We're running LFM2.5-VL-1.6B in production on iOS via your MLX int4 artifact (works great: <1s/image, ~0.9 GB RSS). On Android, XNNPACK is our only option and it's an order of magnitude slower on our workloads.
From your FAQ and the shipped 0.9.3 binaries, the Android runtime already has the Vulkan backend statically linked (verified: backends/vulkan symbols present in libexecutorch.so), and Gemma-4 multimodal runs Vulkan by default on Android since v0.9.1 — so our understanding is the only missing piece is the exported artifact.
Two questions: (1) does LFM2.5's hybrid architecture (gated short-conv blocks + cache pattern) lower cleanly to the Vulkan backend at all? (2) Would you publish a vulkan/ export alongside xnnpack/ and mlx/? Happy to test a nightly on our device fleet and report numbers, or contribute a delegation report if we attempt the export ourselves first (per #1124 precedent). If the honest answer is "we tried and it doesn't lower / it was slower than XNNPACK" — equally valuable, saves us a redundant attempt.
(Not a dupe of #556 — that's the generic GPU exploration; this is a specific model × backend ask.)
We're running LFM2.5-VL-1.6B in production on iOS via your MLX int4 artifact (works great: <1s/image, ~0.9 GB RSS). On Android, XNNPACK is our only option and it's an order of magnitude slower on our workloads.
From your FAQ and the shipped 0.9.3 binaries, the Android runtime already has the Vulkan backend statically linked (verified:
backends/vulkansymbols present inlibexecutorch.so), and Gemma-4 multimodal runs Vulkan by default on Android since v0.9.1 — so our understanding is the only missing piece is the exported artifact.Two questions: (1) does LFM2.5's hybrid architecture (gated short-conv blocks + cache pattern) lower cleanly to the Vulkan backend at all? (2) Would you publish a
vulkan/export alongsidexnnpack/andmlx/? Happy to test a nightly on our device fleet and report numbers, or contribute a delegation report if we attempt the export ourselves first (per #1124 precedent). If the honest answer is "we tried and it doesn't lower / it was slower than XNNPACK" — equally valuable, saves us a redundant attempt.(Not a dupe of #556 — that's the generic GPU exploration; this is a specific model × backend ask.)