Tiel-Coder-35B-A3B — ROCmFP4 FAST GGUF — AMD Ryzen AI Max+ 395 / gfx1151

⚠️ Read this before comparing any number here to a discrete GPU

Every measurement is from an AMD Ryzen AI MAX+ 395 "Strix Halo" — an integrated GPU with unified memory. No discrete VRAM: the Radeon 8060S addresses system RAM through the GTT aperture, shared with the CPU. Capacity is huge; bandwidth (~215 GB/s measured) is the trade.

Requantized from the excellent peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF — the Tiel-Coder fine-tune and the original GGUF conversion are their work, not mine. Go star their repo. The underlying base is ornith-ai/Ornith-1.5-35B-A3B.

FAST is the smallest tier: attention, experts and token embeddings all ROCmFP4, with only the output head lifted to Q6_K. general.file_type = 103 (Q4_0_ROCMFP4_FAST). 17.37 GiB.

⚠️ Honest provenance: this is a requantization of a Q8, not a BF16 build

Their repo publishes GGUFs, not unquantized weights, so this was built from their UD-Q8_K_XL (35.81 GiB) using --allow-requantize — quantization applied on top of quantization. That is lossier than quantizing from BF16/F32, and you should weigh it against a Q4 built from full-precision weights. I used their highest available tier as the source to keep the loss as small as this path allows. Flagging it because you cannot tell from the file.

What's inside

output.weight is never left at 4-bit: every sampled token passes through the lm head, so its error lands directly in the argmax. Verified by exact name — output.weight is a substring of attn_output.weight, so a loose check reports success on a 4-bit head.

Building a runtime that loads these files

qwen35moe and the ROCmFP4 tensor types are both already in the public ROCmFPX fork — no patch needed:

git clone https://github.com/charlie12345/ROCmFPX.git
cd ROCmFPX && git checkout d3ca537
cmake -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-server llama-quantize -j$(nproc)

Verified: a binary built from that clean clone loads these files and generates. Stock upstream llama.cpp will not work — it has no ROCmFP4 tensor types.

Measured — Ryzen AI MAX+ 395, gfx1151, ROCm 7.2.4, full offload

• prompt processing: 1174 tok/s — 1235.8 / 1182.7 / 1173.9 / 1174.0 / 1172.1
• generation: 60.28 tok/s — 61.59 / 60.30 / 60.17 / 60.28 / 60.10
• GTT resident: 18.0 GiB

Method: one fixed 4,000-token prompt from a real source corpus, reused across samples with cache_prompt: false; run 1 discarded as warm-up; median of the 4 settled samples.

vs the Q8_K_XL it was built from

Same box, same method, same prompt:

Half the footprint and faster on both axes — the trade is the requantization loss described above. Quality was not benchmarked; if you need maximum fidelity, use their Q8_K_XL.

Usage

llama-server \
  --model Tiel-Coder-35B-A3B-Q4_0-ROCmFP4-FAST.gguf \
  --host 127.0.0.1 --port 8080 \
  --n-gpu-layers 999 --flash-attn on --fit off \
  --ctx-size 32768 --threads 16 --jinja

Model supports up to 262,144 context. A vision tower (mmproj) is available in the source repo.

Acknowledgements

peculiar-ragdoll — the Tiel-Coder fine-tune and the GGUF conversion this is built from. Their work; I only requantized it.

ornith-ai — Ornith-1.5-35B-A3B, the underlying base model.

ROCmFPX — defines the ROCmFP4 tensor formats and carries the qwen35moe support used here.

llama.cpp — ggml-org and contributors.

AMD ROCm — the compute platform targeted (ROCm 7.2.4, gfx1151).