SuperPUPS/Qwen3.8-Flash-Next-INT8KV

Qwen3.8-Flash-Next — INT8 KV build (int8_per_token_head)

A serve-time release bundle for running Qwen3.8-Flash-Next-INT8x25-AttnBF16 with an INT8 per-token-per-head KV cache, which roughly doubles the KV pool at a small prefill cost.

This bundle contains no weights. The checkpoint is byte-identical to INT8x25-AttnBF16; int8_per_token_head only changes the KV-cache dtype at runtime. It also contains no full engine source — that lives in vllm-qfn-source; this bundle ships the focused INT8 KV delta in patches/.

Engine identity

• Engine: vllm-qfn
• Snapshot: branch feature/qsa-int8-kv, commit e12683d245
• Reported vllm --version: 0.28.1rc1.dev78+gb009560d6.d20260923.cu132
• Full source: SuperPUPS/vllm-qfn-source

Upstream vLLM 0.31 does not accept int8_per_token_head in the QSA path — this is a fork extension.

What it buys you (measured, C4 / INT8x25, 2× CMP 170HX 64 GB, TP2+EP, MTP k=4)

Metric	BF16 KV (default)	INT8 KV (int8_per_token_head)
KV pool @262K	483 850 tok (1.85×)	776 632–796 231 tok (2.96–3.04×)
Bytes / token / GPU	1024	520
Prefill	—	mean tax 4.3–5.1% (worst 16K = 5.1%)
Decode	—	within ±3.1%
concurrent2 wall	7.10–7.25 s	7.41–7.48 s (~4%)

The INT8 KV mode lets the higher-quality INT8x25 checkpoint reach roughly the concurrency of the cheaper INT8x17 sibling (≈3× vs 1.85×), at a ~4–5% prefill cost.

Quality gates (INT8 KV vs C4-BF16 baseline)

Gate	Result
Golden probes (temp=0)	3/3
10-task battery	6/10, identical pass-set to C4-BF16 (0 regressions)
Determinism (8× same prompt)	1/8 unique
Needle-in-a-haystack @128K	HIT (118 593 prompt tokens)
Prefix-cache repeat	identical + fact recall
Pool gate A7 (floor 700K / target 770K)	PASS

Raw artifacts are in evidence/.

Quick start

export VLLM_PLE_CPU_OFFLOAD=1
export VLLM_PLE_MMAP_TABLE=1     # required on <135 GiB RAM hosts
bash launch/run_qsa_int8kv.sh

See launch/ENV.md for the full variable list and rationale.

Files

launch/run_qsa_int8kv.sh    validated INT8 KV launch recipe (sanitized)
launch/ENV.md               env vars + KV flag rationale
patches/qsa-int8.diff       focused INT8 KV delta (engine source)
patches/int8_read_fix.patch read-path scale-layout fix
patches/int8_unit_test.py   standalone GPU unit tests for the INT8 KV path
evidence/*.json|txt         gate evidence (probes, battery, needle, det, prefix, A/B bench, pool)
benchmark_vllm.py           streaming A/B benchmark used for the numbers
CHECKSUMS.sha256            sha256 of every file in this bundle

Caveats

• Requires the patched vllm-qfn engine at e12683d245 (or later on the same branch); plain upstream vLLM 0.31 will reject the flag.
• Needs VLLM_PLE_MMAP_TABLE=1 on hosts with <135 GiB RAM (the model carries a 102.4 GB PLE ngram table).
• The INT8 KV numbers were measured with reasoning_effort: medium and --language-model-only; xhigh also works but was not the measured config.
• zram-based swap was enabled on the reference host. This is host tuning only; it is not required by the model or the mmap path, and it is not a workaround for the upstream vLLM 0.31 pinned-table OOM.

Reproduction

1. Install the vllm-qfn engine from vllm-qfn-source.
2. bash launch/run_qsa_int8kv.sh.
3. python benchmark_vllm.py out.json for the A/B numbers; python patches/int8_unit_test.py for the INT8 KV unit tests.