Pivot to llama.cpp with DFlash speculative decoding on ROCm

- Abandon vLLM/FP8 (Triton compilation hangs, ~0.5 tok/s on Strix Halo)
- llama.cpp v10133 with ROCm0 offload, DFlash draft model
- Main: unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 (~22GB)
- Draft: Alittlehammmer/Qwen3.6-27B-DFlash-GGUF-llama.cpp:Q8_0 (~1.8GB)
- 3 parallel streams, 131072 context/stream
- systemd service with minimal sandboxing (ROCm requires /dev/kfd access)
- Baseline: 7.62 tok/s (threshold 5.0 tok/s)

scripts/serve-llamacpp-dflash.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Serve Qwen3.6-27B with DFlash + MTP speculative decoding via llama.cpp on AMD Strix Halo (gfx1151)
+# Uses llama.cpp built with ROCm support
+# 3 parallel streams, 131072 context each
+set -euo pipefail
+
+SERVER_BIN="/home/dank/llama.cpp/build/bin/llama-server"
+PORT="${LLAMA_PORT:-8080}"
+N_PARALLEL="${LLAMA_N_PARALLEL:-3}"
+CTX_SIZE="${LLAMA_CTX_SIZE:-131072}"
+BATCH_SIZE="${LLAMA_BATCH_SIZE:-4096}"
+DRAFT_N_MAX="${LLAMA_DRAFT_N_MAX:-16}"
+
+# Models via Hugging Face
+TARGET_REPO="unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0"
+DRAFT_REPO="Alittlehammmer/Qwen3.6-27B-DFlash-GGUF-llama.cpp:Q8_0"
+
+echo "Starting llama.cpp server..."
+echo "  Target model: ${TARGET_REPO}"
+echo "  Draft model:  ${DRAFT_REPO}"
+echo "  Parallel streams: ${N_PARALLEL}"
+echo "  Context/stream: ${CTX_SIZE}"
+echo "  Draft tokens: ${DRAFT_N_MAX}"
+echo ""
+
+exec "${SERVER_BIN}" \
+    --hf-repo "${TARGET_REPO}" \
+    --spec-draft-hf "${DRAFT_REPO}" \
+    --spec-draft-n-max "${DRAFT_N_MAX}" \
+    --device ROCm0 \
+    --spec-draft-device ROCm0 \
+    --gpu-layers all \
+    --fit on \
+    --fit-target 512 \
+    --parallel "${N_PARALLEL}" \
+    --ctx-size "${CTX_SIZE}" \
+    --batch-size "${BATCH_SIZE}" \
+    --port "${PORT}" \
+    --host "0.0.0.0" \
+    --threads 16 \
+    --load-mode mlock