Search results

multi-slot

Status: Open.

#537 In Kaden-Schutt/hipfire;

· artenais opened on Jul 20, 2026

Feature: MQ2G256 batched prefill support for Qwen3.5/3.6 (22x prefill gap vs llama.cpp)

Status: Open.

#533 In Kaden-Schutt/hipfire;

· rakaarwaky opened on Jul 17, 2026

Remove noisy 'RDNA2 GEMV variant' eprintln from gemv_hfq4g256_for_arch

Status: Open.

#531 In Kaden-Schutt/hipfire;

· rakaarwaky opened on Jul 17, 2026

fix(cli): recognize all registry model extensions in local catalog

Status: Open.

#526 In Kaden-Schutt/hipfire;

· HUSRCF opened on Jul 12, 2026

Pro W7900 & Pro W7900 Dual Slot benchmark

Status: Open.

#499 In Kaden-Schutt/hipfire;

· HUSRCF opened on Jul 7, 2026

instrument: latency_score drops simds_per_cu (2x error, biases roofline toward Latency)

Status: Open.

#491 In Kaden-Schutt/hipfire;

· Kaden-Schutt opened on Jul 2, 2026

instrument: mem_latency_ns is cache-deflated on Infinity-Cache parts (roofline latency axis untrustworthy)

Status: Open.

#490 In Kaden-Schutt/hipfire;

· Kaden-Schutt opened on Jul 2, 2026

DFlash spec-decode OOMs at max_seq≥24576 on 24GB — VRAM bloat (drafter + verify scratch)

Status: Open.

#486 In Kaden-Schutt/hipfire;

· Kaden-Schutt opened on Jun 30, 2026

gfx1030 (RDNA2) cannot serve default model qwen3.5:9b-mq6 — pre-warm crash: no implementation for GemvMq6G256Prerotated

Status: Open.

#478 In Kaden-Schutt/hipfire;

· rakaarwaky opened on Jun 24, 2026

Benchmarks: gfx1201 (9070 non-xt)

Status: Open.

#475 In Kaden-Schutt/hipfire;

· maxchisto opened on Jun 23, 2026

DFlash from-scratch trainer + MTP-AR drafter — contributor handoff (@fivetide)

Status: Open.

#469 In Kaden-Schutt/hipfire;

· Kaden-Schutt opened on Jun 22, 2026

Serve regression: PR #455 left qwen3.5 DeltaNet/KV bundle migration incomplete → </think> attractor on multi-request serve

Status: Open.

#462 In Kaden-Schutt/hipfire;

· Kaden-Schutt opened on Jun 21, 2026