硬體

AI 晶片、GPU、資料中心、邊緣裝置與硬體安全。

Transformers Community Proposes T5 Attention Fix as Tensor Layout May Send SDPA to Slower Kernel

A Transformers community report on October 12 says T5 positional-bias tensor layout may prevent SDPA from using fused kernels, slowing embedding workloads. A proposed fix includes AMD and NVIDIA tests, but has not been merged, and gains still need validation on real workloads.

2026-10-13

GitHub Repo

Ollama Community Reports Discrepancy in Prompt Throughput Calculation; Cache Hits May Inflate Performance Figures

A report posted to the Ollama community on October 11 says the OpenAI-compatible endpoint may divide the total number of prompt tokens, including cached tokens, by the processing time for uncached tokens, inflating prompt throughput. The report includes a streaming response and reproduction steps, but the fix status and cross-version impact remain unconfirmed.

2026-09-23

GitHub Repo

Transformers Community Reports Mamba2 Grouping Gap in Normalization; Non-Fused Path May Differ from Training Core Computation

An October 12 report to the Transformers community says non-fused Mamba2 RMSNorm does not normalize by configured groups, potentially producing different outputs and gradients from the fused training path. The report includes CPU and CUDA comparisons and a local patch check, but no official fix or full-model quality evaluation.

2026-10-12

GitHub Repo

Ollama Community Proposes Tool Argument Precision Fix After Go Client May Silently Rewrite Large Integers

An October 11 report from the Ollama community says the Go client may round large integers and high-precision decimals when decoding tool calls, then send the altered values back in conversation history. The proposed fix adds streaming and desktop tool tests but has not been merged; the numeric type change also requires compatibility handling.

2026-09-25

GitHub Repo

PyTorch Community Reports CUDA Batched SVD Performance Gap; Matrices Over 32 Dimensions May Run One at a Time

A PyTorch community report on October 10 says CUDA batched SVD may slow sharply when matrix size increases from 32 to 33, possibly because execution switches paths. The report includes measurements across environments and a proposed fix, but the precision trade-offs of an alternative driver and the status of an official fix still need confirmation.

2026-10-11

GitHub Repo

PyTorch Community Tracks XPU Compilation Optimization Gap: reduce-overhead Still Lacks Graph Replay

An October 10 PyTorch development-build report says reduce-overhead mode on Intel XPU does not enable the expected graph capture and replay. Official documentation still limits this optimization to CUDA; XPU integration and performance validation remain outstanding.

2026-10-10

GitHub Repo

ONNX Runtime 1.31.0 Stabilizes Model Package APIs and Strengthens Quantized MoE Inference and Plugin Compatibility Requirements

Microsoft released ONNX Runtime 1.31.0 on October 9, stabilizing the model package and EPContext data interfaces while improving CPU model loading and quantized MoE execution. The release is available on PyPI, but teams using GPU plugins should verify operator schema compatibility and measure performance in their own environments.

2026-10-09

GitHub Repo

PyTorch Development Build Reports CUDA Graph Copy Bug: Deferred-Copy Tensors May Read Stale Data During Replay

Reports filed on October 7 for PyTorch and TorchTitan indicate that copy-on-write tensors materialized during CUDA Graph capture may omit the copy operation, causing incorrect results on replay. The reports include a minimal reproduction and an MoE inference comparison, but the impact on released versions and an upstream fix remain unconfirmed.

2026-10-08

GitHub Repo

Transformers Community Reports Quantized Cache Bug: Batched Token Appends May Exceed Full-Precision Buffer Threshold

A reproduction on the October 7 development build suggests that appending multiple tokens after reusing a prefix may leave more than the configured threshold in the full-precision cache. The report includes a test that requires no weight downloads, but maintainers have not confirmed the issue, and GPU memory and output-quality effects have not been verified.

2026-10-05

Liquid AI Releases d1 Decision Model Weights for Text and Image Classification in a Single Forward Pass

On October 7, Liquid AI released d1-3B and the experimental d1-omni-600M, which return structured decisions without generating answers token by token. Local and Jetson deployment examples are available, but the low-latency figures depend on warm-up conditions, and evaluation results are still primarily based on the company’s own testing.

2026-10-07

GitHub Repo

Ollama Community Reports Dual-Model Loading Crash, GPU Allocation Error During Qwen3-VL Image Encoding

A community report dated October 6 says image requests to Qwen3-VL may repeatedly crash the inference process when another large model is already using both GPUs. The report includes a core-dump stack trace, but maintainers have yet to confirm whether memory pressure caused a null-pointer error.

2026-10-07

GitHub Repo

PyTorch Community Reports Compiler Dropping a Coefficient, Potentially Amplifying 2.14 CUDA Momentum Updates Tenfold

A PyTorch community reproduction on October 5 found that Inductor may drop an addition’s alpha coefficient under a specific combination of clipping and scaling. A draft fix was proposed on October 6 but has not been merged; the impact on full training runs remains to be verified.

2026-10-06