Research: large-model background offloading on the RX 6700 XT (AirLLM eval + V620 sim)#6
Draft
marchah wants to merge 2 commits into
Draft
Research: large-model background offloading on the RX 6700 XT (AirLLM eval + V620 sim)#6marchah wants to merge 2 commits into
marchah wants to merge 2 commits into
Conversation
…tion) Research notes for a future dedicated background-task card. Evaluated AirLLM (rejected: PyTorch/no-Vulkan, sub-1 tok/s) vs llama.cpp --n-cpu-moe + mmap (the path forward). Includes a V620-simulated 12 GiB/10 GiB run of the 35B-A3B MoE (~14-22 tok/s decode, optimistic upper bound), the offload-vs-accuracy note (placement only, no quality impact), and next steps for when the card is installed. No code/scripts changed — doc only. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
CT 120 serves the model (llama-server); Hermes is a separate agent container that consumes its API. Fix the conflation in the goal statement. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This was referenced Jul 4, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Draft / research-only. Adds
rx-6700-xt/BACKGROUND-OFFLOAD-FINDINGS.mdcapturing the evaluation of running models bigger than VRAM on the retired RX 6700 XT as a future dedicated background-task card. No scripts or runtime config changed — doc only.Findings
--n-cpu-moe+mmapis the path — same trick, done better, on the Vulkan stack already in use (works on gfx1031). Streams only active experts for MoE; only touches disk for genuine VRAM+RAM overflow.--n-cpu-moe 24(~11.8 GiB VRAM) + a 10 GiB RAM cap: ~14–22 tok/s decode (vs ~53 full-VRAM). Read as an optimistic upper bound — the V620 is a stronger GPU on faster PCIe, and this is a favorable 3B-active MoE (big-active MoE / dense models will be much slower).Status / next steps
Parked until the card is physically installed and there's a concrete large-model background workload. Next steps (install + render-node split, pick MoE model, new
rx-6700-xt/create-lxc-llamacpp-<model>.sh, benchmark) are listed in the doc.The V620-simulation test was fully reversible; CT 120 (the V620 llama-server LLM runtime — not Hermes, which is a separate agent container that consumes its API) was restored to its normal config (ctx 262144 / parallel 4 / ~30 GiB VRAM) and verified healthy.
🤖 Generated with Claude Code