{
  "model": "llama3_2_1b_instruct_q4gsw",
  "upstream": {
    "model": "meta-llama/Llama-3.2-1B-Instruct",
    "revision": "9213176726f574b556790deb65791e0c5aa438b6",
    "official": true,
    "license": "Llama 3.2 Community License (tokenizer/LICENSE.txt, tokenizer/USE_POLICY.md); Built with Llama",
    "files": {
      "original/consolidated.00.pth": {
        "lfs": "fc17d497df5e4175b3a8acb4f5865b26f7fc1b009b25bef814b95fde10e8a1f3",
        "sha256": "fc17d497df5e4175b3a8acb4f5865b26f7fc1b009b25bef814b95fde10e8a1f3"
      },
      "original/params.json": {
        "blob": "9cd8dbdf2dc6f4d8abb60bdb5ce64f4bec2fdfd9",
        "git_blob_sha1": "9cd8dbdf2dc6f4d8abb60bdb5ce64f4bec2fdfd9"
      },
      "tokenizer.json": {
        "blob": "5cc5f00a5b203e90a27a3bd60d1ec393b07971e8",
        "git_blob_sha1": "5cc5f00a5b203e90a27a3bd60d1ec393b07971e8"
      },
      "tokenizer_config.json": {
        "blob": "4ff488a165e900e5129cda7c20ab32d568d2a475",
        "git_blob_sha1": "4ff488a165e900e5129cda7c20ab32d568d2a475"
      },
      "special_tokens_map.json": {
        "blob": "02ee80b6196926a5ad790a004d9efd6ab1ba6542",
        "git_blob_sha1": "02ee80b6196926a5ad790a004d9efd6ab1ba6542"
      },
      "config.json": {
        "blob": "3e3aaf51a035cb5092d9f6827a0dc074657ba88c",
        "git_blob_sha1": "3e3aaf51a035cb5092d9f6827a0dc074657ba88c"
      },
      "generation_config.json": {
        "blob": "75ae08310d6d23df373ee2644b497192b3cce6d8",
        "git_blob_sha1": "75ae08310d6d23df373ee2644b497192b3cce6d8"
      },
      "LICENSE.txt": {
        "blob": "085b47c1575cb889b7024030e60b78f54f0b8c9e",
        "git_blob_sha1": "085b47c1575cb889b7024030e60b78f54f0b8c9e"
      },
      "USE_POLICY.md": {
        "blob": "ac3c5f21b9779e3da0677d6d3c587778fe3a331e",
        "git_blob_sha1": "ac3c5f21b9779e3da0677d6d3c587778fe3a331e"
      }
    },
    "hub_revision_checked": "9213176726f574b556790deb65791e0c5aa438b6",
    "downloaded_from": {
      "repo": "LIF1014/ptdbench-Llama-3.2-1B-Instruct",
      "revision": "828e5f37c833a02415416260d5a93780fd41d10e",
      "note": "all files verified byte-identical to the official hashes above"
    }
  },
  "versions": {
    "python": "3.12.15+meta",
    "torch": "2.14.1+cpu",
    "torchao": "0.18.0+cpu",
    "executorch": "1.5.1+cpu",
    "executorch_git": "3b60683923245cf472b7323426920e15623ba361",
    "transformers": "5.19.0",
    "huggingface_hub": "2.2.0"
  },
  "export": {
    "api": "python -m executorch.extension.llm.export.export_llm (hydra overrides below)",
    "overrides": [
      "base.model_class=llama3_2",
      "base.checkpoint=<out>/_work/checkpoint.pth",
      "base.params=<out>/_work/params.json",
      "model.use_kv_cache=true",
      "model.use_sdpa_with_kv_cache=true",
      "model.enable_dynamic_shape=true",
      "model.dtype_override=fp32",
      "export.max_seq_length=513",
      "export.max_context_length=8960",
      "export.output_dir=<out>",
      "export.output_name=<out>/llama3_2_1b_instruct_q4gsw.pte",
      "quantization.qmode=4w",
      "quantization.group_size=64",
      "quantization.embedding_quantize='4,64'",
      "backend.vulkan.enabled=true"
    ],
    "checkpoint": "original/consolidated.00.pth (Meta format, bf16; no conversion), params = original/params.json",
    "params": {
      "dim": 2048,
      "n_layers": 16,
      "n_heads": 32,
      "n_kv_heads": 8,
      "vocab_size": 128256,
      "ffn_dim_multiplier": 1.5,
      "multiple_of": 256,
      "norm_eps": 1e-05,
      "rope_theta": 500000.0,
      "use_scaled_rope": true
    },
    "rope": "llama3 scaled RoPE, rope_scale_factor=32 (set by ExecuTorch for model_class llama3_2 when use_scaled_rope)",
    "partitioner": "VulkanPartitioner via backend.vulkan.enabled (WebGPUPartitioner is a pass-through wrapper of it; delegate id VulkanBackend)",
    "inputs": [
      {
        "name": "tokens",
        "shape": [
          1,
          "1..512"
        ],
        "dtype": "int64"
      },
      {
        "name": "input_pos",
        "shape": [
          1
        ],
        "dtype": "int64"
      }
    ],
    "outputs": [
      {
        "name": "logits",
        "shape": [
          1,
          128256
        ],
        "dtype": "float32",
        "note": "last position only"
      }
    ],
    "prefill_max": 512,
    "max_seq_len_metadata": 513,
    "max_context_length": 8960,
    "dtype": "float32 activations",
    "quantization": "4-bit groupwise weight-only linears (torchao IntxWeightOnlyConfig int4, group 64) and 4-bit groupwise embedding (group 64); output head quantized too",
    "kv_cache": "fp32 buffers in the program; fp16 KV is a runtime option of the WebGPU runtime (enable_f16_kv_cache / et_set_kv_f16), not an export flag"
  },
  "files": {
    "llama3_2_1b_instruct_q4gsw.pte": {
      "bytes": 734868736,
      "sha256": "49de6eec3a54f1f45a785152375edafec07cb0fd879a142100340fd2765c40a7"
    },
    "tokenizer/tokenizer.json": {
      "sha256": "79e3e522635f3171300913bb421464a87de6222182a0570b9b2ccba2a964b2b4"
    },
    "tokenizer/tokenizer_config.json": {
      "sha256": "9823dcfdc1121869029da45192238e85cf44f0b232a6d9dc20e4fe6f4242a14e"
    },
    "tokenizer/special_tokens_map.json": {
      "sha256": "6f38c73729248f6c127296386e3cdde96e254636cc58b4169d3fd32328d9a8ec"
    },
    "tokenizer/config.json": {
      "sha256": "2febf68cea25bf4611be02b7536f2488a5ba523bb1134986e3610152abe74fdb"
    },
    "tokenizer/generation_config.json": {
      "sha256": "88effbb63300dbbc7390143fbbdd9d9fa50587b37e8bfd16c8c90d4970a74a36"
    },
    "tokenizer/LICENSE.txt": {
      "sha256": "0b4284c1f87029e67654c7953afa16279961632cf73dcfe33374c4c2f298fa35"
    },
    "tokenizer/USE_POLICY.md": {
      "sha256": "40e2777d7faa6beaf98400654170f414d8ab29b921b5163ad4ea0a1d39894201"
    }
  },
  "graph": {
    "non_delegated_kernels": {},
    "delegates": [
      {
        "id": "VulkanBackend",
        "ops": {
          "et_vk.prepack.default": 2,
          "sym_size.int": 1,
          "et_vk.embedding_q4gsw.default": 1,
          "et_vk.select_as_symint.default": 2,
          "add": 1,
          "aten.slice_copy.Tensor": 2,
          "et_vk.rms_norm.default": 33,
          "et_vk.linear_q4gsw.default": 113,
          "aten.view_copy.default": 128,
          "et_vk.apply_rotary_emb.default": 16,
          "sdpa_with_kv_cache.default": 16,
          "aten.add.Tensor": 32,
          "aten.sigmoid.default": 16,
          "aten.mul.Tensor": 32,
          "aten.select_copy.int": 1
        },
        "inputs": 2,
        "outputs": 1
      }
    ]
  },
  "command": "python recipes/llama-3.2-1b-instruct.py --out out/llama-3.2-1b-instruct"
}