Qwen3.8-27B — Q3_0_ROCMFPX base + agent (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 in two arms with the ROCmFPX fork FP3 presets:

• base — Q3_0_ROCMFPX, 4.44 bpw effective, 14.125 GiB
• agent — Q3_0_ROCMFPX_AGENT, 5.72 bpw effective, 18.198 GiB

MTP layer included (blk.64 with nextn.* tensors, nextn_predict_layers = 1): serve with --spec-type draft-mtp to enable speculative decoding. Built 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).

⚠️ Both GGUFs use GGML type 104 (q3_0_rocmfpx) and are for the ROCmFPX fork of llama.cpp. They will not load in stock llama.cpp (invalid ggml type).

Honest verdict — read this first

On this hybrid architecture these presets are not a true 3-bit. The fork routing protects ALL attention tensors — including the 48 fused attn_qkv of the gated-deltanet layers — with K-quants, so the effective density is 4.44 bpw (base, 14.13 GiB) and 5.72 bpw (agent, 18.20 GiB). Neither arm beats the ROCmFP4 baseline where it matters: base is dominated by STRIX_LEAN on all three axes (worse ppl, −15.6% tg128, +2.2% size — see pugant/Qwen3.8-27B-MTP-Q4_0_ROCMFP4_STRIX_LEAN); agent does buy back ppl (+1.89%/+2.04% vs BF16, better than LEAN's +2.74%/+3.42%) but pays with 5.72 bpw (larger than the FULL preset) and −31.7% tg128 — worse quality-per-byte and quality-per-second than the FP4 presets. This repo is published as reproducible research data (preset comparison under matched conditions) and for the type-104 GGUFs themselves.

⚠️ Critical warnings — read before use

• Fork-specific tensor types (q3_0_rocmfpx, q6_0_rocmfpx, K-quant attention protection). Requires a ROCmFPX fork build with GGML type 104 support — see Runtime. Upstream charlie12345/ROCmFPX loads these files too (any recent build with the custom GGML types — HIP or the Vulkan-only build).
• AMD RDNA 3.5 (gfx1151 / Strix Halo) target. Tested on Radeon 8060S iGPU, not elsewhere.
• FP3 here is a memory-bandwidth play that did not materialize: as the verdict above says, the effective density lands at 4.44–5.72 bpw, not below the FP4 baseline.

Benchmarks

Hardware: AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151, 128 GB unified LPDDR5X), exclusive GPU, Vulkan RADV, 2026-08-23.

Perplexity method: llama-perplexity -c 512 -b 512 -fa on -ngl 999, wikitext-2-en 150k tokens + Italian technical corpus 51.5k tokens. BF16 baseline 6.6409 (en) / 11.7156 (it). The STRIX_LEAN control was reproduced within sigma on the same backend (Vulkan offset +0.34% / +0.49%).

tg/pp = llama-bench -p 512 -n 128 -fa 1 -ngl 999 -r 5, same session; tg/pp deltas vs LEAN in-session; quality deltas vs BF16 historical references, with the LEAN control within sigma.

Notes from the table:

• base is strictly dominated: larger than the baseline (14.125 vs 13.82 GiB), worse perplexity on both corpora, and slower on both axes (−15.6% tg128, −22.3% pp512).
• agent buys back quality (+1.89% / +2.04% vs BF16, the closest of the three to BF16) but pays with 5.72 bpw effective, 18.198 GiB and −31.7% tg128.
• Structural cause below: the 3-bit block type never reaches the attention mass on this hybrid architecture, so the density advantage is spent before it starts. The FP4 STRIX_LEAN preset remains the Pareto pick on gfx1151.

Routing histogram (post-quant verification)

866 tensors per arm, 0 routing errors, 0 NaN/Inf (post-quant verification over every tensor):

In the base arm the Q8_0 tensor is nextn.eh_proj and the two Q4_0_ROCMFP4_FAST tensors are the MTP draft FFNs — the same tensor mapping as the STRIX_LEAN build. The 360 F32 tensors are the small tensors left untouched. Both histograms sum to 866 with no unclassified tensor.

Usage (fork build)

llama-server -m Qwen3.8-27B-Q3_0_ROCMFPX-base.gguf \
  -ngl 999 -fa on --jinja -c 32768 \
  --host 0.0.0.0 --port 1234
# + MTP speculative decoding:
#   --spec-type draft-mtp --spec-draft-n-max 4 \
#   --spec-draft-p-min 0.75 --spec-draft-p-split 0.10

Swap -m for Qwen3.8-27B-Q3_0_ROCMFPX-agent.gguf to serve the agent arm.

Requires a ROCmFPX fork build with GGML type 104 support (recent Vulkan or HIP builds) — see Runtime.

Importance matrix

Same self-produced importance matrix as the STRIX_LEAN build, 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

• Presets: Q3_0_ROCMFPX (base arm) and Q3_0_ROCMFPX_AGENT (agent arm), quantized in the fork convert container.
• The "3.50 bpw nominal" caveat. The preset is nominally 3.50 bpw; the effective density on this architecture is 4.44 bpw (base) and 5.72 bpw (agent). Structural explanation: in the fork's quantize routing the attention branch (category_is_attn_v) covers the fused attn_qkv tensors of all 48 gated-deltanet layers and routes them to K-quants instead of the FP3 block type (Q5_K/Q4_K in base; Q6_K/Q5_K in agent — visible in the histogram above), and the use_more_bits imatrix boost lifts further tensors. With only the FFN mass left for the 3-bit type, the nominal density never materializes.
• Source: BF16 GGUF from unsloth/Qwen3.8-27B-GGUF, SHA-verified at quant time by the STRIX_LEAN pipeline (same source shards).
• imatrix: the self-produced matrix of the STRIX_LEAN pipeline, loaded with 496/496 entries (the LEAN coverage-gate count: all 64 non-MTP layers covered, 0 NaN/Inf).
• Base model: Qwen/Qwen3.8-27B (Apache-2.0). MTP layer included (blk.64, nextn_predict_layers = 1).

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 by the pipeline).
• charlie12345 and all the contributors of ROCmFPX — the llama.cpp fork that defines the ROCmFPX presets (the Q3_0_ROCMFPX family quantized here and the ROCmFP4/STRIX_LEAN baseline), 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.

---

Files in this repo

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.

Kept as a documented NO-GO (not a true 3-bit on this arch); if you want small, use the LEAN above.

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