Fix device placement of constants, shared inputs and delegate outputs by shoumikhin · Pull Request #24088 · pytorch/executorch

Normalize the dim order of parameters before delegation, like constants

shoumikhin

committed

4b22439

Keep a device round-trip when a host op writes the host copy

shoumikhin

committed

6a8207d

Copy a device delegate output to the host when it is used without getitem

shoumikhin

committed

9c26c3e

Return delegate outputs on the device whenever skip_d2h_for_method_outputs is set

shoumikhin

committed

8f40af7

Copy constants that feed a device delegate into device memory

shoumikhin

committed

ccab255

Let a method input feed several delegates on one device without a copy

shoumikhin

committed

73dc9f1

Copy a tensor to a device once when several delegates there read it

shoumikhin

committed

fd1ac9e

Keep in-place results on the host copy of a device delegate output

shoumikhin

committed

9b508c7

Count a host write through a view as a write to the tensor

shoumikhin

committed

ba5a303

Test tensors shared by delegates on different devices

shoumikhin

committed

0160993

Test device placement end to end with CUDA delegates

shoumikhin

committed

b10cc1e

Say when skip_d2h_for_method_outputs still copies a returned output

shoumikhin

committed

e9d8c0e

Ignore host writes that come after a device copy is read

shoumikhin

committed

2ff25a8

Share the device copy of a written output with a later delegate

shoumikhin

committed

a22b904

Check the placeholder, not the input spec, in convert_constant_dim_order_pass

shoumikhin

committed

e92206b

Run the CUDA unit tests when the device placement passes change

shoumikhin

committed

dda74d3

Warn when to_edge changes the layout of a parameter, buffer or constant

shoumikhin

committed

5f40202

Give a delegate whose target device is the CPU no device copies

shoumikhin

committed

825f187

Copy a device input to the host for readers off the device

shoumikhin

committed

d3d3978

Add error_on_per_call_copies to make copies the skip flags keep an error

shoumikhin

committed

ad9e14c

Shorten the device copy notes in the CUDA docs

shoumikhin

committed

2a8bff8

Return an output written on the host from the host

shoumikhin

committed

c367248

Describe every per-call copy left by the skip flags correctly

shoumikhin

committed

4041857

Fix a test comment about how often to_edge converts a constant

shoumikhin

committed

d517e67

Pin auto_functionalized v2 in the host writer tests

shoumikhin

committed

4c56ae2

Name the output in the warning for an output returned from the host

shoumikhin

committed

964a611

Navigation Menu

Repository navigation

Fix device placement of constants, shared inputs and delegate outputs - #24088

Fix device placement of constants, shared inputs and delegate outputs#24088
shoumikhin merged 26 commits into
pytorch:mainfrom
shoumikhin:device-pass-constants

Commits on Oct 26, 2026

Normalize the dim order of parameters before delegation, like constants

Keep a device round-trip when a host op writes the host copy

Copy a device delegate output to the host when it is used without getitem

Return delegate outputs on the device whenever skip_d2h_for_method_outputs is set

Copy constants that feed a device delegate into device memory

Let a method input feed several delegates on one device without a copy

Copy a tensor to a device once when several delegates there read it

Keep in-place results on the host copy of a device delegate output

Count a host write through a view as a write to the tensor

Test tensors shared by delegates on different devices

Test device placement end to end with CUDA delegates

Say when skip_d2h_for_method_outputs still copies a returned output

Ignore host writes that come after a device copy is read

Share the device copy of a written output with a later delegate

Check the placeholder, not the input spec, in convert_constant_dim_order_pass

Run the CUDA unit tests when the device placement passes change

Warn when to_edge changes the layout of a parameter, buffer or constant

Give a delegate whose target device is the CPU no device copies

Copy a device input to the host for readers off the device

Add error_on_per_call_copies to make copies the skip flags keep an error

Shorten the device copy notes in the CUDA docs

Return an output written on the host from the host

Describe every per-call copy left by the skip flags correctly

Fix a test comment about how often to_edge converts a constant

Pin auto_functionalized v2 in the host writer tests

Name the output in the warning for an output returned from the host