[preetha-intel]

Details:

This PR lets the OpenVINO GGUF frontend serve as the translation layer for llama.cpp's ggml-openvino backend, in addition to loading GGUF files natively. llama.cpp hands the frontend a ggml graph through a decoder. The frontend converts that graph into an ov::Model, so the backend no longer needs its own op translators and fusion passes.

To do that, the frontend has to be independent of which decoder feeds it. Each op is now described by named attributes and live operand shapes. Both the native GGUF builder and the llama.cpp cgraph decoder can fill those in, and one translator body serves both.

1. Materialize strided and KV-cache VIEWs generically (acb90e0)

• VIEW materialization now depends only on the view's shapes, byte strides, offset and optional dynamic axis, not on which decoder produced it.
• Supported view kinds:
  • dense crops with a runtime dynamic axis
  • rectangular and strided crops
  • mixed-radix grid views for non-dense layouts such as fused QKV and interleaved projections
  • views of non-dense sources, re-expressed over the source's logical tensor and checked element by element before use
  • a kv_view descriptor for KV-cache reads with a runtime length and first stream
• Unsupported stride layouts are now rejected. Previously the inner-row flatten fallback could read row-crossing views as if they were contiguous, which was silently wrong.
• SSM_CONV indexes its operand shapes from the innermost axis, so a decoder can describe the conv weight with its own rank.
• When a translator throws, the error now reports the failing node and its operand shapes.

2. Describe operations with named attributes instead of op_case (5440037)

• Translators no longer dispatch on a decoder-assigned case number. Each op is described by named attributes and operand shapes, which both decoders supply.
  • VIEW and RESHAPE: slice/reshape descriptions, ggml strides and offset, kv_view, reshape_target, shape_axes and merge_heads.
  • ROPE: rope_mode is one of normal, neox, imrope or vision. The sin/cos tables are now memoized per configuration, replacing the shared graph scaffold.
  • FLASH_ATTN_EXT: it takes ggml's [B, n_head, n_tokens, head_size] operands, with has_mask and f32_attention. K/V produced by a head/token transpose are tiled before the transpose, so stateful-SDPA fusion, MakeStateful and SDPAToPagedAttention still match their patterns.
  • CPY: source_slice, reshape_target and in-place axis updates.
  • Other ops: GET_ROWS, POOL_2D, PERMUTE, CONT, SCALE, ROLL, SOLVE_TRI, ADD, REPEAT, TOP_K and others read explicit attributes and live operand shapes.
• In-place writes (SET_ROWS, CPY, SCALE) now rebind storage aliases in TranslateSession. The shared mask and RoPE scaffolding is removed.
• NodeContext drops get_op_case, restricts by-name inputs to operands, and adds a per-translation cache. The builder API drops its op_case argument.
• Tests and docs are updated to the attribute contract.

3. Translate ops added by the llama.cpp OpenVINO backend (b91ed60)

• Ports the ops the backend used to translate itself, so the frontend owns them:
  • CONV_2D_DW, CONV_3D, CONV_TRANSPOSE_1D/2D, IM2COL_3D
  • MEAN, SUM, DUP (as CONT)
  • the HARDSIGMOID, EXPM1, STEP and ROUND unary ops
• These read ggml's op parameters as typed attributes (conv_params, im2col_params) and live operand shapes, and compute in f32 regardless of device.
• UPSCALE also accepts a fixed target size (upscale_sizes) for single-operand graphs.

Llama.cpp PR enabling OV frontend usage : ravi9/llama.cpp#342