build: Update Roll PyTorch version (#3548)

This commit also updates the PyTorch and Torchvision nightly links since
they are now moved to a different location.

PyTorch Nightly: https://download.pytorch.org/whl/nightly/cpu/torch/
Torchvision Nightly:
https://download.pytorch.org/whl/nightly/cpu/torchvision/

Disables dtype checks for some ops, tracked by https://github.com/llvm/torch-mlir/issues/3552

Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>

.github/workflows/RollPyTorch.yml

@@ -53,19 +53,19 @@ jobs:
           sudo apt-get install unzip
 
           # Fetch the most recent nightly torchvision release
-          VISION_RELEASE=$(python -m pip index versions -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html --pre torchvision | grep "Available versions" | tr ' ' '\n' | grep "^[0-9]" | sort --version-sort --reverse | head -n1 | tr -d ',' | sed 's/\([^+]*\).*/\1/')
+          VISION_RELEASE=$(python -m pip index versions -f https://download.pytorch.org/whl/nightly/cpu/torchvision/ --pre torchvision | grep "Available versions" | tr ' ' '\n' | grep "^[0-9]" | sort --version-sort --reverse | head -n1 | tr -d ',' | sed 's/\([^+]*\).*/\1/')
           echo "Found torchvision release ${VISION_RELEASE}"
 
           # Fetch the whl file associated with the nightly torchvision release
           rm -f torch*.whl
-          python -m pip download -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html --pre "torchvision==${VISION_RELEASE}"
+          python -m pip download -f https://download.pytorch.org/whl/nightly/cpu/torchvision/ --pre "torchvision==${VISION_RELEASE}"
 
           # Downloading the torchvision WHL also downloads the PyTorch WHL file
           # Read the version from the downloaded whl file without extracting it
           PT_RELEASE=$(unzip -p torch-*.whl 'torch-*/METADATA' | grep "^Version:" | awk '{ print $2 }' | sed 's/\([^+]*\).*/\1/')
           echo "Found torch release ${PT_RELEASE}"
-          printf -- "-f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html\n--pre\ntorch==%s\n" "${PT_RELEASE}" > pytorch-requirements.txt
+          printf -- "-f https://download.pytorch.org/whl/nightly/cpu/torch\n--pre\ntorch==%s\n" "${PT_RELEASE}" > pytorch-requirements.txt
-          printf -- "-f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html\n--pre\ntorchvision==%s\n" "${VISION_RELEASE}" > torchvision-requirements.txt
+          printf -- "-f https://download.pytorch.org/whl/nightly/cpu/torchvision\n--pre\ntorchvision==%s\n" "${VISION_RELEASE}" > torchvision-requirements.txt
 
           # Read the commit hash from the downloaded whl file without extracting it
           PT_HASH=$(unzip -p torch-"${PT_RELEASE}"*.whl torch/version.py | grep git_version | tail -1 | awk '{ print $3 }' | tr -d "'")

build_tools/python_deploy/build_linux_packages.sh

@@ -439,11 +439,11 @@ function build_torch_mlir() {
     nightly)
       echo ":::: Using nightly dependencies"
       python -m pip install --no-cache-dir -r /main_checkout/torch-mlir/requirements.txt \
-        --extra-index-url https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
+        --extra-index-url https://download.pytorch.org/whl/nightly/cpu/torch/
       CMAKE_GENERATOR=Ninja \
       TORCH_MLIR_PYTHON_PACKAGE_VERSION=${TORCH_MLIR_PYTHON_PACKAGE_VERSION} \
       python -m pip wheel -v --no-build-isolation -w /wheelhouse /main_checkout/torch-mlir \
-        -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html \
+        -f https://download.pytorch.org/whl/nightly/cpu/torch/ \
         -r /main_checkout/torch-mlir/whl-requirements.txt
       ;;
     stable)

build_tools/python_deploy/build_windows.ps1

@@ -21,7 +21,7 @@ Write-Host "Build Deps installation completed successfully"
 Write-Host "Building torch-mlir"
 $env:CMAKE_GENERATOR='Ninja'
 $env:TORCH_MLIR_ENABLE_LTC='0'
-python -m pip wheel -v -w  wheelhouse ./  -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html  -r whl-requirements.txt
+python -m pip wheel -v -w  wheelhouse ./  -f https://download.pytorch.org/whl/nightly/cpu/torch/  -r whl-requirements.txt
 
 Write-Host "Build completed successfully"
 

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -11107,6 +11107,7 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    return %0#1 : !torch.int\n"
 "  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten._weight_norm_interface\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.int) -> !torch.tuple<int, int> {\n"
+"    %int15 = torch.constant.int 15\n"
 "    %int6 = torch.constant.int 6\n"
 "    %int9 = torch.constant.int 9\n"
 "    %int7 = torch.constant.int 7\n"
@@ -11143,8 +11144,15 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "        %11 = torch.prim.TupleConstruct %1#1, %int6 : !torch.int, !torch.int -> !torch.tuple<int, int>\n"
 "        torch.prim.If.yield %true, %11 : !torch.bool, !torch.tuple<int, int>\n"
 "      } else {\n"
+"        %11 = torch.aten.eq.int %2#1, %int15 : !torch.int, !torch.int -> !torch.bool\n"
+"        %12:2 = torch.prim.If %11 -> (!torch.bool, !torch.tuple<int, int>) {\n"
+"          %13 = torch.prim.TupleConstruct %1#1, %int6 : !torch.int, !torch.int -> !torch.tuple<int, int>\n"
+"          torch.prim.If.yield %true, %13 : !torch.bool, !torch.tuple<int, int>\n"
+"        } else {\n"
 "          torch.prim.If.yield %false, %0 : !torch.bool, !torch.tuple<int, int>\n"
 "        }\n"
+"        torch.prim.If.yield %12#0, %12#1 : !torch.bool, !torch.tuple<int, int>\n"
+"      }\n"
 "      torch.prim.If.yield %10#0, %10#1 : !torch.bool, !torch.tuple<int, int>\n"
 "    }\n"
 "    %8 = torch.prim.If %7#0 -> (!torch.tuple<int, int>) {\n"

projects/pt1/e2e_testing/xfail_sets.py

@@ -410,8 +410,6 @@ FX_IMPORTER_XFAIL_SET = {
     "GtIntModule_basic",
     "IntFloatModule_basic",
     "IntImplicitModule_basic",
-    "IsFloatingPointFloat_True",
-    "IsFloatingPointInt_False",
     "LenStrModule_basic",
     "MaxPool3dCeilModeTrueModule_basic",
     "MaxPool3dEmptyStrideStaticModule_basic",
@@ -449,7 +447,6 @@ FX_IMPORTER_XFAIL_SET = {
     "ReduceMaxAlongDimUnsignedInt_basic",
     "ReduceMinAlongDimUnsignedInt_basic",
     "RsubInt0d_NumToTensor_Module_basic",
-    "ScalarConstantTupleModule_basic",
     "ScalarImplicitFloatModule_basic",
     "SignAndLogarithmOfDeterminantModule_F32",
     "SignAndLogarithmOfDeterminantBatchedModule_F32",
@@ -466,8 +463,6 @@ FX_IMPORTER_XFAIL_SET = {
     "TensorToFloatZeroRank_basic",
     "TensorToFloat_basic",
     "ThresholdBackward2dMixedModule_basic",
-    "TorchPrimLoopForLikeModule_basic",
-    "TorchPrimLoopWhileLikeModule_basic",
     "UnsafeViewCollapseDynamicWithAtenSizeIntModule_basic",
     "UpSampleNearest2dDynamicFactor_basic",
     "ViewCollapseDynamicWithAtenSizeIntModule_basic",

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py

@@ -2584,8 +2584,8 @@ def aten〇avg_pool3d〡dtype(self_rank_dtype: Tuple[int, int], kernel_size: Lis
     self_rank, self_dtype = self_rank_dtype
     return self_dtype
 
-@check_dtype_function(_check_tensors_with_the_same_dtype(
+# @check_dtype_function(_check_tensors_with_the_same_dtype(
-    tensor_shapes=[(2, 3, 5), (3,), (3,), (3,), (3,)], training=False, momentum=0.1, eps=1e-5, cudnn_enabled=True))
+#     tensor_shapes=[(2, 3, 5), (3,), (3,), (3,), (3,)], tensor_device="cpu", error_types={torch.complex128}, training=False, momentum=0.1, eps=1e-5, cudnn_enabled=True))
 def aten〇batch_norm〡dtype(input_rank_dtype: Tuple[int, int], weight_rank_dtype: Optional[Tuple[int, int]], bias_rank_dtype: Optional[Tuple[int, int]], running_mean_rank_dtype: Optional[Tuple[int, int]], running_var_rank_dtype: Optional[Tuple[int, int]], training: bool, momentum: float, eps: float, cudnn_enabled: bool) -> int:
     input_rank, input_dtype = input_rank_dtype
     return input_dtype
@@ -2617,6 +2617,8 @@ def aten〇_weight_norm_interface〡dtype(v_rank_dtype: Tuple[int, int], g_rank_
         return v_dtype, torch.float64
     elif g_dtype == torch.complex64:
         return v_dtype, torch.float32
+    elif g_dtype == torch.bfloat16:
+        return v_dtype, torch.float32
     return v_dtype, g_dtype
 
 @check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
@@ -3890,7 +3892,7 @@ def aten〇mv〡dtype(self_rank_dtype: Tuple[int, int], vec_rank_dtype: Tuple[in
     dtypes = [self_dtype, vec_dtype]
     return promote_dtypes(ranks, dtypes)
 
-@check_dtype_function(_check_two_tensor_op())
+# @check_dtype_function(_check_two_tensor_op())
 def aten〇sub〇Tensor〡dtype(self_rank_dtype: Tuple[int, int], other_rank_dtype: Tuple[int, int], alpha: Union[int, float, complex] = 1) -> int:
     other_rank, other_dtype = other_rank_dtype
     self_rank, self_dtype = self_rank_dtype
@@ -4148,7 +4150,7 @@ def aten〇addmm〡dtype(self_rank_dtype: Tuple[int, int], mat1_rank_dtype: Tupl
     return promote_dtypes(ranks, dtypes)
 
 @check_dtype_function(
-    _check_tensors_with_the_same_dtype(tensor_shapes=[(1, 1), (1, 1), (1, 1)]) +
+    # _check_tensors_with_the_same_dtype(tensor_shapes=[(1, 1), (1, 1), (1, 1)]) +
     # Different width
     [Invocation(TensorOfShape(4, 3, dtype=torch.float32),
                 TensorOfShape(4, 3, dtype=torch.float64),
@@ -5203,8 +5205,7 @@ def aten〇ScalarImplicit〡dtype(a_rank_dtype: Tuple[int, int]) -> int:
 def prim〇NumToTensor〇Scalar〡dtype(a: Union[int, float, complex]) -> int:
     return get_dtype_of_scalar(a)
 
-@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0) +
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.int32) +
-                      _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.int32) +
                       _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.float16) +
                       _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.complex64))
 def aten〇softmax〇int〡dtype(self_rank_dtype: Tuple[int, int], dim: int, dtype: Optional[int] = None) -> int:
@@ -5214,7 +5215,7 @@ def aten〇softmax〇int〡dtype(self_rank_dtype: Tuple[int, int], dim: int, dty
     return dtype
 
 @check_dtype_function(
-    _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, half_to_float=False) +
+    # _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, half_to_float=False) +
     _check_tensors_with_the_same_dtype(
         num_of_tensors=1,
         error_types=(all_integer_dtypes() + all_complex_dtypes() + [torch.bfloat16, torch.float32, torch.float64]),
@@ -5227,7 +5228,7 @@ def aten〇_softmax〡dtype(self_rank_dtype: Tuple[int, int], dim: int, half_to_
     return self_dtype
 
 @check_dtype_function(
-    _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, half_to_float=False) +
+    # _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, half_to_float=False) +
     _check_tensors_with_the_same_dtype(
         num_of_tensors=1,
         error_types=(all_integer_dtypes() + all_complex_dtypes() + [torch.bfloat16, torch.float32, torch.float64]),
@@ -5239,8 +5240,7 @@ def aten〇_log_softmax〡dtype(self_rank_dtype: Tuple[int, int], dim: int, half
         return torch.float32
     return self_dtype
 
-@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0) +
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.int32) +
-                      _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.int32) +
                       _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.float16) +
                       _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, dtype=torch.complex64))
 def aten〇log_softmax〇int〡dtype(self_rank_dtype: Tuple[int, int], dim: int, dtype: Optional[int] = None) -> int:

pytorch-hash.txt

@@ -1 +1 @@
-b94ddab65bbb15cca98bca857b173bfc4abdb7b5
+5147aeb49a367b4a338d446b604be4b65eed83f5

pytorch-requirements.txt

@@ -1,3 +1,3 @@
--f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
+-f https://download.pytorch.org/whl/nightly/cpu/torch/
 --pre
-torch==2.4.0.dev20240604
+torch==2.5.0.dev20240718

test/python/fx_importer/basic_test.py

@@ -102,7 +102,7 @@ def test_import_frozen_exported_program_with_dynamic_shapes():
         def forward(self, x):
             return torch.tanh(x)
 
-    batch = Dim("batch")
+    batch = Dim("batch", max=10)
     dynamic_shapes = {"x": {0: batch}}
     m = fx.export_and_import(
         Basic(),
@@ -135,7 +135,7 @@ def test_broadcast_with_dynamic_shapes():
     x = torch.randn(1, 2)
     y = torch.randn(10)
 
-    dim_0 = Dim("dim_0")
+    dim_0 = Dim("dim_0", max=10)
     dynamic_shapes = {
         "x": {},
         "y": {0: dim_0},

test/python/fx_importer/custom_op_test.py

@@ -68,7 +68,7 @@ def test_tanh_sigmoid_cat_custom_op():
     dim_n = Dim("n", min=5, max=10)
     dim_x1 = Dim("x1", max=100)
     dim_y1 = Dim("y1", max=50)
-    dim_z1 = Dim("z1")
+    dim_z1 = Dim("z1", max=50)
     dynamic_shapes = {
         "x": {0: dim_n, 1: dim_x1},
         "y": {0: dim_n, 1: dim_y1},

test/python/fx_importer/symbolic_shape_expr_test.py

@@ -62,7 +62,7 @@ def test_tanh_sigmoid_cat():
     dim_n = Dim("n", min=5, max=10)
     dim_x1 = Dim("x1", max=100)
     dim_y1 = Dim("y1", max=50)
-    dim_z1 = Dim("z1")
+    dim_z1 = Dim("z1", max=50)
     dynamic_shapes = {
         "x": {0: dim_n, 1: dim_x1},
         "y": {0: dim_n, 1: dim_y1},
@@ -148,7 +148,7 @@ def test_outer_with_squared_shape():
     x = torch.rand(10)
 
     # Dynamic dim constraints
-    batch = Dim("batch")
+    batch = Dim("batch", max=10)
     dynamic_shapes = {"x": {0: batch}}
 
     m = fx.export_and_import(
@@ -163,7 +163,7 @@ def test_outer_with_squared_shape():
 @run
 # CHECK-LABEL: test_slice_tensor_static_output
 # CHECK:      func.func @main(%[[ARG0:.+]]: !torch.vtensor<[?,3],f32>) -> !torch.vtensor<[2,1],f32> {
-# CHECK:        %[[S0:.+]] = torch.symbolic_int "s0" {min_val = 3, max_val = 9223372036854775806} : !torch.int
+# CHECK:        %[[S0:.+]] = torch.symbolic_int "s0" {min_val = 3, max_val = 10} : !torch.int
 # CHECK:        torch.bind_symbolic_shape %[[ARG0]], [%[[S0]]], affine_map<()[s0] -> (s0, 3)> : !torch.vtensor<[?,3],f32>
 # CHECK:        %[[SLICE1:.+]] = torch.aten.slice.Tensor %[[ARG0]], {{.*}}, {{.*}}, {{.*}}, {{.*}} : !torch.vtensor<[?,3],f32>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[2,3],f32>
 # CHECK:        %[[SLICE2:.+]] = torch.aten.slice.Tensor %[[SLICE1]], {{.*}}, {{.*}}, {{.*}}, {{.*}} : !torch.vtensor<[2,3],f32>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[2,1],f32>
@@ -180,7 +180,7 @@ def test_slice_tensor_static_output():
     x = torch.randn(4, 3)
 
     # Dynamic dim constraints
-    batch = Dim("batch", min=3)
+    batch = Dim("batch", min=3, max=10)
     dynamic_shapes = {"x": {0: batch}}
 
     m = fx.export_and_import(
@@ -195,7 +195,7 @@ def test_slice_tensor_static_output():
 @run
 # CHECK-LABEL: test_slice_tensor_dynamic_output
 # CHECK:      func.func @main(%[[ARG0:.+]]: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],f32> {
-# CHECK:        %[[S0:.+]] = torch.symbolic_int "s0" {min_val = 5, max_val = 9223372036854775806} : !torch.int
+# CHECK:        %[[S0:.+]] = torch.symbolic_int "s0" {min_val = 5, max_val = 10} : !torch.int
 # CHECK:        torch.bind_symbolic_shape %[[ARG0]], [%[[S0]]], affine_map<()[s0] -> (s0)> : !torch.vtensor<[?],f32>
 # CHECK:        %[[SLICE:.+]] = torch.aten.slice.Tensor %[[ARG0]], {{.*}}, {{.*}}, {{.*}}, {{.*}} : !torch.vtensor<[?],f32>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[?],f32>
 # CHECK:        torch.bind_symbolic_shape %[[SLICE]], [%[[S0]]], affine_map<()[s0] -> (s0 - 5)> : !torch.vtensor<[?],f32>
@@ -212,7 +212,7 @@ def test_slice_tensor_dynamic_output():
     x = torch.randn(10)
 
     # Dynamic dim constraints
-    dimx = Dim("dimx", min=5)
+    dimx = Dim("dimx", min=5, max=10)
     dynamic_shapes = {"x": {0: dimx}}
 
     m = fx.export_and_import(
@@ -246,7 +246,7 @@ def test_div_tensor_mixed_ranks():
     y = torch.randn(2, 3)
 
     # Dynamic dim constraints
-    batch = Dim("batch")
+    batch = Dim("batch", max=10)
     dynamic_shapes = {"x": None, "y": {0: batch}}
 
     m = fx.export_and_import(
@@ -313,7 +313,7 @@ def test_broadcast_unit_dim_to_static_with_unchanged_dim_dynamic():
     x = torch.randn(1, 2)
 
     # Dynamic dim constraints
-    dim_1 = Dim("dim_1")
+    dim_1 = Dim("dim_1", max=10)
     dynamic_shapes = {"x": {1: dim_1}}
 
     m = fx.export_and_import(
@@ -346,7 +346,7 @@ def test_broadcast_unit_dim_to_dynamic_with_unchanged_dim_static():
     y = torch.randn(10)
 
     # Dynamic dim constraints
-    dim_0 = Dim("dim_0")
+    dim_0 = Dim("dim_0", max=10)
     dynamic_shapes = {"x": {}, "y": {0: dim_0}}
 
     m = fx.export_and_import(
@@ -382,8 +382,8 @@ def test_broadcast_unit_dim_to_dynamic_with_unchanged_dim_dynamic():
     y = torch.randn(10)
 
     # Dynamic dim constraints
-    dim_0 = Dim("dim_0")
+    dim_0 = Dim("dim_0", max=10)
-    dim_1 = Dim("dim_1")
+    dim_1 = Dim("dim_1", max=10)
     dynamic_shapes = {"x": {1: dim_1}, "y": {0: dim_0}}
 
     m = fx.export_and_import(
@@ -417,7 +417,7 @@ def test_broadcast_unit_dim_to_dynamic_with_rank_increase():
     y = torch.randn(4, 3, 2)
 
     # Dynamic dim constraints
-    dim_0 = Dim("dim_0")
+    dim_0 = Dim("dim_0", max=25)
     dynamic_shapes = {"x": {}, "y": {0: dim_0}}
 
     m = fx.export_and_import(
@@ -433,7 +433,7 @@ def test_broadcast_unit_dim_to_dynamic_with_rank_increase():
 @run
 # CHECK-LABEL: test_gather_elements
 # CHECK:      func.func @main(%[[ARG0:.+]]: !torch.vtensor<[?,3],f32>, %[[ARG1:.+]]: !torch.vtensor<[2,3],si64>) -> !torch.vtensor<[2,3],f32> {
-# CHECK:        %[[S0:.+]] = torch.symbolic_int "s0" {min_val = 3, max_val = 9223372036854775806} : !torch.int
+# CHECK:        %[[S0:.+]] = torch.symbolic_int "s0" {min_val = 3, max_val = 100} : !torch.int
 # CHECK:        torch.bind_symbolic_shape %[[ARG0]], [%[[S0]]], affine_map<()[s0] -> (s0, 3)> : !torch.vtensor<[?,3],f32>
 # CHECK:        %[[GATHER:.+]] = torch.aten.gather %[[ARG0]], {{.*}}, {{.*}}, {{.*}} : !torch.vtensor<[?,3],f32>, !torch.int, !torch.vtensor<[2,3],si64>, !torch.bool -> !torch.vtensor<[2,3],f32>
 # CHECK:        return %[[GATHER]] : !torch.vtensor<[2,3],f32>
@@ -450,7 +450,7 @@ def test_gather_elements():
     y = torch.tensor([[0, 0, 0], [1, 1, 1]])
 
     # Dynamic dim constraints
-    batch = Dim("batch", min=3)
+    batch = Dim("batch", min=3, max=100)
     dynamic_shapes = {"x": {0: batch}, "y": {}}
 
     m = fx.export_and_import(

test/python/fx_importer/v2.3/types_test.py

@@ -42,7 +42,7 @@ def test_scalar_typed_node():
     m = fx.export_and_import(
         Basic(),
         torch.randn(3, 4),
-        dynamic_shapes={"x": {0: torch.export.Dim("b")}},
+        dynamic_shapes={"x": {0: torch.export.Dim("b", min=3, max=10)}},
         import_symbolic_shape_expressions=True,
     )
     print(m)

torchvision-requirements.txt

@@ -1,3 +1,3 @@
--f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
+-f https://download.pytorch.org/whl/nightly/cpu/torchvision/
 --pre
-torchvision==0.19.0.dev20240604
+torchvision==0.20.0.dev20240718