ggml: add ggml_can_fuse_subgraph

[am17an]

This PR adds ggml_can_fuse_subgraph which is a less strict extension of ggml_can_fuse. It checks given inputs/outputs of a subgraph, whether all the intermediate tensors can be fused. Putting as draft to iterate on the correct API

[jeffbolznv]

While this does check that the internal nodes aren't used outside of the fusion region, we also need to check that the internal connectivity of the graph is what we expect. IMO this necessarily has to be verbose, because we have to check that all the node->src[] values are what we expect them to be.

The first idea that comes to mind would be to pass in a list of triples, where each triple is a { dst_node, src_idx, src_node }, and verify that nodes[start + dst_node]->src[src_idx] == nodes[start + src_node].

[am17an]

While this does check that the internal nodes aren't used outside of the fusion region, we also need to check that the internal connectivity of the graph is what we expect. IMO this necessarily has to be verbose, because we have to check that all the node->src[] values are what we expect them to be.

I think this is better done at the caller site, which has more context about the fusion. This is just to avoid cases where the write is needed elsewhere we end up fusing it and have a common function for other sanity checks. Maybe a better a name for this function should be is_fusion_candidate. Passing triples like you mentioned is equivalent to building the graph at the caller site, but just not doing equality checks.

[jeffbolznv]

I think this is better done at the caller site

It's probably fine to have it as a separate function, but IMO it can still be common code (as much as any of this can be common code - there will always be special cases we want to handle differently).

[am17an]

Maybe need to walk the tree till we get the non-view parent instead of what I did here

[jeffbolznv]

I guess the most conservative thing would be to check all parents. It seems plausible we'll want to fuse a view of a view in the future.

[ggerganov]

I thought this logic guarantees that view_src is always the "root" source:

llama.cpp/ggml/src/ggml.c

Lines 1630 to 1646 in ad6af8f

	static struct ggml_tensor * ggml_new_tensor_impl(
	struct ggml_context * ctx,
	enum ggml_type type,
	int n_dims,
	const int64_t * ne,
	struct ggml_tensor * view_src,
	size_t view_offs) {
	GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
	GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
	// find the base tensor and absolute offset
	if (view_src != NULL && view_src->view_src != NULL) {
	view_offs += view_src->view_offs;
	view_src = view_src->view_src;
	}

So I don't think we can ever get node->view_src->view_src != NULL. But I guess it does not hurt to have this traversal.

[jeffbolznv]

I guess the most conservative thing would be to check all parents. It seems plausible we'll want to fuse a view of a view in the future.

[jeffbolznv]

I implemented logic to check the graph edges in master...jeffbolznv:llama.cpp:pull_16662_ps4. We can do this as a separate change, after these other things land.

[am17an]

Will merge once CI passes