NVLC is exact entropy coding. It does not quantize, round, retrain, or alter a
BF16 weight. The integrated streaming verifier reconstructed every logical
source shard and compared it byte-for-byte with the immutable upstream
checkpoint.
Compression result
Metric
Value
BF16 tensors
1,199
BF16 weights
27,781,427,952
Raw BF16 weight bytes
55,562,855,904
Canonical NVLC bytes
36,440,103,827
Exact bits per weight
10.493371
Reduction from raw BF16
34.416431%
Per-codec verification
pass
Full integrated inverse verification
pass
The repository contains only the canonical distribution payloads and their
verification manifests. Local training samples, learned-model intermediates,
build logs, and local command paths are intentionally omitted. These omitted
files are not required to decode the checkpoint.
Encoding
Each BF16 word is separated into its exact fields:
[ sign:1 | exponent:8 | fraction:7 ]
The joint sign and exponent symbol is coded with a per-tensor rANS19 model.
After that symbol is decoded, it selects a soft-routed fraction distribution
for the exact seven-bit fraction. A small residual stream preserves the
safetensors header and any non-BF16 spans. Reconstruction combines the exact
fields bitwise.
manifest.json records upstream provenance, artifact sizes and hashes,
compression accounting, and all 18 exactness gates. SOURCE_SHA256SUMS records
the expected hashes of fully reconstructed upstream safetensors shards.
DGX Spark result
The linked implementation derives a kernel-oriented Fixed7/Compact11 runtime
from this canonical model. On one NVIDIA DGX Spark, TP1, C1, with four
sequential 512-input to 256-output requests per repeat:
Runtime
Median output TPS
Change
Stock BF16
4.413039
baseline
NVLC compressed-only
5.957506
+34.997797%
The frozen correctness request matched all 256 output token IDs, all 256
selected-token logprob rows, and the decoded text. The active runtime covered
851 of 851 required text tensors and retained no dense checkpoint or persistent
decoded-weight cache.
The runtime result is narrow: one DGX Spark, language-model only, no MTP,
concurrency one. Median mean TTFT increased from 635.617 ms to 1,255.411 ms in
the sealed workload, while median mean TPOT decreased from 224.768 ms to
163.426 ms.
The NVLC files are not directly loadable by standard Transformers, vLLM, or an
unmodified SGLang installation. Use the linked implementation to verify,
reconstruct, or generate the DGX Spark runtime streams.
Exact reconstruction
Build the host reconstructor from the implementation repository, then run:
The output path must not already exist. Reconstruction is bounded and
streaming, but the complete dense output still requires approximately the
upstream checkpoint size on disk. The compressed candidate runtime does not
depend on this reconstructed output.
Runtime representation
The canonical rANS files are the distribution format. The measured GPU runtime
derives fixed-width, lane-oriented streams once at initialization:
Runtime accounting
Bytes
Change from raw active text weights
Raw active BF16 text weights
53,791,996,928
baseline
Resident codes plus metadata
38,199,875,520
-28.985950%
Streams plus fully charged workspace
40,751,072,704
-24.243242%
No GPU-native runtime cache is included in this model repository because it is
hardware and kernel-layout specific. The implementation produces it from this
canonical model and verifies each operation against the canonical source
identity.
License
The model metadata and this compressed representation are distributed under
the upstream Apache License 2.0. See LICENSE and UPSTREAM_README.md.
