kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-GGUF

### ⚠️ STOCK llama.cpp WILL NOT LOAD THIS MODEL Q8_0_ROCMFPX is a ROCmFPX quant type — it exists only in `charlie12345/ROCmFPX`, not upstream llama.cpp. Ignore the auto-generated "Use this model" commands above. 📦 25.92 GiB, 8.28 bpw · ✅ tools 7/7 in BOTH thinking and non-thinking 🚀 25.07 tok/s with MTP on Ryzen AI MAX+ 395. ⚠️ The 4-bit build is still faster (38.32). ⚠️ Run it WITH the bundled MTP draft head — without it you get ~7.9 tok/s, a third of the speed.

Qwen3.8-27B — ROCmFPX 8-bit (Q8_0_ROCMFPX) GGUF

An 8-bit ROCmFPX quantization of Qwen3.8-27B for AMD gfx1151 (Ryzen AI MAX+ 395 / Strix Halo), built because 128 GB of unified memory makes 8-bit genuinely affordable on this hardware. Quantized from the 51.3 GiB BF16 GGUF — not requantized from a lower-bit build.

### ⚠️ If the MTP draft command crashes on your build The separate-model draft-mtp path has known bugs in the legacy charlie12345/ROCmFPX line — reported on Windows 11 / gfx1151 / HIP SDK 7.2 (five stacked bugs, ending in an h-row width mismatch in the draft's embedding buffer). Fix: build the official repo instead — no patches needed. ``bash git clone https://github.com/ROCmFPX/ROCmFPX.git ` The MTP path was reworked there (unified nembdout row widths, thnextn reset in llmgraphresult::reset(), ctx_other` wired centrally), which covers the whole chain. Reported and verified on that configuration: 24–31 tok/s, coherent output, tool calling working. Legacy-line patch: PR #109. Linux builds on the legacy line are not known to be affected.

Benchmarks — run this WITH the MTP draft head

Ryzen AI MAX+ 395 (gfx1151, 128 GB unified, ROCm 7.2.4), median of 3, warm-up discarded, idle box, shipped serving flags: --spec-type draft-mtp --model-draft mtp-Qwen3.8-27B-Q4_0.gguf --spec-draft-n-max 4 -ngl 999 -fa on -fit off.

⚠️ Without the draft head this model runs at ~7.9 tok/s — MTP is worth 3.2× here. The draft head (mtp-Qwen3.8-27B-Q4_0.gguf) is included in this repo; use it.

Choosing: the 4-bit build is ~1.53× faster and 12 GiB smaller. Take 8-bit for fidelity headroom, not throughput. And if you are running MTP, prefer the AGENT variant — it accepts more draft tokens (0.953 vs 0.911) and is 6.2% faster despite being marginally larger.

llama-server -m Qwen3.8-27B-Q8_0_ROCMFPX.gguf \
  --spec-type draft-mtp --model-draft mtp-Qwen3.8-27B-Q4_0.gguf \
  --spec-draft-ngl 99 --spec-draft-n-max 4 \
  -ngl 999 -fa on -fit off --jinja --ctx-size 32768

Verified

Per-tensor types (audited, 851 tensors)

output.weight Q8_0 · token_embd.weight Q8_0 · bulk TYPE_103 (ROCmFPX 8-bit layout).

🩹 Prompt caching with the MTP draft head — fixed

Reported by a user of this repo: with --spec-type draft-mtp loaded, llama-server disabled prefix caching entirely. Every agentic turn reprocessed the whole prompt. Reproduced here on an 8045-token stable prefix:

279× less prompt processing per turn, with MTP still drafting.

Root cause

The saved speculative state is the MTP boundary — the target model's pre-norm hidden row at the cached prompt's exact end position. Any partial-prefix reuse would leave it describing a position that no longer exists, so the server demanded an exact full-prefix match and otherwise reprocessed cold, erasing its own context checkpoints on the way.

The fix

patches/mtp-prompt-cache-fix.patch (4 files, applies to 2809dc5) captures the speculative boundary inside the context checkpoint (common_prompt_checkpoint::data_spec). create_checkpoint runs between decode batches — exactly where that boundary is valid — so exact state is saved and restored together with the KV, never rebuilt.

⛔ Two approaches were tried first and rejected: rebuilding the boundary from a zero-fill changed the model's output (deterministically, 3/3), and truncating the KV back to the reuse point is impossible here — the bounded rollback window is 4 tokens against the 333 a real turn needs. Exact state restore is the only shape that preserves output.

Verification

Independently gated 10/10: same prompt cold vs warm, temperature 0, byte-identical every run, with the cache genuinely engaged (cache_n=5101, not a vacuous pass). The output hash also matches the unpatched build, so behaviour is unchanged. Fails closed — an unreachable rollback logs reason=spec-checkpoint-missing and cold-reprocesses rather than guessing.

Related upstream

This is the same family as open llama.cpp issues #20225, #19794 and #24055 — checkpoints being invalidated on hybrid/recurrent models. This patch is not upstreamed; it is offered here as-is.

What was NOT measured

• —No perplexity run, and no quality A/B vs BF16 or the 4-bit build. We show 8-bit is slower; we have not demonstrated it is better. If you need proof that 8 bits buys accuracy here, that measurement does not yet exist.
• —An earlier revision of this card quoted 7.92 vs 13.84 tok/s. Those were measured without the MTP draft head and understated both builds; the table above supersedes them.
• —No long-context testing (model supports 131,072).
• —No coding/reasoning benchmark.

Base model licence inherited. Credit for the model goes to Qwen.