]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
vulkan: Support mul_mat_id with f32 accumulators (llama/15337)
authorJeff Bolz <redacted>
Sat, 16 Aug 2025 09:18:31 +0000 (04:18 -0500)
committerGeorgi Gerganov <redacted>
Mon, 18 Aug 2025 16:15:25 +0000 (19:15 +0300)
commit9bcb7e0f5b3201e3b75947043e69526fe5001e02
tree6cc6218c076d3d777d5699a85eabc3a11c1fb74b
parentb9bc4d1ff0b50bed5bdf002d9a3b30b7c2062870
vulkan: Support mul_mat_id with f32 accumulators (llama/15337)

* vulkan: Add missing bounds checking to scalar/coopmat1 mul_mat_id

* vulkan: Support mul_mat_id with f32 accumulators, but they are not hooked up

- There's no explicit way to request f32 precision for mul_mat_id, but there
probably should be, and this gets the code in place for that.
- A couple fixes to check_results.
- Remove casts to fp16 in coopmat1 FA shader (found by inspection).
src/ggml-vulkan/ggml-vulkan.cpp
src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp