bart-large model support #225
Assignees
Labels
Comments
6 tasks
|
@bjacob In iree code, what's the difference between: Here is the link to download all the mlir files: bart-large.default.torch-onnx.mlir bart-large.default.pytorch.torch.mlir bart-large.default.pytorch.linalg.mlir |
|
Sorry I'm unfamiliar with torch-mlir and probably not the best person to help here, but here is my best attempt. Looking into each error specifically, Run without --torchtolinalg get iree issue1bart-large.default.pytorch.torch.mlir:466:12: error: 'linalg.generic' op inferred input/output operand #1 has shape's dimension #1 to be 16, but found 1
%378 = torch.aten.add.Tensor %377, %239, %int1_71 : !torch.vtensor<[1,16,9,9],f32>, !torch.vtensor<[?,?,9,9],f32>, !torch.int -> !torch.vtensor<[?,16,9,9],f32>
^
bart-large.default.pytorch.torch.mlir:466:12: note: see current operation:
%748 = "linalg.generic"(%746, %642, %747) <{indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = [#linalg.iterator_type<parallel>, #linalg.iterator_type<parallel>, #linalg.iterator_type<parallel>, #linalg.iterator_type<parallel>], operandSegmentSizes = array<i32: 2, 1>}> ({
^bb0(%arg1: f32, %arg2: f32, %arg3: f32):
%2708 = "arith.addf"(%arg1, %arg2) <{fastmath = #arith.fastmath<none>}> : (f32, f32) -> f32
"linalg.yield"(%2708) : (f32) -> ()
}) : (tensor<1x16x9x9xf32>, tensor<1x1x9x9xf32>, tensor<1x16x9x9xf32>) -> tensor<1x16x9x9xf32>This is noting that the value
So the size of the dimention indexed by Since this Run with --torchtolinalg get iree issue2bart-large.default.pytorch.linalg.mlir:333:11: error: 'tensor.expand_shape' op expected dimension 0 of collapsed type to be dynamic since one or more of the corresponding dimensions in the expanded type is dynamic
%27 = linalg.generic {indexing_maps = [#map10, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%26 : tensor<?x9x1xf32>) outs(%24 : tensor<?x9x1xf32>) {
^
bart-large.default.pytorch.linalg.mlir:22:3: note: called from
func.func @main_graph(%arg0: tensor<1x9xi64>) -> (tensor<1x9x50265xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>, tensor<1x16x9x64xf32>) {
^
bart-large.default.pytorch.linalg.mlir:333:11: note: see current operation: %254 = "tensor.expand_shape"(%253) <{reassociation = [[0, 1]]}> : (tensor<9xf32>) -> tensor<?x9xf32>
%27 = linalg.generic {indexing_maps = [#map10, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%26 : tensor<?x9x1xf32>) outs(%24 : tensor<?x9x1xf32>) {
^In this error message, the dumped input is already %27 = linalg.generic {indexing_maps = [#map10, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%26 : tensor<?x9x1xf32>) outs(%24 : tensor<?x9x1xf32>) {But the error says that its lowering includes an ill-formed bart-large.default.pytorch.linalg.mlir:333:11: error: 'tensor.expand_shape' op expected dimension 0 of collapsed type to be dynamic since one or more of the corresponding dimensions in the expanded type is dynamic
[...]
bart-large.default.pytorch.linalg.mlir:333:11: note: see current operation: %254 = "tensor.expand_shape"(%253) <{reassociation = [[0, 1]]}> : (tensor<9xf32>) -> tensor<?x9xf32>Indeed, the 1D static shape You could find which pass created it by rerunning this |
|
@Shukla-Gaurav With iree-compiler bump torch-mlir to onnx.resize patch iree-org/iree#17358: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
torchtolinalg pipeline issue
Related iree issues: iree-org/iree#17021
The following two run should generate same issue, but it turns out not
Run without --torchtolinalg get iree issue1
python ./run.py --torchmlirbuild ../../torch-mlir/build --tolerance 0.001 0.001 --cachedir ./huggingface_cache --ireebuild ../../iree-build -f pytorch -g models --mode onnx --report --tests pytorch/models/bart-largeCoresponding command that run:
Use standalone torch-mlir-opt lowering onnx dialect to torch dialect, then use iree-compile lower torch dialect to vm
Option2: Run with
--torchtolinalgget iree issue2python ./run.py --torchmlirbuild ../../torch-mlir/build --tolerance 0.001 0.001 --cachedir ./huggingface_cache --ireebuild ../../iree-build -f pytorch -g models --mode onnx --report --tests pytorch/models/bart-large --torchtolinalgCoresponding command that run:
Use standalone torch-mlir-opt lowering onnx dialect to linalg dialect, then use iree-compile lower linalg dialect to vm
Here is the link to download all the mlir files:
bart-large.default.torch-onnx.mlirbart-large.default.pytorch.torch.mlirbart-large.default.pytorch.linalg.mlirhttps://onnxstorage.blob.core.windows.net/onnxstorage/bugcases/torchtolinalgpipelineissue.zip
The text was updated successfully, but these errors were encountered: