SAM 2's instance now reads wgpu's environment variables, as the render context's does, so a driver such as
Mesa's Dozen (WGPU_ALLOW_UNDERLYING_NONCOMPLIANT_ADAPTER=1) is visible to it, and WGPU_POWER_PREF applies.
An adapter without compute shaders is refused before a device is made on it, so the CPU takes over with one
line instead of a validation error per kernel. The refused adapter is kept for the life of the process, as
the device on it was: dropping it crashed the tests on WSL2's GL driver. Gpu::shared no longer repeats
"SAM 2 compute:" in its errors.
Candle adds a bias on one thread; the Linear layers now add it (and apply GELU
for the MLP in) in one parallel pass. Encode 2.4 s -> 2.1 s, memory attention
2.4 -> 2.3 s; reference tests unchanged. gpu_matmul example: a tiled WGSL
matmul on the RX 7900 GRE reaches 1.9-3.6 TFLOPS against 0.16-0.47 for the CPU.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
JonRC
Kristian MyramandClaude Sonnet 5.5