Kimi K2.5 / K2.6 / K2.7-Code complete tiny random BF16 CPU fixture
0
2 commits
2 linked in READMEs
updated Sep 7, 2026
Untrained independently seeded random weights; no upstream weights or training data. This is a reproducibility fixture, not useful language modeling or production quality evidence.
Upstream moonshotai/Kimi-K2.7-Code@74797c9c62378b951a1f6fcf5c4631024e9b8bef.
Actual loaded class: Kimi_K25ForConditionalGeneration. Complete untied head and real small vision tower/projector.
Parameters: 238656; vision parameters: 9536; semantic tensors: 90.
Complete Kimi_K25ForConditionalGeneration native wrapper with real 1-layer MoonViT, 4-layer DeepseekV3 MLA text (one dense + three MoE layers), 4 routed/top2 + 1 shared expert, q/KV low-rank projections and full untied vocabulary head.
Native Transformers5.16.1, no remote model Python; complete shrunk real vision, text-panel QFS capture plus synthetic vision probes.
No MTP tensors or layers. Byte-level BPE tokenizer: 272 tokens, IDs 0..271.
Two independent builds have identical complete file inventories. Exact semantic reload checks values and dtypes.
Two fresh CPU capture processes, two threads each; honest lane other; sealed 4x64 synthetic panel, 252 scored positions.
Forced replay KL: 0.0; top-1 agreement: 1.0.
The actual NumPy float64 tokenwise artifact is checked for 252 positive zeros, not only a summary metric.
Cache boundary probes and real synthetic vision-input sensitivity are in native-verification.json.
Original packed formats are NOT validated by this BF16 fixture: Published source is nested compressed-tensors pack-quantized symmetric INT4 group32. BF16 random fixture is not a dequantization or test of those tensors. Native 5.16.1 Kimi_K25Config does not hoist nested quantization_config; direct packed loading is unverified and may fail strict missing/unexpected key checks. Original upstream serving stack recommends Transformers >=4.57.1,<5.0.0 with compatible vLLM/SGLang/KTransformers and compressed-tensors INT4 kernels. For QFS CPU use a complete independently verified standard BF16/FP32 checkpoint retaining original config/format provenance; do not strip quantization metadata to load packed tensors.
Captured model/code immutable pin: malaiwah/kimi-k25-tiny-random-bf16@07fec4f0e008a5947066b80cc7cb7c8058e5cfe0.
Use Python 3.12 and requirements-cpu.txt; $QFS is the recorded QFS checkout.
python release.py --stage-only --out NEW_DIR --qfs-root "$QFS" stages privately and refuses repository collisions.
python release.py --resume FAILED_DIR --out FAILED_DIR/recovery-001 --qfs-root "$QFS" reruns all checks without replacing the failed artifacts; it requires the unchanged private model pin and untouched empty private evidence repository.
python release.py --publish --out SAME_DIR publishes only unchanged successful stages and verifies anonymous downloads.
For third-party capture use the exact commands in support.json with an explicit --author.
Custom model code, if applicable, is never implicitly trusted; pass the exact immutable reviewed code pin.
Evidence: https://huggingface.co/datasets/malaiwah/kimi-k25-tiny-cpu-repro-v1
The evidence repository is a receipt bundle, not itself a single QFS root dataset; first/ and repeat/ are roots.
Reproduction is same-machine/same-stack repeatability, not cross-hardware equivalence.
Upstream Modified MIT retained with Kimi K2.7 Code attribution threshold; native Transformers implementation Apache-2.0. Independently generated random weights; no upstream tensors copied. See LICENSE, upstream-code-audit.json and retained source/license evidence. No paid GPU/cloud use.
Recovery of a failed private attempt. Original logs and receipts are retained unchanged under failed-attempt/ in the evidence repository and at the original local stage. Any prior reproduction-verified run is historical evidence only; this stage reruns both builds, native vision/cache/tensor verification, panel generation and both captures.
See recovery.json for the old and recovered immutable pins and retained file hashes.
Kimi K2.5 / K2.6 / K2.7-Code complete tiny random BF16 CPU fixture
0
2 commits
2 linked in READMEs
updated Sep 7, 2026
Untrained independently seeded random weights; no upstream weights or training data. This is a reproducibility fixture, not useful language modeling or production quality evidence.
Upstream moonshotai/Kimi-K2.7-Code@74797c9c62378b951a1f6fcf5c4631024e9b8bef.
Actual loaded class: Kimi_K25ForConditionalGeneration. Complete untied head and real small vision tower/projector.
Parameters: 238656; vision parameters: 9536; semantic tensors: 90.
Complete Kimi_K25ForConditionalGeneration native wrapper with real 1-layer MoonViT, 4-layer DeepseekV3 MLA text (one dense + three MoE layers), 4 routed/top2 + 1 shared expert, q/KV low-rank projections and full untied vocabulary head.
Native Transformers5.16.1, no remote model Python; complete shrunk real vision, text-panel QFS capture plus synthetic vision probes.
No MTP tensors or layers. Byte-level BPE tokenizer: 272 tokens, IDs 0..271.
Two independent builds have identical complete file inventories. Exact semantic reload checks values and dtypes.
Two fresh CPU capture processes, two threads each; honest lane other; sealed 4x64 synthetic panel, 252 scored positions.
Forced replay KL: 0.0; top-1 agreement: 1.0.
The actual NumPy float64 tokenwise artifact is checked for 252 positive zeros, not only a summary metric.
Cache boundary probes and real synthetic vision-input sensitivity are in native-verification.json.
Original packed formats are NOT validated by this BF16 fixture: Published source is nested compressed-tensors pack-quantized symmetric INT4 group32. BF16 random fixture is not a dequantization or test of those tensors. Native 5.16.1 Kimi_K25Config does not hoist nested quantization_config; direct packed loading is unverified and may fail strict missing/unexpected key checks. Original upstream serving stack recommends Transformers >=4.57.1,<5.0.0 with compatible vLLM/SGLang/KTransformers and compressed-tensors INT4 kernels. For QFS CPU use a complete independently verified standard BF16/FP32 checkpoint retaining original config/format provenance; do not strip quantization metadata to load packed tensors.
Captured model/code immutable pin: malaiwah/kimi-k25-tiny-random-bf16@07fec4f0e008a5947066b80cc7cb7c8058e5cfe0.
Use Python 3.12 and requirements-cpu.txt; $QFS is the recorded QFS checkout.
python release.py --stage-only --out NEW_DIR --qfs-root "$QFS" stages privately and refuses repository collisions.
python release.py --resume FAILED_DIR --out FAILED_DIR/recovery-001 --qfs-root "$QFS" reruns all checks without replacing the failed artifacts; it requires the unchanged private model pin and untouched empty private evidence repository.
python release.py --publish --out SAME_DIR publishes only unchanged successful stages and verifies anonymous downloads.
For third-party capture use the exact commands in support.json with an explicit --author.
Custom model code, if applicable, is never implicitly trusted; pass the exact immutable reviewed code pin.
Evidence: https://huggingface.co/datasets/malaiwah/kimi-k25-tiny-cpu-repro-v1
The evidence repository is a receipt bundle, not itself a single QFS root dataset; first/ and repeat/ are roots.
Reproduction is same-machine/same-stack repeatability, not cross-hardware equivalence.
Upstream Modified MIT retained with Kimi K2.7 Code attribution threshold; native Transformers implementation Apache-2.0. Independently generated random weights; no upstream tensors copied. See LICENSE, upstream-code-audit.json and retained source/license evidence. No paid GPU/cloud use.
Recovery of a failed private attempt. Original logs and receipts are retained unchanged under failed-attempt/ in the evidence repository and at the original local stage. Any prior reproduction-verified run is historical evidence only; this stage reruns both builds, native vision/cache/tensor verification, panel generation and both captures.
See recovery.json for the old and recovered immutable pins and retained file hashes.
