frank-default-large: swap to Qwen3.8-27B-Uncensored Q6_K, q8_0/f16 KV

- MODEL_FILE -> Qwen3.8-27B-Uncensored-Q6_K.gguf (22.43 GB, no bundled
  MTP head; repo ships a standalone draft sidecar)
- SPEC_DRAFT_MODEL -> Qwen3.8-27B-Uncensored-draft-Q8_0.gguf (3.16 GB)
- CACHE_TYPE_K=q8_0, CACHE_TYPE_V=f16 (mismatched, FA retained via the
  GGML_CUDA_FA_ALL_QUANTS rebuild of head HIP + worker CUDA at 9b05354ec)
- Split kept at 0.53 for phase 1 (Q6_K frees ~4.7 GiB on the worker at
  this split); re-maximize after measuring draft placement
- KV math: K 1088 B/token + V 2048 B/token = 784 MiB per FA layer

scripts/frank-default-large.env

@@ -15,7 +15,7 @@
 # (HIP runtime collision, see the 2026-08-13 TEMP incident).
 
 MODEL_DIR=/home/dank/models
-MODEL_FILE=Qwen3.8-27B-UD-Q8_K_XL.gguf
+MODEL_FILE=Qwen3.8-27B-Uncensored-Q6_K.gguf
 HEAD_BIN=/home/dank/llama.cpp/build/bin/llama-server
 ALIAS=frank-default-large
 BIND_ADDR=0.0.0.0
@@ -25,15 +25,12 @@ LISTEN_PORT=7770
 RPC_SERVERS=192.168.0.235:50052
 
 # Model split fractions, one per device in order [RPC0, iGPU].
-# 0.53,0.47 = 34 of 65 blocks on the 3090 (blocks 0-33) - the maximum
+# 0.53,0.47 = 34 of 65 blocks on the 3090 (blocks 0-33). Phase-1 value
-# provably-safe split, measured 2026-08-17 (two failed attempts):
+# for the Q6_K swap (2026-08-17); the split will be re-maximized after
-#  - 0.50: 32 blocks, KV 8192 MiB (8 FA) -> healthy, 1477 MiB free.
+# measuring the draft-model placement on the worker. With the Q6_K model
-#  - 0.54: weights 35 blocks but KV jumps to 9216 MiB (9th FA block) ->
+# (layer pool 20334 MB) + q8_0/f16 KV (784 MiB per FA layer) the worker
-#    cudaMalloc OOM by ~740 MiB. The KV-per-device count rounds the
+# frees ~4.7 GiB vs the old model at this split; the 0.53 boundary was
-#    split fraction over 66 layers, independent of the weight boundary;
+# provably safe under the old model and is comfortably safe here.
-#    the flip to 9 FA sits at f ~= 0.538, just below the weight
-#    threshold for block 35. 0.53 keeps 8 FA under every rounding rule.
-# GGUF block scan: FA at 3,7,...,63; linear blocks 431 MB, FA 410 MB.
 TENSOR_SPLIT=0.53,0.47
 
 # Context pool. 262144 total = 1 parallel slot x 262144 per slot
@@ -41,25 +38,26 @@ TENSOR_SPLIT=0.53,0.47
 CTX_SIZE=262144
 PARALLEL=1
 
-# Speculative decoding. Matches production frank-default-large: MTP (bundled
+# Speculative decoding. MTP with the STANDALONE draft head (the Q6_K
-# head, n_max 4 per the published tuning curve) with the ngram-mod fallback.
+# model has no bundled MTP head; the repo ships the draft sidecar) + the
-# Set SPEC_NONE=1 for a raw baseline with no speculation.
+# ngram-mod fallback. Set SPEC_NONE=1 for a raw baseline.
 SPEC_NONE=0
 SPEC_DEFAULT=1
 MTP=1
 SPEC_DRAFT_N_MAX=4
-# Optional standalone MTP draft head (e.g. the Q4_0 ggml-org file). When set,
+# Standalone MTP draft head (Q8_0 MTP-only sidecar from the same repo).
-# it replaces the bundled in-model head via --spec-draft-model.
+SPEC_DRAFT_MODEL=/home/dank/models/Qwen3.8-27B-Uncensored-draft-Q8_0.gguf
-SPEC_DRAFT_MODEL=
 SPEC_NGRAM_MOD_N_MIN=48
 SPEC_NGRAM_MOD_N_MAX=64
 SPEC_NGRAM_MOD_N_MATCH=24
 
-# KV on GPU, unquantized (f16). 16 full-attention layers x 2 (K,V) x 4 heads
+# KV on GPU: K q8_0 + V f16 (mismatched, 16 full-attention layers x 4
-# x 256 dim x 2 bytes = 64 KiB/token -> 262144 tokens = 16.6 GiB total,
+# heads x 256 dim: K = 1088 B/token, V = 2048 B/token -> 3136 B/token ->
-# held in proportion to the full-attention layers each device holds.
+# 784 MiB per FA layer at 262144, 12.25 GiB total). Mismatched K/V with
-# Keep K == V so Flash Attention stays enabled.
+# Flash Attention requires the GGML_CUDA_FA_ALL_QUANTS rebuild (2026-08-17:
-CACHE_TYPE_K=f16
+# head HIP + worker CUDA at 9b05354ec). Do not revert the binaries without
+# also reverting this.
+CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=f16
 
 # Batch sizing: production-verified values (2026-08-14).