wenyi-li
- 196 skills
- 0 followers
- 12 hours ago last updated
- ▌ Pytorch Kernel Optimization · wenyi-liThis skill should be used when optimizing PyTorch models and kernels, including efficient tensor operations, torch.compile, custom autograd/CUDA/Triton extensions, mixed precision, memory and data pipeline tuning, model optimization techniques, CUDA graphs, and profiling.
- ▌
- ▌ Vllm Plugin Fl Setup Flagos · wenyi-li bundleInstall and configure vLLM-Plugin-FL for multiple hardware backends including NVIDIA, Ascend and etc. Use when setting up vllm-plugin-fl, configuring the environment for specific hardware backend, installing dependencies, checking whether dependencies are installed successfully, resolving runtime issues, and launching inference to verify successful model serving. Trigger when the user says things like "setup vllm-plugin-fl", "install vllm-plugin-fl", "configure FL plugin", "set up FlagGems", or "set up FlagCX".
- ▌
- ▌
- ▌
- ▌ Write Triton Dequant Kernel · wenyi-li bundleSkill: Write a Triton Dequant Kernel (int4 / int8 → fp16 / bf16)
- ▌
- ▌
- ▌ Ad Add Fusion Transformation · wenyi-liClaude Code skill (trtllm-agent-toolkit): implement or extend TensorRT-LLM AutoDeploy fusion transforms under transform/library/ in a TensorRT-LLM checkout. Prefer existing kernels and custom ops; use Triton only when no viable existing-kernel path exists. Use ad-graph-dump for AD_DUMP_GRAPHS_DIR workflows. Covers TRT-LLM paths, registry, default.yaml registration, graph validation, tests, and a review checklist — without prescribing profiling tools or throughput targets.
- ▌ Triton Ascend Case Elemwise Concat · wenyi-liSlice+Concat融合算子优化：通过精确切片加载（只load需要部分）和索引计算拼接（避免cat指令）避免中间结果存储和多次内存访问，适用于多输入需要切片后拼接的融合算子场景
- ▌ Triton Ascend Case Index Histogram · wenyi-li直方图统计（histogram）优化：预排序+二分查找降低算法复杂度（O(n×m)→O(n log n + m log n)，性能提升19倍），转换为float32调用Vec Core硬件加速排序，适用于大规模统计类操作（50万+元素）
- ▌ Triton Ascend Elementwise · wenyi-li适用于纯逐元素(element-wise)类算子的优化指南。当算子的核心计算是对张量每个元素独立执行相同操作、无跨元素依赖时应选择此指南，典型算子包括：relu, sigmoid, tanh, gelu, selu, leaky_relu, elu, swish, softplus, hardsigmoid, hardtanh, softsign, exp, log, sqrt, pow, add, mul, sub, div, abs, neg, clamp, cast(类型转换), where, fill, copy 等。也适用于涉及标量广播(broadcast)的运算。不适用于需要跨元素归约(如 sum/mean/max)或矩阵乘法的算子。如果算子同时包含逐元素计算和全局归约（如损失函数 MSELoss、HuberLoss、HingeLoss），应选择 elementwise-reduce-fused 指南。
- ▌ Triton Ascend Case Matmul Swizzle2d · wenyi-li大矩阵乘法Swizzle2D优化：固定核心数启动（grid=20而非所有块）+Swizzle2D块重排（GROUP_SIZE=4）提升缓存局部性，根据M/N比例自适应选择分组方向，适用于大规模矩阵乘法（千万级元素）的Ascend NPU场景
- ▌
- ▌ Triton Ascend Error Fix · wenyi-litriton-ascend 常见错误修复：UB/CBUF溢出、BiShengIR编译失败、语法限制违反、数值正确性、多维索引分解错误、张量连续性
- ▌ Triton Ascend Example Relu · wenyi-liReLU 逐元素算子的完整 Triton Ascend 实现示例。展示向量化逐元素操作的标准模式：1D 分块遍历、mask 边界处理、交错循环。当生成 elementwise 类算子时可参考此示例的代码结构。
- ▌ Triton Ascend Case Reduction Sum Fused · wenyi-liReduction+Elementwise融合算子优化：先逐元素操作再归约，行二次切分+计算重组，grid=40且SUB切分不含尾块时性能最优（47.58us），融合优化逻辑以reduce为主，适用于需要先逐元素计算再reduce的融合场景
- ▌ Triton Ascend Case Reduction Sum Large · wenyi-li大规模归约（sum）非reduce轴很大优化：计算重组减少归约次数，在优先占满UB前提下为reduce轴分配较大切分尺寸（BLOCK_SIZE_N=1024最优685.65us），适用于非reduce轴非常大（6万+）、reduce轴中等（千级）的2D归约场景
- ▌ Triton Ascend API Rules · wenyi-liTriton Ascend hard API restrictions and forbidden syntax. MUST-follow rules that apply to every kernel: forbidden control flow (return/break/continue/lambda/while), tensor slice/index restrictions, scalar conversion rules, BLOCK_SIZE upper bound. Violating any of these produces a compile or runtime error on Ascend.
- ▌ Triton Ascend Debugging · wenyi-liTriton Ascend 调试排查清单和常见错误速查表，包括编译错误、运行时错误、精度问题和性能问题的诊断方法。适用于内核代码生成、出现错误需要定位原因、或需要验证代码正确性的调试场景
- ▌ Triton Ascend Case Reduction Amax Large · wenyi-li非reduce轴很小、reduce轴很大的归约优化：将reduce轴映射到多核（而非常规的非reduce轴），使用原子操作跨线程块归约，通过二次切分避免超UB，适用于极端shape比例（M<<N如16×262144）的归约场景
- ▌ Triton Ascend Case Reduction Amax Small · wenyi-li极小规模归约（amax）优化：单核处理（grid=1）优于多核并行（2.16us vs 3.51us），避免并行化带来的调度开销，适用于数据规模很小（<1000元素）的归约场景
- ▌ Triton Ascend Case Reduction Amin Large · wenyi-li极大规模1D归约（amin）优化：二次切分避免超UB+计算重组减少归约次数，网格数接近AI Core数量（grid=32）、UB用满、无尾块时性能最优（9.61us），适用于极大规模1D数据（400万级元素）的全量归约场景
- ▌ Triton Ascend Case Reduction Amin Small · wenyi-li中等规模1D归约（amin）优化：适中并行度（grid=8时最优2.21us），存在最优平衡点（过小导致单块负载过重、过大引入调度开销），适用于中等规模1D数据（6万级元素）的全量归约场景
- ▌ Triton Ascend Case Reduction Mean Large · wenyi-li大规模reduce最后根轴（mean）行二次切分优化：每个kernel计算多行减少线程块数量、kernel内二次切分避免超UB，grid=40且SUB切分不含尾块时性能最优（16.00us），尾块计算会显著降低性能，适用于非reduce轴中等、reduce轴较大的2D归约场景
- ▌ Flaggems Pr Submit Flagos · wenyi-li bundleThis skill should be used when submitting FlagGems operator PRs, reviewing operator code before submission, preparing operator code for PR, or when the user mentions "提PR", "提交算子", "submit operator", "PR提交", "代码审核", "pre-commit". It automates code review, validates completeness and compliance, runs pre-commit and worktree tests, and directly submits PR to upstream with full description including speedup data.
- ▌
- ▌
- ▌ Kernel Tileir Optimization · wenyi-li bundleOptimize existing Triton kernels for NVIDIA TileIR backend on Blackwell GPUs (sm_100+). Adds TileIR-specific autotune configs: occupancy, num_ctas, TMA descriptors. Covers kernel classification (dot-related, norm-like, elementwise, reduction), type-specific transformations, and PTX-vs-TileIR benchmarking. Triggered by: "optimize for TileIR", "add TileIR configs", "Blackwell optimization", "TMA descriptors", "2CTA mode", "occupancy tuning". Kernels use standard `import triton`; TileIR activates via ENABLE_TILE=1 when nvtriton is installed.
- ▌
- ▌ Tilelang Cuda Optimization · wenyi-liTileLang CUDA 性能优化通用策略、最佳实践和调试技巧汇总。适用于需要提升 TileLang 内核性能、遇到编译/运行错误需要排查、或需要了解 TileLang 平台限制的内核代码生成和优化场景
- ▌ Triton Ascend Optimization · wenyi-liTriton Ascend 性能优化通用策略: BLOCK_SIZE 选择 (1024-2048 for elementwise, must be <65536), grid configuration (use VEC_CORE_NUM / CUBE_CORE_NUM, 2D/3D grid for matmul / conv / reduce, 1D grid + inner loop for elementwise / pointwise), 256B alignment for memory transfers, autotune block-size patterns, fp16 / fp32 precision conversion. Bind via keywords like matmul, elementwise, reduce, block_size, grid, autotune, alignment, fp16, fp32, tile, interleaved-loop, cube-core, vec-core.
- ▌ Triton Kernel Optimization · wenyi-liThis skill should be used when writing or tuning Triton GPU kernels, including autotuning block sizes, coalesced accesses, tiled matmul, fused ops, reductions, flash-attention style kernels, quantization, custom gradients, and profiling.
- ▌
- ▌ Debugging Tilelang Programs · wenyi-li bundleHow to diagnose and fix errors in TileLang programs. Use this skill whenever a TileLang kernel fails to compile, crashes at runtime, produces incorrect results, or shows numerical mismatches. Also use when the user encounters TVM/TIR errors, CUDA codegen failures, shape mismatches, assertion failures, NaN/inf outputs, tilelang.compile errors, or needs to inspect generated CUDA code, use T.print for debugging, run AutoDD for minimal reproduction, or compare against a PyTorch reference. Trigger even for vague complaints like "my kernel doesn't work" or "wrong output" when TileLang is involved. Also trigger for TMA descriptor errors, illegal memory access, or kernel launch failures.
- ▌ Profiling Tilelang Programs · wenyi-li bundleHow to benchmark and profile TileLang GPU kernels for performance measurement. Use this skill whenever the user wants to measure kernel latency, compare TileLang performance against PyTorch/cuBLAS/Triton, use do_bench or get_profiler, calculate TFLOPS or bandwidth, run Nsight Compute or Nsight Systems on a TileLang kernel, understand occupancy or shared memory usage, or figure out why a kernel is slow. Also trigger when the user mentions benchmarking, profiling, timing, throughput, performance numbers, or latency in the context of TileLang or GPU kernels. Even for simple questions like "how fast is my kernel" or "how do I time this".
- ▌ Triton Ascend Case Reduction Prod Small · wenyi-li小规模reduce第一根轴（prod）优化：使用自定义mul函数配合tl.reduce实现连乘（triton无prod接口），最优网格数明显小于AI Core数量（grid=16最优2.15us），过高并行度反而因调度开销降低性能，适用于shape较小（10万级元素）的reduce第一根轴场景
- ▌ Triton Ascend Example Matmul · wenyi-li标准矩阵乘法的完整 Triton Ascend 实现示例。展示 2D 分块(tiling)、K 维循环累加、2D mask 处理、Cube Core 利用等关键模式。当生成 matmul 类算子时可参考此示例的代码结构。
- ▌ Triton Ascend Case Elemwise Broadcast 2d · wenyi-li2D广播除法优化：小维度完整处理不切分（循环外加载复用），通过固定NUM_BLOCKS实现核间并行（40核），核内SUB_M控制粒度平衡UB利用率，适用于broadcast轴大但非broadcast轴小的2D场景
- ▌ Triton Ascend Case Elemwise Broadcast 3d · wenyi-li跨轴3D广播优化（最后一维很小）：采用两阶段kernel策略（先broadcast展开+reshape为2D，再标准多核处理）提升向量化效率，适用于跨轴broadcast且最后一维特别小（<20）导致向量化效果差的场景
- ▌ Triton Ascend Case Reduction Amax Medium · wenyi-li中等规模归约（amax）优化：计算重组（循环内累加、循环外归约）减少归约次数，grid=40等于核数时性能最优（25.73us），适用于非reduce轴中等、reduce轴较大（千万级元素）的2D归约场景
- ▌ Triton Ascend Case Reduction Amin Atomic · wenyi-li原子操作归约（amin）优化：非reduce轴很小时将reduce轴映射多核，提供循环内/外两种原子操作方案（减少存储vs减少竞争），通过二次切分+计算重组提升性能，适用于M<<N（如16×262144）的极端shape比例场景
- ▌ Triton Ascend Case Reduction Amin Medium · wenyi-li大规模2D归约（amin）reduce轴很大优化：在优先占满UB前提下为reduce轴分配较大切分尺寸（BLOCK_SIZE_N=16384最优），减少循环次数但需权衡单次迭代负载，适用于非reduce轴中等、reduce轴很大（50万级元素）的场景
- ▌ Triton Ascend Case Reduction Mean Medium · wenyi-li中等规模reduce第一根轴（mean）优化：计算重组减少归约次数，网格规模略小于AI Core数量且避免尾块时性能最佳（grid=32最优9.98us），适用于reduce第一根轴、两轴均中等（百万级元素）的2D归约场景
- ▌ Triton Ascend Example Softmax · wenyi-liSoftmax 归约算子的完整 Triton Ascend 实现示例。展示三阶段归约模式（求 max → 求 sum(exp) → 归一化）、分块累加、标量累加器精度提升等技巧。当生成 reduce 类算子时可参考此示例的代码结构。
- ▌ Triton Ascend Grid Config · wenyi-liGrid/Block 配置策略，包括核数选择、并行度调优、二次切分和大 shape 算子处理方案。适用于需要确定 kernel 启动参数、优化多核并行效率、或处理超大规模数据的内核代码生成场景
- ▌ Triton Ascend Example Layernorm · wenyi-liLayerNorm 归约算子的完整 Triton Ascend 实现示例。展示两阶段归约模式（统计量计算 → 归一化输出）、标量累加器、分块遍历等技巧。当生成 reduce/normalize 类算子时可参考此示例的代码结构。
- ▌ Triton Ascend Examples Mindspore · wenyi-liMindSpore 框架下 Triton Ascend 内核的集成示例，展示 MindSpore 自定义算子注册、Primitive 定义、tensor 传入传出等标准写法。当目标框架为 mindspore 时应导入此示例作为代码结构参考。
- ▌ Triton Ascend Case Reduction Weighted Swiglu · wenyi-li3D融合算子（Weighted SwiGLU Backward）优化：Reshape降维将前两维合并简化并行策略，行二次切分避免超UB，在优先占满UB前提下为reduce轴分配较大切分尺寸，grid数较大时可能性能更优，适用于3D张量逐元素+reduce融合的场景
- ▌ Triton Ascend Example Double Kernel · wenyi-li双内核调用模式的 Triton Ascend 实现示例。展示在 forward 中先后调用两个 kernel 的标准写法：中间结果缓冲区分配、两次 kernel 启动。适用于需要分阶段计算的融合算子。
- ▌ Triton Ascend Elementwise Reduce Fused · wenyi-li适用于同时包含逐元素计算和全局归约两个阶段的复合算子。典型算子包括：损失函数（MSELoss, HuberLoss, HingeLoss, SmoothL1Loss, CrossEntropyLoss, KLDivLoss, CosineSimilarityLoss, TripletMarginLoss 等）、以及自定义的先逐元素变换再全局聚合的算子。这类算子的计算模式是：第一步对张量每个元素独立执行变换（差值、平方、clamp 等），第二步对变换结果做全局或按维度归约（sum/mean）得到标量或低维结果。与纯 elementwise 或纯 reduce 不同，这类算子需要在同一个 kernel 中融合两个阶段以避免中间结果的额外 GM 读写。
- ▌ Triton Ascend Performance Improvement · wenyi-liTriton Ascend 性能优化实战经验。从批量自适应搜索中提炼的通用优化模式，覆盖 tile 调优方法论、内存加载优化、reduction 优化、隐式广播、多 Pass 合并、数据访问重构等。
- ▌ Triton Ascend Fused Operator Optimization · wenyi-liAscend NPU 上融合算子的深度优化方法论。覆盖性能天花板分析框架、多 Pass 合并策略、数据访问模式重构、Normalization 两阶段决策、NPU 原生算子评估方法论。适用于 elementwise 融合、归一化融合、softmax+topk 融合、matmul+activation 融合等场景。
- ▌ Pypto Case Norm Layernorm · wenyi-li模式 C 示例：2D Norm + Loop — LayerNorm，展示 forward 降维为 2D、kernel 内 sum 归约 + 归一化
- ▌
- ▌
- ▌ Pypto Optimization · wenyi-liPyPTO 性能优化规则与调参顺序。适用于需要优化 tile/loop/归约性能、比较不同 tile 方案、解释同一算子不同 tile 性能差异（尤其 softmax/logsoftmax/reduction/norm/loss）的场景
- ▌ Pypto Case Loss Crossentropy · wenyi-li模式 D 示例：Loss — CrossEntropyLoss，展示多输入 kernel、两段 tile、softmax+gather+sum、标量输出
- ▌
- ▌ Triton Cuda Reduce · wenyi-li归约算子(reduce)优化策略，包含 sum/mean/max/min、softmax、layernorm、logsoftmax 等实现技巧。适用于需要在 CUDA GPU 上实现任意维度归约、规范化层或注意力分数计算的内核代码生成场景
- ▌ Triton Ascend Case Index Put · wenyi-li索引赋值（index_put）优化：批量加载索引数据到UB后循环内通过get_element复用（避免重复访问全局内存），显著降低内存访问延迟，适用于需要在循环中多次访问同一片数据的不规则内存访问场景
- ▌ Triton Ascend Reduce · wenyi-li适用于归约(reduce)类算子和含归约子步骤的复合算子（如归一化）的优化指南。典型算子包括：sum, mean, max, min, prod, argmax, argmin, cumsum, cumprod, softmax, logsoftmax, layernorm, rmsnorm, groupnorm, instancenorm, batchnorm, l1norm, l2norm, frobeniusnorm, var, std, average_pooling, sum_pooling 等。特别重要：当归约维度不是最后一维（如 dim=1 归约 shape=[B,F,D1,D2]），需要正确处理多维索引和两阶段归约。包含 PyTorch normalized_shape 多轴归一化语义说明。不适用于纯逐元素运算或矩阵乘法。如果算子是损失函数（先逐元素计算再全局归约），应选择 elementwise-reduce-fused 指南。
- ▌
- ▌ Triton Cuda Elementwise · wenyi-li逐元素算子(element-wise)优化策略，包括 add/mul/relu/sigmoid/tanh/gelu/exp/log 等操作的向量化实现和融合技巧。适用于实现激活函数、逐元素运算、广播操作等向量模式算子的 CUDA 内核代码生成场景
- ▌ Triton Cuda Grid Config · wenyi-liGrid/Block 配置策略，包括线程块大小选择、SM 占用率优化和大 shape 算子处理方案。适用于需要确定 CUDA kernel 启动参数、优化 GPU 并行效率、或处理超大规模数据的内核代码生成场景
- ▌ Triton Ascend Case Elemwise Cast · wenyi-li大shape类型转换（int8→fp16）优化：通过二次切分（BLOCK_SIZE+TILE_SIZE）提高UB利用率，核数2048时性能最优，适用于shape较大（百万级元素）的elementwise类型转换场景
- ▌ Triton Ascend Case Elemwise Zeros · wenyi-li小shape张量创建（zeros/arange/full）优化：通过减少核数避免多核启动和调度开销，单核处理性能优于多核并行，适用于shape较小（千级元素）的elementwise创建张量场景
- ▌ Triton Ascend Case Matmul Large K · wenyi-li矩阵乘法矩阵乘法 A[M, K] @ B[K, N] = C[M, N]中，大K维度矩阵乘法(K>>M,N)优化：针对M/N较小但K极大(如M=N=256,K=131072)的场景，Split-K切分K维度并行化、Workspace+Reduce替代全局同步，实现显著性能提升
- ▌ Triton Cuda Examples Torch · wenyi-liPyTorch 框架下 Triton CUDA 内核的完整集成示例，包括 vector_add、matmul、layer_norm、softmax 等标准算子实现。适用于需要参考 PyTorch 算子包装方式、torch.autograd.Function 实现模式的 CUDA 内核代码生成场景
- ▌
- ▌ Testing Fwd Bwd Kernels · wenyi-li bundleHow to write and test TileLang kernels that need both forward and backward passes. Use this skill whenever the user is implementing custom operators with gradients, writing attention forward+backward, linear attention fwd+bwd, any op used inside torch.autograd.Function, or debugging gradient mismatches. Also trigger when the user mentions backward pass, gradients, dQ/dK/dV, autograd, ctx.save_for_backward, atomic_add for gradient accumulation, or gradient testing in TileLang context. Use even for questions like "how do I test my TileLang backward kernel" or "my gradients don't match the reference".
- ▌ Tilelang Op Test Design · wenyi-li bundleTileLang-Ascend 算子测试设计技能。支持多种场景：(1) 从 design.md 设计测试配置 (2) 从 examples/{op}/*.py 补充测试 (3) 手动提供算子信息生成测试 (4) 测试覆盖率分析。理解算子实现逻辑后智能判断测试策略。触发：设计算子测试、生成测试用例、补充测试、测试覆盖率不足。
- ▌ Tilelang Torch Profiler · wenyi-li bundleProfiles TileLang and mixed PyTorch+TileLang workloads with torch.profiler. Invoke when users want traces, operator breakdowns, memory profiling, or a lightweight first-pass performance diagnosis.
- ▌ Triton Ascend Attention · wenyi-li适用于注意力(attention)机制类算子的优化指南。当算子的核心计算是 Transformer 风格的注意力运算时应选择此指南，典型算子包括：self_attention, cross_attention, multi_head_attention, flash_attention, scaled_dot_product_attention, causal_attention, masked_attention 等。涵盖 QKV 矩阵乘分块、在线 softmax、因果 mask 处理、Flash Attention 分块策略等关键技巧。不适用于不含注意力结构的普通矩阵乘法或归约运算。
- ▌ Flashinfer · wenyi-li bundleFlashInfer — High-performance kernel library for LLM inference with optimized attention, paged KV-cache, FP8/FP4 quantization
- ▌ Mi300 Cdna3 Architecture · wenyi-li bundleMI300/CDNA3 architecture guide for HIP/Triton optimization—MFMA variants, dual register files, data formats, sparsity, LDS/GWS, and best practices.
- ▌ Tilelang Ascend Tile API · wenyi-liTileLang-Ascend 新增 Ascend 专属 T.tile.xxx 小 API 的端到端开发流程。用户要求新增、封装、暴露、实现或测试 ascend_tile.py 中的 T.tile API / Ascend tile primitive 时必须使用本 skill，尤其适用于需要同时打通 Python 前端、C++ lowering/codegen、Ascend C helper、文档和 CI 测试的任务。
- ▌ Triton Cuda Optimization · wenyi-liTriton CUDA 性能优化通用策略、API 限制说明和调试技巧汇总。适用于需要提升 GPU 内核性能、遇到编译/运行错误需要排查、或需要了解 CUDA 平台限制的内核代码生成和优化场景
- ▌
- ▌
- ▌ Writing Tilelang Kernels · wenyi-li bundleHow to write TileLang GPU kernels from scratch or by adapting existing patterns. Use this skill whenever the user wants to create a new TileLang kernel, implement a GPU operator in TileLang, write a prim_func, set up tiling/blocking, define shared memory and fragment allocations, use T.gemm or T.copy, create a JIT-compiled kernel, or port an algorithm (GEMM, elementwise, reduction, attention, normalization, softmax, convolution) to TileLang. Also use when the user asks about tilelang.jit, T.prim_func, T.Kernel, T.Tensor, T.alloc_shared, T.alloc_fragment, T.Pipelined, T.Parallel, or any TileLang DSL construct for kernel authoring. Even for questions like "how do I start with TileLang" or "show me a basic TileLang example", this skill provides the right templates and workflow.
- ▌ Flaggems Pr Review Flagos · wenyi-li bundleThis skill should be used when reviewing FlagGems operator PRs, performing code review, self-reviewing before submission, or when the user mentions "review PR", "审PR", "代码review", "code review", "review #123", "self-review", "自审", "检查PR", "审查算子", "review operator". It fetches PR diffs, applies FlagGems domain-specific review rules (structural checks, naming, registration, tests, benchmarks), and posts inline review comments directly on GitHub.
- ▌ Choose Tile Size And Work Partitioning · wenyi-li bundleSkill: Choose Tile Size and Work Partitioning
- ▌ Write Tensorrt Plugin Integration Plan · wenyi-li bundleSkill: Write a TensorRT Plugin Integration Plan
- ▌ Triton Ascend Ascend Hardware Constraints · wenyi-liAscend 硬件约束与编译器限制速查。涵盖 CUBE/VEC 存储层级预算计算方法、bishengIR 编译器已知限制、strided access 性能特征。适用于所有 Triton Ascend 算子生成和调试场景。
- ▌ Tilegym Monkey Patch Kernels To Transformers · wenyi-li bundleIntegrate TileGym kernels into Hugging Face `transformers` models by replacing the library's submodule(s) and certain class(es)' implementations, and patching certain class(es)' init/forward/load weight methods prior to instantiating models. Used when the user requires integrating TileGym kernels into `transformers` models.
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌ Pypto Loop View · wenyi-lipypto.loop + pypto.view 的正确写法：view shape 必须是编译期常量，适用于 matmul、norm、elementwise 等所有 loop 场景
- ▌ Pypto Case Norm Batchnorm · wenyi-li模式 C 示例：3D Norm — BatchNorm，展示 3D 降维、连续单轴 sum 多维归约、expand_clone 广播
- ▌ Perf Host Optimization · wenyi-li bundleProfiles and optimizes TensorRT-LLM host/CPU overhead using line_profiler (with nsys support planned). Runs iterative profile-analyze-optimize-validate rounds. Use when GPU utilization is low or optimizing PyExecutor throughput.