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)
+# Tunables for frank-default-large (unsloth Qwen3.8-27B UD-Q8_K_XL, dankclaw)
-# Runtime: llama.cpp llama-server; AMD Strix Halo (gfx1151) 96GB UMA; Ryzen AI MAX+ 395.
+# 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.
+# Context pool. Qwen3.8 native context = 262144 per sequence.
-CTX_SIZE=602100
+# 524288 total / PARALLEL 2 = 262144 per slot, matching the pi client's
-PARALLEL=3
+# contextWindow=262144 for frank-default-large. Model is 29GB; ROCm has
-# Spec: speculative decoding mode.
+# ~70GB headroom, so 2x262k quantized KV fits. Verify at startup.
-#   draft-dflash  - DFlash block-diffusion draft (benchmark winner pre-2026-08-14)
+CTX_SIZE=524288
-#                   BUT: premature EOS under parallel load when a new task
+PARALLEL=2
-#                   reuses a cached prompt (shared draft ctx corruption). Retired.
+# Spec: selected from the 2026-08-14 matrix (bench-spec-matrix.sh).
-#   draft-mtp     - MTP sidecar draft (a4lg/Qwen3.6-27B-MTP-ONLY-Q8_0.gguf)
+# WINNER: draft-mtp alone, n_max=10, p_min off -> ~30 t/s, STABLE, no
-#   ngram-mod     - draftless ngram matching (no draft model, no shared ctx)
+# premature stop.  Layered draft-mtp,ngram-mod was faster on paper (median
-# Benched 2026-08-14: see docs/frank-default-large-benchmark.md for the winner.
+# 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-').
+# Bundled MTP: Qwen3.8 keeps the MTP head in the target GGUF. llama.cpp
-# DFlash: Qwen3.6-27B-DFlash-Q8_0.gguf ; MTP: Qwen3.6-27B-MTP-ONLY-Q8_0.gguf
+# auto-detects it (b10411+). DRAFT_FILE MUST stay EMPTY.
-DRAFT_FILE=Qwen3.6-27B-MTP-ONLY-Q8_0.gguf
+DRAFT_FILE=
-# Draft length per spec type: dflash 8, mtp 2-3.
+# Draft length + sampling floor, from the spec matrix (n_max 2..10, p_min
-SPEC_DRAFT_N_MAX=3
+# 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
+# Backend device from --list-devices. ROCm0 is the production backend.
-# DFlash spec unusable (NONE-op abort) and gen is slower on ngram (6.9 t/s).
 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
+# KV on GPU, quantized (K q8_0, V q8_0). Keep K == V for Flash Attention;
-# build has GGML_CUDA_FA_ALL_QUANTS=OFF, and Flash Attention requires
+# mixed types silently disable FA (long-context decode collapses).
-# 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.
 CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=q8_0
-# Batch sizing: empirically verified 2026-08-13 with caches dropped.
+# Batch sizing (start conservative; tune after verifying memory):
-# batch 8192 / ubatch 2048 loads and serves (61 GiB VRAM peak, ctx 602100).
+# ubatch is the compute-graph driver; larger = more VRAM, faster prefill.
-# ubatch is the compute-graph driver; larger ubatch = more VRAM, faster prefill.
+BATCH_SIZE=2048
-# Re-verified 2026-08-14 during the spec swap (see benchmark doc).
+UBATCH_SIZE=512
-BATCH_SIZE=8192
+# Sampling settings — per Qwen3.8-27B model card thinking mode.
-UBATCH_SIZE=8192
+# Thinking: temp=1.0, top_p=0.95, top_k=20, min_p=0. Clients may override.
-# 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.
 # 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

scripts/serve-llama-frank-large.sh

@@ -1,8 +1,7 @@
 #!/usr/bin/env bash
-# Serve HauhauCS Qwen3.6-27B-Uncensored-HauhauCS-Aggressive:Q8_K_P
+# Serve unsloth Qwen3.8-27B UD-Q8_K_XL on dankclaw with bundled-MTP
-# with DFlash draft speculative decoding (draft-dflash) on dankclaw
+# speculative decoding (draft-mtp). ROCm backend.
-# (AMD Strix Halo gfx1151, ROCm backend).
+# Runtime: llama.cpp llama-server (>= b10411).
-# Runtime: llama.cpp llama-server.
 #
 # Tunables come from frank-default-large.env in this directory unless the
 # FRANK_LARGE_ENV_FILE environment variable overrides that path.
@@ -19,15 +18,15 @@ fi
 
 # --- defaults (used when the env file does not set them) -----------------
 : "${MODEL_DIR:=/home/dank/models}"
-: "${MODEL_FILE:=Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf}"
+: "${MODEL_FILE:=Qwen3.8-27B-UD-Q8_K_XL.gguf}"
-: "${DRAFT_FILE:=Qwen3.6-27B-DFlash-Q8_0.gguf}"
+: "${DRAFT_FILE:=}"        # empty: Qwen3.8 bundles its MTP drafter in-model
 : "${ALIAS:=frank-default-large}"
 : "${HOST:=0.0.0.0}"
 : "${PORT:=7770}"
-: "${CTX_SIZE:=602100}"
+: "${CTX_SIZE:=262144}"   # Qwen3.8 native context (env may raise for parallel≥2)
-: "${PARALLEL:=3}"
+: "${PARALLEL:=2}"
-: "${SPEC_TYPE:=draft-dflash}"
+: "${SPEC_TYPE:=draft-mtp}"
-: "${SPEC_DRAFT_N_MAX:=8}"
+: "${SPEC_DRAFT_N_MAX:=10}"
 : "${SPEC_NGRAM_MOD_N_MIN:=48}"
 : "${SPEC_NGRAM_MOD_N_MAX:=64}"
 : "${SPEC_NGRAM_MOD_N_MATCH:=24}"
@@ -53,16 +52,19 @@ if [[ ! -f "$MODEL" ]]; then
     exit 1
 fi
 
-# Draft-model flags apply only to spec types that use a draft file
+# Draft-model flag applies only when a separate draft file is provided.
-# (draft-dflash, draft-mtp, draft-eagle3). ngram-mod is draftless.
+# Qwen3.8 bundles MTP in-model; leave DRAFT_FILE empty to use bundled MTP.
 SPEC_ARGS=()
-if [[ "$SPEC_TYPE" == *draft-* ]]; then
+if [[ -n "$DRAFT_FILE" ]]; then
     if [[ ! -f "$DRAFT" ]]; then
         echo "missing required file: $DRAFT" >&2
         exit 1
     fi
     SPEC_ARGS+=(--spec-draft-model "$DRAFT")
 fi
+if [[ -n "${SPEC_DRAFT_P_MIN:-}" ]]; then
+    SPEC_ARGS+=(--spec-draft-p-min "$SPEC_DRAFT_P_MIN")
+fi
 
 exec "$LLAMA_SERVER" \
     --model "$MODEL" \