Trouble with llama.cpp vulkan

MLOps1 post from 1 person1d active+00 posts in the last 7 days, 0 the 7 days before (fading)Unverified

Posts per day1 posts, Oct 6 to Oct 21

The posts behind it

1, newest first

GitHub issuesOct 6 vulkan: cm1 int8 mmq uses fp16 mmq wg denoms (tile split breaks when BN differs)## Bug `tc_mmq_cm1_int` in `ggml_vk_load_shaders()` pairs the int8 cm1 warptiles with `l/m/s_mmq_wg_denoms`, which are tuned for the fp16 mmq path, instead of deriving the denoms from the int warptile itself (the `_int_k` family already derives them from its warptile). On Intel...ggml-org/llama.cpp · dosermenOpen ↗

Companies and products named

Named1 post

llama.cpp

About this problem

Evidence

1 post from 1 person in 1 place. Mostly on ggml-org/llama.cpp.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts about llama.cpp with no known issue, sharing the word "vulkan". First post Oct 6, 2026, latest Oct 6, 2026. Unverified: fewer than 3 posts, or one person or place only. Method

Other llama.cpp problems

Same product29 postsRunning llama.cpp locally hits memory and speed limits Same product14 postsTrouble with llama.cpp misc Same product10 postsllama.cpp users keep filing bugs Same product5 postsllama.cpp upgrades break working setups Same product4 postsFine-tuning llama.cpp is costly and unpredictable Same product4 postsllama.cpp runs out of context on real work Same product3 postsTrouble with llama.cpp builds Same product3 postsllama.cpp lock-in makes switching costly Same product2 postsTrouble with llama.cpp flash Same product2 postsllama.cpp runs out of memory

History

MomentumOct 13, 2026new to fading

AddedOct 12, 2026Trouble with llama.cpp vulkan (1 post)