{
  "benchmark_id": "cosyvoice2-titan-20260415",
  "date": "2026-04-15",
  "session": 107,
  "branch": "cosyvoice2-bench-20260415",
  "target_model": "FunAudioLLM/CosyVoice2-0.5B",
  "target_model_release": "master HEAD 2026-03-16",
  "verdict": "deferred_blackwell_sm121_pin_incompatibility",
  "benchmark_run": false,
  "license_passes": true,
  "license_detail": {
    "code": "Apache-2.0",
    "weights": "Apache-2.0",
    "primary_source": "https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B (license tag: apache-2.0)",
    "commercial_use_allowed": true
  },
  "blocker": {
    "type": "infrastructure",
    "detail": "requirements.txt pins incompatible with Blackwell SM_121 (DGX Spark aarch64 / GB10 / CUDA 13)",
    "incompatible_pins": {
      "torch": {
        "pinned_version": "2.3.1",
        "problem": "Blackwell SM_121 requires torch 2.5+ (compute capability registration landed in 2.5 nightly, stabilized 2.6+)",
        "fix": "NGC container nvcr.io/nvidia/pytorch:25.12-py3 ships torch 2.6.0+ with SM_121 kernels; pip downgrade unsafe due to ABI mismatch"
      },
      "tensorrt-cu12": {
        "pinned_version": "10.13.3.9",
        "problem": "CUDA 12 build; Blackwell SM_121 host requires CUDA 13",
        "fix": "Swap to tensorrt-cu13==10.x (aarch64 wheel exists); possible API drift in 10.13.x cu12→cu13 needs source patch"
      },
      "onnxruntime-gpu": {
        "pinned_version": "1.18.0",
        "problem": "Last aarch64 wheel on PyPI was 1.17.x for cu12; 1.18+ aarch64 not published",
        "fix": "Build ORT from source on Titan OR use NGC bundled ORT OR drop ONNX codepath"
      },
      "deepspeed": {
        "pinned_version": "0.15.1",
        "problem": "CUDA extension compilation with cu13 has known issues on 0.15.x",
        "fix": "Bump to 0.16+ (safer)"
      }
    },
    "ecosystem_corroboration": [
      "https://github.com/vllm-project/vllm/issues/36821 — No sm_121 support on aarch64 (DGX Spark)",
      "https://forums.developer.nvidia.com/t/.../349881 — DGX Spark CUDA install pitfalls",
      "https://forums.developer.nvidia.com/t/.../357663 — DGX Spark SM121 software support severely lacking"
    ]
  },
  "estimated_port_cost_hours": {
    "best_case": 1,
    "likely_case": 3,
    "worst_case": 8,
    "assumptions": "Best case: torch/TRT swap works cleanly, ORT codepath can be dropped. Worst case: TRT-LLM optimization paths require source changes, DeepSpeed compile issues, transformers/diffusers interlocking version resolution."
  },
  "time_budget_per_plan": "1 hour (docs/NEXT-SESSION-F5-TTS-BENCH.md §196)",
  "positive_findings_for_future": {
    "streaming_ttfa_claim_ms": 150,
    "languages_supported": 9,
    "languages_list": ["Chinese", "English", "Japanese", "Korean", "German", "Spanish", "French", "Italian", "Russian"],
    "chinese_dialects_plus": 18,
    "zero_shot_cross_lingual_clone": true,
    "model_size_params": "0.5B (fits alongside Gemma 4 26B on Titan)"
  },
  "revisit_triggers": [
    "Chatterbox Samantha quality insufficient on production phone calls",
    "Indic-language TTS needed (Chatterbox is English-only)",
    "Someone upstream ships CosyVoice 2 with cu13/torch 2.6+ pins (monitor requirements.txt)"
  ],
  "samantha_reference_status": "available at services/audio-pipeline/voice-references/samantha_movie_primary.wav — reusable when/if port happens",
  "chatterbox_gemma_impact": {
    "gemma_paused": false,
    "pre_flight_p50_ms": 190,
    "post_flight_p50_ms": "n/a (no install attempted)",
    "drift_ratio": "n/a"
  },
  "decision_rule_applied": {
    "rule": "Plan §132: IF CosyVoice 2 also fails same rule → keep_chatterbox (survey complete, Chatterbox wins)",
    "result": "keep_chatterbox_survey_complete"
  },
  "final_verdict_2026_04_tts_survey": {
    "winner": "Chatterbox (already deployed, MIT, working on Panda :8772)",
    "rejected_candidates": [
      {"name": "Voxtral-4B-TTS-2603", "reason": "encoder weights withheld from open-source release (session 106)"},
      {"name": "F5-TTS v1.1.18", "reason": "weights CC-BY-NC-4.0 non-commercial (session 107)"},
      {"name": "CosyVoice2-0.5B", "reason": "Blackwell SM_121 pin incompatibility, deferred (session 107)"}
    ],
    "untested_remaining": ["Chatterbox-Turbo (if/when released)"]
  }
}