malaiwah/qwen4-exp-tiny-cpu-repro-v1

Qwen4-Exp tiny CPU reproduction receipts

This is an artifact/receipt bundle, not training data and not a single root-format QFS dataset. All four readable synthetic documents are embedded in panel/panel.receipt.json.

Provenance and limitations

This is an independently generated, untrained random checkpoint inspired by Qwen/Qwen3.8-Flash-Next@de4b8e4d43b917e7706784d8bb445c9af86a3540, not a quantization, distillation, behavioral replica, or fine-tune. No source model weights, training dataset, remote Python code, paid GPU, or cloud compute were used. All parameters were constructed by the installed native Transformers class. Construction is FP32 then rounded to the complete BF16 artifact; native integer hash buffers are preserved.

The upstream weights, configuration and card carry Qwen Community License 1.0, including attribution thresholds and commercial Model-as-a-Service/AI Work Assistant terms. Those upstream weights and configuration files are not redistributed here. Our original fixture scripts, tokenizer, synthetic texts and generated random weights are offered under the accompanying MIT license; this is not a relicensing of Qwen's upstream release. Installed Transformers implementation files carry Apache-2.0 headers. Architecture inspiration is not a claim of affiliation or endorsement.

Exact tiny architecture

• Native Qwen4ExpForConditionalGeneration wrapper (qwen4_exp) with qwen4_exp_text, not Qwen3.5.
• Four 64-wide decoder layers: DeltaNet, DeltaNet + PLE, DeltaNet, QSA (zero-based QSA layer 3; one-based PLE layer 2).
• Linear attention: 2 key heads, 6 value heads, dimension 16, causal convolution kernel 4; sigmoid output gating.
• QSA: 4 query/1 KV attention heads of dimension 16; 2-query/1-key indexer heads of dimension 8; compressed microblocks of 4, token budget 8.
• Four gated residual streams (hc_count=4), low-rank mixer bottleneck 8.
• Four routed experts, top-2 plus a real shared expert in every layer; each intermediate size 32.
• Trainable bigram/trigram PLE: two hash heads per order, prime-vocabulary base 67, padded to divisor 8, embedding width 32, two native save shards. No 20-million-entry default survives.
• PLE kernel 4/dilation 3; unlike its native zero initialization, the complete learned convolution receives deterministic normal random initialization to exercise its contribution. 'Learned' means trainable parameterization, not trained values.
• Real vision transformer: depth 1, width 32, intermediate 64, 4 heads, 16 positions, 2×2 spatial/2 temporal patches, merge size 2, output width 64. Image/video forwards are not part of these text-only checks.
• Complete untied 264×64 output head; complete 256-byte tokenizer plus 8 in-range text/modality special IDs. Context limit 256.
• MTP is not implemented by this native class and is absent; no placeholder or unmatched MTP weights.

Measured CPU evidence

Generated from successful stage outputs, not handwritten expected values. Receipt bundle contains independent QFS roots reproduction/first/ and reproduction/repeat/ plus reproduction/comparison/. The repository root is a multi-artifact evidence bundle, not a single QFS root dataset. Both fresh processes use two CPU threads and BF16 with lane other explicitly denoting CPU. Both undergo strict tensor verification; forced numerical replay checks every tokenwise value is positive float64 zero. The native cached single/multi-token and microblock/chunk boundary probes, causal-prefix invariance, learned PLE context sensitivity, limited-history and EOS-reset invariance are separate from the four 64-token synthetic panel. The panel is not a quality benchmark.

Immutable model/tokenizer pin used by the panel: malaiwah/qwen4-exp-tiny-random-bf16@ae3f6ab374b103aabc71418aefebf44579d0685e. Final documentation commits retain those exact model/tokenizer bytes.

{
  "captures": {
    "first": {
      "capture_content_digest": "17ccbb9a64d4bd470148b6f358f4cc77175cc0acda1773a4c629ded73a5a807d",
      "checkpoint_identity_sha256": "3ae4d9c9f4142513f77509f9e9b98487d7ac4711838974dca7978a0caba3248f",
      "cold_run": "cpu-first-358fe6f200e341c0958cd2c1bd7517cb",
      "dataset_sha256": "e323852a4e7503592abe035cdd50f70193728610e2e0398b365b3234c0af8ec7",
      "generation_sanity_probe": {
        "degenerate": false,
        "enforced": false,
        "entropy_nats": 5.573547920546279,
        "expect": null,
        "logit_spread": 0.3623046875,
        "note": "one extra window through the same schedule; its hidden state is discarded and never enters the dataset.",
        "prompt": "The capital of France is",
        "schema": "malaiwah.generation-sanity-probe.v1",
        "status": "recorded",
        "top1_id": 36,
        "top1_probability": 0.004473608316620615,
        "top1_text": "=",
        "top_k": [
          {
            "id": 36,
            "probability": 0.004473608316620615,
            "text": "="
          },
          {
            "id": 201,
            "probability": 0.004434461776218208,
            "text": "\u0005"
          },
          {
            "id": 0,
            "probability": 0.004425809170622273,
            "text": "<pad>"
          },
          {
            "id": 130,
            "probability": 0.004404251448270759,
            "text": "\ufffd"
          },
          {
            "id": 240,
            "probability": 0.004399952520893088,
            "text": "\ufffd"
          }
        ],
        "uniform_entropy_nats": 5.575949103146316,
        "vocab_size": 264
      },
      "head_tensor_content_sha256": "f0e1210bca4f610f646e8224e5950f9f2e154ab274d002b08ade291892a969e4",
      "memory_report": {
        "context_length": 64,
        "decoder_layers": null,
        "device": "cpu",
        "hidden_size": 64,
        "layer_residency": null,
        "model": "malaiwah/qwen4-exp-tiny-random-bf16",
        "peak_resident_weight_bytes": 684552,
        "peak_resident_weight_detail": {
          "buffers": 128,
          "parameters": 684424,
          "total": 684552
        },
        "peak_resident_weight_gb": 0.001,
        "peak_rss_bytes": 371265536,
        "peak_rss_gb": 0.371,
        "resident_weight_note": "the maximum, over the run, of the materialised parameter+buffer bytes. On the CPU path peak_rss_bytes ALSO counts the safetensors mmap page cache, which is evictable and is not held by the schedule; this figure is not confounded by it. On CUDA, peak_cuda_allocated_bytes is the authoritative number because it includes activations and workspace and has no page cache.",
        "rss_units_source": "Linux ru_maxrss",
        "schedule": "window-outer",
        "vocab_size": 264,
        "windows": 4
      },
      "path": "first",
      "pid": 712649,
      "resources": {
        "bytes": {
          "checkpoint_read": 708634,
          "hidden_d2h": 36624,
          "weights_h2d": null
        },
        "checkpoint_bytes": 708634,
        "checkpoint_files": 2,
        "device_name": null,
        "forward_timing": "wall-clock",
        "note": "provenance, not identity: peaks and stage seconds this run measured; forward_sum is device time under cuda-events, wall time otherwise; seal covers head serialisation up to this block's assembly",
        "peak_cuda_allocated_bytes": null,
        "peak_cuda_reserved_bytes": null,
        "peak_resident_weight_bytes": 684552,
        "peak_rss_bytes": 371265536,
        "rss_units_source": "Linux ru_maxrss",
        "seconds": {
          "decode_sum": null,
          "elapsed": 3.15,
          "fill_sum": null,
          "forward_sum": 0.539,
          "identity": 0.008,
          "layer_load_max": null,
          "layer_load_sum": null,
          "layer_loads": null,
          "resident_load": 2.554,
          "seal": 0.006
        }
      },
      "runtime_versions": {
        "cuda_runtime": null,
        "executable": "/tmp/qfs-tiny-fixture-venv/bin/python",
        "huggingface_hub": "1.30.0",
        "machine": "x86_64",
        "numpy": "2.5.3",
        "pid": 712649,
        "platform": "Linux-7.0.0-30-generic-x86_64-with-glibc2.43",
        "python": "3.12.14 (main, Aug 14 2026, 15:34:45) [Clang 22.1.3 ]",
        "safetensors": "0.8.0",
        "stage": "fixture_child_runtime",
        "thread_environment": {
          "MKL_NUM_THREADS": "2",
          "NUMEXPR_NUM_THREADS": "2",
          "OMP_NUM_THREADS": "2",
          "OPENBLAS_NUM_THREADS": "2",
          "VECLIB_MAXIMUM_THREADS": "2"
        },
        "tokenizers": "0.23.2",
        "torch": "2.11.0+cpu",
        "torch_num_interop_threads": 2,
        "torch_num_threads": 2,
        "transformers": "5.16.1"
      },
      "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd"
    },
    "repeat": {
      "capture_content_digest": "17ccbb9a64d4bd470148b6f358f4cc77175cc0acda1773a4c629ded73a5a807d",
      "checkpoint_identity_sha256": "3ae4d9c9f4142513f77509f9e9b98487d7ac4711838974dca7978a0caba3248f",
      "cold_run": "cpu-repeat-c89d9f9065ed4e9daa0be2384815f6bc",
      "dataset_sha256": "8d0847eb0f046bc28b48b09b352c7474b5f2abc56f5b35671bccb0b0173febc3",
      "generation_sanity_probe": {
        "degenerate": false,
        "enforced": false,
        "entropy_nats": 5.573547920546279,
        "expect": null,
        "logit_spread": 0.3623046875,
        "note": "one extra window through the same schedule; its hidden state is discarded and never enters the dataset.",
        "prompt": "The capital of France is",
        "schema": "malaiwah.generation-sanity-probe.v1",
        "status": "recorded",
        "top1_id": 36,
        "top1_probability": 0.004473608316620615,
        "top1_text": "=",
        "top_k": [
          {
            "id": 36,
            "probability": 0.004473608316620615,
            "text": "="
          },
          {
            "id": 201,
            "probability": 0.004434461776218208,
            "text": "\u0005"
          },
          {
            "id": 0,
            "probability": 0.004425809170622273,
            "text": "<pad>"
          },
          {
            "id": 130,
            "probability": 0.004404251448270759,
            "text": "\ufffd"
          },
          {
            "id": 240,
            "probability": 0.004399952520893088,
            "text": "\ufffd"
          }
        ],
        "uniform_entropy_nats": 5.575949103146316,
        "vocab_size": 264
      },
      "head_tensor_content_sha256": "f0e1210bca4f610f646e8224e5950f9f2e154ab274d002b08ade291892a969e4",
      "memory_report": {
        "context_length": 64,
        "decoder_layers": null,
        "device": "cpu",
        "hidden_size": 64,
        "layer_residency": null,
        "model": "malaiwah/qwen4-exp-tiny-random-bf16",
        "peak_resident_weight_bytes": 684552,
        "peak_resident_weight_detail": {
          "buffers": 128,
          "parameters": 684424,
          "total": 684552
        },
        "peak_resident_weight_gb": 0.001,
        "peak_rss_bytes": 371552256,
        "peak_rss_gb": 0.372,
        "resident_weight_note": "the maximum, over the run, of the materialised parameter+buffer bytes. On the CPU path peak_rss_bytes ALSO counts the safetensors mmap page cache, which is evictable and is not held by the schedule; this figure is not confounded by it. On CUDA, peak_cuda_allocated_bytes is the authoritative number because it includes activations and workspace and has no page cache.",
        "rss_units_source": "Linux ru_maxrss",
        "schedule": "window-outer",
        "vocab_size": 264,
        "windows": 4
      },
      "path": "repeat",
      "pid": 713922,
      "resources": {
        "bytes": {
          "checkpoint_read": 708634,
          "hidden_d2h": 36624,
          "weights_h2d": null
        },
        "checkpoint_bytes": 708634,
        "checkpoint_files": 2,
        "device_name": null,
        "forward_timing": "wall-clock",
        "note": "provenance, not identity: peaks and stage seconds this run measured; forward_sum is device time under cuda-events, wall time otherwise; seal covers head serialisation up to this block's assembly",
        "peak_cuda_allocated_bytes": null,
        "peak_cuda_reserved_bytes": null,
        "peak_resident_weight_bytes": 684552,
        "peak_rss_bytes": 371552256,
        "rss_units_source": "Linux ru_maxrss",
        "seconds": {
          "decode_sum": null,
          "elapsed": 2.982,
          "fill_sum": null,
          "forward_sum": 0.543,
          "identity": 0.005,
          "layer_load_max": null,
          "layer_load_sum": null,
          "layer_loads": null,
          "resident_load": 2.385,
          "seal": 0.007
        }
      },
      "runtime_versions": {
        "cuda_runtime": null,
        "executable": "/tmp/qfs-tiny-fixture-venv/bin/python",
        "huggingface_hub": "1.30.0",
        "machine": "x86_64",
        "numpy": "2.5.3",
        "pid": 713922,
        "platform": "Linux-7.0.0-30-generic-x86_64-with-glibc2.43",
        "python": "3.12.14 (main, Aug 14 2026, 15:34:45) [Clang 22.1.3 ]",
        "safetensors": "0.8.0",
        "stage": "fixture_child_runtime",
        "thread_environment": {
          "MKL_NUM_THREADS": "2",
          "NUMEXPR_NUM_THREADS": "2",
          "OMP_NUM_THREADS": "2",
          "OPENBLAS_NUM_THREADS": "2",
          "VECLIB_MAXIMUM_THREADS": "2"
        },
        "tokenizers": "0.23.2",
        "torch": "2.11.0+cpu",
        "torch_num_interop_threads": 2,
        "torch_num_threads": 2,
        "transformers": "5.16.1"
      },
      "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd"
    }
  },
  "checkpoint_tensors": 173,
  "comparison": {
    "comparator": {
      "accumulation_dtype": "float64",
      "bf16_reduced_precision_reduction": false,
      "cublas_workspace_config": null,
      "deterministic_algorithms": true,
      "device": "cpu",
      "estimator_backend": "torch:k6_kld_report._token_kld",
      "head_applied_candidate_tensor_content_sha256": "f0e1210bca4f610f646e8224e5950f9f2e154ab274d002b08ade291892a969e4",
      "head_applied_reference_tensor_content_sha256": "f0e1210bca4f610f646e8224e5950f9f2e154ab274d002b08ade291892a969e4",
      "head_applied_tensor_content_sha256": "f0e1210bca4f610f646e8224e5950f9f2e154ab274d002b08ade291892a969e4",
      "logprob_dtype": "float64",
      "position_block": 128,
      "replay_backend": "numpy:cpu:float32",
      "replay_env": {
        "blas_configuration": "OpenBLAS 0.3.34.106.0  USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell MAX_THREADS=64",
        "blas_name": "scipy-openblas",
        "blas_threads": 2,
        "blas_threads_source": "ctypes:scipy_openblas_get_num_threads64_",
        "blas_version": "0.3.34.106.0",
        "cpu_model": "Intel(R) Xeon(R) CPU           X5570  @ 2.93GHz",
        "library": "numpy",
        "numpy_version": "2.5.3",
        "replay_dtype": "float32"
      },
      "replay_peak_device_bytes": null,
      "short_circuited": true,
      "source_file_hashes_verified": true,
      "tensor_content_digests_verified": true,
      "tf32": false,
      "tool": {
        "entrypoint": "bin/fidelity-dataset compare",
        "note": "the fp64 estimator is k6_kld_report._token_kld, imported not copied, whenever torch is importable; estimator_backend records which path ran."
      },
      "two_pass": true,
      "vocab_chunk": 8192
    },
    "comparison_kind": "reproduction_confirmation",
    "kl": {
      "max": 0.0,
      "mean": 0.0,
      "median": 0.0,
      "p95": 0.0,
      "p99": 0.0,
      "p99_9": 0.0
    },
    "measurement_scope": {
      "contexts": 4,
      "covers_full_panel": true,
      "position_filter": "all",
      "positions_per_context": 63,
      "scored_positions": 252,
      "subset_detail": null
    },
    "metric": {
      "direction": "reference_to_candidate",
      "direction_label": "KL(reference || candidate)",
      "higher_is_better": false,
      "name": "mean_tokenwise_kld",
      "units": "nats",
      "value": 0.0
    },
    "note": "QFS retains its hash-proof result and top-1 value for reproduction confirmations; --force-compute independently runs the estimator and requires its tokenwise KL array to agree exactly. short_circuited=true therefore does not mean replay was skipped.",
    "receipt": "comparison/comparison-receipt.json",
    "receipt_sha256": "466be39f2fcbc48c3757e9ec600950960e69482fedf2ffa7079b5539dd176119",
    "self_compare": {
      "asserted_exact_zero": true,
      "capture_content_digest_equal": true,
      "expected_tokenwise_bytes": 2144,
      "expected_tokenwise_sha256": "896ad5b6b732dce1588636c3af3a8a93af8eb1250df72809f6e26fc29914f65d",
      "force_compute_agreed": true,
      "head_digest_equal": true,
      "weights_identity_equal": true
    },
    "tokenwise": {
      "bytes": 2144,
      "path": "tokenwise-kld.npy",
      "sha256": "896ad5b6b732dce1588636c3af3a8a93af8eb1250df72809f6e26fc29914f65d"
    },
    "top1_agreement": 1.0
  },
  "native_state_roundtrip_equal": true,
  "ngram": {
    "embedding_dimension_per_head": 8,
    "head_vocab_sizes": [
      67,
      71,
      73,
      79
    ],
    "padded_vocab_size": 296,
    "split_ngram_parts": 2
  },
  "parameter_count": 342212,
  "probes": {
    "boundaries": [
      {
        "tokens": 1,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 2,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 3,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 4,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 5,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 7,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 8,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 9,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 15,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 16,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 17,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 63,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 64,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      },
      {
        "tokens": 65,
        "variants": [
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "single-token"
          },
          {
            "all_finite": true,
            "max_abs_logit_error": 0.0,
            "mode": "multi-token"
          }
        ]
      }
    ],
    "cache_atol": 0.02,
    "cache_rtol": 0.02,
    "causal_prefix_max_abs_error": 0.0,
    "elapsed_seconds": 12.35180641617626,
    "ngram": {
      "cached_boundary_exact": true,
      "context_max_abs_feature_change": 0.0472412109375,
      "embedding_nonzero_elements": 2350,
      "eos_reset_invariance": true,
      "limited_history_invariance": true,
      "ple_context_max_abs_change": 0.055419921875,
      "ple_conv_nonzero_elements": 1019,
      "ple_output_max_abs": 0.04345703125
    },
    "peak_rss_kib": 360072,
    "schema": "malaiwah.qwen4-exp-native-probes.v1",
    "scope": "CPU BF16 text-only; real vision weights present but image/video forward not tested",
    "status": "passed",
    "torch": "2.11.0+cpu",
    "torch_num_interop_threads": 2,
    "torch_num_threads": 2
  }
}

Reproduce locally without publication

Use Python 3.12 with requirements-cpu.txt in an isolated environment, with no FLA or causal-conv1d installed. Download this fixture at the immutable revision, not upstream production weights. QFS source identity and actual file hashes are in qfs-source.json; use the recorded checkout (including the exact recorded source bytes).

python build_fixture.py --out /tmp/qwen4-generated
python probe_native.py --model-dir /tmp/qwen4-generated --out /tmp/qwen4-probes.json
python build_panel.py --model-dir /tmp/qwen4-generated --model-repository malaiwah/qwen4-exp-tiny-random-bf16 --model-revision ae3f6ab374b103aabc71418aefebf44579d0685e --out /tmp/qwen4-panel --qfs-root /path/to/quant-fidelity-suite
python verify_cpu.py --model-dir /tmp/qwen4-generated --model-repository malaiwah/qwen4-exp-tiny-random-bf16 --model-revision ae3f6ab374b103aabc71418aefebf44579d0685e --panel-dir /tmp/qwen4-panel --out /tmp/qwen4-reproduction --qfs-root /path/to/quant-fidelity-suite

Every output path must be new. release.py --stage-only --out NEW_DIRECTORY --qfs-root CHECKOUT is the maintainer release operation: it refuses existing repository IDs, generates twice, checks native save/reload and byte equality, runs probes, creates a private immutable Hub pin, builds the sealed panel, runs both strict captures plus replay, then assembles these cards/evidence. release.py --publish --out STAGED_DIRECTORY only accepts an unchanged passed stage, uploads allowlisted package files privately, changes visibility, and performs anonymous forced-download verification. No tokens, model caches, or environment directories are packaged.

Risks and non-claims

This experimental native architecture depends on the exact pinned Transformers release, including native ngram sharded conversion and multi-state cache support. Tiny geometry exercises implementation paths, not production scale or throughput. Cached BF16 equivalence uses declared atol=0.02, rtol=0.02; exact equality is required separately for the two same-shape QFS captures and the native ngram boundary features. Random generations are meaningless. Text-only execution does not validate vision, video, MTP, long-context quality, GPU kernels, training, or compatibility with other dependency versions.