f

Author: roberteg16

ggml/src/ggml-cuda/ggml-cuda-roofline.cpp

@@ -89,11 +89,9 @@ struct dispatch {
 // the node itself); in_storage_ids are the storages it reads from outside the span. The consumer
 // links each in_storage_id to the most recent prior launch whose out_storage_id matches (last
 // writer wins, which also resolves in-place ops whose output aliases an input).
-// Per external source, aligned 1:1 with in_storage_ids, so the consumer can label and classify
-// each operand without re-deriving it from the shared (head-only, for a fused row) geometry record:
-// the tensor name makes weight-vs-input unambiguous (a model weight is named e.g.
-// "blk.0.ffn_gate.weight"; a dynamic input is "inp_pos" / "cache_k_l0" / ...), and the type/shape
-// are the operand's own -- which for a fused span's non-head operands the geometry record omits.
+// One entry per external source (aligned 1:1 with in_storage_ids). Holds each operand's own
+// name/type/shape so the consumer can label and classify it (weight vs dynamic input, by name) --
+// which a fused row's head-only geometry record can't supply.
 struct src_operand {
     std::string name;
     std::string type;             // ggml_type_name