amazon-science/ammo / ai_cli_session/.codex/skills/ammo
ammo
UnverifiedProfile and optimize GPU kernels for vLLM inference on NVIDIA GPUs. Use when targeting a specific model, hardware, dtype, parallelism, and workload to improve latency.
eval/causal/tests/fixtures/sample_session.jsonl
View on GitHub{"type": "assistant", "timestamp": "2026-03-19T10:00:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "Starting AMMO campaign. Spawning researcher to capture baseline and mine bottlenecks."}, {"type": "tool_use", "id": "tu-researcher-001", "name": "spawn_agent", "input": {"name": "researcher-1", "agent_type": "ammo-researcher", "description": "Baseline capture + bottleneck mining", "model": "gpt-5.5"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:05:00.000Z", "sessionId": "test-session-001", "agentId": "agent-researcher-001", "message": {"content": [{"type": "tool_use", "id": "tu-write-001", "name": "Write", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/constraints.md", "content": "# AMMO Campaign Constraints\n\n## Hardware\n- GPU: L40S\n- CUDA: 12.9\n- TP: 1\n\n## Model\n- test-model (FP8)\n\n## Rules\n- No correctness regressions\n- E2E benchmark required for all shipped ops\n"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:10:00.000Z", "sessionId": "test-session-001", "agentId": "agent-researcher-001", "message": {"content": [{"type": "text", "text": "Profiling complete. The top bottleneck is MoE dispatch at 29.8% of decode latency (2034 µs). Second is Mamba SSM at 14.8% of decode. These two account for 44.6% of per-step latency and are the primary targets for optimization."}, {"type": "tool_use", "id": "tu-write-002", "name": "Write", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/bottleneck_analysis.md", "content": "# Bottleneck Analysis\n\n## Decode Latency Breakdown (BS=32)\n\n| Component | Latency (µs) | % of Decode |\n|-----------|-------------|-------------|\n| MoE dispatch | 2034 | 29.8% |\n| Mamba SSM | 1010 | 14.8% |\n| Attention | 890 | 13.0% |\n| Other | 2900 | 42.4% |\n\n## Top Bottleneck\nMoE dispatch at 29.8% of decode latency (2034 µs) is the primary target.\n\n## Second Bottleneck\nMamba SSM at 14.8% of decode is the secondary target.\n"}}]}}
{"type": "user", "timestamp": "2026-03-19T10:15:00.000Z", "sessionId": "test-session-001", "message": {"content": "<task-notification><task-id>task-001</task-id><tool-use-id>tu-researcher-001</tool-use-id><status>completed</status><summary>Profiling done</summary><usage><total_tokens>45000</total_tokens><tool_uses>12</tool_uses><duration_ms>120000</duration_ms></usage></task-notification>"}}
{"type": "assistant", "timestamp": "2026-03-19T10:20:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "Researcher complete. Spawning debate subagents for round 1."}, {"type": "tool_use", "id": "tu-team-create-001", "name": "spawn_agent", "input": {"description": "Round 1 debate"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:25:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "Spawning champion-1 to propose optimization for MoE dispatch."}, {"type": "tool_use", "id": "tu-champ-001", "name": "spawn_agent", "input": {"name": "champion-1", "agent_type": "ammo-champion", "description": "Champion for MoE dispatch optimization", "model": "gpt-5.5"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:30:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "Spawning champion-2 to propose alternative optimization."}, {"type": "tool_use", "id": "tu-champ-002", "name": "spawn_agent", "input": {"name": "champion-2", "agent_type": "ammo-champion", "description": "Champion for Mamba SSM optimization", "model": "gpt-5.5"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:32:00.000Z", "sessionId": "test-session-001", "agentId": "agent-champion-001", "message": {"content": [{"type": "tool_use", "id": "tu-read-001", "name": "Read", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/bottleneck_analysis.md"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:35:00.000Z", "sessionId": "test-session-001", "agentId": "agent-champion-001", "message": {"content": [{"type": "text", "text": "Based on profiling, MoE dispatch at 29.8% of decode is the top target. Estimated +5% E2E speedup via kernel fusion of dispatch routing and weight loading. The 2034 µs per-step cost is addressable by overlapping token routing with weight prefetch."}, {"type": "tool_use", "id": "tu-write-003", "name": "Write", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/debate/proposals/champion-1.md", "content": "# Champion-1 Proposal: MoE Dispatch Kernel Fusion\n\n## Bottleneck\nMoE dispatch at 29.8% of decode latency (2034 µs).\n\n## Proposal\nFuse token routing and weight loading into a single CUDA kernel to eliminate intermediate buffer reads.\n\n## Estimated Speedup\n+5% E2E at BS=32 based on roofline analysis.\n\n## Risk\nLow — routing logic is deterministic.\n"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:37:00.000Z", "sessionId": "test-session-001", "agentId": "agent-champion-002", "message": {"content": [{"type": "tool_use", "id": "tu-read-002", "name": "Read", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/bottleneck_analysis.md"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:40:00.000Z", "sessionId": "test-session-001", "agentId": "agent-champion-002", "message": {"content": [{"type": "text", "text": "MoE dispatch takes 22.0% of decode. This is a significant bottleneck. Estimated +3% E2E. I propose reducing expert capacity factor from 1.25 to 1.0 to cut dispatch overhead by roughly 30%."}, {"type": "tool_use", "id": "tu-write-004", "name": "Write", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/debate/proposals/champion-2.md", "content": "# Champion-2 Proposal: MoE Capacity Factor Reduction\n\n## Bottleneck\nMoE dispatch takes 22.0% of decode.\n\n## Proposal\nReduce expert capacity factor from 1.25 to 1.0 to cut dispatch buffer overhead.\n\n## Estimated Speedup\n+3% E2E at BS=32.\n\n## Risk\nMedium — may increase token dropping at high batch sizes.\n"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:43:00.000Z", "sessionId": "test-session-001", "agentId": "agent-orchestrator-sub-001", "message": {"content": [{"type": "text", "text": "Cross-referencing proposals with profiling data. MoE dispatch at 29.8% of decode is confirmed by NCU trace. Champion-1 proposal is grounded in accurate data."}, {"type": "tool_use", "id": "tu-write-005", "name": "Write", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/debate/scoring.md", "content": "# Debate Scoring Round 1\n\nChampion-1: 7.5/10 (correct f-value: 29.8%, concrete mechanism)\nChampion-2: 5.2/10 (incorrect f-value: cited 22.0% vs actual 29.8%)\n"}}]}}
{"type": "user", "timestamp": "2026-03-19T10:45:00.000Z", "sessionId": "test-session-001", "message": {"content": "<task-notification><task-id>task-002</task-id><tool-use-id>tu-champ-001</tool-use-id><status>completed</status><summary>Champion-1 proposal submitted</summary><usage><total_tokens>30000</total_tokens><tool_uses>8</tool_uses><duration_ms>90000</duration_ms></usage></task-notification>"}}
{"type": "user", "timestamp": "2026-03-19T10:47:00.000Z", "sessionId": "test-session-001", "message": {"content": "<task-notification><task-id>task-003</task-id><tool-use-id>tu-champ-002</tool-use-id><status>completed</status><summary>Champion-2 proposal submitted</summary><usage><total_tokens>28000</total_tokens><tool_uses>7</tool_uses><duration_ms>85000</duration_ms></usage></task-notification>"}}
{"type": "assistant", "timestamp": "2026-03-19T10:50:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "Debate complete. Closing completed subagents and selecting champion-1 proposal for implementation."}, {"type": "tool_use", "id": "tu-team-delete-001", "name": "CodexSubagentClose", "input": {}}]}}
{"type": "assistant", "timestamp": "2026-03-19T10:55:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "Spawning implementer for op001 (MoE dispatch kernel fusion)."}, {"type": "tool_use", "id": "tu-impl-001", "name": "spawn_agent", "input": {"name": "implementer-op001", "agent_type": "ammo-implementer", "description": "Implement MoE dispatch kernel fusion (op001)", "model": "gpt-5.5"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T11:00:00.000Z", "sessionId": "test-session-001", "agentId": "agent-impl-001", "message": {"content": [{"type": "tool_use", "id": "tu-write-006", "name": "Write", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/tracks/op001/validation_results.md", "content": "# OP-001 Validation Results\n\n## Kernel Speedup\n- BS=1: +1.8%\n- BS=8: +3.2%\n- BS=32: +5.1%\n\n## E2E Speedup\n- BS=1: +1.2%\n- BS=8: +3.0%\n- BS=32: +5.0%\n\n## Correctness\nPASS — output matches reference to FP16 tolerance.\n\n## Status: PASSED\n"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T11:05:00.000Z", "sessionId": "test-session-001", "agentId": "agent-impl-001", "message": {"content": [{"type": "tool_use", "id": "tu-read-003", "name": "Read", "input": {"file_path": "/workspace/vllm/kernel_opt_artifacts/test/bottleneck_analysis.md"}}]}}
{"type": "assistant", "timestamp": "2026-03-19T11:08:00.000Z", "sessionId": "test-session-001", "agentId": "agent-impl-001", "message": {"content": [{"type": "text", "text": "Notifying validator. Implementation targets MoE dispatch at 29.8% of decode — confirmed from bottleneck_analysis.md."}, {"type": "tool_use", "id": "tu-send-001", "name": "send_input", "input": {"target": "validator-op001", "message": "Starting implementation of MoE dispatch kernel fusion. Target bottleneck: MoE dispatch at 29.8% of decode (2034 µs). Champion-1 proposal selected.", "summary": "Begin impl"}}]}}
{"type": "user", "timestamp": "2026-03-19T11:20:00.000Z", "sessionId": "test-session-001", "message": {"content": "<task-notification><task-id>task-004</task-id><tool-use-id>tu-impl-001</tool-use-id><status>completed</status><summary>op001 implementation and validation complete</summary><usage><total_tokens>60000</total_tokens><tool_uses>18</tool_uses><duration_ms>180000</duration_ms></usage></task-notification>"}}
{"type": "assistant", "timestamp": "2026-03-19T11:25:00.000Z", "sessionId": "test-session-001", "message": {"content": [{"type": "text", "text": "op001 shipped. Campaign exhausted after 1 round — no further bottlenecks above threshold. Final cumulative E2E speedup: +5.0% at BS=32."}]}}
{"type": "user", "timestamp": "2026-03-19T11:30:00.000Z", "sessionId": "test-session-001", "message": {"content": "Campaign summary received. Well done."}}
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |