hipfire 0.4.3 and 0.4.3.1: up to 9.5x A3B prefill on the Radeon R9700, then a patch for 27B MTP and faster decode

hipfire, the Rust inference engine for AMD GPUs, shipped two releases on 24 October. v0.4.3 went out at 05:24 UTC and the v0.4.3.1 patch followed at 20:37 UTC. Install the patch, not 0.4.3.

0.4.3: the new mq4rs file does most of the work

The headline is prefill on the Radeon AI PRO R9700 (gfx1201) for Qwen3.6-35B-A3B, a mixture-of-experts model. The project’s table compares 0.4.2.1 on the old mq4r file with 0.4.3 on the new qwen3.6:35b-a3b-mq4rs file:

Prefill tok/s	0.4.2 (mq4r)	0.4.3 (mq4r)	0.4.3 (mq4rs)
pp512	3,292	3,332	11,589
pp4096	2,886	3,737	19,828
pp16384	1,957	3,561	18,545

That is 3.5x at 512 tokens and 9.5x at 16,384. The notes are careful about what this means. Most of the gain comes from the new quant file, not the engine. The two versions ran their own default configs, which differ in KV cache format and prefill chunk size, so the middle column is “what a user gets from upgrading without changing the model file”. And fast mq4rs prefill is gfx1201 only: on other GPUs the file needs HIPFIRE_MOE_SYM_IU4=0, which the notes say is about 5x slower. Users on the old mq4r file are told to pull mq4rs.

0.4.3 also lists the first Flash-Next sub-4-bit tiers, LFM2 support and new kernel-authoring tools.

0.4.3.1: one regression fixed, one decode speedup

The patch fixes a load failure: Qwen3.8-27B with --spec mtp “failed to load on v0.4.3”, because the sealed 27B plan missed the MTP head. The notes trace it to a commit in the 0.4.3 cycle and the fix to another.

It also moves A3B mq4rs decode on gfx1201 to the PM4 replay route by default, with output “bit-exact against the HIP route”:

Decode tok/s, R9700	0.4.3	0.4.3.1	Change
tg128, ctx 128	157	232	+48%
tg128, ctx 8K	146.6	217.0	+48%
1024-token reasoning run	154	225.0	+46%

Prefill is unchanged, and again this does not apply to other GPUs. replay.backend = "hip" turns it off. A third fix makes a failed verified-artifact cache write print one warning instead of silently re-hashing the model on every request.

To pin the patch:

curl -fsSL https://hipfire.dev/install.sh | bash -s -- --tag v0.4.3.1

Why it matters

Local inference on AMD cards still depends on engines tuned per chip, and hipfire’s notes show both sides of that: large gains on one card and one quant, and an explicit warning that other cards do not get them. All numbers are the project’s own runs on its test cards. Our earlier coverage: hipfire 0.4.2 and 0.4.0.