[PR #126] [MERGED] Add ability for aimdo to write VBAR contents back to RAM on eviction #127

New issue

Closed

opened 2026-10-26 22:24:50 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:24:50 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-aimdo/pull/126
Author: @rattus128
Created: 10/20/2026
Status: ✅ Merged
Merged: 10/21/2026
Merged by: @rattus128

Base: master ← Head: prs/vbar-writeback

📝 Commits (4)

15af197 Add whole-tensor descriptor writes through the transfer queue
62e1be7 Propagate VBAR failures through allocation APIs
d96e0e7 Let owners complete writeback before VBAR eviction
ff46043 Test whole-tensor pageout and failed writeback retention

📊 Changes

17 files changed (+415 additions, -80 deletions)

View changed files

📝 comfy_aimdo/host_buffer.py (+19 -0)
📝 comfy_aimdo/model_vbar.py (+32 -8)
📝 src-cuda/dispatch.c (+1 -0)
📝 src-hip/dispatch.c (+1 -0)
📝 src-posix/xfer-file-plat.c (+18 -6)
📝 src-win/xfer-file-plat.c (+18 -7)
📝 src/gpu_abi.h (+1 -0)
📝 src/gpu_dispatch.h (+3 -0)
📝 src/hostbuf-file-reader.c (+42 -16)
📝 src/malloc-graph.c (+22 -5)
📝 src/model-vbar.c (+68 -24)
📝 src/plat.h (+3 -2)
📝 src/pyt-cu-plug-alloc-async.c (+16 -6)
📝 src/vrambuf.c (+6 -2)
📝 src/xfer-file.c (+16 -4)
📝 src/xfer-file.h (+5 -0)
➕ tests/vbar-pageout-cuda.py (+144 -0)

📄 Description

This extends comfy-aimdo to be able to dynamically manage writable data (like KV caches).

_{🔄 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/Comfy-Org/comfy-aimdo/pull/126 **Author:** [@rattus128](https://github.com/rattus128) **Created:** 10/20/2026 **Status:** ✅ Merged **Merged:** 10/21/2026 **Merged by:** [@rattus128](https://github.com/rattus128) **Base:** `master` ← **Head:** `prs/vbar-writeback` --- ### 📝 Commits (4) - [`15af197`](https://github.com/Comfy-Org/comfy-aimdo/commit/15af1979fabb0e08ea341ea056a61d91b9df5a7b) Add whole-tensor descriptor writes through the transfer queue - [`62e1be7`](https://github.com/Comfy-Org/comfy-aimdo/commit/62e1be7010c192ffc295474af278ecc77ec5dd45) Propagate VBAR failures through allocation APIs - [`d96e0e7`](https://github.com/Comfy-Org/comfy-aimdo/commit/d96e0e76a942e0558c79e18518523f561b1bbc0c) Let owners complete writeback before VBAR eviction - [`ff46043`](https://github.com/Comfy-Org/comfy-aimdo/commit/ff46043debb94eee94e8c9d9aac940dba549c1bf) Test whole-tensor pageout and failed writeback retention ### 📊 Changes **17 files changed** (+415 additions, -80 deletions) <details> <summary>View changed files</summary> 📝 `comfy_aimdo/host_buffer.py` (+19 -0) 📝 `comfy_aimdo/model_vbar.py` (+32 -8) 📝 `src-cuda/dispatch.c` (+1 -0) 📝 `src-hip/dispatch.c` (+1 -0) 📝 `src-posix/xfer-file-plat.c` (+18 -6) 📝 `src-win/xfer-file-plat.c` (+18 -7) 📝 `src/gpu_abi.h` (+1 -0) 📝 `src/gpu_dispatch.h` (+3 -0) 📝 `src/hostbuf-file-reader.c` (+42 -16) 📝 `src/malloc-graph.c` (+22 -5) 📝 `src/model-vbar.c` (+68 -24) 📝 `src/plat.h` (+3 -2) 📝 `src/pyt-cu-plug-alloc-async.c` (+16 -6) 📝 `src/vrambuf.c` (+6 -2) 📝 `src/xfer-file.c` (+16 -4) 📝 `src/xfer-file.h` (+5 -0) ➕ `tests/vbar-pageout-cuda.py` (+144 -0) </details> ### 📄 Description This extends comfy-aimdo to be able to dynamically manage writable data (like KV caches). --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-26 22:24:50 +00:00

bsakaguchi referenced this issue

2026-10-26 22:24:50 +00:00

[PR #127] Retry undocumented CUDA NOT_READY from cuMemSetAccess #128

bsakaguchi closed this issue

2026-10-26 22:24:51 +00:00