ytfh44
- 15 skills
- 0 followers
- 6 hours ago last updated
- ▌ Gpu Code Optimizer · ytfh44Load this skill and follow it when planning, reviewing, or carrying out performance optimization for GPU compute, resource, or runtime-state workloads, especially when deciding which specialized optimization skills to load.
- ▌ Gpu Compiler Runtime · ytfh44Load this skill and follow it when optimizing compiled graphs, fusion, graph breaks, recompilation, GPU graphs, or multi-GPU runtime behavior in systems such as PyTorch Inductor, torch.compile, JAX/XLA, or Triton.
- ▌ Gpu Kernel Execution · ytfh44Load this skill and follow it when optimizing GPU kernel thread mapping, coalesced memory access, tiling, shared memory, registers, occupancy, synchronization, or atomic operations.
- ▌ Gpu Numerical Safety · ytfh44Load this skill and follow it when optimizing GPU code that may change floating-point evaluation order, precision, boundary behavior, NaN or Inf propagation, determinism, or other program semantics.
- ▌ Gpu Persistent State · ytfh44Load this skill and follow it when designing runtime GPU state that survives across kernels, steps, requests, or sessions, especially when growth, mutation, snapshots, branching, checkpoint placement, ownership, or reconstruction semantics differ across state objects.
- ▌ Gpu Reductions Scans · ytfh44Load this skill and follow it when optimizing GPU reductions, scans, prefix operations, recurrences, streaming state, tile-local partials, or their boundary semantics.
- ▌ Gpu Memory Scheduling · ytfh44Load this skill and follow it when jointly scheduling GPU compute, allocation, mapping, transfers, prefetch, offload, rematerialization, barriers, or reclamation to minimize critical-path stalls and bound memory pressure.
- ▌ Gpu Training Autodiff · ytfh44Load this skill and follow it when optimizing GPU training paths, backpropagation, saved tensors, recomputation, gradient reductions, or mixed-precision forward and backward execution together.
- ▌ Gpu Memory Fusion Layout · ytfh44Load this skill and follow it when reducing GPU global-memory traffic, intermediate tensor materialization, layout conversions, redundant copies, or when designing safe kernel and operator fusion.
- ▌ Gpu Performance Evidence · ytfh44Load this skill and follow it when establishing a GPU performance baseline, analyzing profiler data, roofline results, or hardware counters, classifying bottlenecks, or validating evidence for a claimed speedup.
- ▌ Gpu State Reuse Eviction · ytfh44Load this skill and follow it when identifying reusable GPU state, defining identity and validity, sharing or copy-on-write, admission, retention, invalidation, or eviction under capacity and performance constraints.
- ▌ Gpu Optimization Validation · ytfh44Load this skill and follow it when turning GPU optimizations into reviewable patches, designing guards and fallbacks, running representative benchmarks, recording failure cases, or performing final validation.
- ▌ Gpu Memory Tiering Migration · ytfh44Load this skill and follow it when placing or moving GPU resources across device memory, peer devices, host memory, storage, or remote tiers, including residency, prefetch, offload, replication, migration, and oversubscription.
- ▌ Gpu Resource Lifetime Allocation · ytfh44Load this skill and follow it when optimizing GPU resource lifetimes, allocation reuse, transient aliasing, pooling, workspace capacity, materialization, rematerialization, or peak live memory across a task graph.
- ▌ Gpu Virtual Memory Fragmentation · ytfh44Load this skill and follow it when diagnosing GPU memory allocatability failures, internal or external fragmentation, virtual-to-physical mapping, sparse or page-backed resources, virtual contiguity, compaction, stitching, or page-granularity tradeoffs.