PyTorch proposal would cache shared pipeline activations per rank
A rank-local cache would serve reused residuals and return one summed gradient, replacing custom stage subclasses.
By tensorA rank-local cache would serve reused residuals and return one summed gradient, replacing custom stage subclasses.
By tensorA proposal would let memory-heavy pipeline stages park saved activations on spare peer GPU memory without changing the compute schedule.
By tensorThe author says an integer overflow in tensor byte-size math can cause an out-of-bounds memory access, and asked maintainers for an advisory and credit.
By tensorAn AMD proposal would run mixture-of-experts grouped GEMMs as one oneDNN call on Zen CPUs, with about a 2x gain on published expert shapes.
By tensorAn automated check says a new pre-shared key lands in compiler artifacts and can be ignored on older backends.
By tensorCPU COO add_ applies a view's storage offset twice, producing wrong results or a heap abort.
By tensorA reported CPU path writes float32-sized zeros into bfloat16 and float16 MKLDNN buffers, crashing the process.
By tensorA crafted file can make the default weights-only loader return a tensor filled with leftover memory, without running pickle code.
By tensorA proposed conversion change would move default numeric results, and some CPU complex32 fills would return infinity instead of raising.
By tensorCPU and other compiled fusion kernels cannot sit in an FX graph today, so the fx_wrapper path fails outside Triton.
By tensorNative float32 and float64 GPU paths land for multi-dimensional binning, with shared input checks and a CPU edge-case fix.
By tensorChanges aimed at the E2B and E4B variants deliver large gains in the prompt-processing phase of local inference.
By tensorA change to serialize cold-cache downloads in torch.hub still needs packaging, docs, and timeout work before merge.
By tensorPlain sm_120 targets alone leave NVFP4 and block-scaled MMA unavailable on new NVIDIA GPUs.
By tensorA new torch.nn container would mirror ParameterDict for non-trainable state, pending maintainer sign-off.
By tensorPer-sequence cell bitmaps cut host-side list rebuild cost in multi-sequence decode while matching split-cache perplexity.
By tensorDistros can link one shared library instead of two static bundled copies, enabling offline builds and faster security updates.
By tensorA security fix makes torch.export.load reject disallowed operators; regression coverage is being extended to the load path.
By tensorOptional bypass of the speculative decoding p_min check lifts tokens per second by roughly 20 to 40 percent on tested hardware.
By tensorA new torch.library.register_aot_impl path aims to compile custom DSL kernels into PyTorch’s ahead-of-time Inductor stack.
By tensorWhen users pass -t -1, the runtime sizes worker threads to physical cores rather than SMT logical cores.
By tensorUnlocked raw pointers in invalidation and cache updates can outlive a concurrent dynamo reset.
By tensorFileSystemReader could map local .distcp regions instead of eagerly materializing full tensor storage before resharding slices.
By tensorForward-only 4-bit support forces a costly upcast, blocking native low-precision fine-tuning on modern NVIDIA GPUs.
By tensorAn integer overflow lets torch.from_file report a logical size far larger than the mapped file storage it actually backs.
By tensorText inference already works on mainline; Unsloth-shipped MoonViT-3d mmproj files still fail the CLIP loader.
By tensorNative PyTorch FP8 concat replaces a defensive workaround, with layout fixes aimed at future kernels.
By tensorNon-contiguous batch dimensions on the out tensor make the CPU path step wrong and can corrupt the heap; CUDA is fine.
By tensorReview of pending allocator work also flags null returns that can produce CUDA tensors backed by address zero.
By tensorHybrid attention and delta-net models such as Qwen3.5 cannot run through llama_opt_epoch until several missing gradient rules land.
By tensorUnvalidated tensor ops in GRAPH_COMPUTE handling can drive PAD_REFLECT_1D past the destination buffer on reachable servers.
By tensorA precompile import path can stage a checked runtime-cache bundle so GPU kernel launches skip first-use compile and autotune.
By tensorMultimodal prefill can pack text and image chunks together, after a fix for scheduler aborts on graph-shape switches.
By tensorThe work also covers a large-context softmax crash and a data race that may affect the full model.
By tensorFuzzing found two paths where malformed model tensors trigger undefined behavior before validation finishes.
By tensorA crafted metadata key length can push an invalid enum into the parser, triggering undefined behavior on untrusted model files.
By tensorUncapped nested range loops in chat templates can pin llama-server before it finishes starting.
By tensorA fill-order versus stride-order mixup in copy_strided lowering overruns CPU buffers under torch.compile with dynamic shapes.
By tensorCompiled functions under fullgraph mode lost in-place list and dict updates performed by methods on classes defined inside the region.
By tensorA proposed change would process output channels concurrently and cut redundant writes on common Fold shapes.
By tensor