iree
iree copied to clipboard
iree-org
[hf-model] error: expected sizes to be non-negative, but got -1
What happened?
For the given IR
module {
func.func @main_graph(%arg0: !torch.vtensor<[?,?,?,?],f32>, %arg1: !torch.vtensor<[11,1,1,384],f32>, %arg2: !torch.vtensor<[?,?,?,?],f32>, %arg3:!torch.vtensor<[11,1,100,384],f32>, %arg4: !torch.vtensor<[?,?,?],f32>) -> !torch.vtensor<[11,1,?,384],f32> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.6.0"} {
%136 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<11x1x1x384xf32>} : () -> !torch.vtensor<[11,1,1,384],f32>
%137 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<11x1x100x384xf32>} : () -> !torch.vtensor<[11,1,100,384],f32>
%138 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<11x384x32x54xf32>} : () -> !torch.vtensor<[11,384,32,54],f32>
%139 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1> : tensor<2xsi64>} : () -> !torch.vtensor<[2],si64>
%none = torch.constant.none
%219 = torch.operator "onnx.Shape"(%arg0) : (!torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[4],si64>
%220 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__1> : tensor<si64>} : () -> !torch.vtensor<[],si64>
%221 = torch.operator "onnx.Gather"(%219, %220) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[4],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[],si64>
%223 = torch.operator "onnx.Shape"(%arg0) : (!torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[4],si64>
%224 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__3> : tensor<si64>} : () -> !torch.vtensor<[],si64>
%225 = torch.operator "onnx.Gather"(%223, %224) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[4],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[],si64>
%270 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__23> : tensor<si64>} : () -> !torch.vtensor<[],si64>
%271 = torch.operator "onnx.Div"(%221, %270) : (!torch.vtensor<[],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[],si64>
%274 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__24> : tensor<si64>} : () -> !torch.vtensor<[],si64>
%275 = torch.operator "onnx.Div"(%225, %274) : (!torch.vtensor<[],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[],si64>
%283 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__27> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%285 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__28> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%287 = torch.operator "onnx.Concat"(%283, %285) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[2],si64>
%302 = torch.operator "onnx.Cast"(%287) {torch.onnx.to = 7 : si64} : (!torch.vtensor<[2],si64>) -> !torch.vtensor<[2],si64>
%303 = torch.operator "onnx.Concat"(%139, %302) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[2],si64>, !torch.vtensor<[2],si64>) -> !torch.vtensor<[4],si64>
%304 = torch.operator "onnx.Resize"(%138, %none, %none, %303) {torch.onnx.coordinate_transformation_mode = "half_pixel", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.mode = "cubic", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[11,384,32,54],f32>, !torch.none, !torch.none, !torch.vtensor<[4],si64>) -> !torch.vtensor<[?,?,?,?],f32>
%305 = torch.operator "onnx.Shape"(%304) : (!torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[4],si64>
%306 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__33> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%307 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__34> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%308 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__35> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%309 = torch.operator "onnx.Slice"(%305, %307, %308, %306) : (!torch.vtensor<[4],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[2],si64>
%310 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__36> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%311 = torch.operator "onnx.Concat"(%309, %310) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[2],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[3],si64>
%312 = torch.operator "onnx.Reshape"(%304, %311) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[?,?,?,?],f32>, !torch.vtensor<[3],si64>) -> !torch.vtensor<[?,?,?],f32>
%313 = torch.operator "onnx.Transpose"(%312) {torch.onnx.perm = [0 : si64, 2 : si64, 1 : si64]} : (!torch.vtensor<[?,?,?],f32>) -> !torch.vtensor<[?,?,?],f32>
%314 = torch.operator "onnx.Mul"(%271, %275) : (!torch.vtensor<[],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[],si64>
%315 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__37> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%316 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__38> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%317 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__39> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%318 = torch.operator "onnx.Unsqueeze"(%314, %317) : (!torch.vtensor<[],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1],si64>
%319 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__40> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%320 = torch.operator "onnx.Concat"(%315, %316, %318, %319) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[4],si64>
%321 = torch.operator "onnx.Reshape"(%313, %320) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[?,?,?],f32>, !torch.vtensor<[4],si64>) -> !torch.vtensor<[?,?,?,?],f32>
%322 = torch.operator "onnx.Concat"(%136, %321, %137) {torch.onnx.axis = 2 : si64} : (!torch.vtensor<[11,1,1,384],f32>, !torch.vtensor<[?,?,?,?],f32>, !torch.vtensor<[11,1,100,384],f32>) -> !torch.vtensor<[11,1,?,384],f32>
return %322: !torch.vtensor<[11,1,?,384],f32>
}
}
{-#
dialect_resources: {
builtin: {
__1: "0x080000000200000000000000",
__3: "0x080000000300000000000000",
__23: "0x080000001000000000000000",
__24: "0x080000001000000000000000",
__27: "0x080000000000000000000000",
__28: "0x080000000000000000000000",
__33: "0x080000000000000000000000",
__34: "0x080000000000000000000000",
__35: "0x080000000200000000000000",
__36: "0x08000000FFFFFFFFFFFFFFFF",
__37: "0x080000000B00000000000000",
__38: "0x080000000100000000000000",
__39: "0x080000000000000000000000",
__40: "0x080000008001000000000000"
}
}
#-}
getting error as
error: expected sizes to be non-negative, but got -1
%322 = torch.operator "onnx.Concat"(%136, %321, %137) {torch.onnx.axis = 2 : si64} : (!torch.vtensor<[11,1,1,384],f32>, !torch.vtensor<[?,?,?,?],f32>, !torch.vtensor<[11,1,100,384],f32>) -> !torch.vtensor<[11,1,?,384],f32>
during iree-flow-canonicalization post CSE.
command:
iree-compile tt.mlir --iree-hal-target-backends=llvm-cpu --iree-llvmcpu-target-cpu=host -o abc.vmfb
IREE version: IREE compiler version 3.1.0rc20241217 @ 362b554894c46021d32749bf01c9c4410f8cbbc4
model: From HF top 1000 most downloaded models(hf_yolos-small-finetuned-license-plate-detection)
dump with '--mlir-print-ir-after-all --mlir-print-ir-before-all --mlir-disable-threading --mlir-elide-elementsattrs-if-larger=4'
Steps to reproduce your issue
- Go to '...'
- Click on '....'
- Scroll down to '....'
- See error
What component(s) does this issue relate to?
Compiler
Version information
No response
Additional context
No response
TLDR: the input IR is invalid.
%304 = torch.operator "onnx.Resize"(%138, %none, %none, %303) {torch.onnx.coordinate_transformation_mode = "half_pixel", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.mode = "cubic", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[11,384,32,54],f32>, !torch.none, !torch.none, !torch.vtensor<[4],si64>) -> !torch.vtensor<[?,?,?,?],f32>
becomes:
%28 = torch.vtensor.literal(dense<0> : tensor<1xsi64>) : !torch.vtensor<[1],si64>
%29 = torch.vtensor.literal(dense<0> : tensor<1xsi64>) : !torch.vtensor<[1],si64>
%30 = torch.prim.ListConstruct %28, %29 : (!torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.list<vtensor>
%int0_4 = torch.constant.int 0
%31 = torch.aten.cat %30, %int0_4 : !torch.list<vtensor>, !torch.int -> !torch.vtensor<[2],si64>
%int4 = torch.constant.int 4
%none_5 = torch.constant.none
%false = torch.constant.bool false
%32 = torch.aten.to.dtype %31, %int4, %false, %false, %none_5 : !torch.vtensor<[2],si64>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[2],si64>
%33 = torch.prim.ListConstruct %3, %32 : (!torch.vtensor<[2],si64>, !torch.vtensor<[2],si64>) -> !torch.list<vtensor>
%int0_6 = torch.constant.int 0
%34 = torch.aten.cat %33, %int0_6 : !torch.list<vtensor>, !torch.int -> !torch.vtensor<[4],si64>
%none_7 = torch.constant.none
%false_8 = torch.constant.bool false
%true = torch.constant.bool true
%str = torch.constant.str "cubic"
%int0_9 = torch.constant.int 0
%int2 = torch.constant.int 2
%35 = torch.aten.select.int %34, %int0_9, %int2 : !torch.vtensor<[4],si64>, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%36 = torch.aten.item %35 : !torch.vtensor<[1],si64> -> !torch.int
%int3 = torch.constant.int 3
%37 = torch.aten.select.int %34, %int0_9, %int3 : !torch.vtensor<[4],si64>, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%38 = torch.aten.item %37 : !torch.vtensor<[1],si64> -> !torch.int
%39 = torch.prim.ListConstruct %36, %38 : (!torch.int, !torch.int) -> !torch.list<int>
%40 = torch.aten.__interpolate.size_list_scale_list %2, %39, %none_7, %str, %false_8, %none_7, %false_8 : !torch.vtensor<[11,384,32,54],f32>, !torch.list<int>, !torch.none, !torch.str, !torch.bool, !torch.none, !torch.bool -> !torch.vtensor<[?,?,?,?],f32>
which simplifies to:
%22 = torch.prim.ListConstruct %int0, %int0 : (!torch.int, !torch.int) -> !torch.list<int>
%23 = torch.shape.calculate {
%100 = torch.aten.__interpolate.size_list_scale_list %13, %22, %none, %str_3, %false, %none, %false : !torch.vtensor<[11,384,32,54],f32>, !torch.list<int>, !torch.none, !torch.str, !torch.bool, !torch.none, !torch.bool -> !torch.vtensor<[11,384,0,0],f32>
torch.shape.calculate.yield %100 : !torch.vtensor<[11,384,0,0],f32>
} shapes {
%100 = torch.prim.ListConstruct %int11, %int384, %int0, %int0 : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
torch.shape.calculate.yield.shapes %100 : !torch.list<int>
} : !torch.vtensor<[11,384,0,0],f32>
The arguments to interpolate are invalid, %22 here is a list of (0,0), resulting in an invalid shape with 0s in ranks 2 and 3.
It is semantically incorrect in pytorch also:
>>> a = torch.ones([11,384,32,54])
>>> b = torch.nn.functional.interpolate(a, (0,0), None, "bicubic", False, None, False)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/dan/shark-ai/.env/lib/python3.12/site-packages/torch/nn/functional.py", line 4594, in interpolate
return torch._C._nn.upsample_bicubic2d(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Input and output sizes should be greater than 0, but got input (H: 32, W: 54) output (H: 0, W: 0)
>>>
This means either the model export in onnx or the lowering in the convert-torch-onnx-to-torch pass is incorrect
This reproducer is likely not sufficient to narrow down the issue. It looks more like an issue with shape inference, and not knowing the source for %arg0 may be problematic.
I'd recommend starting with the full model to debug this issue.
Some context https://discord.com/channels/689900678990135345/1398431144210333818