Instantiating Slang defined struct that contains Tensor in Python causes problems

[StreamVersus]

About: Linux, Python3.13, latest SlangPY, tested on Vulkan and cpu - same issues

Struct for testing:

import slangpy;

struct Weights
{
    ITensor<float, 2> weights;
}

Three distinct cases:

1. Using numpy array as data:

device = spy.create_device(include_paths=[pathlib.Path(__file__).parent])
layers = spy.Module.load_from_file(device, "layers.slang")

# ================================================================
# Helpers
# ================================================================

def tensor(data: np.ndarray) -> spy.Tensor:
    return spy.Tensor.from_numpy(device, data)

def randn(*shape) -> np.ndarray:
    return np.random.randn(*shape).astype(np.float32)

weights = layers.Weights(
    weights = randn(VOCAB_SIZE, EMBED_DIM)
)

causes:

RuntimeError: numpy array shape dim 1 does not match the expected shape (-1 != 64)

-1 is dimensions that it expects, it might happen if check assumes tensor is static/we try to replace it, but here its uninitialized

2. Using tensor as data:

weights = layers.Weights(
    weights = tensor(randn(VOCAB_SIZE, EMBED_DIM)) # added tensor() wrap to first one
)

[ERROR] (rhi) slang: error[E41014]: type cannot be packed for dynamic dispatch
   --> /home/stream/projects/venv/lib/python3.13/site-packages/slangpy/slang/tensor.slang:286:15
    |
286 | public struct Tensor<T, let D : int> : ITensor<T, D>
    |               ^^^^^^ type 'Tensor' contains fields that cannot be packed into ordinary bytes for dynamic dispatch.
----'

Traceback (most recent call last):
  File "/home/stream/projects/kernels/training_test.py", line 33, in <module>
    weights = layers.Weights(
        weights = tensor(randn(VOCAB_SIZE, EMBED_DIM))
    )
  File "/home/stream/projects/venv/lib/python3.13/site-packages/slangpy/core/struct.py", line 204, in __call__
    return self.__getattr__("__init")(*args, **kwds)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
RuntimeError: m_rhi_command_encoder->finish(rhi_command_buffer.writeRef()) failed with error: -2147467259 (SLANG_FAIL)

RHI messages:
slang: error[E41014]: type cannot be packed for dynamic dispatch
   --> /home/stream/projects/venv/lib/python3.13/site-packages/slangpy/slang/tensor.slang:286:15
    |
286 | public struct Tensor<T, let D : int> : ITensor<T, D>
    |               ^^^^^^ type 'Tensor' contains fields that cannot be packed into ordinary bytes for dynamic dispatch.
----'


/project/src/sgl/device/command.cpp:891 in function finish

3. Using compile-time sized arrays instead of ITensor simply causes freezes on big arrays, probably due to being allocated in other type of memory(on cpu it should be stack, causing stack overflows), it probably should fail with actual error, works on small sizes

It happens with any type of Tensor(ITensor, RW version, just Tensor, DiffTensor), and currently i see no way to actually make struct around Tensor.

[jkwak-work]

This issue is filed from a question on Discord, [link]

[jhelferty-nv]

@ccummingsNV any thoughts on this one? Is this expected to work?

[jhelferty-nv]

Discussed with Chris, this will need some debugging to understand what's going on with the default constructor. It sounds like we hadn't intended to make it work like that.