File size: 5,956 Bytes
dd4b081 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 | {
"audit_kind": "static source review; no model or workflow executed by fixture authoring agent",
"effects": {
"dynamic_eval_exec_in_runtime_source": [],
"execution_time": [
"Initializers use torch random normal and zero operations.",
"Forward uses ordinary CPU-capable PyTorch operations; CPU-only proof is produced only by release execution.",
"HF/QFS dynamic-code loader may fetch and verify explicitly selected immutable Hub source; this is loader activity, not runtime-source network behavior."
],
"filesystem_calls_in_runtime_source": [],
"import_time": [
"Imports installed torch/transformers modules; transitive dependency import effects are not independently sandboxed.",
"Obtains module logger through transformers.utils.logging.get_logger.",
"Defines classes/functions and evaluates type annotations and can_return_tuple decorator.",
"Appends Spark2_5RMSNorm to transformers.pytorch_utils.ALL_LAYERNORM_LAYERS; global normalization registry mutation retained from upstream."
],
"network_calls_in_runtime_source": [],
"scope_exclusion": "Build/verification/publication scripts live in the evidence repository and are separately invoked tools, not auto_map imports. The model repo has only the two runtime .py files; original upstream .py bytes are preserved with an inert .py.txt suffix. release.py performs intended subprocesses and Hub mutations only when Main/user executes it.",
"subprocess_calls_in_runtime_source": []
},
"imports": {
"dependencies": [
"torch",
"transformers"
],
"optional_accelerator_imports": [],
"relative": [
"configuration_spark"
],
"standard_library": [
"math",
"typing (reviewed compatibility import)"
]
},
"known_upstream_semantics": [
"Rotary positions are taken from cache_position; arbitrary per-example position_ids are not newly supported by this fork. Fixture uses contiguous unpadded positions.",
"Eager attention only, as upstream; no FlashAttention/SDPA claim."
],
"license": {
"file_evidence": "LICENSE contains full Apache 2.0 and Copyright 2026 XHToken; configuration_spark.py carries Apache header; modeling_spark.py has no original per-file header but is covered by repository LICENSE",
"metadata": "README.md frontmatter license apache-2.0",
"retention": [
"LICENSE",
"NOTICE",
"upstream/LICENSE",
"upstream/configuration_spark.py.txt in model; upstream/configuration_spark.py in evidence",
"upstream/modeling_spark.py.txt in model; upstream/modeling_spark.py in evidence"
],
"spdx": "Apache-2.0",
"upstream_notice": "No NOTICE file in pinned repository sibling inventory."
},
"patches": [
{
"change": "Import Unpack from Python3.12 typing instead of removed transformers.processing_utils re-export; annotations only, no model math changes.",
"file": "modeling_spark.py",
"installed_api": "transformers/processing_utils.py imports typing_extensions but no longer exports Unpack",
"mathematical_change": false
},
{
"change": "Add pinned source attribution and Apache notice.",
"file": "modeling_spark.py",
"mathematical_change": false
},
{
"change": "Mask helper input_embeds renamed inputs_embeds; obsolete cache_position keyword removed. Transformers5.16 infers mask offsets from Cache and position_ids.",
"file": "modeling_spark.py",
"installed_api": "transformers/masking_utils.py create_causal_mask and create_sliding_window_causal_mask signatures",
"mathematical_change": false
},
{
"change": "Replace Transformers4 tied-key list with Transformers5 target/source mapping lm_head.weight -> model.embedding.weight.",
"file": "modeling_spark.py",
"installed_api": "transformers/modeling_utils.py get_expanded_tied_weights_keys",
"mathematical_change": false
},
{
"change": "Use self.lm_head(hidden_states) for both tied and untied configurations; declared tying aliases the original embedding parameter, retaining F.linear operands without duplicate invented head weights. Enables actual output-module pre-hook.",
"file": "modeling_spark.py",
"mathematical_change": false,
"proof": "verify_native.py checks pointer alias and actual hooked-head logits bit-exact equality to upstream F.linear."
}
],
"preserved_equations": [
"GQA eager attention with FP32 softmax",
"three sliding layers then one full layer",
"per-attention-type partial RoPE and theta",
"headwise sigmoid output gating",
"GELU(gate_proj(x))*up_proj(x) MLP",
"RMSNorm and FP32 residual stream",
"legitimately tied output vocabulary head"
],
"runtime_files": {
"configuration_spark.py": "02218597240490f490659b184052db940b08163ff9c3af7b7e59323f6f963722",
"modeling_spark.py": "ba15bd14cca26322c9461ee2cf129423b59a5122826449947f7bb1acf5d9f2a6"
},
"schema": "malaiwah.spark2-5-upstream-code-audit.v1",
"upstream": {
"repository": "XHToken/Spark-X2.5-4B",
"revision": "5e10fcc0286756aebf7c41dc52c1e42d95c70281"
},
"upstream_files": {
"LICENSE": {
"bytes": 11338,
"sha256": "525c465ab952779d2d360a9ff36f7f49ae018a23d80302ff839a850a77784c36"
},
"README.md": {
"bytes": 24346,
"sha256": "7aac2c2e00d2c8ea608254c9085152c853282d93dad0312c62d7ba328bc6a5be"
},
"config.json": {
"bytes": 2035,
"sha256": "767161ece8ed2891e44345afdf98d29c1cacb964116cb6db7d20dc1c757490ad"
},
"configuration_spark.py": {
"bytes": 4329,
"sha256": "02218597240490f490659b184052db940b08163ff9c3af7b7e59323f6f963722"
},
"modeling_spark.py": {
"bytes": 19418,
"sha256": "9cf0d1ad2b54b9f7088792779ddd8b5cb4d4fe63bfea054da7f2dd16362bcaf4"
}
},
"verification_entrypoint": "verify_native.py; results only exist after stage execution"
}
|