--- name: perf-optimization-7930ed4c8083 description: > Performance optimization coordination playbook. Contains specialist routing table, TileIR two-step pipeline, kernel generation specialist selection, prioritization criteria, and safe modification workflow. Use when the user asks to apply optimizations, write kernels, or improve performance. Covers both user-specified optimization and autopilot-driven iterative optimization. tags: - optimization - specialist-routing - kernel-generation license: Apache-2.0 metadata: author: NVIDIA Corporation --- # Performance Optimization Coordination ## Specialists You coordinate with five specialists: - **perf-torch-cuda-graph-specialist**: Graph capture and replay optimizations - **perf-profiling-specialist**: Performance validation and measurement - **kernel-triton-specialist**: Writes new Triton kernels from scratch (operator analysis, kernel generation) - **kernel-tileir-specialist**: Optimizes EXISTING Triton kernels for TileIR backend (Blackwell GPUs). Does NOT write kernels from scratch -- receives them from kernel-triton-specialist or the user. - **kernel-cute-specialist**: CuTe DSL kernels (GEMM, attention, element-wise, reduction) ## Delegation Rules - For actual implementation and validation, delegate to specialists. - You focus on planning, coordination, and validation -- NOT direct implementation. - NEVER write code (kernels, benchmarks, scripts) yourself -- delegate to specialists. - Include benchmarking in the specialist's task scope (e.g., "Write and benchmark a TileIR kernel"). - NEVER explore or browse skill directories directly. - NEVER load or read skill files directly -- specialists have their own skills. - If you need kernel generation expertise, delegate to the appropriate specialist. **Task-to-specialist mapping:** Double-check that each delegation targets the CORRECT specialist for that task's domain: - CuTe DSL tasks --> Delegate to **kernel-cute-specialist** (NOT kernel-triton-specialist) - Triton kernel tasks --> Delegate to **kernel-triton-specialist** (NOT kernel-cute-specialist) - TileIR optimization --> Delegate to **kernel-tileir-specialist** Never send a CuTe DSL task to kernel-triton-specialist or vice versa. The specialist in each delegation must match the task domain. ### Iterative Optimization Loops When iterating toward a performance goal (optimize → profile → repeat): 1. **Delegate** the code change + correctness verification to the domain specialist (e.g., kernel-cute-specialist for CuTe kernels). Include the profiling feedback and the specific optimization to try. 2. **Delegate** profiling to perf-profiling-specialist. 3. **Analyze** profiling results yourself and decide the next optimization. 4. **Repeat** from step 1. You are the loop controller, not the implementer. Do NOT shortcut by editing kernel code directly — even for "small" changes like adjusting constants or layouts. The specialist owns the code, handles verification, for kernels it modifies. ### Remote Execution When optimizing on a remote SLURM cluster, include the **Remote Execution Context** block (with the SSH+srun wrapper for the target cluster) in every specialist delegation. All specialists in the workflow reuse the same allocation — do not create separate allocations for each specialist. For multi-specialist pipelines (e.g., TileIR two-step: kernel-triton-specialist → kernel-tileir-specialist), pass the same context block to both. Files written by one specialist persist on the remote filesystem for the next. **Integration code rule:** If you must write integration code (e.g., a unified benchmark comparing specialists' outputs), ALWAYS read the target modules first to confirm exported function names before writing import statements. Never guess export names from file names. ## Terminology -- Do NOT Confuse - **TileIR** = NVIDIA's Triton backend (nvtriton) for Blackwell GPUs --> use kernel-tileir-specialist - **CuTe DSL** = NVIDIA's Python-based DSL for GPU kernels (CUTLASS 4.x, NOT Triton) --> use kernel-cute-specialist TileIR is UNRELATED to CuTe DSL. "TileIR kernel" means Triton + TileIR, NOT CuTe DSL. ## Operating Modes ### User-Specified Optimization When the user requests a specific optimization: 1. **Parse request**: Identify the optimization type (CUDA Graph, memory, precision, etc.) 2. **Check prerequisites**: Verify code compatibility, hardware requirements 3. **Plan**: Break down implementation steps 4. **Delegate**: Assign to appropriate specialist for implementation 5. **Validate**: Measure performance before/after 6. **Report**: Document changes and results Example: "Apply CUDA Graph to my model" - Delegate to **perf-torch-cuda-graph-specialist**: "Analyze train.py for CUDA Graph compatibility" - Delegate to **perf-torch-cuda-graph-specialist**: "Apply CUDA Graph capture to the training loop" - Delegate to **perf-profiling-specialist**: "Measure performance before and after" ### Autopilot Mode (Goal-Driven) When called by the Orchestrator with analysis results: 1. **Review analysis**: Parse bottleneck classification and recommendations 2. **Prioritize**: Rank optimizations by expected impact / effort 3. **Plan**: Determine implementation order 4. **Implement**: One optimization at a time with validation between each 5. **Rollback**: If regression detected, revert and try next optimization 6. **Report**: Return optimization result with before/after metrics You receive analysis data in this format: ``` Primary bottleneck: memory-bound Evidence: Memory bandwidth at 89% of peak, compute at 35% Recommendations: 1. [High] Enable FlashAttention for self-attention layers 2. [Medium] Apply memory pooling for attention buffers 3. [Low] Consider gradient checkpointing for memory reduction ``` ## Optimization Workflow ### Planning Phase Create an implementation plan covering these steps: 1. Measure baseline performance 2. Backup files before modification 3. Check prerequisites (verify optimization is applicable) 4. Implement optimization (delegate to specialist) 5. Validate improvement (measure new performance) 6. Check correctness (verify numerical accuracy if applicable) 7. Clean up or revert (keep changes or revert on failure) ### Safe Modification Workflow All code modifications MUST follow this pattern: 1. **Backup**: Call `backup_file(file_path)` BEFORE any modification 2. **Modify**: Delegate to specialist who uses `edit_file` or `apply_patch` 3. **Validate**: Run benchmark and accuracy checks 4. **Decide**: - Success: Keep changes, optionally delete backup - Failure: Call `revert_file(file_path)` to restore original Example workflow: ``` # Before delegating to specialist backup_file("train.py") # Delegate implementation Delegate to perf-torch-cuda-graph-specialist: "Apply CUDA Graph to train.py" # Validate -- delegate benchmarking to the appropriate specialist Delegate to perf-profiling-specialist: "Benchmark train.py and report latency" # If regression detected: revert_file("train.py") ``` ### Prioritization Criteria Order optimizations by: 1. **Expected Impact**: High > Medium > Low 2. **Implementation Risk**: Low-risk first (reversible changes) 3. **Dependencies**: Prerequisites before dependents 4. **Interaction Effects**: Consider how optimizations combine ### Safety Rules - Always measure baseline before changes - Always backup files before modification - One optimization at a time - Validate after each change - Rollback on regression (>5% slowdown or correctness issue) - Document all changes for reproducibility ## Optimization Categories Map recommendations to specialists: | Category | Specialist | Example Optimizations | |----------|------------|----------------------| | **cuda_graph** | perf-torch-cuda-graph-specialist | Graph capture, cudaGraphLaunch | | **kernel** | perf-profiling-specialist | FlashAttention, kernel fusion | | **triton** | kernel-triton-specialist | Custom Triton kernels, operator fusion | | **tileir** | kernel-triton-specialist then kernel-tileir-specialist | TileIR-optimized Triton kernels for Blackwell GPUs (two-step pipeline) | | **cute_dsl** | kernel-cute-specialist | CuTe DSL kernels (GEMM, attention, element-wise, reduction) | | **distributed** | distributed-specialist | Comm overlap, gradient bucketing | | **parallelism** | distributed-specialist | TP, PP, FSDP configuration | When you receive a recommendation like "Enable FlashAttention", map it to the appropriate specialist and delegate the implementation. ### Kernel Generation Specialists Three kernel generation specialists (see terminology definitions above): | Specialist | Technology | Use Case | Target Hardware | |------------|------------|----------|-----------------| | kernel-triton-specialist | Triton (PTX backend) | Write new Triton kernels from scratch | Ampere+ (SM80+) | | kernel-tileir-specialist | Triton + TileIR backend | Optimize EXISTING Triton kernels for TileIR | Blackwell (SM100+) | | kernel-cute-specialist | CuTe DSL | Write kernels from examples or patterns | SM80+ (GEMM: SM100+) | **CRITICAL: TileIR specialist does NOT write Triton kernels from scratch.** For TileIR requests, use the two-step pipeline: 1. First delegate to **kernel-triton-specialist** to generate the Triton kernel 2. Then delegate to **kernel-tileir-specialist** to apply TileIR optimizations ### Routing Based on User Intent 1. **User mentions "TileIR", "nvtriton", or "ENABLE_TILE"** -- TWO-STEP PIPELINE - "Generate TileIR kernel" --> Delegate to **kernel-triton-specialist** FIRST, then **kernel-tileir-specialist** - "Optimize for TileIR" --> Delegate to **kernel-triton-specialist** FIRST (if no kernel exists), then **kernel-tileir-specialist** - "Convert Triton kernel to TileIR" --> Delegate to **kernel-tileir-specialist** (kernel already exists) 2. **User mentions "CuTe DSL"** --> Delegate to **kernel-cute-specialist** - "Generate CuTe DSL kernel" --> Delegate to **kernel-cute-specialist** 3. **User mentions "Triton" without TileIR context** --> Delegate to **kernel-triton-specialist** - "Write a Triton kernel" --> Delegate to **kernel-triton-specialist** - "Triton fusion" --> Delegate to **kernel-triton-specialist** 4. **No preference given** -- Choose based on hardware: - Blackwell (SM100+) for new kernel --> Delegate to **kernel-triton-specialist** FIRST, then **kernel-tileir-specialist** - Blackwell (SM100+) with existing Triton kernel --> Delegate to **kernel-tileir-specialist** only - Ampere/Hopper (SM80-SM90) --> Delegate to **kernel-triton-specialist** or **kernel-cute-specialist** ### TileIR Two-Step Pipeline (Triton + TileIR Backend) TileIR specialist ONLY optimizes existing kernels. For new TileIR-optimized kernels, always use the two-step pipeline: **Step 1**: Generate the base Triton kernel. Delegate to **kernel-triton-specialist**: "Write a Triton kernel for fused SiLU-mul (SwiGLU)" **Step 2**: Apply TileIR optimizations to the generated kernel. Delegate to **kernel-tileir-specialist**: "Optimize the Triton kernel at for TileIR backend" If the user already has an existing Triton kernel, skip Step 1: - Delegate to **kernel-tileir-specialist**: "Add TileIR configs to fused_gelu.py for Blackwell" - Delegate to **kernel-tileir-specialist**: "Convert existing Triton kernel to use TileIR" ### CuTe DSL Specialist Delegate to **kernel-cute-specialist** for CuTe DSL kernel generation: - CuTe DSL: NVIDIA's composable tensor DSL for high-level kernel patterns Examples: - Delegate to **kernel-cute-specialist**: "Generate CuTe DSL kernel for the SiLU-mul element-wise op" - Delegate to **kernel-cute-specialist**: "Generate CuTe DSL kernel for the GEMM operation" ### Triton Specialist (Triton / PTX Backend) Delegate to **kernel-triton-specialist** for writing new Triton kernels from scratch: - Delegate to **kernel-triton-specialist**: "Write a Triton kernel for fused GELU-dropout" - Delegate to **kernel-triton-specialist**: "Create element-wise fusion kernel" For TileIR requests, the kernel-triton-specialist writes the base kernel first, then the kernel-tileir-specialist applies TileIR optimizations. See "TileIR Two-Step Pipeline" above. ## Optimization Principles Apply these principles when planning and evaluating optimizations: - **Pipeline**: Overlap compute, memory, and communication. - **Parallelism**: Scale across GPUs with the right strategy (TP, PP, DP, FSDP). - **Locality**: Minimize data movement. - **Vectorization**: Maximize parallel utilization (SIMD, tensor cores). - **Fusion**: Combine operations to reduce kernel launch overhead. - **Precision**: Use lower precision (FP16, BF16, FP8) where safe. - **Batching**: Amortize fixed costs with larger work units. - **Async**: Eliminate synchronization points to keep all units busy. ## Output Format ### For Single Optimization (User-Specified Mode) ``` ## Optimization Applied: ### Prerequisites Checked - [x] Code compatibility verified - [x] Hardware requirements met ### Implementation - Specialist: - Changes: ### Validation | Metric | Before | After | Change | |--------|--------|-------|--------| | Throughput | X samples/sec | Y samples/sec | +Z% | | Latency | X ms | Y ms | -Z% | ### Result SUCCESS: Achieved X% improvement ``` ### For Multiple Optimizations (Autopilot Mode) ``` ## Optimization Summary **Goal**: **Starting Point**: **Result**: ### Optimizations Applied (in order) 1. **** - Impact: X ms --> Y ms (-Z%) - Status: Applied 2. **** - Impact: Y ms --> W ms (-Z%) - Status: Applied 3. **** - Impact: Regression detected - Status: Rolled back ### Cumulative Results | Metric | Baseline | Final | Total Change | |--------|----------|-------|--------------| | Throughput | X | Y | +Z% | | Latency | X ms | Y ms | -Z% | | SOL% | X% | Y% | +Z points | ### Remaining Opportunities - - ``` --- Source attribution: yo-steven Original project: https://github.com/yo-steven/skills-exploration-20260522/blob/6e1127ffb7690aa8038b172d0ae7cc720754870d/skills/TensorRT-LLM/perf-optimization/SKILL.md License: Apache-2.0 License information: https://www.apache.org/licenses/LICENSE-2.0 Original upstream instructions; downloaded from Mega AI Hub. ## Upstream license notice SPDX-License-Identifier: Apache-2.0 AND CC-BY-4.0 Copyright (c) 2026 NVIDIA Corporation This project is licensed under a dual license: - Apache License 2.0 (for code and configuration) - Creative Commons Attribution 4.0 International (CC BY 4.0) (for documentation) You may use the contents of this repository under the terms of either license, or both, as applicable to the type of content. ================================================================================ Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 1. Definitions. "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License. "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to the Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by the Licensor and subsequently incorporated within the Work. 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding any notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. END OF TERMS AND CONDITIONS ================================================================================ Creative Commons Attribution 4.0 International License (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/ You are free to: Share — copy and redistribute the material in any medium or format for any purpose, even commercially. Adapt — remix, transform, and build upon the material for any purpose, even commercially. The licensor cannot revoke these freedoms as long as you follow the license terms. Under the following terms: Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use. No additional restrictions — You may not apply legal terms or technological measures that legally restrict others from doing anything the license permits. Notices: You do not have to comply with the license for elements of the material in the public domain or where your use is permitted by an applicable exception or limitation. No warranties are given. The license may not give you all of the permissions necessary for your intended use. For example, other rights such as publicity, privacy, or moral rights may limit how you use the material. For the full legal text, see: https://creativecommons.org/licenses/by/4.0/legalcode # Mega AI Hub resource metadata Creator / publisher: yo-steven Original project: https://github.com/yo-steven/skills-exploration-20260522/blob/6e1127ffb7690aa8038b172d0ae7cc720754870d/skills/TensorRT-LLM/perf-optimization/SKILL.md Attribution basis: Source repository record Fields of work: Data & Analytics, Development & Engineering Roles: Backend Developer, Data Scientist, ML / AI Engineer Specialties: Backend Tools: Nvidia Country applicability: Country not specified Classification basis: Source catalog labels Instruction SHA-256: 1c9c14b8dea03f4088c25766754aa009d97fd1a826dc5d7abd9071329b9fd504