MiniV TurboQuant + DFlash Follow-up

Public appendix · 2026-07-08 · separated prefill/decode/draft metrics

Measured negative · integrated build

The combined TurboQuant+DFlash binary runs, but DFlash is not viable for this 27B target/drafter pair.

Qwen/Qwopus3.6 27B Q6_K with Qwen3.6-27B-DFlash Q8_0 produced 0% draft-token acceptance across n_max=2/4/8/12. Every DFlash case was slower than TurboQuant baseline using turbo4/turbo3 KV cache.

JSON dataset Summary CSV Back to DFlash pilot

Scope

Hardware: Corsair AI Workstation 300 / MiniV, AMD Ryzen AI MAX 385, Radeon 8050S RADV Vulkan, 64GB UMA.
Build: llama.cpp TurboQuant fork + experimental DFlash backport, version 9681 (8c2b1b3f8).
Context: 8,192 tokens · completion cap: 256 tokens.

baseline_tq

n_max0

decode9.347 tok/s

wall9.109 tok/s

accepted/generated—

acceptance—

dflash_tq_n2

n_max2

decode7.004 tok/s

wall6.849 tok/s

accepted/generated0 / 507

acceptance0.0%

dflash_tq_n4

n_max4

decode5.395 tok/s

wall5.304 tok/s

accepted/generated0 / 1010

acceptance0.0%

dflash_tq_n8

n_max8

decode2.421 tok/s

wall2.402 tok/s

accepted/generated0 / 2004

acceptance0.0%

dflash_tq_n12

n_max12

decode2.337 tok/s

wall2.319 tok/s

accepted/generated0 / 2982

acceptance0.0%

Per-case metrics

Case	n_max	Status	Load sec	Prefill tok/s	Decode tok/s	Wall tok/s	Acceptance
baseline_tq	0	OK	74.4	93.7	9.347	9.109	—
dflash_tq_n2	2	OK	80.1	82.5	7.004	6.849	0.0%
dflash_tq_n4	4	OK	80.2	82.8	5.395	5.304	0.0%
dflash_tq_n8	8	OK	78.1	77.3	2.421	2.402	0.0%
dflash_tq_n12	12	OK	72.1	77.3	2.337	2.319	0.0%

Conclusion

Do not promote this path to the MiniV production router. Larger draft windows only increased rejected-draft overhead. Baseline TurboQuant remains the correct production lane for this target.