Introducing Hy4 Preview
Introducing Hy4 Preview New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, 1.56TB on Hugging Face. This is a big size…
Daily edition · AI
The daily ledger
TILens turns technical updates into a focused daily brief: official releases, trusted reporting, and practitioner analysis, deduplicated and organized by topic.
ggml-org/llama.cpp: b10687
opencl: use a better matmul path on two Adreno GPU generations (#27640) opencl: default the Adreno xmem F16xF32 GEMM on for X2E kernel_mul_mm_f16_f32_l4_lm is the slowest matmul this backend has on Adreno: on the X2-90…
pytorch/pytorch: viable/strict/1788042194
[MPS] Add complex64 support to linalg.lu_factor (unblocks complex det…
pytorch/pytorch: viable/strict/1788040196: Fix static launcher 16-bit float scalars (#184065)
Triton reports fp16 and bf16 scalar arguments as distinct 16-bit types, but the static launcher mapped both to the 32-bit float ABI code. CUDA and HIP therefore wrote four-byte values into slots that kernels read as…
ggml-org/llama.cpp: b10686
metal : assert shared memory padding (#27951) metal : assert shared memory padding cont : add ref Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/43867822 macOS/iOS: macOS…
pytorch/pytorch: viable/strict/1788035047
[BE] Skip bmm_shared_a autotune tests when Triton is unavailable (#19…
ggml-org/llama.cpp: b10685
metal : add remaining fa-vec tunings for M4 Pro (#27915) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/43865923 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon…
ggml-org/llama.cpp: b10684
sycl: make --fit respect --fit-target better (#27629) improve the --fit algorithm to take into account the actual peak required VRAM for a given context size on a SYCL backend. This includes both properly accounting for…
pytorch/pytorch: viable/strict/1788032998
[pipelining] Annotate schedule actions with mark_kernels for CUDA gra…
pytorch/pytorch: trunk/68dedf079d7c18b7d28ba3a1356aca12f6203cfb
[pipelining] Annotate schedule actions with mark_kernels for CUDA gra…
ggml-org/llama.cpp: b10683
vulkan: combine duplicated fastdiv functions, rename the one optimizing small divs (#27526) vulkan: combine duplicated fastdiv functions, rename the one optimizing small divs remove one more fastdiv Website:…
ggml-org/llama.cpp: b10682
metal : add fa-vec tunings for M1 Max (#27932) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/43859747 macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon (arm64,…
ggml-org/llama.cpp: b10681
vulkan: Change mul_mat_id to pad K rather than N (#27925) The N padding is needed for mul_mat, but not mul_mat_id. For mul_mat_id, we indirect the row index through a shared memory lookup table which avoids any OOB row…
pytorch/pytorch: trunk/58a938a422d785b35430328d5e718e7fb0e7c07e
[ROCm] Guard native BMM outer-product launches against the HIP work-i…
pytorch/pytorch: viable/strict/1788028080
[ROCm] Guard native BMM outer-product launches against the HIP work-i…
pytorch/pytorch: viable/strict/1788025796: [MPS] Fix linear backward SIGABRT with a 1-D weight (#187989)
Fixes #187988 What linear() accepts a 1-D weight (shape [in_features]): it contracts and drops the last input dimension. mps_linear (forward) already handles this by unsqueezing the weight to [1, in_features]. The…
pytorch/pytorch: trunk/1782e8ec673e736ffda3414fd921f3c1b1b62002: [MPS] Fix linear backward SIGABRT with a 1-D weight (#187989)
Fixes #187988 What linear() accepts a 1-D weight (shape [in_features]): it contracts and drops the last input dimension. mps_linear (forward) already handles this by unsqueezing the weight to [1, in_features]. The…
pytorch/pytorch: viable/strict/1788024140: Revert "[c10d] Make nccl2 the default NCCL backend (#192281)"
This reverts commit 4078ac4. Reverted #192281 on behalf of https://github.com/pytorch-auto-revert due to Reverted automatically by pytorch's autorevert, to avoid this behaviour add the tag autorevert: disable (comment)
pytorch/pytorch: trunk/59cfa7a7f76e0f9d9a08e3264a7e203d155680ee
Revert "[distributed] Warn pipeline users about eager NCCL communicat…
pytorch/pytorch: trunk/8b802259223ccdc4df77e910dfa081441ae98aa7: [c10d] Set local rank in backend test workers (#195103)
This change was prepared with AI assistance and reviewed by a human. MultiProcessTestCase workers inherit LOCAL_RANK from their parent process. A stale value can cause eager NCCL2 initialization to bind every worker to…
Loading more...
Showing 1 day · 100 items available