This is a huge backend change that essentially started with rewriting
the concurrency handling for processes and blew up to a refactor of the
entire application. In short these are the improvements:
**Better state and life cycle management:**
Life cycle management of processes has always been the trickiest part of
the code. Juggling mutex locks between multiple locations to reduce race
conditions was complex. Too complex for my feeble brain to build a
simple mental model around as llama-swap gained more features. All of
that has been refactored. Most of the locks are gone, replaced with a
single run() that owns all state changes. There is one place to start
from now to understand and extend routing logic.
The improved life cycle management makes it easier to implement more
complex swap optimization strategies in the future like #727.
**Collation of requests:**
llama-swap previously handled requests and swapping in the order they
came in. For example requests for models in this order ABCABC would
result in 5 swaps. Now those requests are handled in this order AABBCC.
The result is less time waiting for swap under a high churn request
queue. This fixes#588#612.
A possible future enhancement is to support a starvation parameter so
swap can be forced when models have been waiting too long.
**Shared base implementation for groups and swap matrix:**
During the refactor it became clear that much of the swapping logic was
shared between these two implementations. That is not surprising
considering the swap matrix was added many moons after groups. Now they
share a common base and their specific swap strategies are implemented
into the swapPlanner interface.
Requests for bespoke or specific swapping scenarios is a common theme in
the issues. Now users can implement whatever bespoke and weird swapping
strategy they want in their own fork. Just ask your agent of choice to
implement swapPlanner. I'll still remaining more conservative on what
actually lands in core llama-swap and will continue to evaluate PRs if
the changes is good for everyone or just one specific use case.
**AI / Agentic Disclosure:**
I paid very close attention to the low level swap concurrency design and
implementation. It's important to keep that essential part reliable,
boring and no surprises. Backwards compatibility was also maintained,
even the one way non-exclusive group model loading behaviour that people
have rightly pointed out be a weird design decision.
With the underlying swap core done the web server, api and UI sitting on
top were largely ported over with Claude Code and Opus 4.7 in multiple
phases. If you're curious I kept the changes in docs/newrouter-todo.md.
I did several passes to make sure things weren't left behind.
However, even frontier LLMs at the time of this PR still make small
decisions that don't make a lot of sense. They get shit wrong all the
time, just in small subtle way.
That said, there's likely to be some new bugs introduced with this
massive refactor. I'm fairly confident that there's no major
architectural flaws that would cause goal seeking agents to make dumb,
ugly code decisions.
For a little while the legacy llama-swap will be available under
cmd/legacy/llama-swap. The plan is to eventually delete that entry point
as well as the proxy package.
On a bit of a personal note, this PR is exciting and a bit sad for me. I
hand wrote much of the original code and this PR ultimately replaces
much of it. While the old code served as a good reference for the agent
to implement the new stuff it still a bit sad to eventually delete it
all.
Add a comprehensive performance monitoring system that collects CPU, memory, swap, load average, network IO, and GPU stats. Provides both a REST API for the UI and a Prometheus /metrics endpoint.
Backend changes:
- New internal/perf package with configurable interval-based stats collection
- GPU monitoring via LACT (Unix socket) and nvidia-smi fallback on Linux
- Ring buffer (internal/ring) for time-series stat storage
- Prometheus /metrics endpoint with all system and GPU metrics
- Moved LogMonitor to internal/logmon package
- New PerformanceConfig for hot-reloadable monitoring settings
- REST /api/performance endpoint replacing SSE streaming
UI changes:
- New Performance page with real-time charts for CPU, memory, GPU, and network
- Reusable PerformanceChart component
- LLAMA_SWAP_URL environment variable support
- Improved capture dialog display
Other:
- Example Grafana dashboard for Prometheus metrics
- monitor-test standalone binary
- Config schema and example updates
fixes#596
- matrix.go change logic to consider any proxy.Process not in
StateStopped or StateShutdown
- process.StopImmediately, and Stop() which called it had a subtle bug
where it only handled state transitions from StateReady to
StateStopping. StateStarting -> StateStopping was ignored completely.
fix: #670
I pointed Opus 4.7 (high effort) at proxy.ProcessGroup to identify any
race conditions in the swapping code. It found a race condition where
there is a small window in the fast path for routing a request to a
loaded model. There is a very small window where:
- model M1 is loaded and ready for requests
- a request, R1, for M1 comes in
- a request, R2, for M2 comes in almost immediately after
- R1 acquires the lock, sees M1 is loaded (fast path), releases the lock
`[race window]` and the request is ready to be forwarded
- the race window occurs between the release of the lock and the request
being forwarded
- the lock is released so requests can be handled concurrently
- R2 comes in within the `[race window]`, acquires the lock, triggers a
model swap to M2. stopping M1
- R1 is forwarded to a model that is unloaded or in the process of
shutting down creating an error response
In deployed systems the race window is very small and doesn't happen
often. However with #635 and PR #656 I though this deserved a bit more
attention. It is not concluded that this race is the cause of #635 but
the race is likely to happen more often under sustained or high load.
AI Note: Opus 4.7 x-high effort took about an hour to write the original
patch. With the pattern discovered the fix to matrix.go was very quick.
GLM 5.1 using the previous established patterns was able to easily write
the fix for ProcessGroup.StopProcesses().
Supersedes: #656
Updates: #277, #635
Add a new swap matrix to supersede groups for running concurrent models.
The matrix uses a solver that picks the lowest cost evictions to make a
requested model available. This simple approach along with a very basic
DSL grammar can enable very complex swapping scenarios.
- add DSL parser for set expressions with & (AND), | (OR), (), +ref
- add MatrixConfig structs, validation, and topological sort for +ref
- add MatrixSolver with cost-minimizing swap decisions
- add Matrix runtime integrating solver with Process lifecycle
- integrate matrix into ProxyManager with if-branches at all endpoints
- update config.example.yaml and config-schema.json with matrix schema
- config enforces groups XOR matrix (cannot use both)
fixes#643
Benson Wong
Benson Wong