kingjones777/Apodex-1.1-mini-ROCmFP4-COHERENT-GGUF

Apodex-1.1-mini (COHERENT) — ROCmFP4 for AMD Strix Halo (gfx1151)

I built this 4-bit ROCmFP4 quantization of apodex/Apodex-1.1-mini — a Qwen3.5-derived hybrid MoE (40 layers of linear attention with full attention every fourth layer, 256 routed experts / 8 active, 262K context) — on my Strix Halo box for the ROCmFPX runtime.

The file

ftype	102 — Q4_0_ROCMFP4_COHERENT
size	19,846,167,808 bytes (18.48 GiB)
bpw	4.58
architecture	qwen3_5_moe (hybrid linear-attention + full-attention MoE)
tensors	733
context	262,144
experts	256 routed, 8 active per token + shared expert
token embedding	Q6_K
output.weight	Q6_K (protected)
sha256	f41f9ff7d61eebd345538bd6fbeb406e6405561538e7bf8c8dbba396cde97d5e

Type histogram, read from the finished file:

Q4_0_ROCMFP4 x430, F32 x301, Q6_K x2

What this build type is — and what it protects

Q4_0_ROCMFP4_COHERENT (ftype 102) quantizes the body to ROCmFP4 — my 4-bit UE4M3-scale layout — and deliberately protects the two tensors that hurt coherence most when crushed to 4-bit: the token embedding table and the LM head both stay at Q6_K. Because this model's head is untied, a 4-bit output.weight would degrade the logits of every single token; so I quantized with --output-tensor-type q6_K and confirmed the head landed at Q6_K by exact-name read-back.

tie_word_embeddings is false on Apodex, so output.weight is a real standalone tensor. I verified the head types by reading the finished file back by exact tensor name (token_embd.weight and output.weight — exact match, not substring).

Text-only, trunk-only — stated up front

• No vision tower. The upstream repo is multimodal; this GGUF carries the language model only. No mmproj is included.
• No MTP head. My converter's MTP merge path dies on this checkpoint's de-fused mtp.layers.0.mlp.experts.* tensors (KeyError: 'model.layers.0.mlp.experts.0.down_proj.weight' in the Qwen2Moe merge loop), so I converted with --no-mtp. The 40-layer trunk is complete (733 tensors); speculative MTP drafting is not available from this file. I do not publish what I have not verified.

How I built it

1. Manifest gate: pulled apodex/Apodex-1.1-mini file list from the HF API with ?blobs=true and recorded the real shard bytes (15 safetensors shards, 71,903,869,048 bytes total — never the index total_size).
2. Downloaded and byte-verified all 28 files against that manifest.
3. Converted with convert_hf_to_gguf.py from my rocmfpx-dspark-halo tree (4eca07e), --outtype bf16 --no-mtp → 733 tensors, 69,376,638,528 bytes. The checkpoint ships fused expert tensors (mlp.experts.gate_up_proj / mlp.experts.down_proj); this converter's Qwen2Moe base ingests the fused layout directly.
4. Quantized with the same tree's llama-quantize at 16 threads and --output-tensor-type q6_K. Dry-run estimate 18,916.30 MiB; the real file landed within ~11 MiB of it.

Measured on my box — partial-offload functional check, stated plainly

amd-halo: AMD Ryzen AI Max+ 395 (Strix Halo, gfx1151), ROCm 7.13.0, 125 GiB unified memory. At test time this box was serving 8 live llama-server seats holding ~107 GiB of unified memory, leaving me ~16 GiB. A full -ngl 999 --no-mmap load does not fit in that headroom, so I functionally checked with partial offload instead:

offload	17 / 40 layers on ROCm0, rest CPU-mmap (-ngl 17 -c 2048)
generation (server-reported)	1.183 t/s over 64 tokens (54.1 s)
prompt processing	19 tokens in 14.75 s
MemAvailable	16.5 GiB before load → 7.0 GiB after

Greedy, port 8497, -t 16, --jinja. These t/s numbers are limited by streaming the CPU-resident layers, not by the ROCm path — treat them as load-and-generate proof with the server's own timing, not as the speed you will get on an idle box. Full-offload throughput: not measured (would require freeing the seats — I don't touch my live seats).

Sample output (greedy, prompt "Explain in one clear sentence what granite rock is primarily made of."):

<think> Thinking Process: 1. Analyze the Request: * Task: Explain what granite rock is primarily made of. * Constraint: Use exactly one clear sentence. … (native thinking-mode generation)

⚠️ Stock llama.cpp will not load this file

Q4_0_ROCMFP4_COHERENT is a custom tensor format that exists only in the ROCmFPX fork of llama.cpp.

llama-server -m apodex-1.1-mini-Q4_0_ROCMFP4_COHERENT.gguf -dev ROCm0 -fa on -ngl 999 -c 8192   # on a box with the memory for it

Not measured

No benchmark sweeps, no context sweeps, no perplexity, no full-offload throughput — per my build discipline this is the 3-tier publish set and one functional check per tier.

Provenance & license

Converted and quantized from apodex/Apodex-1.1-mini (Apache 2.0). This quantized build is released under the same Apache 2.0 license. The ROCmFPX runtime is a third-party fork; its own terms apply to the runtime, not to these weights.

Corrected speed — full GPU offload

My first published number (1.183 tok/s) was measured with PARTIAL offload on a box already serving 8 models — that was my harness's fault, not the model's. Re-measured on an idle Ryzen AI MAX+ 395 (gfx1151, ROCm 7.2.4), full -ngl 999 offload (), 32K ctx:

63.30 tok/s (128-token generation, greedy).

All my quants of Apodex-1.1-mini

build	what it is	size	tok/s (full GPU offload)
STRIX_LEAN	my leaner 4-bit tier, Q6_K head — smallest of my 4-bit builds, the one most people want	17.46 GiB	64.87
COHERENT	my 4-bit ROCmFP4 tier with the Q6_K-protected head — the balance I run day to day	18.48 GiB	63.30
Q8_0	straight 8-bit ROCmFPX — highest fidelity I publish	33.36 GiB	45.05
Q8_0-AGENT	8-bit ROCmFPX with the agent-tuned tensor set — for tool-calling work where precision matters	33.90 GiB	32.54

All measured by me on a Ryzen AI MAX+ 395 (Strix Halo, gfx1151, ROCm 7.2.4) with the whole model on GPU (-ngl 999), 128-token greedy generation. A dash means I haven't measured that one yet — I won't put a number in a card I didn't measure.

Base model: Apodex/Apodex-1.1-mini