malaiwah/spark2-5-tiny-cpu-repro-v1

Spark2.5 tiny random CPU fixture

Complete untrained Spark2_5ForCausalLM with independently seeded random BF16 weights. This is a reproducibility fixture, not a useful language model, distilled model, quality benchmark, or production registry measurement. No upstream weights, training data, paid GPU or cloud rentals were used.

Architecture, code and license

Source: XHToken/Spark-X2.5-4B at 5e10fcc0286756aebf7c41dc52c1e42d95c70281. The complete text causal model preserves three sliding layers followed by one full attention layer, headwise sigmoid output gates, GELU-gated MLPs and the genuinely tied vocabulary head. Hidden64, intermediate128, Q4/KV1, head16, sliding-window32, context256, rotary dimensions16 (sliding) /4 (full), theta10000/5000000. The real byte BPE tokenizer has256 byte tokens plus four special tokens, IDs0..259. There are no vision modules, MTP modules or invented output-head weights.

Complete text-only Spark2_5ForCausalLM; 3 sliding/1 full attention, partial RoPE dimensions 16/4, headwise sigmoid gates, GELU gated MLP, tied vocabulary head. Random BF16 only; no learned capabilities, million-token context, accelerator, training, quantized-format, or production-size evidence.

This is custom Python, not an installed native Transformers architecture. The reviewed runtime is an Apache-2.0 derivative; LICENSE, NOTICE, original files under upstream/ and upstream-code-audit.json retain pinned source identity. Patches migrate the mask API and tied-weight mapping to Transformers5.16.1 and route the exact same tied linear operation through the real output-head module. No attention, RoPE, gate or MLP equations are replaced. The upstream root format is BF16, not FP8/MXFP4; this tiny measurement proves neither full-size performance nor accelerator or quantized runtime support.

Actual measurements

• Parameters: 157,504; semantic tensors: 35; seed: 20260907.
• Two independent build subprocesses produce identical complete file inventories.
• Save/reload and QFS load check exact values and dtypes, with no missing-key overrides.
• Loader report: {"conversion_errors": {}, "conversion_errors_visible": true, "error_msgs": [], "mismatched_keys": [], "missing_keys": [], "observed": true, "unexpected_keys": []}.
• Actual head pre-hook is exercised; output equals upstream tied F.linear exactly.
• Actual sliding/full masks at queries31/32/33/64 and partial RoPE at0/31/32/127/255 are probed.
• Six cold/cached schedules cross sliding32, panel64 and context128/256 boundaries. Exact repetition and finite states are required; cached/cold BF16 differences are measured and bounded, not claimed zero. See native-verification.json.
• Two fresh CPU captures use two threads, BF16, lane other, sealed4x64 panel, all252 shifted positions. Strict verify/validate run for each capture.
• Forced --self-compare --force-compute replay: metric 0.0, KL {"max": 0.0, "mean": 0.0, "median": 0.0, "p95": 0.0, "p99": 0.0, "p99_9": 0.0}, top1 1.0. The actual tokenwise artifact must contain exactly252 positive float64 zeros.
• Captured weights and runtime code pin: malaiwah/spark2-5-tiny-random-bf16@dd4b081b0a6b042771cbfdaedf82ef345e59f5e9. QFS checkout HEAD d95b4e96d19e37c46e84e5fe54cc3e66c5316782; executed source hashes are recorded separately.

Capture weights are local; explicit immutable Hub code verification may use network. This is not an entirely offline workflow. Comparison requires neither GPU nor custom code. The evidence repository is a receipt bundle: sealed roots are first/ and repeat/, with comparison/, panel, strict reports, runtime/resource measurements and command logs.

Reproduce

Install Python3.12 and pip install -r requirements-cpu.txt. Review the custom source and audit before consenting. Download the workflow scripts and original upstream/ sources from the evidence repository; the model repository intentionally contains only the two executable runtime Python files. With QFS at $QFS:

python build_fixture.py --out /tmp/spark-rebuilt
python verify_native.py --model-dir /tmp/spark-rebuilt --model-repository malaiwah/spark2-5-tiny-random-bf16 --model-revision dd4b081b0a6b042771cbfdaedf82ef345e59f5e9 --qfs-root "$QFS" --out /tmp/spark-native.json
python build_panel.py --model-dir /tmp/spark-rebuilt --model-repository malaiwah/spark2-5-tiny-random-bf16 --model-revision dd4b081b0a6b042771cbfdaedf82ef345e59f5e9 --qfs-root "$QFS" --out /tmp/spark-panel
python verify_cpu.py --model-dir /tmp/spark-rebuilt --model-repository malaiwah/spark2-5-tiny-random-bf16 --model-revision dd4b081b0a6b042771cbfdaedf82ef345e59f5e9 --panel-dir /tmp/spark-panel --qfs-root "$QFS" --out /tmp/spark-repro

For original or your own same-architecture weights, keep checkpoint/config bytes unchanged, use your actual immutable weights pin and a tokenizer-bound sealed panel, and explicitly select this reviewed runtime fork. Set AUTHOR to your own attribution:

python "$QFS/engines/tools/hf_capture.py" --model "$MODEL_DIR" --weights-repository "$WEIGHTS_REPO" --model-revision "$WEIGHTS_SHA" --trust-remote-code --code-repository malaiwah/spark2-5-tiny-random-bf16 --code-revision dd4b081b0a6b042771cbfdaedf82ef345e59f5e9 --panel "$PANEL_DIR" --panel-role final --panel-tokenizer-root "$MODEL_DIR" --role root --lane other --device cpu --dtype bfloat16 --schedule window-outer --dataset-id "$DATASET_ID" --dataset-name "$DATASET_NAME" --repository "$EVIDENCE_REPO" --author "$AUTHOR" --out "$OUTPUT_DIR"

All SHAs must be full immutable40-hex commits. Original4B weights need substantially more RAM; no original4B CPU throughput or numerical result is asserted. support.json records exact argument-vector recipes and prerequisites. Publisher only: release.py --stage-only --out NEW_DIR --qfs-root "$QFS" requires write access to the fixed repository IDs and refuses collisions. It stages privately. release.py --publish --out NEW_DIR releases only a successful unchanged stage and verifies anonymous downloads; it cannot overwrite these published repositories.