{
  "artifact": {
    "name": "Qwen3.8-Flash-Next-AWQ-g32",
    "proposed_repository": "leoncca/Qwen3.8-Flash-Next-AWQ-g32",
    "proposed_tag": "v1-formal-r6-da62ba7b",
    "schema_version": 1
  },
  "sources": {
    "base_model": {
      "repository": "Qwen/Qwen3.8-Flash-Next",
      "revision": "f5d08274bafd880402bd16f5e3e6c514136ec06c"
    },
    "ple": {
      "repository": "Qwen/Qwen3.8-Flash-Next-FP8",
      "revision": "bcd9f01ddc9cff2316eb84281bebcd5b058bddce"
    },
    "license": {
      "filename": "LICENSE",
      "name": "Qwen Community License 1.0",
      "sha256": "a0dc422560841fd68e06d974907f8b4c709bca44a67daad2b528437bdf676c08"
    }
  },
  "quantization": {
    "method": "AWQ",
    "weight_bits": 4,
    "activation_dtype": "float16",
    "group_size": 32,
    "zero_point": true,
    "checkpoint_format": "gemm",
    "layout": "per-expert asymmetric routed-expert projections",
    "quantized_projection_names": [
      "gate_proj",
      "up_proj",
      "down_proj"
    ],
    "protected": [
      "vision",
      "ple",
      "attention",
      "linear_attention",
      "router",
      "shared_expert",
      "embeddings",
      "lm_head",
      "hyper_connections",
      "mtp"
    ],
    "quantizer": "gptqmodel:7.3.2",
    "calibration_record_count": 684,
    "calibration_corpus_sha256": "d2fbb16c5dcfaf26b713fd44d090cc8a3f2bbb31f1713f179e149f1f196b3dbb",
    "processor_audit_sha256": "1ca4205795e83f3cf341cb3d8eb10ef7169bb4934a4435878af7bef8507a8c05",
    "selective_bypass_selected_pairs": 93,
    "selective_bypass_target_manifest_sha256": "9d4c2372bb065d2f9aadce457d98b0314c3844e1989e8fc6e36ce2a7088770c9",
    "selective_bypass_source_report_sha256": "69d421dd1a57cf04ea55c533b20f042e9dc6a99e2b6f489c54aae3d2039b05a1",
    "expert_coverage": {
      "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,
      "selective_augmentation_pairs": 93,
      "runtime_zero_coverage_fallback_pairs": 0,
      "native_route_matrix_sha256": "6c987b6d2a1506978f38bc0b4fa167d205250ba8f0629f05adddc880a26828b1"
    },
    "group_size_rationale": {
      "expert_intermediate_width": 640,
      "validated_tensor_parallel_size": 4,
      "local_intermediate_width": 160,
      "g32_divides_local_width": true,
      "g128_divides_local_width": false,
      "g128_runtime_effective_group_size": 32,
      "native_g32_checkpoint_scale_and_zero_point_metadata_gib_approx": 8.79,
      "g128_checkpoint_scale_and_zero_point_metadata_gib_approx": 2.20,
      "native_g32_checkpoint_metadata_increase_gib_approx": 6.59,
      "tp4_runtime_metadata_increase_vs_expanded_g128": false,
      "decision": "Store native g32 so checkpoint groups align exactly with the TP4 expert partition and do not depend on load-time g128-to-g32 expansion.",
      "quality_claim_boundary": "Native g32 was quality-validated; no matched full-checkpoint g32-versus-g128 quality comparison was performed."
    }
  },
  "ple_splice": {
    "dtype": "float8_e4m3fn",
    "indexed_tensor_count": 129,
    "indexed_payload_bytes": 51200245762,
    "shard_count": 33,
    "scale_value": 0.00019931793212890625,
    "requantized": false
  },
  "checkpoint": {
    "index_filename": "model.safetensors.index.json",
    "base_awq_index_sha256": "da62ba7bfef57f8117f35fc78bf5daf14c0ab5eb1b8173b4d712c1f4d294819b",
    "index_sha256": "f066a0154a9101b359c3a4d4fa6a83fb12b7126a9cf0f19cb772611d08cc07ee",
    "indexed_tensor_count": 222771,
    "indexed_payload_bytes": 137042968666,
    "index_metadata_total_size": 137073585730,
    "index_metadata_total_size_contract": "AWQ shard file bytes plus indexed PLE tensor bytes plus complete QSA scale shard bytes",
    "physical_weight_shard_bytes": 138133206482,
    "physical_model_shard_bytes": 138133209122,
    "stored_tensor_count_in_all_selected_shards": 223827,
    "unindexed_tensor_count_in_reused_ple_shards": 1056,
    "awq_shard_count": 10,
    "ple_shard_count": 33,
    "qsa_kv_scale_shard_count": 1,
    "qsa_kv_scale_filename": "model-kvscales.safetensors",
    "qsa_kv_scale_tensor_count": 24,
    "qsa_kv_scale_file_sha256": "c89b08bacde7d869342190b2f95046a83af37080241ae00ccae3a3d6888852fd",
    "carrier_completion_sha256": "70fb852c73620edbf75858ab36c519005a5b4cec6597d2493c1355b4b901f077",
    "host_side_final_audit_sha256": "b854f56010a564d7db9636c9977154f10c4ac2280f852899802ad7145cf85475",
    "host_side_final_audit_status": "PASS"
  },
  "runtime_provenance": {
    "fp16_kv_quality_source_commit": "c9af1be6e11011e398e8f89808e62d43fa453031",
    "safetensors_index_assignment_commit": "1f8c110fb044a7ac3767e3b790d4efc936797dc9",
    "awq_tp_group_alignment_commit": "ad4bfbec41b4ffb20877e2895f07957685d6438b",
    "awq_tp_group_alignment_effect_for_g32_tp4": "inactive: effective group g32 and repeat factor 1",
    "qsa_e4m3_core_pr": "https://github.com/1CatAI/1Cat-vLLM/pull/447",
    "qsa_e4m3_core_commit": "58542ea69d43c6fc66242d218fbddbfdae7cd3b9",
    "qsa_e4m3_bitcast_pr": "https://github.com/1CatAI/1Cat-vLLM/pull/452",
    "qsa_e4m3_bitcast_commit": "6e42fb899767b4ac0769acaff9156314fb193d3a",
    "qsa_e4m3_scale_hoist_pr": "https://github.com/1CatAI/1Cat-vLLM/pull/453",
    "qsa_e4m3_scale_hoist_commit": "10a7bd0b36ff4684432b6e235013b7ad817a3534",
    "combined_performance_test_commit": "526570c756e456a867aab86ea3ffd345195a082b",
    "qwen4exp_multimodal_enablement_pr": "https://github.com/1CatAI/1Cat-vLLM/pull/458",
    "qwen4exp_multimodal_enablement_commit": "dd68b9e8f44d248ad7082b871243a66c5fc1696a"
  },
  "publication": {
    "internal_build_files_excluded": [
      "final-checkpoint-audit.json",
      "fp8-ple-splice.json",
      "quant_log.csv",
      "quantization-provenance.json"
    ],
    "public_replacements": [
      "MODEL_PROVENANCE.json",
      "EXPERT_COVERAGE.json",
      "VALIDATION.json",
      "PUBLIC_RELEASE_AUDIT.json",
      "SHA256SUMS"
    ],
    "config_sanitization": {
      "gpu_profile": "single-rtx-3090",
      "offload_to_disk_path_removed": true,
      "runtime_quantization_fields_changed": false
    },
    "weights_repacked": false,
    "model_index_changed": true,
    "model_index_change": "Added 24 checkpoint-bound QSA K/V scale assignments without changing AWQ or PLE weight payloads."
  }
}