All issues

--spec-mtp-strict-qwen does not preserve exact greedy output for Qwen3.8/qwen4exp on ROCm gfx1151

#110

· arigas1982-maker opened

on Sep 3, 2026

Separate-model draft-mtp for Gemma 4 crashes on Windows/gfx1151 - 5 stacked bugs, root cause: n_embd_pre_norm width mismatch (1024 vs 5376)

#108

· BEEugene opened

on Sep 1, 2026

Missing GGML_TYPE_Q2_0 (and GGML_FTYPE_MOSTLY_Q2_0) — gap vs llama.cpp upstream at type ID 42

#102

· ghazni101 opened

on Aug 30, 2026

Feature Request: prebuild ubuntu binaries

#100

· deepblue96 opened

on Aug 27, 2026

[BUG] SIGABRT (exit 134): DeepSeek-V4-Flash-180B STRIX_LEAN MTP + draft-mtp crashes with "failed to remove sequence" in common_context_seq_rm (p1=-1)

#92

· fjvicioso-lang opened

on Aug 21, 2026

Port of nathan strix-halo FA prefill stack: +59% deep prefill, fixes f16-KV 64k crashes, silent-CPU-fallback trap

#86

· aic0d3r opened

on Aug 18, 2026

Nemotron-H MoE: expert NVFP4 scale2 tensors not loaded — repetitive garbage output

#85

· poad42 opened

on Aug 17, 2026

Upstreaming the ROCmFP4/ROCmFPX tensor types: adoption data, AMD Lemonade results, and an offer to help

#84

· kingjones30 opened

on Aug 17, 2026

MTP (draft-mtp) disables the server prompt cache on ROCmFPX; stock llama.cpp with the same upstream fix does not

#67

· Lafunamor opened

on Aug 14, 2026

Field report: 13 ROCmFP4 builds on gfx1151, hybrid-attention data point, and build_vit per-layer masks

#65

· kingjones30 opened

on Aug 11, 2026

Eval bug: Escha-W2 does not work with vision encoder from qwen3.6 35b-a3b

#63

· pritam-dey3 opened

on Aug 10, 2026

Converter silently picks wrong half of stacked gate_up when n_embd == 2*n_ff (model loads, emits garbage)

#49

· kingjones30 opened

on Jul 31, 2026