github/AMD-SHARK-Studio
1
1
Fork 0
mirror of https://github.com/nod-ai/AMD-SHARK-Studio.git synced 2026-04-03 03:00:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

Add mlir_type for torch_model_list.csv (#1428)

Browse Source 
- Enable stablehlo/tosa mlir output for torch model
- Add BERT stablehlo support
This commit is contained in:
Chi_Liu
2023-05-15 10:23:54 -07:00
committed by GitHub
parent a9bae00606
commit a0097a1ead
4 changed files with 30 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
shark/shark_importer.py
View File

@@ -191,6 +191,7 @@ class SharkImporter:
dir=tempfile.gettempdir(),
model_name="model",
golden_values=None,
mlir_type="linalg",
):
if self.inputs == None:
print(
@@ -204,6 +205,7 @@ class SharkImporter:
tracing_required,
func_name,
save_dir=artifact_path,
mlir_type=mlir_type,
)
# TODO: Make sure that any generic function name is accepted. Currently takes in the default function names.
# TODO: Check for multiple outputs.

2
shark/torch_mlir_utils.py
View File

@@ -21,7 +21,7 @@ import io
mlir_type_mapping_dict = {
"linalg": torch_mlir.OutputType.LINALG_ON_TENSORS,
"mhlo": torch_mlir.OutputType.STABLEHLO,
"stablehlo": torch_mlir.OutputType.STABLEHLO,
"tosa": torch_mlir.OutputType.TOSA,
}

3
tank/generate_sharktank.py
View File

@@ -50,6 +50,7 @@ def save_torch_model(torch_model_list, local_tank_cache, import_args):
tracing_required = row[1]
model_type = row[2]
is_dynamic = row[3]
mlir_type = row[4]
tracing_required = False if tracing_required == "False" else True
is_dynamic = False if is_dynamic == "False" else True
@@ -121,6 +122,7 @@ def save_torch_model(torch_model_list, local_tank_cache, import_args):
tracing_required=tracing_required,
dir=torch_model_dir,
model_name=torch_model_name,
mlir_type=mlir_type,
)
# Generate torch dynamic models.
if is_dynamic:
@@ -129,6 +131,7 @@ def save_torch_model(torch_model_list, local_tank_cache, import_args):
tracing_required=tracing_required,
dir=torch_model_dir,
model_name=torch_model_name + "_dynamic",
mlir_type=mlir_type,
)

47
tank/torch_model_list.csv
View File

@@ -1,23 +1,24 @@
model_name, use_tracing, model_type, dynamic, param_count, tags, notes
efficientnet_b0,True,vision,False,5.3M,"image-classification;cnn;conv2d;depthwise-conv","Smallest EfficientNet variant with 224x224 input"
efficientnet_b7,True,vision,False,66M,"image-classification;cnn;conv2d;depthwise-conv","Largest EfficientNet variant with 600x600 input"
microsoft/MiniLM-L12-H384-uncased,True,hf,True,66M,"nlp;bert-variant;transformer-encoder","Large version has 12 layers; 384 hidden size; Smaller than BERTbase (66M params vs 109M params)"
bert-base-uncased,True,hf,True,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
bert-base-cased,True,hf,True,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
google/mobilebert-uncased,True,hf,True,25M,"nlp,bert-variant,transformer-encoder,mobile","24 layers, 512 hidden size, 128 embedding"
alexnet,False,vision,True,61M,"cnn,parallel-layers","The CNN that revolutionized computer vision (move away from hand-crafted features to neural networks),10 years old now and probably no longer used in prod."
resnet18,False,vision,True,11M,"cnn,image-classification,residuals,resnet-variant","1 7x7 conv2d and the rest are 3x3 conv2d"
resnet50,False,vision,True,23M,"cnn,image-classification,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
resnet101,False,vision,True,29M,"cnn,image-classification,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
squeezenet1_0,False,vision,True,1.25M,"cnn,image-classification,mobile,parallel-layers","Parallel conv2d (1x1 conv to compress -> (3x3 expand | 1x1 expand) -> concat)"
wide_resnet50_2,False,vision,True,69M,"cnn,image-classification,residuals,resnet-variant","Resnet variant where model depth is decreased and width is increased."
mobilenet_v3_small,False,vision,True,2.5M,"image-classification,cnn,mobile",N/A
google/vit-base-patch16-224,True,hf_img_cls,False,86M,"image-classification,vision-transformer,transformer-encoder",N/A
microsoft/resnet-50,True,hf_img_cls,False,23M,"image-classification,cnn,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
facebook/deit-small-distilled-patch16-224,True,hf_img_cls,False,22M,"image-classification,vision-transformer,cnn",N/A
microsoft/beit-base-patch16-224-pt22k-ft22k,True,hf_img_cls,False,86M,"image-classification,transformer-encoder,bert-variant,vision-transformer",N/A
nvidia/mit-b0,True,hf_img_cls,False,3.7M,"image-classification,transformer-encoder",SegFormer
mnasnet1_0,False,vision,True,-,"cnn, torchvision, mobile, architecture-search","Outperforms other mobile CNNs on Accuracy vs. Latency"
resnet50_fp16,False,vision,True,23M,"cnn,image-classification,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
bert-base-uncased_fp16,True,fp16,False,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
bert-large-uncased,True,hf,True,330M,"nlp;bert-variant;transformer-encoder","24 layers, 1024 hidden units, 16 attention heads"
model_name, use_tracing, model_type, dynamic, mlir_type, param_count, tags, notes
efficientnet_b0,True,vision,False,linalg,5.3M,"image-classification;cnn;conv2d;depthwise-conv","Smallest EfficientNet variant with 224x224 input"
efficientnet_b7,True,vision,False,linalg,66M,"image-classification;cnn;conv2d;depthwise-conv","Largest EfficientNet variant with 600x600 input"
microsoft/MiniLM-L12-H384-uncased,True,hf,True,linalg,66M,"nlp;bert-variant;transformer-encoder","Large version has 12 layers; 384 hidden size; Smaller than BERTbase (66M params vs 109M params)"
bert-base-uncased,True,hf,True,linalg,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
bert-base-cased,True,hf,True,linalg,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
google/mobilebert-uncased,True,hf,True,linalg,25M,"nlp,bert-variant,transformer-encoder,mobile","24 layers, 512 hidden size, 128 embedding"
alexnet,False,vision,True,linalg,61M,"cnn,parallel-layers","The CNN that revolutionized computer vision (move away from hand-crafted features to neural networks),10 years old now and probably no longer used in prod."
resnet18,False,vision,True,linalg,11M,"cnn,image-classification,residuals,resnet-variant","1 7x7 conv2d and the rest are 3x3 conv2d"
resnet50,False,vision,True,linalg,23M,"cnn,image-classification,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
resnet101,False,vision,True,linalg,29M,"cnn,image-classification,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
squeezenet1_0,False,vision,True,linalg,1.25M,"cnn,image-classification,mobile,parallel-layers","Parallel conv2d (1x1 conv to compress -> (3x3 expand | 1x1 expand) -> concat)"
wide_resnet50_2,False,vision,True,linalg,69M,"cnn,image-classification,residuals,resnet-variant","Resnet variant where model depth is decreased and width is increased."
mobilenet_v3_small,False,vision,True,linalg,2.5M,"image-classification,cnn,mobile",N/A
google/vit-base-patch16-224,True,hf_img_cls,False,linalg,86M,"image-classification,vision-transformer,transformer-encoder",N/A
microsoft/resnet-50,True,hf_img_cls,False,linalg,23M,"image-classification,cnn,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
facebook/deit-small-distilled-patch16-224,True,hf_img_cls,False,linalg,22M,"image-classification,vision-transformer,cnn",N/A
microsoft/beit-base-patch16-224-pt22k-ft22k,True,hf_img_cls,False,linalg,86M,"image-classification,transformer-encoder,bert-variant,vision-transformer",N/A
nvidia/mit-b0,True,hf_img_cls,False,linalg,3.7M,"image-classification,transformer-encoder",SegFormer
mnasnet1_0,False,vision,True,linalg,-,"cnn, torchvision, mobile, architecture-search","Outperforms other mobile CNNs on Accuracy vs. Latency"
resnet50_fp16,False,vision,True,linalg,23M,"cnn,image-classification,residuals,resnet-variant","Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)"
bert-base-uncased_fp16,True,fp16,False,linalg,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
bert-large-uncased,True,hf,True,linalg,330M,"nlp;bert-variant;transformer-encoder","24 layers, 1024 hidden units, 16 attention heads"
bert-base-uncased,True,hf,False,stablehlo,109M,"nlp;bert-variant;transformer-encoder","12 layers; 768 hidden; 12 attention heads"
1 model_name use_tracing model_type dynamic mlir_type param_count tags notes
2 efficientnet_b0 True vision False linalg 5.3M image-classification;cnn;conv2d;depthwise-conv Smallest EfficientNet variant with 224x224 input
3 efficientnet_b7 True vision False linalg 66M image-classification;cnn;conv2d;depthwise-conv Largest EfficientNet variant with 600x600 input
4 microsoft/MiniLM-L12-H384-uncased True hf True linalg 66M nlp;bert-variant;transformer-encoder Large version has 12 layers; 384 hidden size; Smaller than BERTbase (66M params vs 109M params)
5 bert-base-uncased True hf True linalg 109M nlp;bert-variant;transformer-encoder 12 layers; 768 hidden; 12 attention heads
6 bert-base-cased True hf True linalg 109M nlp;bert-variant;transformer-encoder 12 layers; 768 hidden; 12 attention heads
7 google/mobilebert-uncased True hf True linalg 25M nlp,bert-variant,transformer-encoder,mobile 24 layers, 512 hidden size, 128 embedding
8 alexnet False vision True linalg 61M cnn,parallel-layers The CNN that revolutionized computer vision (move away from hand-crafted features to neural networks),10 years old now and probably no longer used in prod.
9 resnet18 False vision True linalg 11M cnn,image-classification,residuals,resnet-variant 1 7x7 conv2d and the rest are 3x3 conv2d
10 resnet50 False vision True linalg 23M cnn,image-classification,residuals,resnet-variant Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)
11 resnet101 False vision True linalg 29M cnn,image-classification,residuals,resnet-variant Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)
12 squeezenet1_0 False vision True linalg 1.25M cnn,image-classification,mobile,parallel-layers Parallel conv2d (1x1 conv to compress -> (3x3 expand | 1x1 expand) -> concat)
13 wide_resnet50_2 False vision True linalg 69M cnn,image-classification,residuals,resnet-variant Resnet variant where model depth is decreased and width is increased.
14 mobilenet_v3_small False vision True linalg 2.5M image-classification,cnn,mobile N/A
15 google/vit-base-patch16-224 True hf_img_cls False linalg 86M image-classification,vision-transformer,transformer-encoder N/A
16 microsoft/resnet-50 True hf_img_cls False linalg 23M image-classification,cnn,residuals,resnet-variant Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)
17 facebook/deit-small-distilled-patch16-224 True hf_img_cls False linalg 22M image-classification,vision-transformer,cnn N/A
18 microsoft/beit-base-patch16-224-pt22k-ft22k True hf_img_cls False linalg 86M image-classification,transformer-encoder,bert-variant,vision-transformer N/A
19 nvidia/mit-b0 True hf_img_cls False linalg 3.7M image-classification,transformer-encoder SegFormer
20 mnasnet1_0 False vision True linalg - cnn, torchvision, mobile, architecture-search Outperforms other mobile CNNs on Accuracy vs. Latency
21 resnet50_fp16 False vision True linalg 23M cnn,image-classification,residuals,resnet-variant Bottlenecks with only conv2d (1x1 conv -> 3x3 conv -> 1x1 conv blocks)
22 bert-base-uncased_fp16 True fp16 False linalg 109M nlp;bert-variant;transformer-encoder 12 layers; 768 hidden; 12 attention heads
23 bert-large-uncased True hf True linalg 330M nlp;bert-variant;transformer-encoder 24 layers, 1024 hidden units, 16 attention heads
24 bert-base-uncased True hf False stablehlo 109M nlp;bert-variant;transformer-encoder 12 layers; 768 hidden; 12 attention heads