{
  "schema_version": 1,
  "artifact_index_sha256": "f066a0154a9101b359c3a4d4fa6a83fb12b7126a9cf0f19cb772611d08cc07ee",
  "base_awq_checkpoint_index_sha256": "da62ba7bfef57f8117f35fc78bf5daf14c0ab5eb1b8173b4d712c1f4d294819b",
  "host_side_checkpoint_audit": {
    "status": "PASS",
    "indexed_tensor_count": 222747,
    "carrier_checkpoint_completion_sha256": "70fb852c73620edbf75858ab36c519005a5b4cec6597d2493c1355b4b901f077"
  },
  "awq_calibration_expert_coverage": {
    "status": "PASS_WITH_SELECTIVE_AUGMENTATION",
    "report_filename": "EXPERT_COVERAGE.json",
    "natural_unique_expert_ids": "512/512",
    "natural_layer_expert_pairs": "24568/24576",
    "natural_layer_expert_coverage_fraction": 0.9996744791666666,
    "fully_naturally_covered_layers": "45/48",
    "zero_natural_token_pairs": 8,
    "below_128_natural_token_pairs": 93,
    "selectively_augmented_layer_expert_pairs": 93,
    "runtime_zero_coverage_fallback_pairs": 0,
    "calibration_records": 684,
    "calibration_active_tokens": 202750,
    "native_route_matrix_sha256": "6c987b6d2a1506978f38bc0b4fa167d205250ba8f0629f05adddc880a26828b1",
    "selective_bypass_target_manifest_sha256": "9d4c2372bb065d2f9aadce457d98b0314c3844e1989e8fc6e36ce2a7088770c9"
  },
  "awq_fp16_kv_quality": {
    "status": "ACCEPTED",
    "contract": {
      "hardware": "4x NVIDIA Tesla V100-PCIE-32GB",
      "tensor_parallel_size": 4,
      "mtp": 0,
      "activation_dtype": "float16",
      "kv_cache_dtype": "float16",
      "enforce_eager": true,
      "prefix_caching": false
    },
    "results": {
      "basic_generation": "4/4",
      "needle_retrieval_1k_to_128k": "6/6 exact",
      "heldout_tool_selection": "10/12",
      "repeat_exact_stability": "6/6 cases over two repeats",
      "gsm8k_five_shot_subset": "29/32",
      "humaneval_mbpp_functional_subset": "9/10",
      "ifeval_strict": "3/5 prompts; 9/12 instructions",
      "ifeval_loose": "3/5 prompts; 9/12 instructions"
    },
    "acceptance_marker_sha256": "47ceb8120dede77f057d9a1ab62196e6dd476c8e5a8512257028069e12c48a81"
  },
  "calibrated_qsa_e4m3_kv_quality": {
    "status": "ACCEPTED",
    "scale_overlay_bundled": true,
    "scale_filename": "model-kvscales.safetensors",
    "merged_index_entry_count": 24,
    "scale_file_sha256": "c89b08bacde7d869342190b2f95046a83af37080241ae00ccae3a3d6888852fd",
    "scale_tensor_count": 24,
    "contract": {
      "hardware": "4x NVIDIA Tesla V100-PCIE-32GB",
      "tensor_parallel_size": 4,
      "mtp": 0,
      "activation_dtype": "float16",
      "main_qsa_kv_cache_dtype": "fp8_e4m3",
      "qsa_indexer_cache_dtype": "float16",
      "enforce_eager": true,
      "prefix_caching": false,
      "gpu_memory_utilization": 0.89
    },
    "calibration": {
      "release_distribution_processed_tokens": 4130597,
      "coverage": [
        "Chinese",
        "English",
        "code",
        "multi-turn dialogue",
        "tool use",
        "1K context",
        "16K context",
        "64K context",
        "128K context"
      ],
      "scale_min": 0.01710728369653225,
      "scale_max": 0.080636166036129,
      "maximum_saturation_ratio": 0,
      "report_sha256": "0612b360e005c8c3cb9332a8cd38d1c13b834f14831584514b9b1775422a9a31"
    },
    "results": {
      "heldout_tool_selection": "10/12; zero correctness regressions versus route-neutral FP16",
      "heldout_tool_exact_stability": "4/6 cases over two repeats; route-neutral FP16 was 3/6",
      "needlebench_128k": "6/6 exact; zero correctness regressions versus route-neutral FP16",
      "matched_first_tokens": "18/18",
      "gsm8k_five_shot_subset": "29/32; zero flips versus FP16",
      "humaneval_mbpp_functional_subset": "9/10; identical to FP16",
      "ifeval_strict": "3/5 prompts; 9/12 instructions; identical to FP16",
      "ifeval_loose": "3/5 prompts; 9/12 instructions; identical to FP16",
      "selected_block_recall_mean": 0.9959058567,
      "selected_block_recall_min": 0.9918218198,
      "qsa_output_cosine_min": 0.9985972106,
      "qsa_output_relative_l2_max": 0.0529616075
    },
    "acceptance_marker_sha256": "0ed39d23b078d449052f58bf254030cc5c6658e4e9d5f2853926a6ffe4e70d8d"
  },
  "image_video_validation": {
    "status": "ACCEPTED_WITH_PENDING_QWEN4EXP_ENABLEMENT",
    "qwen4exp_enablement_in_current_release": false,
    "qwen4exp_enablement_pr": "https://github.com/1CatAI/1Cat-vLLM/pull/458",
    "qwen4exp_enablement_commit": "dd68b9e8f44d248ad7082b871243a66c5fc1696a",
    "production_accepted": false,
    "contract": {
      "hardware": "4x NVIDIA Tesla V100-PCIE-32GB",
      "tensor_parallel_size": 4,
      "mtp": 0,
      "activation_dtype": "float16",
      "enforce_eager": true,
      "prefix_caching": false,
      "gpu_memory_utilization": 0.89,
      "maximum_model_length": 32768,
      "maximum_sequences": 1,
      "video_frames": 16,
      "repeats": 2
    },
    "scope": {
      "image_cases": 5,
      "video_cases": 4,
      "requests_per_kv_arm": 18,
      "maximum_observed_prompt_tokens": 3182
    },
    "visual_payload_integrity": {
      "tensor_count": 333,
      "bitwise_equal_to_reference": "333/333",
      "mismatch_count": 0,
      "aggregate_sha256": "cb57bb7af556cab0c09f457e72fe0894b42d7db08437478a62beb760b69c984c"
    },
    "results": {
      "fp16_kv_semantic_score": "18/18",
      "e4m3_calibrated_kv_semantic_score": "18/18",
      "successful_http_requests_per_arm": "18/18",
      "runtime_or_request_errors": 0,
      "repeat_exact_stability_per_arm": "9/9 cases",
      "cross_arm_exact_answers": "16/18",
      "cross_arm_whitespace_normalized_answers": "18/18",
      "e4m3_calibrated_scale_gate": "24/24"
    },
    "capacity_observation": {
      "fp16_kv_tokens": 34859,
      "e4m3_kv_tokens": 59234,
      "ratio": 1.699247826960039,
      "formal_performance_claim": false
    },
    "limitations": [
      "Validation requires Qwen4Exp model-integration PR #458 until it is included in a 1Cat-vLLM release.",
      "This was a compact regression set rather than a broad multimodal benchmark.",
      "Long-context multimodal composition was not tested."
    ],
    "internal_report_sha256": "9432f9978a480e90c9d134ab7b52cce57c1dc0f294b1872a424504baf1a56e0d"
  },
  "calibrated_qsa_e4m3_kv_performance": {
    "status": "ACCEPTED",
    "contract": {
      "hardware": "4x NVIDIA Tesla V100-PCIE-32GB",
      "tensor_parallel_size": 4,
      "mtp": 0,
      "activation_dtype": "float16",
      "gpu_memory_utilization": 0.89,
      "contexts": [1024, 4096, 16384, 32768, 65536, 131072],
      "concurrencies": [1, 4, 8],
      "output_tokens": 256,
      "scored_repeats": 3,
      "chunked_prefill_tokens": 8192,
      "cuda_graphs": "FULL_AND_PIECEWISE",
      "prefix_caching": false
    },
    "capacity": {
      "fp16_kv_tokens": 310480,
      "e4m3_kv_tokens": 569579,
      "ratio": 1.834511079618655
    },
    "matrix": {
      "grid_cells": 18,
      "matched_ok_cells": 15,
      "fp16_ok_cells": 15,
      "e4m3_ok_cells": 17,
      "fp16_scored_requests": 174,
      "e4m3_scored_requests": 210,
      "request_failures": 0
    },
    "matched_16k_and_longer_geometric_deltas_pct": {
      "prefill_throughput": -3.42101961847342,
      "decode_throughput": -2.5445261709874822,
      "server_e2e_elapsed": 3.1165392932944957
    },
    "caveats": [
      "FP16 P1024 and generic E4M3 G6 P256 XQA use different reduction structures.",
      "Bitcast decoding and scale hoisting were tested together rather than isolated.",
      "The result supports a capacity gain with a small long-context latency cost, not a general speedup claim."
    ],
    "analysis_sha256": "bcf263b3a623b15eff953148740ec22b98f0f7a779eb648f4dd321e958227bfb",
    "acceptance_marker_sha256": "c0f3acc6f1e7e9c4b7f5450a95035d86513a9f0aa45f1967951f205197b413b7"
  },
  "not_accepted": [
    "TP8",
    "MTP",
    "prefix caching",
    "256K context",
    "generic Transformers loading",
    "upstream vLLM",
    "Qwen4Exp multimodal startup without the pending enablement change",
    "production deployment"
  ],
  "huggingface_upload_authorized": false,
  "huggingface_upload_performed": false
}