vmlinux/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-ROCmFPX-GGUF

Qwen3.6 27B Architect Polaris2 Fable B F451 — MTP ROCmFPX GGUF

Text-only ROCmFPX/ROCmFP4 GGUF builds of nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451, with the checkpoint's embedded one-layer MTP draft model preserved.

These are experimental AMD Strix Halo builds. The custom ROCmFPX tensor types require charlie12345/ROCmFPX; they will not load in an unmodified stock llama.cpp build.

Files

File	Size	Recipe	Intended use
Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-ROCmFP8.gguf	28.19 GB	Q8_0_ROCMFPX, no imatrix	Recommended quality-first build
Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-iMatrix-ROCmFP4.gguf	16.88 GB	Q4_0_ROCMFP4_STRIX, F16 embedding, Q6_K output, narrative/general imatrix	Fast build
Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-iMatrix-ROCmFP8_QUALITY_Q6.gguf	25.28 GB	Q6_0_ROCMFPX_AGENT, Q8 protection for sensitive tensors, narrative/general imatrix	Experimental quality-biased Q6 build

All three files retain 64 main transformer blocks plus the embedded blk.64 MTP block. The source checkpoint's vision tensors were intentionally omitted; these GGUFs are for text generation.

Checksums

179ec314e293cbce80620ecc5eeba75d8566d9d3fa1aab3169dffb524f530187  Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-ROCmFP8.gguf
0acfe339d09a9f0864312eaac91ad3145cc30afc3ad9044c735c9224ae1b6fb4  Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-iMatrix-ROCmFP4.gguf
2255f94eaa4bdb856b799eff7198a1d0e56ded53429c1c15523e8af441d9819a  Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-iMatrix-ROCmFP8_QUALITY_Q6.gguf

Working ROCmFP8 router preset

This is the current text-only llama.cpp router preset used on a 128 GB Strix Halo system. It gives one slot the full native 262,144-token context, uses F16 KV, enables cross-turn checkpoint reuse, and runs the embedded MTP head with nmax=5.

The chat template is froggeric's chat_template.jinja v21.3. Download it separately and replace the example paths below.

[Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-ROCmFP8]
model = /path/to/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-MTP-ROCmFP8.gguf
chat-template-file = /path/to/chat_template.jinja
chat-template-kwargs = {"preserve_thinking": true}
ctx-size = 262144
parallel = 1
flash-attn = on
batch-size = 2048
ubatch-size = 256
threads = 16
threads-batch = 16
cache-type-k = f16
cache-type-v = f16
checkpoint-every-n-tokens = 256
cache-reuse = 256
cache-ram = 65536
mmap = 0
metrics = 1
spec-type = draft-mtp
spec-draft-device = Vulkan0
spec-draft-ngl = all
spec-draft-n-max = 5
spec-draft-p-min = 0.0
spec-draft-p-split = 0.10
reasoning = on
reasoning-format = deepseek
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.0
repeat-penalty = 1.0

cache-ram = 65536 and F16 KV are intentionally memory-hungry settings for a 128 GB unified-memory machine. Reduce the reuse-cache ceiling and/or use Q8 KV on smaller systems. Temperature 0.6 is the precise/coding profile; 1.0 is a useful starting point for general or creative work.

Validation notes

• Native 262,144-token context allocation passed for every artifact with both Q8/Q8 and F16/F16 KV.
• ROCmFP8 passed a deterministic 131,072-token planted-fact recall probe, 4/4 facts recovered.
• ROCmFP8 measured 12–14 tok/s in the original controlled MTP tests. With the router profile above, a short routed smoke measured 21.9 tok/s and a 4.46-token mean accepted MTP draft length.
• ROCmFP4 measured approximately 21–23 tok/s in short MTP tests on this host.
• These are directional single-host measurements, not benchmark medians.

Host: AMD Strix Halo / RADV GFX1151, 128 GB unified memory. Runtime: ROCmFPX commit 61f2f2d7bc4955e9bca821095ef69125837133b5 (llama.cpp build 162). The build checkout also contained a pre-existing MoE-only local diff (1783321e5a89eadc541ce2905874e5ed1172223204f257b5e569278c4f34004f); this dense 27B model has no expert tensors affected by that code path.

Build summary

• Source revision: nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451@d92178a
• One checked BF16 text+MTP GGUF was converted from the pinned source.
• ROCmFP8 was quantized directly from BF16 without an imatrix.
• ROCmFP4 and ROCmFP8_QUALITY_Q6 use the same narrative/general importance matrix: 511 × 2,048-token chunks, 496/496 quantizable tensor coverage. The calibration corpus used public-domain Project Gutenberg texts with whole-document holdouts.

Upstream and attribution

• Source merge and model work: NightMedia — Qwen3.6-27B Architect Polaris2 Fable B F451
• Base architecture: Qwen — Qwen3.6-27B
• Custom quantization/runtime: charlie12345/ROCmFPX
• Upstream llama.cpp: ggml-org/llama.cpp
• Runtime and ROCmFP4 recipe reference: Plunderstruck — Qwen3.6-27B MTP ROCmFP4 GGUF
• Qwen3.6 chat template: froggeric/Qwen-Fixed-Chat-Templates
• Public-domain calibration texts: Project Gutenberg

The model weights and this derivative packaging retain the upstream Apache-2.0 license. Please consult the linked source model card for its full merge lineage and intended-use notes.