Search results

[Bug] storage_size (TP) overestimates allocation of fused linear layers like QKV / QKVZ

Status: Open.

#193 In turboderp-org/exllamav3;

· mratsim opened on Apr 16, 2026

Unexpected high VRAM usage on turboderp/gemma-4-31B-it-exl

Status: Open.

#191 In turboderp-org/exllamav3;

· cmoncure opened on Apr 14, 2026

[Bug] Prefill failure with image on Qwen3.5

Status: Open.

#190 In turboderp-org/exllamav3;

· mratsim opened on Apr 14, 2026

[BUG] Autosplit overpacks VRAM by discarding KV Cache placeholders per-layer

Status: Open.

#187 In turboderp-org/exllamav3;

· Danielcurly opened on Apr 8, 2026

[Support Request] NVIDIA-Nemotron-3-Super-120B-A12B-BF16

Status: Open.

#179 In turboderp-org/exllamav3;

· blackcat1402 opened on Mar 29, 2026

RuntimeError: AVX2 is required for tensor-parallel inference using native backend

Status: Open.

#177 In turboderp-org/exllamav3;

· Koesn opened on Mar 26, 2026

[Support Request] Please support mistralai/Mistral-Small-4-119B-2603

Status: Open.

#176 In turboderp-org/exllamav3;

· ncoder-ai opened on Mar 26, 2026

Feedback wanted: native MLA baseline with optional flash-attn acceleration

Status: Open.

#175 In turboderp-org/exllamav3;

· lesj0610 opened on Mar 26, 2026

How to slide window when context length is beyond setting of context length?

Status: Open.

#172 In turboderp-org/exllamav3;

· blackcat1402 opened on Mar 22, 2026

[Feature / Bug-fix] Replace id_to_piece with sliding-window detokenization to eliminate U+FFFD emission

Status: Open.

#171 In turboderp-org/exllamav3;

· lesj0610 opened on Mar 17, 2026

Convert.py on Qwen3.5 35B gives CUDA erro

Status: Open.

#169 In turboderp-org/exllamav3;

· niceblue88 opened on Mar 14, 2026

Stepfun has double BOS?

Status: Open.

#164 In turboderp-org/exllamav3;

· Ph0rk0z opened on Mar 6, 2026