[PR #8073] [CLOSED] llama: Fix the KV cache quants q4_0 and q8_0 lead server abort in large context chat. #17883

Closed
opened 2026-04-16 06:17:57 -05:00 by GiteaMirror · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/8073
Author: @mengqin
Created: 12/12/2024
Status: Closed

Base: mainHead: main


📄 Description

This fix include two parts which cause the problem, first the context switch need a correct operation for the q4_0 to f32 temp cache tensor. So added new function in ggmal_cuda_cpy and ggml_cuda_dup to process with it.
The second part happened in ggml_compute_forward_dup operation which also need a correct way to process with q4_0 to f32 and q8_0 to f32, add them too.

This is a fix of https://github.com/ollama/ollama/issues/7938


🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/ollama/ollama/pull/8073 **Author:** [@mengqin](https://github.com/mengqin) **Created:** 12/12/2024 **Status:** ❌ Closed **Base:** `main` ← **Head:** `main` --- ### 📄 Description This fix include two parts which cause the problem, first the context switch need a correct operation for the q4_0 to f32 temp cache tensor. So added new function in ggmal_cuda_cpy and ggml_cuda_dup to process with it. The second part happened in ggml_compute_forward_dup operation which also need a correct way to process with q4_0 to f32 and q8_0 to f32, add them too. This is a fix of https://github.com/ollama/ollama/issues/7938 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
GiteaMirror added the pull-request label 2026-04-16 06:17:57 -05:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#17883