All issues

Search results

gfx1030 - qwen3.8 + dflash issue

Status: Open.

#771 In warpfront/hipfire;

· zues75 opened on Sep 19, 2026

dense TP=2 decode crashes HSA_STATUS_ERROR_INVALID_ALLOCATION on tied/aliased-lm_head dense Qwen (qwen3.5-4b); separate-lm_head (3.8-27b) works

Status: Open.

#770 In warpfront/hipfire;

· sklarsa opened on Sep 17, 2026

Dflash hurts perfomance with Qwen 3.8 on Strix Halo

Status: Open.

#754 In warpfront/hipfire;

· alucryd opened on Sep 11, 2026

feat: DeepSeek-V4.1-Flash implementation plan — MQ3/MQ2 weights and SSD-streamed Engram

Status: Open.

#743 In warpfront/hipfire;

· fivetide opened on Sep 10, 2026

serve: nondeterministic model-load hang — CPU spin at random layer, GPU idle, frozen IO counters (gfx1100, qwen3.8:27b mq4r, 59669d9)

Status: Open.

#739 In warpfront/hipfire;

· nadaboulaares opened on Sep 9, 2026

Create a new release version

Status: Open.

#733 In warpfront/hipfire;

· HUSRCF opened on Sep 8, 2026

Qwen3.8-27B native MTP crashes when Q8 batched GEMM exceeds MAX_BATCH=64.

Status: Open.

#732 In warpfront/hipfire;

· Sergey7210 opened on Sep 7, 2026

Memory preflight OOM guard is unconditionally enabled — refuses valid loads on discrete-GPU dev boxes

Status: Open.

#696 In warpfront/hipfire;

· ghazni101 opened on Sep 4, 2026

DFlash decode collapses below AR past ~10k ctx: verify-block target attention costs B×L per cycle

Status: Open.

#693 In warpfront/hipfire;

· Kaden-Schutt opened on Sep 3, 2026

JetSpec: Parallel Tree Drafting

Status: Open.

#684 In warpfront/hipfire;

· SlobCoder opened on Sep 2, 2026

EP MoE generate fails after a successful tp=4 load: 'EP arch mismatch (expected dense Qwen TP)' (ornith-1.5:35b-a3b-mq4r, master)

Status: Open.

#683 In warpfront/hipfire;

· Kaden-Schutt opened on Sep 2, 2026

Publish Gemma 4 E2B-it typed MQ4/MQ6 + PLE-Q8 artifact

Status: Open.

#672 In warpfront/hipfire;

· HUSRCF opened on Sep 1, 2026