QUANT_PROVENANCE.json

{
  "base_model": "microsoft/harrier-oss-v1-0.6b",
  "base_revision": "f9b9dc8d367d443f2479d27aa5d8d2850c0774ee",
  "variant": "fp8",
  "toolchain": "llm-compressor==0.13.0 (compressed-tensors==0.18.0), torch==2.11.0+cu130, vLLM==0.26.0 for serving/eval",
  "scheme": "FP8_DYNAMIC (compressed-tensors preset): static per-channel FP8 E4M3 weight scales computed directly from the weights, dynamic per-token FP8 E4M3 activation quantization at inference. No calibration forward pass is strictly required for this scheme, but one was run over the pinned calibration manifest (calibration/manifest.json, 448 samples) for parity with the NVFP4 export path.",
  "modules_quantized": [
    "196 nn.Linear modules across all 28 Qwen3Model decoder layers: self_attn.{q,k,v,o}_proj and mlp.{gate,up,down}_proj (7 Linear modules x 28 layers)"
  ],
  "modules_full_precision": [
    "model.embed_tokens (embedding table, tied to lm_head)",
    "all RMSNorm modules (input_layernorm, post_attention_layernorm, model.norm, q_norm, k_norm)",
    "lm_head (tied to embed_tokens; unused for embedding output)",
    "the sentence-transformers pooling (last-token) and normalize (L2) heads — not part of the transformers module tree, applied post-hoc, always full precision"
  ],
  "contract_test": {
    "module_inventory_file": "quantization/fp8_module_inventory.json",
    "total_modules": 427,
    "quantized_modules": 196,
    "never_quantize_substrings": ["embed_tokens", "lm_head", "norm"],
    "violations": [],
    "verdict": "PASS"
  },
  "ignore_list": ["lm_head", "re:.*embed_tokens.*", "re:.*norm.*"],
  "quantize_script": "quantization/quantize_fp8.py"
}