QVAC-19119 test(llm-llamacpp): overlay qvac-fabric for Qwen3VL multi-tile batching

[yingying0906]

What problem does this PR solve?

This PR is for CI testing only and will not be merged.

1. Qwen3.5 lacked multi-tile support. Images were always processed as a single tile regardless of resolution, losing detail on high-resolution inputs.
2. Even where multi-tiling existed, the vision encoder ran N sequential forward passes, one per tile. On GPU backends this is inefficient; a single batched dispatch over all tiles produces identical embeddings with far less overhead.

How does it solve it?

Implements multi-tile image preprocessing for Qwen3.5 (using the qwen3vl encoder path it shares with Qwen3VL), and replaces the N sequential vision encoder forward passes with a single batched forward pass over all tiles (batch dim = N tiles).

A new --image-tile-mode flag controls the dispatch mode:

Mode	Behaviour
batched	Single GPU forward pass over all tiles
sequential	Original N sequential passes (for benchmarking)
baseline	Single-tile dyn_size path, matches current master behaviour (default)

Preliminary benchmark (Qwen3.5-2B Q4_K_M, F16 mmproj, 1920x1280 image, 4 tiles):

MacBook (M5 Pro, Metal):

Variant	Encode (ms)	Prompt eval (ms)	Decode (tok/s)
Baseline (master)	3,503	5,104	63.6
Sequential	1,723	3,258	62.4
Batched	1,762	3,297	63.0

Samsung S25 Ultra (Adreno 830, OpenCL):

Variant	Encode (ms)	Prompt eval (ms / tok)	Prompt eval (tok/s)
Baseline (master)	57,288	199,552 / 2418 tok	12.1
Sequential	24,991	139,200 / 2322 tok	16.7
Batched	15,865	117,811 / 2322 tok	19.7

Breaking changes

None. To revert: delete packages/llm-llamacpp/vcpkg/ports/qvac-fabric/ and remove the VCPKG_OVERLAY_PORTS line from CMakeLists.txt.

🤖 Generated with Claude Code