TL;DR 🎯 A 284B MoE in 103 GB (2.90 bpw) that we verified against the full-precision official API on identical protocols: 🟢 90.8% token-identical, 🟢 240/240 paired-QA parity, 🟢 deep-derivation parity (SymPy-checked), 🟢 0.0% confident-wrong-answer rate. Tuned end-to-end for AMD Strix Halo (Ryzen AI Max+ 395, gfx1151, 115 GB unified RAM) on mainline llama.cpp — no fork. Every flag below is a measurement, not a vibe. 📏

What's different from the original unsloth UD-IQ3_XXS? The three attention tensor families that dominate the per-token byte budget (attn_q_b, attn_output_a, attn_output_b — 51.9% of all bytes read per token, more than the 256 experts combined!) are requantized Q8_0 → Q6_K from the clean source. Measured effect: +23.8% generation on long-form output, and draft acceptance jumping 44% → 63% (less attention noise keeps the speculative drafter and the target in agreement). 🚀

📦 Install (Strix Halo / gfx1151)

bash

# 1️⃣ One-time: raise the unified-memory GPU ceiling (the single biggest win: ×2.8 !)
#    Default GTT = half your RAM. On a 128 GB box you want ~110 GiB:
echo 'GRUB_CMDLINE_LINUX_DEFAULT="$GRUB_CMDLINE_LINUX_DEFAULT amdgpu.gttsize=112640 ttm.pages_limit=28835840"' \
  | sudo tee /etc/default/grub.d/99-amdgpu-gtt.cfg
sudo update-grub && sudo reboot

# 2️⃣ Build llama.cpp (mainline!) for ROCm/HIP
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151 -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

# 3️⃣ Download this repo (model + the DSpark drafter from unsloth's repo)
hf download Kevletesteur/DeepSeek-V4-Flash-0731-StrixHalo-Verified-GGUF
hf download unsloth/DeepSeek-V4-Flash-0731-GGUF dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf
#    → requantize the drafter to Q3_K_S (8.5 GB) with llama-quantize, or use it as-is (10.9 GB)

# ✅ Verify integrity BEFORE first launch (we learned this the hard way):
sha256sum -c <(echo "538ec6932c50c6993f8a2cb18dc138e0fb5f6a7c2659e53a7566f1dc8ebb4344  DeepSeek-V4-Flash-0731-UD-IQ3_XXS-q6kattn.gguf")

⚙️ Launch (every flag measured — see "why" table below)

bash

GGML_CUDA_DISABLE_GRAPHS=1 LLAMA_NO_FUSE_HC_POST=1 \
./build/bin/llama-server \
  -m DeepSeek-V4-Flash-0731-UD-IQ3_XXS-q6kattn.gguf \
  --no-mmap -fa on -ctk q8_0 -ctv q8_0 \
  -ngl 99 -t 12 -b 2048 -ub 512 -c 262144 \
  --n-cpu-moe 16 \
  --model-draft DSpark-draft-Q3_K_S.gguf --spec-type draft-dspark \
  --spec-draft-n-max 2 --spec-draft-type-k q8_0 --spec-draft-type-v q8_0 \
  --temp 1.0 --top-p 0.95 --min-p 0.0

flag	why (measured)
`--no-mmap`	🔴 mandatory on unified memory — with mmap the model exists twice (GPU buffer + page cache): 51 min of thrashing vs 40 s load
`GGML_CUDA_DISABLE_GRAPHS=1`	+3.0% — a 256-expert MoE re-routes every token; HIP graph re-instantiation costs more than launch amortization saves
`LLAMA_NO_FUSE_HC_POST=1`	+1.5% — this specific micro-fusion loses on gfx1151
`-ctk/-ctv q8_0`	on THIS quant, f16 KV is −20.6% (on the previous quant it was +21%… settings do not transpose — re-measure in your config)
`--n-cpu-moe 16`	frees 33 GB of GTT for 0.1% cost — and prefill improves 2.4× (the constraint is memory pressure, not compute)
`--spec-draft-n-max 2`	⚠️ the optimum depends on context depth: 3 wins at short context, 2 wins from ~16k on (+3-4%). Long-context agents → 2. There is a hard cliff at n≥4 (RDNA3 MMVQ `max_batch=4` for these expert types)
`--temp 1.0 --top-p 0.95`	the values sourced from DeepSeek + unsloth (lower temp is faster via draft acceptance, but unsourced)
speed you should see	~20-27 t/s short ctx (task-dependent!), ~14-16 t/s @16k, ~11 t/s @150k · prefill ~130 t/s

🧨 Errors to avoid (each one cost us hours-to-days)

🥇 Don't evaluate quality through raw /completion. The auto-injected <think> can close instantly at temp 0 → the model answers reflexively with 10-char reasoning traces and looks broken. We "discovered" a factual-recall deficit that was 100% this artifact. Use /v1/chat/completions for anything quality-related.
🐉 Never use a 2-bit drafter. Draft hit-rate collapses to 9-15% and speculation then loses 40% vs no drafter at all. Q3_K_S (3.44 bpw) minimum. "A degraded drafter only costs acceptance" — true, and acceptance IS the entire gain.
⛓️ Never combine draft-dspark with ngram modes. Syntactically accepted, destroys DSpark (+101% → +13% on tool calls). Ngram modes alone lose up to −46%.
📏 Don't transpose optima across contexts, quants or models. Batch size swept at 32k and deployed at 256k cost us ×3.7. KV f16 flipped from +21% to −20.6% between two quants.
🧊 Wait for GTT release after killing a server before starting the next (the driver frees buffers after the process dies) — poll mem_info_gtt_used, don't sleep 5.
🎲 Speculative decoding here is quality-neutral but not bit-exact (batched verification reorders float reductions; near-tie argmax can flip). For token-paired benchmarks, bench with the drafter off.
💾 Use --slot-save-path + the save/restore API around restarts: 25 ms instead of ~6 min of re-prefill for a 30k-token session (×14,000).

🔬 How this was verified (summary)

Paired protocols, identical serving path both sides, instruments validated on synthetic cases before any model run (they caught 9 real bugs — see the harness repo):

probe	result vs full-precision API
🧬 token-level teacher forcing (17,929 positions, 4 domains × EN/FR)	90.8% top-1 agreement · original's token outside local top-20: 0.04%
🧮 240-item exact-answer bank (McNemar, 6 families)	240/240 = 240/240, zero discordance (gap bounded ≲1.2%)
📐 60 multi-step derivations, SymPy oracle	parity
🛠️ structured tool-call sequences	25/25 both sides
🚨 confident-wrong-answers P(answers ∧ wrong)	0.0%
🇫🇷 French	more token-faithful than English (88.1% vs 83.0% on physics)

Also measured: Q5_K attention has better wikitext perplexity and worse real behavior (failed temp-0 probes, one repetition-loop) — our third demonstration that PPL ≠ quality.

🔁 Reproduce

recipe/requant_attention.sh + recipe/tensor_types_q6kattn.txt rebuild this file from unsloth's UD-IQ3_XXS shards in ~15 min (129 tensors requantized, 1,199 copied bit-identical, every tensor explicitly pinned — unpinned tensors silently fall into mixture heuristics).

🧪 Full methodology, harness & negative results

👉 Kevletesteur/quant-vs-api-parity-harness — the complete benchmark toolkit (gates, teacher forcing, SymPy oracles, calibration), plus NEGATIVE_RESULTS.md: what did not work, with numbers — three heavy engineering efforts abandoned on measurement, and the 9 instrument traps we fell into so you don't have to. 🪤

📜 Provenance & integrity

deepseek-ai/DeepSeek-V4-Flash-0731 (MIT) → unsloth/DeepSeek-V4-Flash-0731-GGUF (UD-IQ3_XXS, imatrix) → this repo (attention families requantized; experts untouched).

sha256  538ec6932c50c6993f8a2cb18dc138e0fb5f6a7c2659e53a7566f1dc8ebb4344
size    103,159,533,696 bytes

Benchmarked on: AMD Ryzen AI Max+ 395 (gfx1151/RDNA 3.5), ROCm 7.1, 115.3 GiB unified, llama.cpp mainline HIP. Measured, documented, and calibrated — enjoy. 🚀