The daily ledger

ggml-org/llama.cpp: b10729

metal : add fa-vec tunings for M1 Ultra (#28088) metal : add fa-vec tunings for M1 Ultra metal : move M1 Ultra tunings after M1 Max section metal : remove duplicate blank line Website: https://llama.app Attestations:…

ggml-org/llama.cpp: b10728

CUDA: XOR swizzle flash attn K,V smem fp16 tiles (#25635) CUDA: XOR swizzle flash attn K,V smem fp16 tiles Signed-off-by: ynankani ynankani@nvidia.com Fix use 64bit generic pointer instead of 32bit shared pointer…

ggml-org/llama.cpp: b10727

metal : add concat support for quantized types (#28116) Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731 Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/44252653 macOS/iOS:…

ggml-org/llama.cpp: b10726

AVX2: Speed up large batch size prompt processing of IQ models (#27402) Batched gemm for grid IQ quants Style updates and a bit more performance Clean up comments Move code around Vectorize IQ panel decode, lower…

pytorch/pytorch: viable/strict/1788220221: Update third_party/kineto submodule to 9dfb9de (#195453)

Includes the following commits: [Kineto] CUPTI PM Sampling -- API layer unit tests (pytorch/kineto#1551) 9dfb9de [Kineto] Make Kineto PM Sampling config options device agnostic (pytorch/kineto#1550) 92167e8 Compact PE…

pytorch/pytorch: viable/strict/1788218631: [MPS] Fix conv backward to preserve input memory format (#174241)

Summary Fixes convolution_backward on MPS to preserve the original input tensor's memory format, rather than deriving it from grad_output. When convolution_backward is called with a channels_last input but contiguous…

pytorch/pytorch: viable/strict/1788216615: [inductor] Enable BF16 index_add lowering on SM90+ and ROCm (#195192)

The Inductor index_add decomposition retained a stale OSS-only BF16 fallback after Triton added BF16 atomic_add support. This enables the decomposition on ROCm and on NVIDIA SM90+, allowing index_add to participate in…

pytorch/pytorch: viable/strict/1788213019

Teach the triage skill to flag issues affecting an upcoming release (…

ggml-org/llama.cpp: b10724

kv-cache : optimize restoring non-contiguous cells (#27991) kv cache : batch state restore scatter reads per contiguous run When restoring state into non-contiguous destination cells (e.g. a prompt-cache snapshot into a…

pytorch/pytorch: viable/strict/1788211462: Refactor scaled matmul meta validation (#195408)

Human Note Small reactor to enable the later commits Agent note This extracts the common scaled-matmul shape and recipe validation from the structured meta implementation without changing its contract. The following…

TimesFM-3: A zero-shot foundation model for multivariate forecasting

Data Management

pytorch/pytorch: viable/strict/1788209416: [dynamo] Migrate DictBuiltinVariable.fromkeys to tp_methods (#195181)

Summary Part of the VariableTracker migration tracked in #195165. Move DictBuiltinVariable.fromkeys from an ad-hoc call_method branch onto the declarative tp_methods table. ListBuiltinVariable is unchanged: its…

ggml-org/llama.cpp: b10723

opencl: tune the quant paths for Intel Xe-LP GPUs to improve its TG and PP performance (#26438) opencl: Q4_K/Q5_K mul_mv N_DST 4->8 on Intel for 2x activation reuse opencl: Q4_K mul_mm 8x8 tile fot Intel opencl: Q5_K…

pytorch/pytorch: viable/strict/1788208014: Revert "Port 3 distributed/_shard tests to Intel GPU. (#189337)"

This reverts commit 797f285. Reverted #189337 on behalf of https://github.com/jansel due to Broke linux-jammy-py3.10-gcc11-debug-distributed (comment)

pytorch/pytorch: viable/strict/1788205825: [BE] Extend dropout support to complex numbers (#195373)

Summary nn.functional.dropout raised NotImplementedError for complex inputs on CPU (and CUDA), while silently "working" on MPS. Root cause: _dropout_impl allocated the Bernoulli mask (noise) in the input's dtype and…

ggml-org/llama.cpp: b10721

webgpu : avoid crash when offset is not multiple of 4 in WebGPU ggml_backend_tensor_get() implementation (#28045) webgpu : avoid crash when offset is not multiple of 4 in WebGPU ggml_backend_tensor_get() implementation…

ggml-org/llama.cpp: b10720

ROCm: add radix TOP_K for long rows (#27466) ROCm: add radix TOP_K for long rows Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/44178618 macOS/iOS: macOS Apple Silicon…

pytorch/pytorch: viable/strict/1788204145

[sparse] Fix hardcoded "cuda" device in semi-structured bitmask conve…

ggml-org/llama.cpp: b10719

metal : add fa-vec tunings for M1 (#28078) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/44172949 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon (arm64, KleidiAI…

Improving our alignment and security efforts