tensor
AI & ML desk
PyTorch Inductor corrupts heap on dynamic-shape index_fill path
A fill-order versus stride-order mixup in copy_strided lowering overruns CPU buffers under torch.compile with dynamic shapes.
PyTorch Dynamo silently drops closure mutations in local class methods
Compiled functions under fullgraph mode lost in-place list and dict updates performed by methods on classes defined inside the region.
PyTorch CPU fold path eyed for parallel col2im speedups
A proposed change would process output channels concurrently and cut redundant writes on common Fold shapes.
Transformers adds early Gemma 4 logits and hidden-state support
A broad patch set enables boolean control over retained logits and exposes last hidden states for the model family.
PyTorch Inductor split-scan sizing bug can silently corrupt GPU memory
Coordinate-descent autotune can pick an R0_BLOCK smaller than workspace allocation assumes, overrunning the buffer under torch.compile.
llama.cpp rejects Apptainer tools support over host access risks
The proposed server runtime skipped container checks and would let remote clients read host files or pull arbitrary images.
PyTorch split rewrite breaks vLLM QK-norm RoPE fusion
A functionalization change that turns multi-output splits into slices leaves vLLM’s pattern matcher with nothing to fuse under Inductor graph partition.
llama.cpp gains opt-in codec to cut recurrent checkpoint RAM
An optional compression path can roughly halve host memory for recurrent-state context checkpoints in the llama.cpp server, remaining off by default.
llama.cpp Vulkan backend adds Intel Xe flash attention kernels
New kernels for Xe-LPG Plus, Xe2, and Xe3 target faster long-context token generation on recent Intel GPUs.
PyTorch Inductor returns NaNs on zero-size interpolate
Compiled F.interpolate skips eager's size checks and reads past empty buffers, handing callers silent garbage instead of an error.
llama.cpp adds SYCL graph record and replay for Intel GPUs
Opt-in graphs on the oneAPI backend show modest decode gains in early Arc tests, with timeouts still under review.
llama.cpp fixes M-RoPE batch position buffer overread
AddressSanitizer exposed a heap read past an undersized position buffer when batches auto-generate multi-dimensional rotary embeddings.
PyTorch AI PR review can treat attacker filenames as system text
Untrusted paths from fork pull requests were fed into the model as trusted hook context, enabling prompt injection without model cooperation.
PyTorch batch-norm stats update overruns undersized tensors
The CPU path indexes running mean and variance by channel count without checking buffer length, causing heap out-of-bounds access.
PyTorch inductor op builds out-of-bounds CPU tensor views
Unchecked offsets in _reinterpret_tensor let callers produce views that read past storage and trigger heap buffer overflows.
llama.cpp M-RoPE path overreads batch positions past API docs
Callers who sized the position array to the documented n_tokens still hit a multi-kilobyte overread and silent corruption on multimodal decode.
Transformers gains DeepSeek-V4.1-Flash text backbone
The addition brings compressed sparse attention, n-gram memory tables, and baked-in activation quantization to the library’s DeepSeek line.
llama.cpp maintainer blocks new IQ2_NL and IQ3_NL types
Johannes Gaessler rejects a pull request adding CPU quantization formats, citing maintenance burden and machine-generated code.
Ollama ROCm on Strix Halo returns prior answers after context overflow
On gfx1151 hardware, truncated over-window prompts can replay another request's completion instead of generating a new one.
PyTorch Inductor to auto-route small decode GEMMs via swap_ab
NVGEMM will match vLLM and SGLang on M<=64 shapes instead of losing to cuBLAS by default.
llama.cpp GGUF loader accepts tensors whose padded size wraps to zero
Alignment math can overflow so a crafted model file bypasses size checks and understates its true footprint.
PyTorch Inductor CPU tiled reductions can overflow the heap
Full-width vector stores on tail blocks of 2D-tiled reductions corrupt memory or silently rewrite neighboring output rows.
Transformers begins FSDP2 and expert parallelism work
The project is adding a two-dimensional device mesh so fully sharded data parallel can run with tensor and expert parallelism.
PyTorch CUDA cumsum reads past buffer past 2^30 elements
A ones-tensor repro shows an out-of-bounds global load in the GPU prefix-sum path once the length exceeds roughly a billion floats.
PyTorch feature request seeks decorator API for custom autograd ops
A proposed torch.autograd.op pattern would pair forward math with backward derivatives without class boilerplate.
llama.cpp adds GLM-5.3-Flash MoE model support
The 320B mixture-of-experts model lands with known decode overhead and no working vision path yet.
llama.cpp speeds AMD RDNA3/4 inference with Vulkan int8 coopmat
A new cooperative-matrix matmul path in the Vulkan backend lifts prompt and token throughput on Radeon RX 7900-class GPUs.
llama.cpp WebUI query param enables CSRF tool runs
Crafted links can auto-submit prompts that invoke enabled server tools, including shell execution when confirmation is waived.
llama.cpp fuses DeepSeek-V4 hyper-connections on Vulkan
Two-node RPC tests show decode and prefill roughly 1.8x faster with lower intermediate memory use.
llama.cpp adds experimental SM120 CUTLASS MoE prefill
New CUDA kernels target MXFP4 and NVFP4 on SM120 GPUs, while maintainers push for MMVQ refactoring before deeper integration.
PyTorch adds experimental torch.compile for complex tensors
Complex dtypes can now go through the compiler via a real-valued decomposition path that must be opted in explicitly.
llama.cpp server gains image-text embeddings for Qwen3-VL
Multimodal embedding and rerank endpoints now support Qwen3-VL models for RAG pipelines that mix pictures and prose.
PyTorch 2.14 RFC would make CUDA 13.2 the default PyPI build
The proposal keeps CUDA 13.0 as a stable alternate and CUDA 12.6 as legacy while pushing Blackwell-capable 13.2 to pip install torch.
llama.cpp MoE GPU cache nearly doubles decode on dual Vulkan cards
A host-offloaded expert-weight LRU cache kept hot experts in VRAM and lifted Qwen MoE decode from about 8 to nearly 19 tokens per second on two RX 6950 XTs.
llama.cpp doubles cold prefill with direct PLE table reads
A direct-read path for lazy PLE tables roughly halves cold-cache prefill on Windows and keeps memory flat on Apple Silicon.
PyTorch device capability API breaks non-CUDA heuristics
CUDA returns a version tuple while XPU returns a dict, so shared autoheuristic metadata never matches on those backends.
llama.cpp Docker images lack SemVer tags, miss stable builds
Maintainers will change release packaging so GHCR containers ship with proper version tags when a stable build is cut.
PyTorch fixes silent out-of-bounds read in compiled kernels
Strided and offset tensor paths in the compiler could read past valid memory without raising an error.
torch.compile over-reads offset views in Triton kernels
PyTorch can silently return wrong results when compiled kernels skip writing masked output regions.
PyTorch MPS causal attention leaks future tokens in half precision
On Apple Silicon, scaled_dot_product_attention with is_causal=True can silently ignore the causal mask for float16 and bfloat16.