Open source
Fixes and bug reports for the open source AI stack I build on.
5 merged pull requests, 13 open, 1 bug report confirmed by maintainers and 3 bug reports. Contributions by Josh Berns / MASHBOX Labs.
llama.cpp ggml-org/llama.cpp
- Merged ggml-org/llama.cpp#30737
The CPU backend computes f32 gelu_quick and geglu_quick exactly, with SIMD versions for AVX512, AVX2, SSE2, SVE, NEON, VXE and RVV and a new test.
Result: Qwen2-VL-2B worst image token cosine vs transformers 0.52 → 0.998; Apple M5 gelu_quick 1 thread 5.20 → 2.26 ms
- Merged ggml-org/llama.cpp#30736
Image preprocessing for Qwen-VL family models matches transformers: no letterbox bars and the same token grid rounding.
Result: Qwen3.5-2B worst image token cosine vs transformers 0.04 → 0.998 (640x240 text image); token counts now match transformers
- Merged ggml-org/llama.cpp#30108
ModernBERT encoders use exact GELU, as transformers does, instead of the tanh approximation.
Result: FFN activation max error vs exact GELU 5.8e-2 → 7.4e-6
- Merged ggml-org/llama.cpp#29692
Reranker heads apply the activation from the model config (gelu, silu or tanh) instead of a fixed one per architecture.
Result: granite-embedding-reranker-english-r2 (silu): max score error vs transformers 0.23 → 0.001
- Merged ggml-org/llama.cpp#29627
ModernBERT rerankers use the pooling set in their config (mean or CLS), so CLS-pooled rerankers score correctly.
Result: granite-embedding-reranker-english-r2: max score error vs transformers 1.94 → 0.23
Gemma 4 images are resized straight to the target size, as in transformers, with no black bars.
Result: gemma-4-E2B-it image embedding cosine vs transformers 0.60-0.95 → 0.998-1.000 on non-square images
InternVL puts the [CLS] token first, so every image patch gets its own position embedding.
Result: InternVL3-1B image embedding cosine vs the original code 0.86-0.89 → 0.99998
JAX / XLA openxla/xla · jax-ml/jax
- Confirmed bug openxla/xla#51365
Reported wrong results with no error on TPU from lax.conv_general_dilated with lhs_dilation=3 (a stride-3 transposed 1-D convolution), with a minimal repro.
“Indeed, this looks wrong. Thanks for reporting this. I'm working on a fix! I will let you know once it lands.” beckerhe, XLA team member, Oct 22, 2026
- Open jax-ml/jax#41850
Pallas TPU kernels that use signed integer division can be exported from a CPU host; the TPU generation is queried only for bf16 sign.
- Open jax-ml/jax#41851
The TPU pipelining guide states that pallas_call supports buffer_count 1 or 2 and points to emit_pipeline for more.
- Reported openxla/xla#51481
Reported that f32 lax.ragged_dot with Precision.HIGHEST fails to compile on TPU v6e (scoped VMEM out of memory) for a large grouped matmul, with a minimal repro; an older libtpu compiled the same call but returned wrong values.
MaxDiffusion AI-Hypercomputer/maxdiffusion
The Z-Image text encoder loads under transformers 5.x by reading rope_theta from rope_parameters.
Wan VAE weights load in the dtype the config sets (vae_weights_dtype), so the VAE runs in float32 as configured.
Result: Wan VAE decoder PSNR vs diffusers 47.5 dB on main → 131.6 dB together with #514
LoRA merges into bfloat16 weights sum in float32 and round once, so the full LoRA delta is kept.
Result: Wan2.2 kernel delta retained 0.42-0.73 → 0.999-1.001
The Wan VAE mid-block self-attention attends over spatial positions, matching the PyTorch Wan VAE.
Result: VAE matches diffusers on real Wan2.1 weights to max abs diff 2e-5 or less
Hugging Face XiaomiMiMo/MiMo-V2.6-Flash-MOPD
The model code splits the fused qkv_proj per shard, matching the layout of the released weights.
Kimodo (NVIDIA) nv-tlabs/kimodo
- Open nv-tlabs/kimodo#54
Separated classifier-free guidance skips the constraint pass when no constraint is set.
Result: 1.24x faster for 1 prompt, 1 sample; 1.44x for 8 samples
SocialNavSUB LARG/SocialNavSUB
- Open LARG/SocialNavSUB#4
Fixes a NameError that stopped the VLM evaluation on the first multiple-select question of every sample.
PyMeshLab cnr-isti-vclab/PyMeshLab
The Linux wheels link their bundled libGLU to libGL.so.1 again instead of libOpenGL.so.0, which is not bundled, so all plugins load on systems without libOpenGL.so.0.
Result: Plugins loaded in clean Ubuntu 22.04, Ubuntu 24.04, Debian 12 and Fedora 40 containers: 49 of 58 → 58 of 58; pytest 8 failures → all pass
Apple Metal Apple Feedback Assistant
- Reported FB25287280
Reported wrong results from bf16 simdgroup matrix multiply-accumulate in the Metal compiler on Apple M5, with a minimal repro.
- Reported FB25287334
Reported a bf16 accumulation bug in MetalPerformancePrimitives matmul2d on Apple M5, with a minimal repro.
Blender blender/blender
The OBJ importer loads the texture maps it uses as data (specular, roughness, metallic, sheen and normal maps) as Non-Color instead of sRGB, as the FBX importer does, with a new import test.