dankclaw: switch frank-default-large spec to DSpark sidecar draft (magnitudedev Qwen3.8-27B-DSpark Q8_0, n_max 7)

scripts/frank-default-large.env

@@ -12,17 +12,21 @@ PORT=7770
 # ~70GB headroom, so 2x262k quantized KV fits. Verify at startup.
 CTX_SIZE=524288
 PARALLEL=2
-# Spec: 3-layer chain, chosen by the user 2026-08-14 (aesthetic/short-term
+# Spec: DSpark sidecar draft (magnitudedev/Qwen3.8-27B-DSpark-GGUF Q8_0,
-# experiment): draft-mtp first, then ngram-mod, then ngram-map-k4v as
+# 1.36B, block_size 7, Markov confidence head).  Switched from bundled-MTP
-# fallbacks.  On real prose this benches ~equal to draft-mtp alone
+# 2026-08-15 at the user's request.  DSpark is the semi-autoregressive
-# (10.4 vs ~10 t/s @16k) - the ngram layers only pay off on self-similar text.
+# extension of DFlash (DeepSeek, arXiv 2607.05147): parallel diffusion
-# Bundled MTP drafter; llama.cpp auto-detects it (b10411+).  DRAFT_FILE empty.
+# backbone + sequential Markov head + confidence-scheduled verification.
-# n_max=5 is the realistic-text peak (n4 12.5, n6 10.6, n10 8.6).
+# n_max clamps to the trained block_size 7 (server warning if set higher).
-# Occasional slow rounds possible when fallbacks misfire; revert to
+# DSpark extends DFlash (which had a shared-draft-context premature-EOS bug)
-# draft-mtp (alone) if the assistant reports stalls.
+# but adds per-sequence Markov/confidence machinery; parallel safety on this
-SPEC_TYPE=draft-mtp,ngram-mod,ngram-map-k4v
+# box is still pending the two-16k-prompt test 2026-08-15. Revert to
-DRAFT_FILE=
+# bundled-MTP or set PARALLEL=1 if the parallel test fails.
-SPEC_DRAFT_N_MAX=5
+# DRAFT_FILE names the sidecar GGUF in MODEL_DIR (downloaded 2026-08-15,
+# sha256 b007a76b2ce57c1a3ecf36046543aebf60a763861624647d653f16d336c781d2).
+SPEC_TYPE=draft-dspark
+DRAFT_FILE=Qwen3.8-27B-DSpark-Q8_0.gguf
+SPEC_DRAFT_N_MAX=7
 SPEC_DRAFT_P_MIN=
 SPEC_NGRAM_MOD_N_MIN=48
 SPEC_NGRAM_MOD_N_MAX=64

scripts/frank-default-large.service

@@ -1,5 +1,5 @@
 [Unit]
-Description=llama.cpp Qwen3.8-27B UD-Q8_K_XL MTP+ngram-mod+ngram-map-k4v (frank-default-large)
+Description=llama.cpp Qwen3.8-27B UD-Q8_K_XL + DSpark sidecar draft (frank-default-large)
 After=network.target
 
 [Service]

scripts/serve-llama-frank-large.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
-# Serve unsloth Qwen3.8-27B UD-Q8_K_XL on dankclaw with bundled-MTP
+# Serve unsloth Qwen3.8-27B UD-Q8_K_XL on dankclaw with a DSpark
-# speculative decoding (draft-mtp). ROCm backend.
+# sidecar draft model (draft-dspark). ROCm backend.
 # Runtime: llama.cpp llama-server (>= b10411).
 #
 # Tunables come from frank-default-large.env in this directory unless the
@@ -19,7 +19,7 @@ fi
 # --- defaults (used when the env file does not set them) -----------------
 : "${MODEL_DIR:=/home/dank/models}"
 : "${MODEL_FILE:=Qwen3.8-27B-UD-Q8_K_XL.gguf}"
-: "${DRAFT_FILE:=}"        # empty: Qwen3.8 bundles its MTP drafter in-model
+: "${DRAFT_FILE:=}"        # set: DSpark/DFlash/MTP/... sidecar GGUF in MODEL_DIR
 : "${ALIAS:=frank-default-large}"
 : "${HOST:=0.0.0.0}"
 : "${PORT:=7770}"
@@ -53,7 +53,8 @@ if [[ ! -f "$MODEL" ]]; then
 fi
 
 # Draft-model flag applies only when a separate draft file is provided.
-# Qwen3.8 bundles MTP in-model; leave DRAFT_FILE empty to use bundled MTP.
+# Qwen3.8-UD bundles MTP in-model; leave DRAFT_FILE empty for bundled MTP,
+# or set it to a DSpark/DFlash sidecar GGUF for a dedicated draft.
 SPEC_ARGS=()
 if [[ -n "$DRAFT_FILE" ]]; then
     if [[ ! -f "$DRAFT" ]]; then