Skip to content

hipfire 0.4.7

Latest

Choose a tag to compare

@github-actions github-actions released this 15 Nov 20:57
· 21 commits to master since this release

hipfire 0.4.7

0.4.7 is a bug-fix patch release on top of 0.4.6. It backports fixes from beta for bugs that 0.4.6 users can hit. It removes nothing and changes no default numerics: the KLD pins are byte-identical to 0.4.6. The full list, with the beta commit for every entry, is in the CHANGELOG.

Highlights

Thank you to everyone who sent a fix or a report. Each item names the author.

  • Flash-Next EP2 loads again with the official binaries (gfx1201). Expert parallelism on two R9700s aborted every load with thread 'hipfire-qwen4-ep-rank1' has overflowed its stack: peer ranks ran on 2 MiB worker threads. Every rank worker now gets an 8 MiB stack, so the RUST_MIN_STACK=8388608 workaround is no longer needed.
  • Flash-Next cold start on gfx1201 no longer compiles kernels. The gfx1201 kernel pack now carries the modules the default Flash-Next route needs (the A8-trunk MQ6 check at load, the lane WMMA GEMMs and prefill tiles, the wide-pass gate/up, and the dead-store-skipping routed MoE down), so a cold hipfire run no longer calls hipcc.
  • Vision keeps the conversation (#1079, reported by @cunzai97). An image no longer erases the history: images stay at their message position, the Qwen3.5/3.8-VL generator renders the whole conversation, and several images in history are accepted.
  • Cancelled vision turns roll back (#1124, #1138, @fivetide). A Qwen3.5-VL, dots.ocr or LFM2-VL turn cancelled after it finished decoding no longer stays resident.
  • The request seed is honoured (#1126, #1140, @fivetide). The Glimmer, LFM2-MoE, MiniMax, Cohere2-MoE and DeepSeek4 singleton routes seeded from the wall clock (DeepSeek4's heterogeneous route from a fixed seed) and ignored the wire seed.
  • Correct finish reasons and fail-closed tool calls (@fivetide).
    • A DeepSeek V4 DSML tool_calls block that parses only partly, or has stray text between tags, is a malformed-DSML error instead of merged, dropped or empty tool calls (HIPF-320, HIPF-378).
    • Cohere2-MoE reports finish_reason: "length" when max_tokens ends the turn (HIPF-379).
    • A Glimmer singleton that stops at the KV cache cap reports length, and tokens matches the stream (HIPF-277).
  • Loading fixes.
    • A runtime that cannot reserve GPU address space no longer fails the load with hipMemAddressReserve: out of memory after the weights upload: automatic KV selection proves a reservation works before choosing VMM (#1058, reported by @kwilkins-82).
    • Configs in the transformers 5 layout get their rope_theta instead of running theta 10000 (HIPF-206, @fivetide).
    • The non-codec loaders refuse a stale ragged payload of any row-packed dtype instead of decoding garbage (HIPF-317, @fivetide).
    • Qwen3.5-family MoE decode no longer fails AR hipGraph capture (hipError 906) when a layer falls back to host top-K; such a model decodes eager (#1149, reported by @ghazni101).
  • MI300X / gfx906 (#1136).
    • Qwen3.5/3.6 MQ4G256 trunks wider than 256 (for example Qwen3.8-27B .mq4r) no longer produce garbage on gfx94x.
    • The default Q8 DeltaNet recurrence compiles on gfx906 / gfx94x again.
    • A Qwen3.5/3.6 model whose weight format has no wave64 GEMV is refused at load instead of failing the first request.
  • Native Windows (#1143, #1144, #1145, reported by @azfarh95). scripts/windows-check.ps1 writes report.txt again; hipfire fit sees the GPUs and host RAM; the daemon warns when it runs at Idle or BelowNormal priority, and hipfire diag shows the priority class.

Known issues

  • Cancel resets the conversation (#1062). Cancelling a request mid-generation resets the live conversation, so the next request re-prefills the whole prompt.
  • Garbage after out-of-memory recovery (#1049). After an out-of-memory recovery at very long context, later prefills in the same process can produce garbage. Restart the daemon to recover.
  • Tool-call re-prefill on some templates. The qwen3_5_moe chat template drops earlier reasoning, so the first request after a run of tool turns re-prefills cold on models like ornith-1.5:35b-a3b. Unchanged from 0.4.6.
  • EP2 C8 can differ from C1 at a near-tie token (gfx1201, by design). Unchanged from 0.4.6.

Install

Binary archives (gfx1100, gfx1151, gfx1201) and kernel packs (gfx906, gfx942, gfx1100, gfx1151, gfx1201) are attached to this release. The host needs ROCm; the HIP runtime is not bundled. install.sh --tag v0.4.7 verifies the pack against the sources and compiles locally on any mismatch.

Toolchain: the official binaries are built with rustc 1.99.1.