[ DIGEST // 2026-11-02 ]

5 highest-signal merged PRs from that day's run.

today's pick · read the full story [ROCm][Perf][Muse Spark] FlyDSL decode monokernel for DSpark verify steps on MI355X (--monokernel)
vllm-project/vllm·performance·+8985-3
GPU kernel launchesPerformance optimizationCommand-line flagsBenchmarking and accuracy evaluation
9/10
[Core] Extensible KV cache
vllm-project/vllm·feature·+4698-143
CUDA graph captureVirtual memory managementLazy memory allocationGPU memory profiling
8/10
Audio from video in jinja
huggingface/transformers·feature·+1288-336
Multimodal data processingAutomatic audio extraction from videoTemplate-based input processingPyTorch tensor handling
8/10
common : support deferred tools and tool_reference content parts
ggml-org/llama.cpp·feature·+220-24
Deferred executionJSON-based API interactionFunction objects and parametersUnit testing in C++
8/10
Add Flint model support.
huggingface/transformers·feature·+1321-0
Class inheritanceAutomatic model registrationCheckpoint serialization and deserializationUnit testing
8/10