GLM-4.7-Flash ROCmFP4 STRIX (GGUF) — AMD Ryzen AI Max+ 395 / Strix Halo / gfx1151
First public ROCmFP4 quant of the full GLM-4.7-Flash weights for AMD Ryzen AI Max+ 395
(gfx1151 / Radeon 8060S).
The existing ROCm-flavoured GLM-4.7 quants on the Hub are all of the pruned
REAP-23B-A3B variant. This is the full 30B model, not a pruned one.
⚠️
Not compatible with upstream llama.cpp. Requires the
charlie12345/ROCmFPX fork built with HIP + ROCmFP4 kernels.
Files
| File | Size | Notes |
|---|
GLM-4.7-Flash-Q4_0_ROCMFP4_STRIX.gguf | 14.93 GiB | 4.28 BPW (quantize report) |
Base model:
zai-org/GLM-4.7-Flash
BF16 source:
unsloth/GLM-4.7-Flash-GGUF
BF16/ 2 shards, 55.79 GiB total (16.00 BPW)
Hardware / stack (validated)
- Ryzen AI Max+ 395, gfx1151, 128 GB unified
- ROCm 7.2.4
- Fork:
charlie12345/ROCmFPX @ b41ce12
Build recipe
1llama-quantize \
2 GLM-4.7-Flash-BF16-00001-of-00002.gguf \
3 GLM-4.7-Flash-Q4_0_ROCMFP4_STRIX.gguf \
4 Q4_0_ROCMFP4_STRIX
Q4_0_ROCMFP4_STRIX (type 105) is the Strix Halo attention-K/V quality recipe.
Dry-run predicted 15278.74 MiB @ 4.28 BPW and the output matched exactly.
Serving
1llama-server --host 127.0.0.1 --port 8080 \
2 --model GLM-4.7-Flash-Q4_0_ROCMFP4_STRIX.gguf \
3 -dev ROCm0 -ngl 999 -fa on --no-mmap \
4 --ctx-size 65536 --parallel 1 -b 2048 -ub 1024 -t 16 --poll 50 --jinja \
5 --reasoning-format deepseek --chat-template-kwargs '{"enable_thinking":false}'
⚠️ --chat-template-kwargs '{"enable_thinking":false}' matters. With thinking on, reasoning
consumes the token budget and answers get truncated at ordinary max_tokens values. Also use the
chat endpoint — raw /completion with a bare instruction makes this model emit repetition loops.
Measured
Against the UD-Q4_K_XL GGUF of the same model on the same machine, chat endpoint, thinking off:
| this quant | UD-Q4_K_XL |
|---|
| size | 14.93 GiB | 16.32 GiB |
| quality battery | 22 / 24 | 22 / 24 |
| tok/s (ROCm) | 56.1 | 48.1 |
The battery is 24 items: 8 code tasks graded by executing the generated function against
assertions, 8 long-tail factual questions, 4 multilingual, 4 maths. Both quants scored 22/24 —
parity — though they fail different items, which is sampling noise rather than quantization damage.
ngram-map-k adds roughly 1.05× on top and is worth enabling for input-grounded work
(code edits, RAG, summarisation).
Note on layout variants
Q4_0_ROCMFP4_FAST (type 103, the "single-scale speed layout") and Q4_0_ROCMFP4_FAST_COHERENT
(104) were both built and benchmarked against this one, same model, same flags, sizes within
0.1 GiB:
| layout | base | + ngram |
|---|
| STRIX (105) | 55.4 | 57.3 |
| FAST (103) | 53.9 | 55.0 |
| FAST_COHERENT (104) | 53.9 | 56.7 |
FAST is ~3% slower than the quality recipe despite its name, on this architecture and hardware.
STRIX is both the faster and the higher-quality choice here, which is why it is the one published.
Runtime knobs (micro-batch 512/2048, 32 threads, q8_0 KV cache, --poll 0) were all
neutral-to-negative — quantizing the KV cache changing nothing indicates the bottleneck is weight
bandwidth rather than KV.
License
Follow the base model (zai-org/GLM-4.7-Flash) license terms.
Other public builds of this model
Compiled from Hugging Face repository metadata — file sizes, shipped files, quant variant as named by each repo. No third-party build was run or benchmarked here, so this table makes no speed or quality claim about any of them. It is here so you can see the size and format options at a glance and pick what fits your hardware.
Base model: zai-org/GLM-4.7-Flash. Generated from Hub metadata; download counts move over time.
Acknowledgements
This build would not exist without the work below. Please star and follow these
projects — the quantisation format used here is their engineering, not mine.
ROCmFPX — maintained by
charlie12345 / caf
The ROCmFP4 / ROCmFPX tensor formats (ggml types 100–106) exist only in this fork.
Every ROCmFP4 file in this repository was produced with its
llama-quantize, and
runs on its runtime. The fork also credits collaborators
ciru-ai,
Tom Turney,
PlunderStruck and
Aydan S., and acknowledges AMD for hardware support.
Licensed MIT, based on upstream llama.cpp.
llama.cpp — ggml-org and contributors
The inference engine, GGUF format and conversion tooling everything here is built on.
AMD ROCm
The compute platform these builds target — ROCm 7.2.4 on gfx1151 / Radeon 8060S.
Base model authors — see base_model in the metadata above; all model weights,
licences and capabilities are theirs. This repository contributes quantisation and
measurement only.
If you use these files, please credit ROCmFPX alongside this repository.
