ggml : WebGPU add TRANSPOSE and RESHAPE to supported ops

[danbev]

This commit adds support for the TRANSPOSE and RESHAPE operations in the
ggml webgpu backend.

Co-authored-by: Diego Devesa [email protected]

[slaren]

This is still a bug somewhere, it should not be hidden by disabling the test.

[danbev]

This is still a bug somewhere, it should not be hidden by disabling the test.

My intention was to not to disable the test, but instead for WebGPU, set flash attention to off. My reasoning here was that the default was previously off unless I'm mistaken, and that the recent change to enable flash attention by default might be what is causing this test to start to fail. I was more curious if this would allow the test to pass.

[slaren]

The intention is to enable flash attention only if the backend supports it. If doing that check causes the backend to crash, then that indicates a problem somewhere, and should not be hidden.

Try this change to fix it instead:

diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index 32f1e304e..4e3f152a7 100644
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -1062,6 +1062,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_NONE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_RESHAPE:
             return true;
         case GGML_OP_CPY:
         case GGML_OP_SET_ROWS:

[CISC]

It should also be added here (and return true (not that it matters (yet) as the return is not checked)):

llama.cpp/ggml/src/ggml-webgpu/ggml-webgpu.cpp

Lines 611 to 614 in ea8412f

	case GGML_OP_NONE:
	case GGML_OP_VIEW:
	case GGML_OP_PERMUTE:
	return false;