All issues

Search results

Weight slab causing wasted vram use for qwen35

Status: Open.

#313 In turboderp-org/exllamav3;

· stevelikesrhino opened on Sep 1, 2026

bc_dsa.py imports triton-only kernels unconditionally, so exl3 1.4.x cannot import on a triton-less machine

Status: Open.

#309 In turboderp-org/exllamav3;

· pb2025ai-creator opened on Aug 29, 2026

convert_model crashes with multiple devices (-d 0,1) in the Triton autotuner: 'NoneType' object is not a mapping

Status: Open.

#308 In turboderp-org/exllamav3;

· UHHHHHHHHHHHHHH opened on Aug 27, 2026

CUDA illegal memory access in coop_autotune.cu on consumer Blackwell (sm_120, 36 SMs) under WSL2 — reproducible on trivial input

Status: Open.

#307 In turboderp-org/exllamav3;

· rmerien35 opened on Aug 24, 2026

Now reading weights from disk?

Status: Open.

#304 In turboderp-org/exllamav3;

· Ph0rk0z opened on Aug 22, 2026

[Feature] Emit a versioned uniform fused-QKV EXL3 object for Qwen full-attention blocks

Status: Open.

#297 In turboderp-org/exllamav3;

· malaiwah opened on Aug 20, 2026

Case VU#369611: ExLlamaV3

Status: Open.

#296 In turboderp-org/exllamav3;

· laurie-tyz opened on Aug 19, 2026

[Performance] TP prefill significantly slower than autosplit on 2x RTX 3080 (sm_86)

Status: Open.

#295 In turboderp-org/exllamav3;

· sorasoras opened on Aug 19, 2026

[Feature Request] Add DFlash2 support

Status: Open.

#294 In turboderp-org/exllamav3;

· timothysu opened on Aug 18, 2026

exllamav3_ext build failuare

Status: Open.

#292 In turboderp-org/exllamav3;

· thatname opened on Aug 18, 2026

DFlash draft cache allocates full-size KV for all-SWA layers — blocks 131k speculative decoding on 12 GB; partial patch included

Status: Open.

#285 In turboderp-org/exllamav3;

· msad12 opened on Aug 14, 2026

# [Bug] Silent output corruption when using layer split and large KV cache

Status: Open.

#282 In turboderp-org/exllamav3;

· pmsouzacampos opened on Aug 12, 2026