triton-analyzer

---

name: triton-analyzer description: Analyze Triton (torch.compile fused) kernels for roofline efficiency. Use when orchestrator needs Triton category analysis. model: claude-opus-4-7-high

Triton Analysis Subagent

Analyze Triton (torch.compile / inductor) fused GPU kernels for roofline efficiency. Renders P-items from the per-category findings the analyzer script has already grouped and gated.

---

Context Passing

When invoked by the orchestrator, you will receive the following context:

Required context provided by orchestrator:

• output_dir: Base analysis output directory
• prefix: Command prefix from <output_dir>/cache/cmd_prefix.txt — contains a template with {CMD} placeholder; substitute {CMD} with the actual command
• comparison_scope: standalone (default) or comparative

Input files (pre-computed by orchestrator):

1. <output_dir>/category_data/triton_ops.csv - Filtered Triton operations (includes call_stack column for architecture context)
2. <output_dir>/metadata/triton_metadata.json - Hardware specs

Output file you must write:

• <output_dir>/category_findings/triton_findings.md

---

Error Handling

If category data files are missing:

1. Write a findings file noting: "No Triton operations found in trace"
2. Return gracefully

If analysis script fails:

1. Write a findings file with Status: ERROR
2. CRITICAL: Do NOT manually analyze the raw CSV data
3. CRITICAL: Do NOT provide any bottleneck findings

---

Language Guidelines

Use vendor-agnostic terminology:

• "GPU kernels" not "CUDA kernels"
• "Triton fused kernels" or "torch.compile fused kernels" for the category
• Focus on operation semantics, not vendor implementation details

---

Analysis Workflow

Step 1: Run Analysis Script

<prefix> python3 \
  TraceLens/Agent/Analysis/category_analyses/triton_analysis.py \
  --output-dir <output_dir>
  --comparison_scope <comparison_scope>