frank8465/Swift-1.5-Qwen3.8-Flash-Next-GPTQ-W4A8-Radiance-AMD-R9700

Swift-1.5-Qwen3.8-Flash-Next — GPTQ W4A8 (radiance .rad)

GPTQ 4-bit quantization of ukisai/Swift1.5-Qwen3.8-Flash-Next (335.3 GiB BF16 → 113.7 GiB) for the radiance engine on AMD RDNA4 (gfx1201). The .rad format loads in radiance only — not GGUF, not safetensors.

File	qwen4exp-swift15-w4dn8-i8-hc8m.rad (rename freely; the engine reads by content)
Size	122,071,973,888 bytes (113.69 GiB), 2.95× vs BF16
SHA256	09dd9306e859f5343d9cb77ee6736636ab9df50f3ed2153350dfbb35be5e6948 — the deployed file and this upload are the same bytes
Engine	stilldeadcode/radiance@sha256:4c3f484564bb2872784e46b0a59cbd71becc9c1e746e301a3596ffae253c322e (pin 1.0.13)
Proof hardware	2× Radeon AI PRO R9700, TP=2 (--tp-wire wht6), ~190 t/s decode at 44.5k context

vs the engine author's quant

StillDeadcode/swift1.5-qwen3.8-next-flash-fp8-iq4r-moe quantizes the same checkpoint with the same GPTQ rule. Two recipe decisions separate them:

	theirs	ours
calibration tokens	1,000,000	10,000,000
BF16-protected experts / layer	16	24
mean KL vs BF16 (their card / our gate)	0.0789	0.0131
size	122,013,646,848 B	122,071,973,888 B (+55.6 MiB, +0.048 %)

Six more protected experts per layer and a 10× calibration pass buy a 6× lower KL for 55 MiB. The KL is the honest axis: at their 92.9 % top-1 parity sits with ordinary Q4 quants; at 97.9 % we sit with Q6 tier.

Quant recipe

• All 49,152 routed experts: GPTQ 4-bit (w4nl64a8h: w4 weights, nl codebook, a8 activations), act-order, damping 0.05.
• Calibration: 10,000,000 tokens captured by serving the BF16 checkpoint (engine 1.2.0, RADIANCE_CALIB_DIR + RADIANCE_CALIB_TOKENS) over mixed code/prose/tool traffic — 10× the usual 1M-token pass. The corpus is published (anonymized): frank8465/swift15-gptq-w4-calib-corpus.
• 24 experts per layer kept whole BF16 (the protected set), derived from the energy share of the calibration Grams (RADGRAM2). The plugin takes one encoding per expert table, so protection is whole-expert — no per-matrix splits.
• Trunk / lm_head: int8. HC mixtures: fp8. PLE n-gram table: fp8 fixed-scale (95.4 GiB BF16 → 47.7 GiB).
• MTP spec-decode head rides in the file; Froggeric v22.5 chat template vendored in scripts/.

Accuracy evidence

Measured against the BF16 reference on the same engine pin, production shape (MNBT 2048, fp8 KV), 50,235 scored positions on a pool disjoint from calibration:

metric	value
mean KL divergence	0.01314 (gate bar 0.03)
top-1 agreement	97.92 %
perplexity ratio	1.0052

The full gate JSON, every scored position (.rows), a second gate run at MNBT 3072 (Δ ≈ 0.0015 = the float-order noise band), the engine's rad-info census, and the 3072 run log are in artifacts/.

Run it

podman run --rm --name radiance \
  --device /dev/kfd --device /dev/dri --group-add keep-groups \
  --ulimit memlock=-1:-1 \
  -v /path/to/models:/models:ro \
  stilldeadcode/radiance@sha256:4c3f484564bb2872784e46b0a59cbd71becc9c1e746e301a3596ffae253c322e \
  --model /models/qwen4exp-swift15-w4dn8-i8-hc8m.rad \
  --tp 2 --tp-wire wht6 \
  --placement expert_tiered --host-pool-mib 40960 \
  --vram-weights-mib 26071 --vram-kv-mib 3543 \
  --max-model-len 200000 --max-num-seqs 3 --max-num-batched-tokens 2048 \
  --num-speculative-tokens 2 --ngram-placement ram \
  --override-chat-template /models/qwen38-froggeric-v22.5.jinja

Sizing: experts live in a pinned host pool + VRAM cache; the n-gram table pins 47.7 GiB of RAM — budget ~132 GiB pinned (LimitMEMLOCK=infinity). --max-num-batched-tokens 2048 is a kernel law for disk-tiered w4, not a tuning knob. A prompt over --max-model-len is refused, never truncated.

Files

File	What
qwen4exp-swift15-w4dn8-i8-hc8m.rad	the container
artifacts/kld-gate-mnbt2048.json (+ .rows)	production-shape gate: stats + every position
artifacts/kld-gate-mnbt3072.json (+ run log)	second gate at MNBT 3072
artifacts/rad-info-*.txt	engine's own tensor/encoding census
scripts/	recipe, convert driver, protected-expert derivation, KLD gate, fleet probes, chat template

License chain

Qwen/Qwen3.8-Flash-Next                  license: qwen-community-1.0
  └─ ukisai/Swift1.5-Qwen3.8-Flash-Next  license: swift-open-license-1.0
       └─ this .rad                      GPTQ w4a8, 10M-token self-calibration

This repository carries the direct base's license; the root model's qwen-community-1.0 acceptable-use terms apply downstream. Calibration and scoring corpus: frank8465/swift15-gptq-w4-calib-corpus (anonymized fleet slices + public tiny-codes/wikitext, CC BY-SA 4.0).