Qwen3.8-27B — Q4_0_ROCMFP4_STRIX_LEAN (Strix Halo / gfx1151)

TL;DR

Qwen3.8-27B — the dense hybrid-attention Qwen release (48 gated-delta-net layers + 16 full-attention layers, full_attention_interval = 4, native 262K context, qwen35 GGUF arch) — quantized to Q4_0_ROCMFP4_STRIX_LEAN (4.34 BPW effective, 13.82 GiB). MTP layer included (blk.64 with nextn.* tensors, nextn_predict_layers = 1): serve with --spec-type draft-mtp to enable speculative decoding. Tuned for AMD Strix Halo (gfx1151) on the ROCmFPX fork family — we serve and benchmark these files on our lab runtime (full source: pugant/strix-nebulosa; upstream: charlie12345/ROCmFPX).

⚠️ This GGUF is for the ROCmFPX fork of llama.cpp. It will not load in stock llama.cpp (invalid ggml type).

⚠️ Critical warnings — read before use

• Fork-specific tensor types (q4_0_rocmfp4, K/V protection, Q5_K token embeddings). Requires a ROCmFPX fork build — see Runtime. Upstream charlie12345/ROCmFPX loads these files too (any build with the custom GGML types — HIP or the Vulkan-only build, both ship the deltanet kernels: gated_delta_net.comp, ssm_scan.comp, ssm_conv.comp).
• AMD RDNA 3.5 (gfx1151 / Strix Halo) target. Tested on Radeon 8060S iGPU, not elsewhere.
• FP4 is a memory-bandwidth play on this class of hardware (RDNA 3.5 has no FP4 silicon): smaller weights → fewer bytes per token on the 270 GB/s unified LPDDR5X.

Benchmarks

Hardware: AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151, 128 GB unified LPDDR5X), exclusive GPU window (no other services), 2026-08-15. Method: plain = llama-bench -ngl 999 -fa on -p 512 -n 128; MTP = llama-server --spec-type draft-mtp --spec-draft-n-max 2 --spec-draft-p-min 0.75 --spec-draft-p-split 0.10 + server timing (timings.predicted_per_second), 2 runs after warm-up.

MTP n-max sweep (this file, p_min 0.75, 2 runs per prompt, same window)

Per-position acceptance: on prose, positions 3+ are rarely accepted (0.21/0.09 at n-4) → prose plateaus at ~18–19 tok/s for any n-max. On deterministic/structured content acceptance stays high deep into the draft: n-4 best task (0.98, 0.93, 0.91, 0.86), n-6 (0.99, 0.90, 0.86, 0.83, 0.82, 0.80) — mean accepted length 6.2 — which is where 43.7 tok/s (3.2× plain decode) comes from. Recommendation: n-max 4 as default, n-max 6 for code/structured/RAG workloads. (Unlike the Qwen3.6-35B-A3B base where n-4 was dominated — optimal n-max is model-specific.)

Head-to-head vs Qwen3.6-27B (same preset, same window, same method)

Qwen3.8-27B wins the MTP match everywhere: +25–35% deterministic / +8–11% prose on ROCm, +20% deterministic on Vulkan, prose tied. Plain decode is identical and bandwidth-bound.

Notes:

• Plain decode of both dense 27B models is bandwidth-bound and identical (~13.8 tok/s on ROCm): the deltanet/attention mix does not change the ceiling; MTP is the speed lever (as with Qwen3.6-27B).
• On dense 27B models ROCm beats Vulkan on decode (13.8 vs 9.0 tok/s, +53%) — the opposite of the fp4 MoE-35B class where the Vulkan fork wins tg (81.6 vs 71.2). Pick the backend by model class: dense → ROCm, MoE fp4 → Vulkan fork. pp512: ROCm +11%.
• MTP acceptance (n-max 2, p_min 0.75): 0.81–0.99, mean accepted length 2.3–2.9 tokens, per-position (pos-1 0.85–0.99, pos-2 0.39–0.95) — the pos-2 draft of Qwen3.8 is unusually strong on predictable content (0.93–0.95), which is where the deterministic-content win comes from.
• MTP speedup over plain (ROCm): ~2.0× deterministic, ~1.35× prose.

Usage (fork build)

llama-server -m Qwen3.8-27B-Q4_0_ROCMFP4_STRIX_LEAN.gguf \
  -ngl 999 -fa on --jinja -c 32768 \
  --host 0.0.0.0 --port 1234
# + MTP speculative decoding (n-max 4 balanced; 6 for structured/code):
#   --spec-type draft-mtp --spec-draft-ngl all --spec-draft-n-max 4 \
#   --spec-draft-p-min 0.75 --spec-draft-p-split 0.10

DFlash2 block-diffusion drafter (optional, our lab build)

Besides the built-in MTP layer, this model works with the external DFlash2 drafter (incoai/Qwen3.8-27B-DFlash2-GGUF, 1.9B, Q4_K_M ≈ 1.1 GB) via our lab build (full source included under rocmfpx/) — port of upstream PR #27342; usage notes in rocmfpx/docs/dflash2.md.

llama-server -m Qwen3.8-27B-Q4_0_ROCMFP4_STRIX_LEAN.gguf \
  -ngl 999 -fa on --jinja -c 16384 \
  --spec-type draft-dflash \
  --spec-draft-model Qwen3.8-27B-DFlash2-Q4_K_M.gguf \
  --spec-draft-ngl all --spec-draft-n-max 7 --spec-draft-p-min 0.75

Route by workload (Vulkan RADV, temp 0, single stream, p_min 0.75):

n-max 5 is the best single-drafter compromise (det 52.2/39.5, prose 17.5/15.9). The DFlash2 acceptance stays ≥0.5 per position up to position 7 on structured content, where MTP collapses after position 1; on free prose the acceptance is on par with MTP but the block-diffusion round has a fixed cost, so MTP wins.

Importance matrix

The quant was built with an importance matrix calibrated on an interleaved corpus (agentic-coding traces from ProCreations/grug-think-v3-10k, Italian prose, real code). The matrix is published for reuse: pugant/Qwen3.8-27B-imatrix (works with any llama.cpp llama-quantize --imatrix).

Quantization details

• Type: Q4_0_ROCMFP4_STRIX_LEAN — fork preset (4.38 BPW nominal): FP4 single-scale blocks for most matrices, attention K/V protected, Q5_K token embeddings. Effective 4.34 BPW on this architecture (13.82 GiB, 866 tensors, MTP included: nextn.eh_proj → q8_0, draft FFNs → q4_0_rocmfp4_fast).
• Source: BF16 GGUF from unsloth/Qwen3.8-27B-GGUF (2 shards, SHA256-verified against the HF LFS OIDs before quantization).
• imatrix: self-produced. Unsloth's calibration matrix for Qwen3.8 is not published (referenced in their quant metadata, absent from the repo). We generated our own with llama-imatrix (fork, GPU offload, 256 chunks × 512 ctx, ~10 min) over a 1.7 MB composite, round-robin-interleaved corpus: ~55% agentic-coding transcripts (tool-calls + code), ~28% Italian technical prose, ~17% real source code.
• imatrix coverage gate (passed): 496 tensor entries — exactly the same count unsloth's imatrix references for this model (metadata quantize.imatrix.entries_count = 496). All 64 non-MTP layers covered (attention + gated-deltanet SSM tensors), 0 NaN/Inf, all entry shapes consistent. By design no imatrix can cover: MTP layer blk.64 (not executed in a plain forward pass), ssm_conv1d (no observer hooks), token_embd/output — these take the unweighted path, identical to unsloth's behaviour on Qwen3.6.
• Base model: Qwen/Qwen3.8-27B (Apache-2.0). The unsloth GGUFs are text-only (no vision tensors); the vision tower is available separately (mmproj-F16.gguf in their repo, untested here).

Acknowledgements

• Alibaba / Qwen team for the base model Qwen/Qwen3.8-27B (Apache-2.0) and for training and open-sourcing the MTP layer that makes speculative decoding possible on this class of dense models.
• Unsloth for the BF16 GGUF release used as quantization source (SHA-verified) and for their calibration-matrix practice, which we replicated with a self-produced imatrix after finding theirs unpublished for this model.
• charlie12345 and all the contributors of ROCmFPX — the llama.cpp fork that defines the ROCmFP4 family, the STRIX_LEAN preset, MTP speculative serving, and the Vulkan kernels for gated-delta-net / SSM ops that this hybrid architecture requires.
• kyuz0 (Donato Capitella) for the amd-strix-halo-toolboxes — the Dockerfiles (ROCm and Vulkan fork builds) used for quantization and serving in this work.
• The Strix Halo / gfx1151 community (strix-halo-guide, RADV bench threads) whose cross-backend benchmarking methodology informed our measurement protocol.

License

Apache-2.0 (inherited from the base model). See LICENSE.

Runtime

Benchmark environment: one bare-metal AMD Strix Halo (Ryzen AI MAX+ 395, 128 GB) — full dated configuration and measurement policy: BARE-METAL.md

Requires a ROCmFPX fork build (custom tensor types — stock llama.cpp refuses the file). Recommended: our lab build (pugant/strix-nebulosa, main) — reasoning budget and persistent prompt cache on every model; drafter features where the model ships one: see the engine section of its README.

The canonical model of the lab: dual-drafter routing was measured here (agentic +19%); MTP layer included.

Everything here is experimental and provided as-is, at your own risk.