUS_ID
diff --git a/examples/pytorch_timm.yaml b/examples/pytorch_timm.yaml
@@ -10,8 +10,9 @@ defaults:
 experiment_name: pytorch_timm
 
 backend:
-  model: timm/mobilenetv3_large_100.ra_in1k
   device: cuda
+  device_ids: 0
+  model: timm/mobilenetv3_large_100.ra_in1k
 
 launcher:
   device_isolation: true
@@ -20,14 +21,18 @@ benchmark:
   input_shapes:
     batch_size: 1
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
-      CUDA_VISIBLE_DEVICES: 0
-      CUDA_DEVICE_ORDER: PCI_BUS_ID
diff --git a/examples/tgi_llama.yaml b/examples/tgi_llama.yaml
@@ -10,25 +10,31 @@ defaults:
 experiment_name: tgi_llama
 
 backend:
+  device: cuda
+  device_ids: 0,1
+  device_map: true
   model: TheBloke/Llama-2-7B-AWQ
   quantization_scheme: awq
   sharded: false
-  device: cuda
 
 benchmark:
   input_shapes:
     batch_size: 1
     sequence_length: 256
   new_tokens: 1000
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
-      CUDA_VISIBLE_DEVICES: 0
-      CUDA_DEVICE_ORDER: PCI_BUS_ID
diff --git a/examples/trt_llama.yaml b/examples/trt_llama.yaml
@@ -10,23 +10,27 @@ defaults:
 experiment_name: trt_llama
 
 backend:
-  model: NousResearch/Llama-2-7b-hf
   device: cuda
+  model: NousResearch/Llama-2-7b-hf
 
 benchmark:
   input_shapes:
     batch_size: 1
     sequence_length: 64
   new_tokens: 128
 
+# hydra/cli specific settings
 hydra:
   run:
+    # where to store run results
     dir: runs/${experiment_name}
   sweep:
+    # where to store sweep results
     dir: sweeps/${experiment_name}
   job:
+    # change working directory to the run directory
     chdir: true
     env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
       OVERRIDE_BENCHMARKS: 1
-      CUDA_VISIBLE_DEVICES: 0
-      CUDA_DEVICE_ORDER: PCI_BUS_ID