Not a smart man, but i think llama.cpp is broken
retained text
i have a 9070xt and B1356 lemonade build is the last Fast Rocm Builds.
Lemonade B1357 Speeds
|model |size|params|backend |ngl|threads|n\_batch|type\_k|type\_v| fa|test|t/s|
|:-|:-|:-|:-|:-|:-|:-|:-|:-|:-|:-|:-|
|qwen35 27B IQ3\_S - 3.4375 bpw | 9.72 GiB| 27.32 B|ROCm | 99|12|512| q8\_0| q8\_0| 1|pp512|460.03 ┬▒ 4.80|
|qwen35 27B IQ3\_S - 3.4375 bpw | 9.72 GiB| 27.32 B|ROCm | 99|12|512| q8\_0| q8\_0| 1|pp1024|484.04 ┬▒ 2.60|
|qwen35 27B IQ3\_S - 3.4375 bpw | 9.72 GiB| 27.32 B|ROCm | 99|12|512| q8\_0| q8\_0| 1|pp2048|495.34 ┬▒ 0.82|
B1356 Speeds
| model | size | params | backend | ngl | threads | n_batch | type_k | type_v | fa | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | ------: | -----: | -----: | --: | --------------: | -------------------: |
| qwen35 27B IQ3_S - 3.4375 bpw | 9.72 GiB | 27.32 B | ROCm | 99 | 12 | 512 | q8_0 | q8_0 | 1 | pp512 | 1049.13 ┬▒ 70.54 |
| qwen35 27B IQ3_S - 3.4375 bpw | 9.72 GiB | 27.32 B | ROCm | 99 | 12 | 512 | q8_0 | q8_0 | 1 | pp1024 | 1185.39 ┬▒ 55.71 |
| qwen35 27B IQ3_S - 3.4375 bpw | 9.72 GiB | 27.32 B | ROCm | 99 | 12 | 512 | q8_0 | q8_0 | 1 | pp2048 | 1213.68 ┬▒ 29.57 |
| qwen35 27B IQ3_S - 3.4375 bpw | 9.72 GiB | 27.32 B | ROCm | 99 | 12 | 512 | q8_0 | q8_0 | 1 | pp4096 | 1203.99 ┬▒ 2.05 |
i built llama.cpp over the top of the fast lemonade B1356 Rocm ,,, and it goes back to slow.
hoping that someone that knows what theyre doing can identify and fix the problemdocument fetched 2026-11-01 00:38 · canonical https://reddit.com/r/LocalLLaMA/comments/1xjy973/not_a_smart_man_but_i_think_llamacpp_is_broken