{
  "id": "cuda/triton-asm-unsupported",
  "signature": "RuntimeError: Triton compilation failed: unsupported instruction 'mma.sync.aligned.m16n8k16.row.col.f16.f16.f16'",
  "signature_zh": "运行时错误：Triton编译失败：不支持的指令 'mma.sync.aligned.m16n8k16.row.col.f16.f16.f16'",
  "regex": "Triton compilation failed: unsupported instruction",
  "domain": "cuda",
  "category": "build_error",
  "subcategory": null,
  "root_cause": "A Triton kernel uses a PTX instruction (e.g., mma.sync) that is not supported by the target GPU architecture, often due to an older GPU or incorrect compute capability.",
  "root_cause_type": "generic",
  "root_cause_zh": "Triton内核使用了目标GPU架构不支持的PTX指令（例如mma.sync），通常是由于GPU较旧或计算能力不正确。",
  "versions": [
    {
      "version": "Triton 2.1.0",
      "introduced": null,
      "deprecated": null,
      "removed": null,
      "behavior_change": null,
      "status": "active"
    },
    {
      "version": "PyTorch 2.2.0",
      "introduced": null,
      "deprecated": null,
      "removed": null,
      "behavior_change": null,
      "status": "active"
    },
    {
      "version": "NVIDIA T4 (sm_75)",
      "introduced": null,
      "deprecated": null,
      "removed": null,
      "behavior_change": null,
      "status": "active"
    }
  ],
  "os_specific": {},
  "dead_ends": [
    {
      "action": "",
      "why_fails": "The error is hardware-limited, not software.",
      "fail_rate": 0.7,
      "condition": "",
      "sources": []
    },
    {
      "action": "",
      "why_fails": "Triton uses its own JIT compiler, independent of TensorExpr.",
      "fail_rate": 0.9,
      "condition": "",
      "sources": []
    }
  ],
  "workarounds": [
    {
      "action": "Run the kernel on a GPU with compute capability >= 8.0 (Ampere or newer). Alternatively, disable Triton by setting environment variable `TORCHDYNAMO_USE_TRITON=0` to fall back to CUDA kernels.",
      "success_rate": 0.8,
      "how": "Run the kernel on a GPU with compute capability >= 8.0 (Ampere or newer). Alternatively, disable Triton by setting environment variable `TORCHDYNAMO_USE_TRITON=0` to fall back to CUDA kernels.",
      "condition": "",
      "sources": []
    }
  ],
  "workarounds_zh": [
    "Run the kernel on a GPU with compute capability >= 8.0 (Ampere or newer). Alternatively, disable Triton by setting environment variable `TORCHDYNAMO_USE_TRITON=0` to fall back to CUDA kernels."
  ],
  "transition_graph": {
    "leads_to": [],
    "preceded_by": [],
    "frequently_confused_with": []
  },
  "official_doc_url": "https://triton-lang.org/main/index.html",
  "official_doc_section": null,
  "error_code": null,
  "verification_tier": "ai_generated",
  "confidence": 0.86,
  "fix_success_rate": 0.8,
  "resolvable": "partial",
  "first_seen": "2024-06-12",
  "last_confirmed": "2024-06-01",
  "last_updated": "2024-06-01",
  "evidence_count": 1,
  "tags": [],
  "locale": "en",
  "aliases": []
}