dankclaw: switch frank-default-large to unsloth Qwen3.8-27B UD-Q8_K_XL

- rebuild llama.cpp stock to b10433 (9b05354ec) for bundled-MTP support
- MODEL_FILE -> Qwen3.8-27B-UD-Q8_K_XL.gguf; DRAFT_FILE empty (MTP bundled)
- SPEC: draft-mtp, n_max=10, p_min off (~30 t/s stable) - matrix winner
- CTX 524288 / parallel 2 = 262144/slot (matches pi contextWindow)
- serve script: bundled-MTP (no --spec-draft-model), optional --spec-draft-p-min

scripts/frank-default-large.env

@@ -1,51 +1,51 @@
-# Tunables for frank-default-large (HauhauCS Qwen3.6-27B Q8_K_P, DFlash+ngram-mod, dankclaw)
-# Runtime: llama.cpp llama-server; AMD Strix Halo (gfx1151) 96GB UMA; Ryzen AI MAX+ 395.
+# Tunables for frank-default-large (unsloth Qwen3.8-27B UD-Q8_K_XL, dankclaw)
+# Runtime: llama.cpp llama-server (b10433+, rebuilt 2026-08-14 for bundled-MTP
+# support); AMD Strix Halo (gfx1151) 96GB UMA; Ryzen AI MAX+ 395.
 MODEL_DIR=/home/dank/models
+MODEL_FILE=Qwen3.8-27B-UD-Q8_K_XL.gguf
 ALIAS=frank-default-large
 HOST=0.0.0.0
 PORT=7770
-# 3 parallel slots x 200700 ctx each = 602100 total context pool.
-CTX_SIZE=602100
-PARALLEL=3
-# Spec: speculative decoding mode.
-#   draft-dflash  - DFlash block-diffusion draft (benchmark winner pre-2026-08-14)
-#                   BUT: premature EOS under parallel load when a new task
-#                   reuses a cached prompt (shared draft ctx corruption). Retired.
-#   draft-mtp     - MTP sidecar draft (a4lg/Qwen3.6-27B-MTP-ONLY-Q8_0.gguf)
-#   ngram-mod     - draftless ngram matching (no draft model, no shared ctx)
-# Benched 2026-08-14: see docs/frank-default-large-benchmark.md for the winner.
+# Context pool. Qwen3.8 native context = 262144 per sequence.
+# 524288 total / PARALLEL 2 = 262144 per slot, matching the pi client's
+# contextWindow=262144 for frank-default-large. Model is 29GB; ROCm has
+# ~70GB headroom, so 2x262k quantized KV fits. Verify at startup.
+CTX_SIZE=524288
+PARALLEL=2
+# Spec: selected from the 2026-08-14 matrix (bench-spec-matrix.sh).
+# WINNER: draft-mtp alone, n_max=10, p_min off -> ~30 t/s, STABLE, no
+# premature stop.  Layered draft-mtp,ngram-mod was faster on paper (median
+# 133 t/s) but is non-deterministic (10-148 t/s) and shows a degenerate
+# instant-EOS (tok=0) -> the same premature-stop class that retired DFlash.
+# Rejected as production-unsafe.
 SPEC_TYPE=draft-mtp
-# Draft model file (only used when SPEC_TYPE contains 'draft-').
-# DFlash: Qwen3.6-27B-DFlash-Q8_0.gguf ; MTP: Qwen3.6-27B-MTP-ONLY-Q8_0.gguf
-DRAFT_FILE=Qwen3.6-27B-MTP-ONLY-Q8_0.gguf
-# Draft length per spec type: dflash 8, mtp 2-3.
-SPEC_DRAFT_N_MAX=3
+# Bundled MTP: Qwen3.8 keeps the MTP head in the target GGUF. llama.cpp
+# auto-detects it (b10411+). DRAFT_FILE MUST stay EMPTY.
+DRAFT_FILE=
+# Draft length + sampling floor, from the spec matrix (n_max 2..10, p_min
+# off/0.8; p_min 0.8 did not help draft-mtp and slowed it slightly).
+SPEC_DRAFT_N_MAX=10
+SPEC_DRAFT_P_MIN=
 SPEC_NGRAM_MOD_N_MIN=48
 SPEC_NGRAM_MOD_N_MAX=64
 SPEC_NGRAM_MOD_N_MATCH=24
-# Backend device from --list-devices. ROCm wins on generation; Vulkan renders
-# DFlash spec unusable (NONE-op abort) and gen is slower on ngram (6.9 t/s).
+# Backend device from --list-devices. ROCm0 is the production backend.
 DEVICE=ROCm0
-# Dense model (Qwen3.6-27B, not MoE): no MoE CPU offload. Full offload.
+# Dense model: full offload. (Not MoE.)
 N_GPU_LAYERS=all
-# KV on GPU, quantized (K q8_0, V q8_0). V MUST match K: this llama.cpp
-# build has GGML_CUDA_FA_ALL_QUANTS=OFF, and Flash Attention requires
-# K type == V type. Mixed q8_0/f16 silently disables FA, collapsing
-# long-context decode to ~1 t/s at 200k ctx. q8_0 V halves V memory too.
+# KV on GPU, quantized (K q8_0, V q8_0). Keep K == V for Flash Attention;
+# mixed types silently disable FA (long-context decode collapses).
 CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=q8_0
-# Batch sizing: empirically verified 2026-08-13 with caches dropped.
-# batch 8192 / ubatch 2048 loads and serves (61 GiB VRAM peak, ctx 602100).
-# ubatch is the compute-graph driver; larger ubatch = more VRAM, faster prefill.
-# Re-verified 2026-08-14 during the spec swap (see benchmark doc).
-BATCH_SIZE=8192
-UBATCH_SIZE=8192
-# Sampling settings — per HauhauCS Qwen3.6-27B model card.
-# Thinking-general: temp=1.0, top_p=0.95, top_k=20, min_p=0.
-# Clients may override per-request; these are the server-side defaults.
+# Batch sizing (start conservative; tune after verifying memory):
+# ubatch is the compute-graph driver; larger = more VRAM, faster prefill.
+BATCH_SIZE=2048
+UBATCH_SIZE=512
+# Sampling settings — per Qwen3.8-27B model card thinking mode.
+# Thinking: temp=1.0, top_p=0.95, top_k=20, min_p=0. Clients may override.
 # WARNING: do not name this variable TEMP. The ROCm HIP runtime reads the
 # TEMP environment variable (JIT kernel-cache dir); TEMP=1.0 made the HIP
-# runtime SEGV (NULL unref) during model load on Aug 13 2026.
+# runtime SEGV during model load on Aug 13 2026.
 TEMPERATURE=1.0
 TOP_K=20
 TOP_P=0.95