triton-cuda-optimization

triton-cuda-optimization is a skill for Claude Code, Codex from mindspore-ai/akg. It costs 66 tokens per session (1,768 once invoked), scanned A, original, Apache-2.0.

A general guide to optimizing Triton kernels on CUDA GPUs. It covers block size, warps, pipeline stages, memory access, operation fusion, occupancy, numerical stability, and API limits.

In plain words
What is it for?
Use it to tune kernel launch parameters, improve memory behavior, balance GPU occupancy, combine suitable operations, and avoid Triton CUDA limitations.
Why use it?
GPU performance depends on several settings that interact with one another. This guide collects the main tuning and debugging considerations in one place.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one.

Good fit Use it to tune kernel launch parameters, improve memory behavior, balance GPU occupancy, combine suitable operations, and avoid Triton CUDA limitations.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/mindspore-ai/akg/triton-cuda-optimization
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add mindspore-ai/akg --skill triton-cuda-optimization
Clone the repo
git clone --depth 1 https://github.com/mindspore-ai/akg

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for triton-cuda-optimization

README.md
[![agentmods](https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-cuda-optimization/github.svg)](https://agentmods.dev/skills/mindspore-ai/akg/triton-cuda-optimization)
Your own site
<a href="https://agentmods.dev/skills/mindspore-ai/akg/triton-cuda-optimization"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-cuda-optimization/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for triton-cuda-optimization

Your own site · 80×15
<a href="https://agentmods.dev/skills/mindspore-ai/akg/triton-cuda-optimization"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-cuda-optimization.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 66 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,768 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector pass 7 Sept 2026
How audits are shown
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00066 $0.01768
Opus 5 $0.00033 $0.00884
Sonnet 5 $0.00013 $0.00354
Haiku 4.5 $0.00007 $0.00177

Measured 9d ago against content hash 3deaf984d615, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-12, from the pricing page.

Security

Grade A, and why

triton-cuda-optimization scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

akg_agents/python/akg_agents/op/resources/skills/triton-cuda/guides/triton-cuda-optimization/SKILL.md · 191 lines

How it starts

The opening of the file, as written. The whole thing — 191 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Triton CUDA 性能优化指南

1. 性能优化策略

1.1 块大小选择

  • 原则: 平衡并行度与资源占用
  • 建议: 使用 2 的幂次(256, 512, 1024)
  • GPU 考量: 需要足够多的 warp 来隐藏延迟

1.2 Warp 和 Stage 调优

CUDA 后端特有的两个重要参数:

  • num_warps: 每个 block 的 warp 数量(每个 warp = 32 个线程)

    • 小 BLOCK_SIZE:使用较少的 warp (2-4)
    • 大 BLOCK_SIZE:使用较多的 warp (4-8)
    • MatMul:通常使用 4-8 个 warp
  • num_stages: 软件流水线级数

    • 更多的 stage 可以更好地隐藏内存延迟
    • 但会占用更多共享内存
    • 通常 2-5 之间选择
@triton.autotune(
    configs=[
        triton.Config({'BLOCK_SIZE': 1024}, num_warps=4, num_stages=3),
        triton.Config({'BLOCK_SIZE': 512}, num_warps=2, num_stages=4),
    ],
    key=['n_elements'],
    restore_value=['output_ptr'],  # 必须:列出所有输出指针参数名
)

1.3 内存访问优化

  • 合并访问 (Coalesced Access): 同一 warp 内的线程应访问连续内存地址
  • 2D数据: 优先使用 tl.make_block_ptr 配合 boundary_check
  • 步幅设计: 仔细设计 stride 参数,错误设置会严重影响性能
  • 数据布局: 保持内存访问的连续性和局部性

1.4 算子拆分策略

  • 复杂算子: 拆分为多个简单 kernel,避免单个 kernel 过于复杂
  • 融合策略: 适度融合以减少全局内存读写(如 fused attention)
  • 平衡: CUDA 后端融合通常比 NPU 更有效,但仍需注意 register pressure

1.5 Occupancy 优化

GPU 利用率(Occupancy)是性能的关键指标:

  • 寄存器使用: 减少每个线程的寄存器使用量,增加并发 block 数
  • 共享内存: 合理使用共享内存,不超过硬件限制
  • Block 大小: 选择能整除 SM 最大线程数的 block 大小

2. 数值稳定性

2.1 防溢出处理

Softmax 数值稳定化:

# 减去最大值防止 exp 溢出
max_val = tl.max(scores, axis=0)
scores = scores - max_val
p = tl.exp(scores)  # CUDA 后端直接使用 tl.exp

2.2 防负值开方

# 方差计算前确保非负
variance = tl.maximum(variance, 0.0)
std = tl.sqrt(variance + eps)

2.3 精度提升

  • 使用 float32 进行累加: 即使输入是 float16/bfloat16
  • 最后再转换: 计算完成后再转回目标精度
  • TF32: Ampere+ GPU 上可使用 TF32 加速 MatMul
accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
# ... 累加计算 ...
result = tl.cast(accumulator, output_dtype)

3. API 使用限制

3.1 禁止使用的语法

禁止使用: return, break, continue, lambda

Triton 内核是一次性执行完整逻辑,不支持提前返回或跳转语句。

# 错误:使用 return
@triton.jit
def kernel(ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    if pid >= n:
        return  # 编译错误!

# 正确:使用 mask
@triton.jit
def kernel(ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offsets = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offsets < n
    data = tl.load(ptr + offsets, mask=mask, other=0.0)
    # ... 所有代码都在同一层级执行

Read the full file on GitHub · 191 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 9d ago First seen · 191 lines · 66 tokens per session scan A 3deaf984d615

Subscribe to this mod's changes

triton-cuda-optimization is a skill published in the GitHub repository mindspore-ai/akg (259 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 66 tokens to every session and 1,768 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.

Related

Other skills, from other repositories

agent-platform-rag-engine-management

Manage and query Agent Platform RAG Engine Corpora and retrieve grounded contexts using the Google GenAI SDK. Use when listing RAG corpora or files, inspecting a corpus, retrieving contexts, or generating content grounded in a RAG corpus. Do not use for standard database queries (use SQL/Spanner skills), Google…

google/skills · 85 tokens

agent-platform-model-registry

Agent Platform Model Registry Management. Use when you need to upload, list, describe, update, or delete machine learning models (and their versions) in the Agent Platform Model Registry. Don't use for model training, model deployment to endpoints, or managing non-Agent Platform models.

google/skills · 60 tokens

foundry-config-setup

Resolve missing setup caused by a hardcoded Foundry project endpoint or model in a sample. Use when a sample fails because it uses a placeholder/hardcoded projectendpoint (for example "https://your-project.services.ai.azure.com") or a hardcoded model instead of reading them from the environment.

microsoft/agent-framework · 65 tokens

google-cloud-solution-agentic-analytics-spark-knowledge-catalog

Discovers requirements and generates guidance to design and deploy a governed, secure agentic-analytics solution for data that's distributed across Google Cloud, other cloud providers, or on-premises. Data that's outside Google Cloud (such as data from Databricks, Snowflake, Salesforce, SAP, or Oracle systems) is…

google/skills · 138 tokens

training-check

Interactively monitor training metrics from the current Codex session, periodically checking WandB or fallback logs for NaN, divergence, plateaus, and broken runs.

wanshuiyin/Auto-claude-code-research-in-sleep · 35 tokens

nemo-automodel-launcher-config

Configure NeMo AutoModel job launches for interactive runs, Slurm clusters, and SkyPilot cloud execution.

NVIDIA/skills · 30 tokens