Local overrides to the pinned runner for production chat:
- --reasoning on --reasoning-format deepseek (thinking to
message.reasoning_content, matching the retired v3 service so the
fleet's reasoning clients work)
- --reasoning-budget -1 (unrestricted thinking)
- --chat-template-kwargs '{"preserve_thinking":"true"}' (preserve
thinking tokens in the conversation history across turns)
- batch 4096 / ubatch 2048 (larger than the runner's 2048/512; unified
memory has headroom)
The Kairic IU4 lane (--kairic-edge) and the MTP draft fast path are
unaffected.
The pinned runner sets LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1, but the
exact-Q8 argmax fast path fails closed on any non-greedy request. The
model card's own chat instructions override sampling per request with
Qwen live values (temp 0.7, top-p 0.8, top-k 20, min-p 0,
presence_penalty 1.5), which the fast path rejects (400). Production
chat forces the fast path to 0; the Kairic IU4 lane (--kairic-edge) and
the MTP draft fast path are unaffected.
Serves jcbtc/Qwen3.8-27B-IU4-Kairic-Edge (four-artifact payload) with
the qualified pinned-runner profile: 262K ctx, 1 slot, b2048/ub512,
t16/tb32, f16 KV, 32 ctx checkpoints, 8 GiB prompt cache with idle-slot
persistence, native MTP4, deterministic sampler, reasoning off.
The serve script replicates the release pinned runner verbatim; the only
differences are safe bind vars (no bare HOST env var; HIP collision rule)
and this machine's split 7.15 SDK LD_LIBRARY_PATH layout.
Rollback: revert this commit and redeploy; v3 model files and the v3
binary are untouched.
Restore the CIRU ROCmFPX setup that DFlash2 replaced:
- Restore jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge v2 runtime
(native depth-4 MTP, PromptForge DualView, ROCm 7.15)
- Archive the DFlash2 Vulkan deployment files in .AGENTTRASH
- DFlash2 was retired the same day it shipped (2026-08-21) for speed
- Revert docs (deployment.md, index.md, AGENTS.md); mark dflash2 page archived
- Restored scripts are byte-identical to the pre-swap CIRU versions
Replace the retired CIRU ROCmFPX runtime with upstream llama.cpp PR #27342
DFlash2 on the Vulkan backend:
- Target: unsloth Qwen3.8-27B-UD-Q8_K_XL.gguf (31.4 GB, already on disk)
- Drafter: z-lab Qwen3.8-27B-DFlash2-Q4_K_M.gguf (block size 8, top-k 16)
- Runtime: /home/dank/llama.cpp-dflash2/build-dflash2/bin/llama-server
(GGML_VULKAN=ON; HIP lacks TOP_K so the stack runs on Vulkan0/RADV)
- spec: --spec-type draft-dflash --spec-draft-n-max 8 (clamps to 7)
Validated 2026-08-21 on scratch port 52626: mean acceptance 5.27,
health + reasoning completion OK.
Archive the retired CIRU .env/.serve script in .AGENTTRASH.
Update deployment.md, index.md, AGENTS.md; add docs/dflash2-large-vulkan.md.
llama.cpp splits -c across -np slots; to give each of the 2 slots the full
262144 native context, set CTX_SIZE=524288. KV ~32 GiB; fits the 96 GiB
unified pool. Update env default, serve-script fallback, and the deployment
doc line.
- Restore the pre-2026-08-17 Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
Q6_K_P + ngram-mod setup as frank-default-small on Dankserver (port 7770,
upstream llama.cpp b10121, MoE experts on CPU via --n-cpu-moe 36).
- Rename the CIRU Qwen3.8-27B ActiveFPX setup from frank-default-small to
frank-default-large (Dankclaw, port 52625); bump to CTX_SIZE=262144
(native max) and PARALLEL=2 (96 GiB unified pool).
- Retire the Qwen3.8-27B IQ4_XS FastMTP frank-default-large (Dankserver) to
.AGENTTRASH.
- deploy.sh: SERVICE[dankserver]=frank-default-small,
SERVICE[dankclaw]=frank-default-large; HEALTH_WAIT[dankserver]=600.
Frank