GeoRefine Qwen3.8-27B Context — Linux CUDA v2

A lossless BF16 weight package with a matched Linux CUDA serving runtime. The complete reference/package storage ratio is 1.530696×. The installed batched runtime has measured token-generation and end-to-end speedups over its dense BF16 reference on an RTX PRO 6000 Blackwell 96GB.

Release scope: Linux x86-64/CUDA. Windows and macOS Metal are unqualified. Results below cover the named artifacts and configurations; they are not a universal serving or cross-platform certificate.

The parent checkpoint is Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, with 27,781,427,952 parameters. The package retains the source weights, vision and MTP components, configuration, tokenizer and other checkpoint assets. Its storage format is bitexact-predictive-context-package-v1 (BCTX/PPCX); it uses no weight quantization, pruning or ablation. Ordinary transformers.from_pretrained cannot load these codec files directly. Restore the checkpoint or use the matched runtime below. Preserve the upstream model license and notices.

Measured Linux results

The installed runtime is georefine-context-fast-runtime==0.1.1+cu129, source 6e24ede1489fa1061c9a435cbc66d5987f1762b7. The measured entry point is glc_serve.context_batch_server, using TBE21-v2 CUDA kernels.

These are actual Linux measurements, replacing the older card's blanket “Not proven” entries for the old reference adapter. They do not qualify every entry point named “fast” in the source tree.

The Blackwell test used 48 vCPU, 180GiB nominal host RAM, a 1,024-token context limit and 64 output tokens. Measured prompts averaged about 68 tokens. Each concurrency cell included three sequential requests per worker after excluded warmup: 192 requests at 64 users and 384 at 128. The reported ratios use the complete cell-wide statistics, not minima from independently synchronized trials. Decode rate is (output_tokens - 1)/(last_token_time - first_token_time); end-to-end throughput uses all output tokens divided by the full client window.

The candidate used natural MTP (spec_k=1, KV-only confirmation); the dense BF16 reference used spec_k=0. These speedups cover the combined codec/runtime/batching/MTP serving path. They are not isolated codec-only speedups or an independent Hugging Face parent-runtime comparison. Whole-process GPU memory is not the stored-weight compression ratio.

A100 40GB resident-worker evidence

A separate full-model Linux run on one visible A100-SXM4-40GB recorded 512/512 identical comparison rows over eight requests, 15.7556 aggregate tokens/s, 17.9645 mean per-request decode tokens/s and 0.6212 s TTFT p99. Its GPU process peak was 40,133,197,824 bytes. The test used a two-GPU VM for its 170GiB host RAM, with one GPU visible to the worker; this is not a two-GPU inference result or a claim about single-GPU VM billing.

External startup was 906.861 s / 908.570 s for the two stages. This exceeded the original 600-second cold-start criterion and was separately accepted for a resident worker. Requests reuse a loaded process; restart or eviction repeats startup. The separate compiler-observer profile failed process classification and remains recorded as failed. These functional measurements do not certify that observer/lifecycle profile.

Pinned artifacts and evidence

The wheel was reproduced twice with the same checksum; all 234 runtime manifest entries were checked. See measured results and receipt hashes, benchmark configuration, and release checksums. The original broader four-platform release matrix remains unqualified and is not represented as passed by this Linux-only release.

Download, verify and restore

hf download Tsotchke-Corporation/Qwen3.8-27B-GeoRefine-Context \
  --revision v2-linux-20261101 --include 'package/*' 'runtime/linux-cuda-v2/*' \
  --local-dir ./qwen38-context
python -c 'import hashlib,pathlib; p=pathlib.Path("qwen38-context/package/manifest.json"); assert hashlib.sha256(p.read_bytes()).hexdigest()=="bbac23fdeba3020f8951d7a0a005016985c40add08979f37a85df2f22ccbacee"'
python ./qwen38-context/package/decoder/predictor_package.py verify ./qwen38-context/package
python ./qwen38-context/package/decoder/predictor_package.py restore \
  ./qwen38-context/package ./qwen38-restored-bf16

Use Python 3.10 or newer, NumPy and a C++17 compiler for package tools. Allow space for both the compressed package, the approximately 55.6GB restored checkpoint and working files. Keep caches outside package/.

Install the measured Linux CUDA runtime

Use a fresh environment. The measured profile uses Python 3.10, PyTorch 2.9.1+cu129, TorchVision 0.24.1+cu129, Transformers 5.5.0, and a CUDA 12.9 toolkit with nvcc, C++17 compiler and Python development headers.

python3.10 -m venv .venv-georefine
. .venv-georefine/bin/activate
python -m pip install --extra-index-url https://download.pytorch.org/whl/cu129 \
  ./qwen38-context/runtime/linux-cuda-v2/georefine_context_fast_runtime-0.1.1+cu129-py3-none-any.whl

For the Blackwell 96GB/large-host-RAM configuration, inspect the launcher plan first:

python ./qwen38-context/runtime/linux-cuda-v2/serve_blackwell.py \
  --package ./qwen38-context/package --parent ./qwen38-restored-bf16 \
  --build-root "$PWD/.scratch/georefine-context-build"

Add --execute to launch on the matching Linux CUDA environment. This launcher delegates to glc_serve.context_batch_server; it is not the older serial reference adapter. The A100 resident-worker measurements use a different bounded profile and are not covered by this Blackwell launcher. Loading includes Context-to-TBE21 preparation; prepared native code caches were used for the startup measurements. No Windows, Metal, universal hardware-speed, isolated codec-only speed, or independent HF-runtime parity claim is made here.