Issues

MoE Q4 sorted-pairs path sums non-finite expert outputs: intermittent all-NaN logits on GB10 (sm_121)

#1025

· AlexIannu opened

on Sep 10, 2026

FYI: DeepSeek v4.1 Flash weights are available

#1023

· fry69 opened

on Sep 10, 2026

ROCm build fails: host-side `rsqrtf()` in `rocm/ds4_rocm_deepseek4_vision.cuh`

#1019

· neomantra opened

on Sep 10, 2026

GLM-5.3-Flash FP8 Optimized Inference Support

#1015

· zzyu17 opened

on Sep 9, 2026

Model residency in memory switch / toggle

#1009

· shpati opened

on Sep 9, 2026

Server: multimodal sessions lose all KV reuse on a tail token mismatch — no common-prefix rewind, no disk-KV fallback, no continued disk stores

#1007

· graymount opened

on Sep 9, 2026

DSpark with the Vision-Exp support model on Metal: ~half the accepted tokens per cycle of the 0731 pair, a wash in ds4-bench and a net loss through ds4-server at every confidence (M5 Max)

#1002

· nazeshinjite opened

on Sep 9, 2026

DSpark corrupts output on the aprojq4-dense-attention branch (GB10/CUDA): intra-word syllable doubling, confined to the full-accept path

#1001

· adamlawi opened

on Sep 7, 2026

GLM (and soon Qwen) tool-call syntax gets no decode-time protection; DSML does

#999

· titan550 opened

on Sep 7, 2026

GLM-5.2 Q4: prefill fails at ≥64K context on M3 Ultra 512GB (KV payload staging failed)

#998

· wbllcbbg-ai opened

on Sep 7, 2026

glm 5.3 flash metal not optimized?

#989

· badelon opened

on Sep 6, 2026

ROCm strix-halo build error

#980

· ngkz opened

on Sep 5, 2026

Navigation Menu

All issues

MoE Q4 sorted-pairs path sums non-finite expert outputs: intermittent all-NaN logits on GB10 (sm_121)

FYI: DeepSeek v4.1 Flash weights are available

ROCm build fails: host-side `rsqrtf()` in `rocm/ds4_rocm_deepseek4_vision.cuh`

GLM-5.3-Flash FP8 Optimized Inference Support

Model residency in memory switch / toggle

Server: multimodal sessions lose all KV reuse on a tail token mismatch — no common-prefix rewind, no disk-KV fallback, no continued disk stores

DSpark with the Vision-Exp support model on Metal: ~half the accepted tokens per cycle of the 0731 pair, a wash in ds4-bench and a net loss through ds4-server at every confidence (M5 Max)

DSpark corrupts output on the aprojq4-dense-attention branch (GB10/CUDA): intra-word syllable doubling, confined to the full-accept path

GLM (and soon Qwen) tool-call syntax gets no decode-time protection; DSML does

GLM-5.2 Q4: prefill fails at ≥64K context on M3 Ultra 512GB (KV payload staging failed)

glm 5.3 flash metal not optimized?

ROCm strix-halo build error

Search results