Two halves of the same problem: nothing stopped a model being placed on a node that could not hold it, and nothing ever removed the weights of a model that had been deleted from the catalog. Placement now has a disk check, shaped exactly like the memory one it sits beside, with one addition: it is skipped on a node that already holds the weights. A node holding the file needs no room for it, and without that the node best placed to run a model — because it already has it — is the one most likely to be rejected, since holding the file is what used the space. Reclaiming is the dangerous half, so the whole design is about what it will not do. llmgw deletes only files it fetched itself, and that is a calculation rather than a heuristic: the cache name is a hash of the source, so the set of wanted files can be computed from the catalog exactly. Anything without that name was put there by an operator — scp, a shared mount, a copy by hand — and is none of this system's business however unreferenced it looks. A model that took a week to copy must not be removable by a command someone ran to tidy up. The node checks every path again before unlinking it, and that duplication is the point. The control plane knows the catalog; the node knows its disk. A control plane that is wrong, newer than the agent, or tampered with must not be able to talk a node into deleting something, so the agent refuses anything outside its own cache directory or without the cache name — including a path that tries to traverse out of it. Listing is the default and deleting takes --collect. The summary says "asked to remove", because each node decides, and one that refuses a path it cannot prove it fetched is this working rather than failing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
llmgw
A single-binary LLM gateway for clusters with multiple InfiniBand fabrics and heterogeneous GPUs.
It serves LLMs running on your own GPU servers (vLLM / SGLang) behind one OpenAI-compatible endpoint. Design document: llmgw-plan.md.
Getting started
Starting from nothing — no packages installed, no drivers, no nodes — follow docs/install.md: it builds a control plane, adds a node and serves a model, and says plainly which parts llmgw does not install for you.
| Guide | For |
|---|---|
| docs/install.md | Standing up a cluster from scratch |
| docs/node-onboarding.md | Adding nodes to one that exists |
| docs/onboarding.md | Giving people accounts and keys |
Status
Releases track the roadmap phases; see CHANGELOG.md.
| Version | Phase | Scope | State |
|---|---|---|---|
| v0.1.0 | 0 | Skeleton and sim-engine | released |
| v0.2.0 | 1 | Static router (filter/score/pick, streaming proxy) | released |
| v0.3.0 | 1.5 | Admission control (priority bands) | released |
| v0.4.0 | 2 | Agent: GPU/IB discovery, engine supervision | released |
| v0.5.0 | 3 | Control plane: catalog, placement, reconciler, ctl | released |
| v0.6.0 | 4 | Multi-node placement and fabric awareness | released |
| v1.0.0 | 5 | Tenancy, prefix affinity, benchmarking | planned |
As of v0.6.0 a cluster runs unattended: agents report their hardware, the control plane places replicas — including ones larger than a single machine — models load when someone asks for them and stop when nobody does, and keys, budgets and rate limits decide who may use any of it.
Before v1.0.0: routing profiles selectable per request, spillover to a fallback pool, tracing, and reference documentation. One thing is worth saying plainly: multi-node replicas are proven on a simulated cluster and on real hardware without a fabric. The InfiniBand, NVML and NCCL paths are written against the formats and failure modes those tools actually have, but they have not yet run on a machine that has them.
Building
make build # bin/llmgw
make test # unit tests
make race # with the race detector
Operating a cluster
llmgw ctl nodes # capacity, health, what is running where
llmgw ctl models # catalog and replica counts
llmgw ctl replicas --model llama # replicas and their shards
llmgw ctl explain-placement llama # why a model has not been placed
llmgw ctl drain gpu-a-01 # move work off a node, one replica at a time
llmgw ctl cordon gpu-a-01 # stop new placement, leave what runs
ctl talks to the admin port (:9090 by default), not the client-facing one.
Add --json to any command for machine-readable output.
explain-placement is the one worth knowing about: "unschedulable" on its own
says nothing about which of six constraints bit, so it prints every candidate
considered and the reason it was rejected.
Running the router
# Two fake engines and a router in front of them.
./bin/llmgw sim-engine --listen :8001 --model llama &
./bin/llmgw sim-engine --listen :8002 --model llama &
./bin/llmgw server \
--endpoint e1=http://127.0.0.1:8001=llama \
--endpoint e2=http://127.0.0.1:8002=llama
curl -N localhost:8080/v1/chat/completions \
-H 'content-type: application/json' \
-H 'x-session-id: demo' \
-d '{"model":"llama","messages":[{"role":"user","content":"hello"}],"stream":true}'
For anything real use a config file; deploy/examples/server.yaml
is commented throughout. Metrics, health and pprof are on the admin port (:9090
by default), deliberately separate from the client-facing one.
Requests carry a priority band (x-llmgw-band). Under load, low bands wait at
the gateway while high bands keep flowing, so the cluster degrades from the
bottom up instead of everywhere at once. Measured effect: interactive p99 stays
within 1.24x of idle while 500 concurrent batch requests run, against 58x with
admission off.
Routing profiles select how endpoints are scored: chat favours session
affinity, batch favours the shortest queue, long-context favours free KV
cache. Pick one per model in the config, or per request with the
x-llmgw-profile header.
sim-engine
Simulates the inference-engine behaviours the router cares about, so the router can be developed and tested without GPUs. What it models:
- continuous batching (bounded number of concurrent sequences)
- KV cache block accounting, queueing once blocks run out (head-of-line blocking included)
- automatic prefix caching (chained block hashes plus an LRU) and its effect on time to first token
- per-token latency growing with batch size
- dropping a request and freeing its blocks when the client disconnects
/metricsusing vLLM's metric names
./bin/llmgw sim-engine --listen :8000 --model llama-3.3-70b
curl -N localhost:8000/v1/chat/completions \
-H 'content-type: application/json' \
-d '{"messages":[{"role":"user","content":"hello"}],"max_tokens":8,"stream":true}'
Tune load and latency behaviour with --max-num-seqs, --num-gpu-blocks,
--prefill-tps, --decode-tps, --batch-contention and
--prefix-cache-blocks. For chaos testing use --fail-rate, and --seed for
deterministic runs.
The engine metric contract
For the router to make routing decisions an engine must expose these three metrics (plan §7.6). sim-engine publishes them under vLLM's real names:
vllm:num_requests_running
vllm:num_requests_waiting
vllm:kv_cache_usage_perc
Adding support for a new engine means satisfying that contract. Metric names drift between versions, so the name mapping is configurable.
Module path
gitea.mynodes.xyz/ilker/llmgw. To change it:
make rename NEW=git.example.com/team/llmgw
Language
Code, comments, tests, commit messages and documentation are in English.