Which context sizes have actually passed?
A context setting reserves space. A long-input test has to count the tokens actually consumed. A parity test adds a third requirement: compare the same tokens or scores against an identified reference.
- 32K · Inkling 2 Mini Q8_0. A component replay at 32,768 cache positions matched its reference. The report explicitly excludes completed 32K model output. 32K record
- 32K · Qwen3.8 dense. The 32K-capable session bundle passed a 512-token prefix plus 65-token suffix check. That particular pass consumed 577 tokens. Session record
- 64K · Inkling. Capacity compiled and seeded boundary cases passed. The sliding-cache report marks full 64K replay pending. 64K record
- 128K · Qwen3.8 dense. Six 131,072-token runs completed on an earlier artifact. Its shorter 4,000-row full-vocabulary comparison matched llama.cpp. Full 128K bit-exact parity remains open.
- 128K · North Mini Code. Two 131,072-token runs reproduced the same input, first-logit and output-token hashes. That establishes repeatability on CKE. Independent long-context parity and task quality remain open. Published boundaries
Run one model at a time
On a host sized for Qwen3.8 27B, supply the model and oracle GGUF paths from the test catalog. Start with 32K. The same output directory allows the harness to resume completed work. These are potentially multi-hour runs.
export V8_QWEN38_MODEL=/absolute/path/to/model.gguf
export V8_QWEN38_GGUF="$V8_QWEN38_MODEL"
.venv/bin/python version/v8/scripts/run_long_context_certification_v8.py \
--models qwen38_27b --contexts 32768,65536,131072 \
--threads 16 --repetitions 2 --output-dir "$HOME/cke-results/qwen38-long"Read the result before claiming parity
Check model hash, host, actual prompt and generated token counts, repeats, peak memory, prompt rate, generation rate and oracle availability. The harness has a separate parity-context option (default 2,048). A capacity sweep does not automatically compare full logits at all three lengths. Keep missing oracle rows, skipped rows and shorter comparisons explicit. Harness options
Compare against llama.cpp at 32K
Use a built, compatible llama.cpp oracle and the exact GGUF used by CKE. Request one matching context and parity context per invocation. Use a fresh output directory for each comparison.
.venv/bin/python version/v8/scripts/run_long_context_certification_v8.py \
--models qwen38_27b --contexts 32768 --parity-context 32768 \
--token-id 1 --parity-new-tokens 8 --no-quality \
--output-dir "$HOME/cke-results/qwen38-parity-32k"Read summary.json and the per-row numerical-parity.json. The harness compares the next few steps after a repeated-token prefix. Compare top1_exact_rows with bit_exact_rows: the same winning token can have different scores. This fixture does not measure recall from a real document.
Linux limit at larger sizes. At this revision the parity harness passes all input IDs as one comma-separated command-line argument. Even single-digit IDs at 128K exceed the single-argument limit on typical 4 KiB-page Linux hosts. The capacity phase uses a token file. Full-length oracle checks need a file-input path through the comparison tools before advertising a reliable 128K parity command. A missing report is an open result.
Repeat the oracle sample at 64K
Use the 32K command above with --contexts 65536 --parity-context 65536, keep --token-id 1, and select a new output directory. With this one-digit fixture the prompt argument is 131,071 bytes. On the documentation host we checked the Linux argument boundary alone: 32K and 64K fit; the 128K argument failed before a child process started. This was an operating-system check, not a model parity run. For 128K use the capacity sweep and label the oracle comparison pending.