whisper : allocate encoder results in dedicated buffer

[ggerganov]

fix #1959

Allocate the encoder results tensors (embd_conv and embd_enc) in dedicated backend buffer. Copy the results into these tensors and used them in follow-up graphs

[ggerganov]

With the wstate.embd_conv now pre-allocated, this view still triggers the assert:

Assertion failed: (tensor_alloc->offset == SIZE_MAX), function ggml_gallocr_init_tensor, file ggml-alloc.c, line 739.

Not sure why

[slaren]

The code that stores the allocation location of leafs didn't take into account pre-allocated leafs. This should fix it:

diff --git a/ggml-alloc.c b/ggml-alloc.c
index 8ac1d3e..60b86c2 100644
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -701,8 +701,13 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
         galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
-        galloc->leaf_allocs[i].leaf.offset = hn->offset;
-        galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
     }

     // reallocate buffers if needed

[ggerganov]

With just this change applied on master the code also works (i.e. without pre-allocating the tensors). But if I understand correctly, it is technically not correct because nothing guarantees that the data in these tensors would not be overwritten by some ops. Since they are currently at the end of the computation graphs it seems to produce correct results, but the concern is that this is not very future-proof if we expand the graphs in the future - is that correct?

What if I use ggml_set_output(ctx0, wstate.embd_conv); in whisper_build_graph_conv() to guarantee that the data would not be overwritten?

[slaren]

Yes, with ggml_set_output it would guarantee that the tensor is never overwritten.

[ggerganov]

Would pre-allocating the tensors as proposed in this PR have any advantage over the ggml_set_output option? Since with this PR we now have to perform extra copy of the data between the graph calls, which otherwise is not needed.

[slaren]

I don't think there would be any significant advantage. Maybe it would allow using ggml_gallocr_reserve here instead of ggml_gallocr_alloc_graph, but the difference would be minimal:

whisper.cpp/whisper.cpp

Lines 496 to 502 in 66df44b

	// since there are dependencies between the different graphs,
	// we need to allocate them instead of only reserving to get the correct compute buffer size
	if (!ggml_gallocr_alloc_graph(alloc, get_graph())) {
	// failed to allocate the compute buffer
	WHISPER_LOG_ERROR("%s: failed to allocate the compute buffer\n", __func__);
	return false;
	}

[ggerganov]

Picked the ggml-alloc commit directly on master