Repository navigation
feat(inference): house 7B tier — every 16 GiB node serves qwen2.5:7b (v0.3.67) - #365
Merged
Merged
Conversation
92% of the fleet's chat asks for qwen2.5:7b (@profullstack/ai-providers), and only two machines served it, so the whole fleet queued on them while 14 c0mpute nodes sat idle with the 1.5B. The house llama.cpp now runs Qwen2.5 7B Instruct Q4_K_M (official Qwen GGUF, two shards, both pinned to a revision and SHA-256) on boxes with twice min_ram_gb (16 GiB default), needing 8 GiB free to start and 10 GB free disk to install; 8 GiB boxes keep the 1.5B, 4 GiB the 0.5B. - HouseModel.parts: further shards of a split GGUF, downloaded beside the first (llama-server loads them from the first); ready stamp and installed() cover every shard; other tiers' files (and shards) are removed on switch - model shards use the resumable, retrying downloader (gigabytes) - HouseModel.disk_gb: per-model disk need on top of the config floor Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ThreatCrush Security Scan14 finding(s) HIGH/CRITICAL: 2 | MEDIUM: 8 | LOW: 4
Snippets are redacted; ThreatCrush never prints matched credential material. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why: 92% of fleet chat (~1,000 of ~1,100 jobs/day) asks for
qwen2.5:7b, and only bonita + vienna served it, so all of it queued on two machines (concurrency 1 each) while the other c0mpute nodes ran the 1.5B or nothing. The Infernet house manifest already carries qwen2.5:7b (8 GB floor); the c0mpute house llama.cpp had no 7B tier.What:
Qwen/Qwen2.5-7B-Instruct-GGUF(revision bb5d59e0), two shards each pinned by SHA-256, served asqwen2.5:7b, on boxes with 2×min_ram_gb(16 GiB default). Start gate 8 GiB free RAM, install gate 10 GB free disk. 8 GiB boxes keep the 1.5B, 4 GiB the 0.5B. Same opt-outs (C0MPUTE_INFERENCE_AUTO=0).HouseModel.parts(split GGUF shards, loaded by llama-server from the first), ready stamp +installed()cover every shard, other tiers' files removed on switch; resumable downloader for the shards;HouseModel.disk_gb.Verified:
cargo test -p c0mpute-core --lib inference8/8 on vienna. End to end on vienna with the exact pins: both shard hashes match, llama-server b12061 loads the split file,/v1/models=qwen2.5:7b, coherent answer (0.5 tok/s on vienna at load ~20; the worker only bids under its load limit).Note: the hosting e2e tests fail on master independently (they fail at v0.3.65 too).
🤖 Generated with Claude Code