fix(metal): mirror dual-use vit.norm weights to device - #7
Merged
richiejp merged 1 commit intoAug 12, 2026
Merged
Conversation
The fused depth path (Engine::depth_native_fused -> DinoBackbone:: build_feats_graph) applies the ViT final layernorm as ggml graph ops (ggml_extend.hpp `layernorm`: ggml_norm -> ggml_mul(vit.norm.weight) -> ggml_add(vit.norm.bias)). But offload_weights classified vit.norm.weight and vit.norm.bias as host-read tensors and kept them host-only, so on a GPU backend these graph leaves had no device buffer (t->buffer == NULL). On CPU this was harmless (the CPU backend reads t->data directly). On Metal it segfaults: ggml_metal_get_buffer_id() guards t == NULL but not buffer == NULL, so encoding the norm's MUL/ADD operand dereferences buffer->context at NULL+0x60 (EXC_BAD_ACCESS). Reproduced on Apple M4 / macOS 26.6 with depth-anything-base-q4_k.gguf; GGML_METAL_FUSION_DISABLE=1 only moved the crash from ggml_metal_op_norm to ggml_metal_op_bin. These two weights are genuinely dual-use: graph operands in the fused path AND host `->data` reads in the unfused / multi-view path (layernorm_host). Mirror them onto the device like every other weight so the graph has a valid buffer, and preserve their host originals in a new host_tensors_ map exposed via ModelLoader::host_tensor(), which the two layernorm_host sites now use. vit.pos_embed / vit.camera_token stay host-only (they only ever feed host-computed graph inputs). Verified on depth-anything-base-q4_k.gguf: - CPU output bit-identical (min=0.7395 max=2.4074, unchanged). - Metal no longer crashes and produces a correct depth map. Fixes localai-org#3. Assisted-by: Claude:opus-4.8
localai-org-maint-bot
force-pushed
the
cron/pr4-without-generated-assets
branch
from
August 12, 2026 09:04
ba63157 to
7ac14bb
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Supersedes #4 at @mudler's request, preserving @Jinshin2534's original fix while omitting the generated
depth.pfmanddepth.pngartifacts.The branch is based on current
masterand contains only changes to:src/dino_backbone.cppsrc/model_loader.cppsrc/model_loader.hppVerification:
git diff --check origin/master...HEADA focused CMake build could not be run on this cron host because
cmakeis not installed.