## Motivation
Currently, when a runner fails, the master retries the instance. Most of
the time, this causes a loop over failure. Retries need backoff and a
cap.
## Changes
- src/exo/worker/main.py: Before creating a runner, check an exponential
backoff timer per instance. After EXO_MAX_INSTANCE_RETRIES failures,
send DeleteInstance to permanently remove the instance. Record attempts
on Shutdown; reset on InstanceDeleted.
- src/exo/utils/keyed_backoff.py: Add attempts() method to query retry
count
- src/exo/shared/constants.py: Add EXO_MAX_INSTANCE_RETRIES = 3.
## Why It Works
The worker gates CreateRunner tasks behind a KeyedBackoff, adding
exponential delay (2s base, 30s cap) between retries. After 3 failures
the worker sends DeleteInstance, stopping retries entirely. The backoff
resets when the instance is deleted, so a fresh placement starts clean.
---------
Co-authored-by: Evan <evanev7@gmail.com>
## Motivation
Nemotron Cascade and Nano failing at long decodes.
## Changes
Fixed upstream, just change pyproject and uv lock here.
## Test Plan
### Automated Testing
Tested with a reproduce script upstream
## Motivation
Addresses #1816
## Changes
Update on min prefix cache > min_prefix_hit_length **and** hit ratio >
_MIN_PREFIX_HIT_RATIO_TO_UPDATE
min_prefix_hit_length = max(1000, system prompt length) -> system
prompts must match exactly.
## Test Plan
### Manual Testing
Test on OpenCode and Claude Code
## Motivation
<!-- Why is this change needed? What problem does it solve? -->
<!-- If it fixes an open issue, please link to the issue here -->
## Changes
<!-- Describe what you changed in detail -->
## Why It Works
<!-- Explain why your approach solves the problem -->
## Test Plan
### Manual Testing
<!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB,
connected via Thunderbolt 4) -->
<!-- What you did: -->
<!-- - -->
### Automated Testing
<!-- Describe changes to automated tests, or how existing tests cover
this change -->
<!-- - -->
## Motivation
<!-- Why is this change needed? What problem does it solve? -->
<!-- If it fixes an open issue, please link to the issue here -->
## Changes
<!-- Describe what you changed in detail -->
## Why It Works
<!-- Explain why your approach solves the problem -->
## Test Plan
### Manual Testing
<!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB,
connected via Thunderbolt 4) -->
<!-- What you did: -->
<!-- - -->
### Automated Testing
<!-- Describe changes to automated tests, or how existing tests cover
this change -->
<!-- - -->
adds a path option to the /state endpoint, allowing you to query
subfields of state without grabbing the whole blob
## test plan
poking around in the api
This PR builds on https://github.com/exo-explore/exo/pull/1677 to enable
custom prompts sent from Firefox `browser.ml.chat` to EXO dashboard
using URL parameters in sidebar for summary and other browser
interactions. See "Summarize page" example below.
## Summary
- Parse `?q=<encoded prompt>` URL parameter on page load and auto-submit
it as a chat message
- Clean up the URL with `history.replaceState` to prevent re-submission
on refresh
- Defer auto-send until both cluster state and model list are loaded so
model auto-selection works correctly
## Context
Firefox's built-in AI sidebar (`about:config: browser.ml.chat.enabled`)
integrates with chat providers by appending the user's prompt as
`?q=<URL-encoded prompt>`. Previously the exo dashboard ignored this
parameter. Users can now configure `http://localhost:52415` as a Firefox
AI chatbot provider.
See: https://support.mozilla.org/en-US/kb/ai-chatbot
## Technical notes
- Frontend-only change in `dashboard/src/routes/+page.svelte`
- Uses a Svelte `$effect` that reacts to `pendingFirefoxQuery`, `data`
(cluster state), and `models.length` — fires exactly once when all three
are ready
- If no model is selected, `handleAutoSend` auto-picks the best
available model; if no model fits memory, a toast is shown
- If a model is selected but not running, the message is queued until
the model loads
## Testing
```
http://localhost:52415/?q=Hello+worldhttp://localhost:52415/?q=Summarize+this+page%3A+%5Bpage+title%5D+%5Bpage+url%5D
```
<img width="2056" height="1329" alt="image"
src="https://github.com/user-attachments/assets/74463eb4-ca1a-400d-806a-c19ba93147b9"
/>
## Motivation
Adding a custom model from the Hub tab shows "Added" toast but the model
doesn't appear in the All tab. You have to add it a second time for it
to work. Also, the "All" button in the model picker sidebar is too small
to read comfortably.
## Changes
**Race condition fix (`src/exo/api/main.py`):**
- Call `add_to_card_cache(card)` directly in `add_custom_model()` after
sending the `ForwarderCommand`, before the API response returns
**Sidebar sizing
(`dashboard/src/lib/components/FamilySidebar.svelte`):**
- Increased sidebar min-width from 72/64px to 80/72px
- Increased "All" icon from `w-5 h-5` to `w-6 h-6`
- Increased all sidebar labels from 9px to 11px
## Why It Works
`POST /models/add` sends a `ForwarderCommand(AddCustomModelCard)` and
returns immediately. The frontend then calls `GET /models` which reads
from `_card_cache`. But the cache was only updated by the worker event
handler after the event round-trips through the master — a race the
frontend almost always loses. By updating the cache directly in the API
handler, `GET /models` immediately reflects the new model. The worker's
later `add_to_card_cache` call is idempotent (dict key assignment).
## Test Plan
### Manual Testing
<!-- Hardware: any Mac -->
- Open model picker → Hub tab → add a custom model → verify it appears
in All tab on the first attempt
- Verify sidebar "All" button and other labels are visually larger and
readable
### Automated Testing
- `uv run basedpyright` passes with 0 errors
- `uv run ruff check` passes
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Motivation
Replace confusing EXO_MODELS_DIR/EXO_MODELS_PATH with clearer
multi-directory support, enabling automatic download spillover across
volumes.
## Changes
- EXO_MODELS_DIRS: colon-separated writable dirs (default always
prepended, first with enough space wins)
- EXO_MODELS_READ_ONLY_DIRS: colon-separated read-only dirs (protected
from deletion)
- select_download_dir(): picks writable dir by free space
- resolve_existing_model(): unified lookup across all dirs
- is_read_only_model_dir(): path-based read-only detection instead of
hardcoded flag
- Updated coordinator, worker, model cards, tests
## Why It Works
Default dir always included so zero-config behavior is unchanged. Disk
space checked at download time for automatic spillover. Read-only status
derived from path, not hardcoded.
## Test Plan
### Manual Testing
- No env vars set → identical behavior
- EXO_MODELS_DIRS=/Volumes/SSD/models → downloads to external storage
- EXO_MODELS_READ_ONLY_DIRS=/mnt/nfs → models found, deletion blocked
### Automated Testing
- 4 new tests in test_xdg_paths.py (prepend, default-only, overlap,
empty read-only)
- Existing tests updated to patch new constants
## Motivation
During master reelection, `_elect_loop` called `worker.shutdown()` (fire
& forget) then immediately created and started a new Worker.
This caused the old runner subprocess's Metal/GPU teardown to race with
the new worker's startup, resulting in `IOConnectUnmapMemory failed:
kr=0xe00002bc` errors and a full node hang requiring `^C`. Same issue
existed for `DownloadCoordinator`.
## Changes
- Added `anyio.Event`-based `_stopped` signal to `Worker` and
`DownloadCoordinator`, set at the end of their `run()` finally blocks
- Added `wait_stopped()` async method to both classes
- Updated `_elect_loop` to `await wait_stopped()` after calling
`shutdown()` on the old Worker and DownloadCoordinator before creating
replacements
## Why It Works
The old Worker's task group contains the RunnerSupervisor tasks, whose
finally blocks join the runner subprocess (with 5s timeout + SIGTERM +
SIGKILL escalation). By awaiting `wait_stopped()`, we guarantee the old
runner process has fully exited — including GPU memory cleanup — before
a new Worker can start and potentially access the GPU. This eliminates
the race without changing the shutdown mechanics themselves.
## Test Plan
### Manual Testing
Hardware: M4 Pro Mac Mini 24GB + M3 Ultra Mac Studio 96GB, connected via
Thunderbolt
**Repro steps:**
1. Start exo on two nodes with a model sharded across both (e.g.
`Josiefied-Qwen3-14B-abliterated-v3-4bit`)
2. Wait for "runner ready" on both
3. `kill -9` the master node
4. Observe the surviving node's re-election behavior
**Before fix (original crash):**
```
[ 11:02:39.0896AM ] Runner supervisor shutting down
[ 11:02:39.0905AM ] bye from the runner
[ 11:02:39.1052AM ] Stopping Worker
IOConnectUnmapMemory failed: kr=0xe00002bc
IOConnectUnmapMemory failed: kr=0xe00002bc
IOConnectUnmapMemory failed: kr=0xe00002bc
IOConnectUnmapMemory failed: kr=0xe00002bc
^C[ 11:03:45 ] ← hung for over a minute, required manual kill
```
**After fix (clean re-election):**
```
[ 12:15:22.4703PM ] runner loaded
[ 12:15:24.1672PM ] runner ready
[ 12:15:33.5393PM ] Waiting for other campaign to finish
[ 12:15:36.5409PM ] Node elected Master
[ 12:15:36.5413PM ] Unpausing API
```
No `IOConnectUnmapMemory` errors, no hang, no `^C` needed.
### Automated Testing
- No existing tests cover the `_elect_loop` re-election path; this is an
integration-level flow requiring a live router/election/worker stack
- All existing tests pass (307/308, 1 pre-existing Rust binding failure)
- basedpyright: 0 errors, ruff: all checks passed
---------
Co-authored-by: Evan <evanev7@gmail.com>
## Motivation
<!-- Why is this change needed? What problem does it solve? -->
<!-- If it fixes an open issue, please link to the issue here -->
## Changes
<!-- Describe what you changed in detail -->
## Why It Works
<!-- Explain why your approach solves the problem -->
## Test Plan
### Manual Testing
<!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB,
connected via Thunderbolt 4) -->
<!-- What you did: -->
<!-- - -->
### Automated Testing
<!-- Describe changes to automated tests, or how existing tests cover
this change -->
<!-- - -->
missed in #1747, issue #1798.
### the issue
we didn't set the memory poll rate after failling the macmon preflight,
only after failing the followups - as we never ran macmon if preflight
failed, we never hit the followup errors etc.
### testing
requires testing on an m5 pro, but the core issue is solved.
## Motivation
PR #1747 fixed macmon support for M5 Pro/Max by pinning the
`swiftraccoon/macmon` fork in `flake.nix`. This works when running from
source (via Nix) but the distributed macOS `.app` build was still broken
on M5 Pro/Max because it was bundling the wrong macmon.
The error on M5 Pro/Max:
```
macmon preflight failed with return code -6: thread 'main' panicked at src/sources.rs:394:41
```
## Changes
- Removed `macmon` from `brew install` in `build-app.yml` — this was
installing the upstream `vladkens/macmon` which doesn't support M5
Pro/Max
- Added a new step that resolves the pinned macmon fork from the Nix dev
shell (same `swiftraccoon/macmon` at rev `9154d23` already defined in
`flake.nix`) and adds it to `$GITHUB_PATH`
- Added a safety `brew uninstall macmon` to ensure no Homebrew macmon
can shadow the pinned version
## Why It Works
PyInstaller bundles macmon via `shutil.which("macmon")`. Previously this
found the Homebrew (upstream) binary. Now it finds the Nix-overlayed
fork that has M5 Pro/Max support, because `$GITHUB_PATH` prepends the
Nix store path before the PyInstaller step runs.
## Test Plan
### Manual Testing
<!-- Hardware: M5 Pro -->
- Trigger a macOS build and verify the bundled macmon is the pinned fork
- Run the built `.app` on M5 Pro/Max and confirm macmon preflight
succeeds
### Automated Testing
- Existing CI build workflow will validate that the macmon binary is
found and bundled correctly
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Motivation
Support cancelling image generation, similar to existing support for
cancelling text generation
## Changes
- Dashboard (app.svelte.ts): Wire up AbortController for both
generateImage and editImage API calls. On abort, show "Cancelled"
instead of an error. Clean up the controller in finally.
- Pipeline runner (pipeline/runner.py): Introduce a cancel_checker
callback and NaN-sentinel cancellation protocol for distributed
diffusion:
- _check_cancellation() - only rank 0 polls the cancel callback
- _send() - replaces data with NaN sentinels when cancelling, so
downstream ranks detect cancellation via _recv_and_check()
- _recv() / _recv_like() wrappers that eval and check for NaN sentinel
- After cancellation, drains any pending ring recv to prevent deadlock
- Skips partial image yields and final decode when cancelled
- Image runner (runner/image_models/runner.py): Deduplicate the
ImageGeneration and ImageEdits match arms into a shared
_run_image_task() method. Thread a cancel_checker closure (backed by the
existing cancel_receiver + cancelled_tasks set) into generate_image().
- Plumbing (distributed_model.py, generate.py): Pass cancel_checker
through the call chain.
## Why It Works
- Rank 0 is the only node that knows about task-level cancellation. When
it detects cancellation, it sends NaN tensors instead of real data.
Higher-order ranks detect the NaN sentinel on recv, set their own
_cancelling flag, and propagate NaN forward
- A drain step after the loop prevents the deadlock case where the last
rank already sent patches that the first would never consume.
- For single-node mode, the loop simply breaks immediately on
cancellation.
## Test Plan
### Automated Testing
New tests in src/exo/worker/tests/unittests/test_image
## Summary
DeepSeek V3.2 (`DeepseekV32ForCausalLM`) is already supported by exo's
inference engine (architecture whitelisted in `model_cards.py`, DSML
encoding added in #1548), but **doesn't work out of the box** due to two
bugs:
### Bug 1: `warmup_inference` passes empty model ID
`warmup_inference()` in `generate.py` accepts `model_id: ModelId` as a
parameter but creates `TextGenerationTaskParams(model=ModelId(""), ...)`
instead of using it. Since `_needs_dsml_encoding()` checks
`"deepseek-v3.2" in task_params.model.lower()`, the empty string never
matches → falls back to `tokenizer.apply_chat_template()` →
**ValueError** because V3.2 has no Jinja chat template.
**Fix:** `model=ModelId("")` → `model=model_id` (one line).
### Bug 2: `_needs_dsml_encoding` limited to tool calling
`_needs_dsml_encoding()` returns `True` only when `task_params.tools` is
present or tool messages exist in `chat_template_messages`. For warmup
and regular chat requests without tools → `return False` → Jinja
fallback → **ValueError**.
Unlike V3.1 (which has a `.jinja` chat template file that transformers
picks up automatically), V3.2 **has no Jinja template at all** — it uses
Python-based DSML encoding for all message types.
**Fix:** For V3.2, always return `True` — DSML encoding handles all
message types.
### Catalog cards
Added inference model cards for:
- `mlx-community/DeepSeek-V3.2-8bit`
- `mlx-community/DeepSeek-V3.2-4bit`
Parameters taken from model `config.json` on HuggingFace, storage sizes
from HF API. Capabilities include `thinking_toggle` (related: #1456).
## Notes
- The model ID string matching approach (`"deepseek-v3.2" in
model.lower()`) is acknowledged tech debt — see #1371 for the planned
architecture-based approach.
## Test plan
- [x] Start exo with DeepSeek V3.2 model → warmup should complete
without crash
- [x] Send a regular chat message (no tools) → should get a response
- [x] Send a chat message with tools → should work as before
- [x] V3.2 cards should appear in the dashboard model catalog
---------
Co-authored-by: user <user@m1.note>
Co-authored-by: Ryuichi Leo Takashige <leo@exolabs.net>
Co-authored-by: Evan <evanev7@gmail.com>
## Motivation
Warmup occasionally had issues; e.g. #1748 and #1793 because we were
using a standard stream_generate, all of which are issues that are
resolved in the wrapper function mlx_generate.
**Enabling peers to be discovered in environments where mDNS is
unavailable (SSH sessions, headless servers, Docker).**
## Motivation
Exo discovers peers exclusively via mDNS, which works great on a local
network but breaks once you move beyond a single L2 broadcast domain:
- SSH sessions on macOS — TCC blocks mDNS multicast from non-GUI
sessions (#1488)
- Headless servers/rack machines — #1682 ("DGX Spark does not find other
nodes")
- Docker Compose — mDNS is often unavailable across container networks;
e.g. #1462 (E2E test framework) needs an alternative
Related works:
#1488 (working implementation made by @AlexCheema and closed because SSH
had a GUI workaround),
#1023 (Headscale WAN then closed due to merge conflicts),
#1656 (discovery cleanup, open).
This PR introduces an optional bootstrap mechanism for peer discovery
while leaving the existing mDNS behavior unchanged.
## Changes
Adds two new CLI flags:
- `--bootstrap-peers` (env: `EXO_BOOTSTRAP_PEERS`) — comma-separated
libp2p multiaddrs to dial on startup and retry periodically
- `--libp2p-port` — fixed TCP port for libp2p to listen on (default:
OS-assigned). Required when bootstrap peers, so other nodes know which
port to dial.
8 files:
- `rust/networking/src/discovery.rs`: Store bootstrap addrs, dial in
existing retry loop
- `rust/networking/src/swarm.rs`: Thread `bootstrap_peers` parameter to
`Behaviour`
- `rust/networking/examples/chatroom.rs`: Updated call site for new
create_swarm signature
- `rust/networking/tests/bootstrap_peers.rs`: Integration tests
- `rust/exo_pyo3_bindings/src/networking.rs`: Accept optional
`bootstrap_peers` in PyO3 constructor
- `rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi` : Update type stub
- `src/exo/routing/router.py`: Pass peers to `NetworkingHandle`
- `src/exo/main.py` : `--bootstrap-peers` CLI arg +
`EXO_BOOTSTRAP_PEERS` env var
## Why It Works
Bootstrap peers are dialed in the existing retry loop — the same path
taken by peers when mDNS-discovered. The swarm handles connection, Noise
handshake, and gossipsub mesh joining from there.
PeerId is intentionally not required in the multiaddr, the Noise
handshake discovers it.
Docker Compose example:
```yaml
services:
exo-1:
environment:
EXO_BOOTSTRAP_PEERS: "/ip4/exo-2/tcp/30000"
exo-2:
environment:
EXO_BOOTSTRAP_PEERS: "/ip4/exo-1/tcp/30000"
```
## Test Plan
### Manual Testing
<details>
<summary>Docker Compose config</summary>
```
services:
exo-node1:
build:
context: .
dockerfile: Dockerfile.bootstrap-test
container_name: exo-bootstrap-node1
hostname: exo-node1
command: ["-q", "--libp2p-port", "30000", "--bootstrap-peers", "/ip4/172.30.20.3/tcp/30000"]
environment:
- EXO_LIBP2P_NAMESPACE=bootstrap-test
ports:
- "52415:52415"
networks:
bootstrap-net:
ipv4_address: 172.30.20.2
deploy:
resources:
limits:
memory: 4g
exo-node2:
build:
context: .
dockerfile: Dockerfile.bootstrap-test
container_name: exo-bootstrap-node2
hostname: exo-node2
command: ["-q", "--libp2p-port", "30000", "--bootstrap-peers", "/ip4/172.30.20.2/tcp/30000"]
environment:
- EXO_LIBP2P_NAMESPACE=bootstrap-test
ports:
- "52416:52415"
networks:
bootstrap-net:
ipv4_address: 172.30.20.3
deploy:
resources:
limits:
memory: 4g
networks:
bootstrap-net:
driver: bridge
ipam:
config:
- subnet: 172.30.20.0/24
```
</details>
Two containers on a bridge network (`172.30.20.0/24`), fixed IPs,
`--libp2p-port 30000`, cross-referencing `--bootstrap-peers`.
Both nodes found each other and established a connection then ran the
election protocol.
### Automated Testing
4 Rust integration tests in `rust/networking/tests/bootstrap_peers.rs`
(`cargo test -p networking`):
| Test | What it verifies | Result |
|------|-----------------|--------|
| `two_nodes_connect_via_bootstrap_peers` | Node B discovers Node A via
bootstrap addr (real TCP connection) | PASS |
| `create_swarm_with_empty_bootstrap_peers` | Backward compatibility —
no bootstrap peers works | PASS |
| `create_swarm_ignores_invalid_bootstrap_addrs` | Invalid multiaddrs
silently filtered | PASS |
| `create_swarm_with_fixed_port` | `listen_port` parameter works | PASS
|
All 4 pass. The connection test takes ~6s
---------
Signed-off-by: DeepZima <deepzima@outlook.com>
Co-authored-by: Evan <evanev7@gmail.com>
## Motivation
The Nemotron model family in the model picker sidebar was displaying as
"Nemotron" with a generic checkmark icon. Since these models are
NVIDIA's Nemotron models, the category should be branded as "NVIDIA"
with the official NVIDIA logo, consistent with how other families are
branded (e.g., "llama" → "Meta", "gpt-oss" → "OpenAI").
## Changes
- **FamilySidebar.svelte**: Added `nemotron: "NVIDIA"` to the
`familyNames` mapping so the sidebar displays "NVIDIA" instead of
"Nemotron"
- **FamilyLogos.svelte**: Added the NVIDIA "eye" logo as an inline SVG
for the `nemotron` family, matching the `viewBox="0 0 24 24"` /
`fill="currentColor"` pattern used by all other brand logos
- **ModelPickerModal.svelte**: Added `"nemotron"` to the `familyOrder`
array so NVIDIA appears in a consistent position in the sidebar
## Why It Works
The model picker derives categories from the `family` field in TOML
model cards. Nemotron models already have `family = "nemotron"`, but the
three UI components (display name, logo, sort order) lacked explicit
entries for it, causing fallback behavior (auto-capitalized name,
checkmark icon, alphabetical sorting). Adding explicit entries for all
three aligns NVIDIA with the existing brand pattern.
## Test Plan
### Manual Testing
<!-- Hardware: N/A - dashboard UI change only -->
- Built dashboard successfully (`npm run build`)
- Verified the NVIDIA logo renders in the sidebar alongside existing
brand logos
### Automated Testing
- No test changes needed — this is a purely cosmetic dashboard change
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
updates macmon to an upstream fork that fixes m5 max issues.
might see if the upstream version gets merged before we release.
---------
Co-authored-by: Alex Cheema <alexcheema123@gmail.com>
## Motivation
Move back to an MLX LM to improve the Qwen 3.5 experience.
## Test Plan
### Manual Testing
Seems to loop less from testing, no speed regressions.
## Motivation
Some users (primarily in China) are unable to access HuggingFace.co. HF
supports the HF_ENDPOINT env variable, and we also support it, but there
is no way to easily do that from the app currently.
## Test Plan
### Manual Testing
hf-mirror works
empty field works
google.com endpoint fails
a fix from pr #1643 which is now stale - should make prefill more
consistent on very slow machines
## testing
qwen-3.5-35b-a3b loads normally
gpt-oss-120b-mxfp4-q8 loads normally
## Motivation
Batch generation reports incorrect statistics, as mlx lm never clears
the original stats, meaning they get polluted over time.
The dashboard also seems considerably slower than bench statistics.
We also have a large discrepancy between B=1 batch generating and
mlx_generate.
Extracting logprobs is massively expensive, causing up to a 25% slowdown
compared to pure batching.
```
[ 12:02:01.1240AM | INFO ] step overhead: 3.49ms (next=12.49ms total=15.99ms)
[ 12:02:02.1600AM | INFO ] step overhead: 3.23ms (next=13.01ms total=16.24ms)
[ 12:02:03.2228AM | INFO ] step overhead: 3.28ms (next=13.38ms total=16.66ms)
[ 12:02:04.2798AM | INFO ] step overhead: 3.25ms (next=12.84ms total=16.10ms)
[ 12:02:05.3152AM | INFO ] step overhead: 3.18ms (next=12.61ms total=15.79ms)
[ 12:02:06.3522AM | INFO ] step overhead: 3.41ms (next=12.83ms total=16.25ms)
[ 12:02:07.3987AM | INFO ] step overhead: 3.38ms (next=13.14ms total=16.52ms)
[ 12:02:08.4537AM | INFO ] step overhead: 1.84ms (next=19.44ms total=21.28ms)
```
## Changes
1. Report stats ourselves instead of using mlx lm's stats for batch
generation (they use perf_counter anyway).
2. Adjust exo bench to match
3. Improve logprobs extraction speed by 10x, improving tps for dashboard
& any requests for logprobs
4. Use an SSE comment to align the speed to the real numbers at the end
of generation
5. Patch mlx for several optimizations given our assumptions and use
cases (e.g. use vllm style RoPE).
6. Switch MLX LM version to latest main, including support for Nemotron
Super and some Qwen3.5 fixes.
## Why It Works
1. Exo bench no longer reports polluted stats
2. Exo bench now handles the reported per-request stats rather than the
aggregate stats
3. The decode speed now jumps back to a real number at the end of the
generation
4. Large batch speedup for rotating KV cache models + 1:1 matching cache
with vllm
## Test Plan
### Manual Testing
Needs testing on OpenCode and CC
Needs eval testing
### Automated Testing
Only going to show the performance optimization difference after the
accurate reporting:
**GPT OSS 20B MXFP4 Q8 (large change)**
Before:
<img width="2466" height="1534" alt="image"
src="https://github.com/user-attachments/assets/88b50637-fca2-4db4-9413-b9eee6e2057e"
/>
<img width="2410" height="1240" alt="image"
src="https://github.com/user-attachments/assets/21e5c76a-2f5f-44d2-8953-121b3ebdbd68"
/>
After:
<img width="2476" height="1472" alt="image"
src="https://github.com/user-attachments/assets/fec5cfbd-fff8-430a-b12e-a329410107a2"
/>
<img width="2454" height="1236" alt="image"
src="https://github.com/user-attachments/assets/0400344b-a4a6-42c0-a9dd-4ee91ade714a"
/>
**Qwen 3.5 35B A3B 8bit (No change)**
Before:
<img width="2414" height="1396" alt="image"
src="https://github.com/user-attachments/assets/e75f0b38-df5d-49fd-ab90-bc1667d981b3"
/>
After:
<img width="2346" height="1234" alt="image"
src="https://github.com/user-attachments/assets/eabfb59c-851f-4d88-b927-e1e699a75cc6"
/>
**Llama 3.2 1B Instruct 4bit (small change)**
Before:
<img width="2516" height="1220" alt="image"
src="https://github.com/user-attachments/assets/c2873655-acff-4536-8263-fb8aea33db80"
/>
After:
<img width="2566" height="1370" alt="image"
src="https://github.com/user-attachments/assets/15f95c75-1c2f-4474-85a2-88c4d0a32543"
/>
## Motivation
Running exo in offline mode on a machine where a model has never been
downloaded causes a crash
## Changes
- `download_shard` now catches `FileNotFoundError` from
`fetch_file_list_with_cache` and returns a `not_started` progress
instead of propagating the exception
## Why It Works
A status query should never crash its caller. By returning
`not_started`, the coordinator's existing offline guard (`if
self.offline:` at line 198) is reached and emits a graceful
`DownloadFailed` event. This also eliminates the warning spam on startup
where the status iterator catches the same exception for every
predefined model.
## Test Plan
### Manual Testing
- Start exo with `--offline` on a machine with no cached models, request
a model via the API — should get a graceful failure instead of a crash
## Motivation
Custom model cards were only saved locally on the node that handled the
API request.
## Changes
- Added AddCustomModelCard and DeleteCustomModelCard commands
- Added CustomModelCardAdded and CustomModelCardDeleted events
- Added custom_model_cards field to cluster State
- Master handles new commands by emitting corresponding events
- Workers persist model cards to disk and update the in-memory cache on
event receipt
- Separated fetch_from_hf (pure fetch) from disk persistence (now
handled by event-sourcing layer)
- Exposed add_to_card_cache() helper for the worker to update the cache
## Why It Works
- Follows the existing event-sourcing pattern: API → Command → Master →
Event → all Workers
- Every node applies the same events, so custom model cards are
consistent across the cluster
## Test Plan
### Manual Testing
Add/delete a custom model via the API on one node, verify it
appears/disappears on all nodes
### Automated Testing
Existing tests cover apply() logic; new event types follow the same
discriminated-union pattern
## Motivation
When placing a model instance across the cluster, the master previously
only considered available RAM. This meant it could pick a node that
hasn't downloaded the model yet, even when another node already has it
(or is further along in downloading it).
## Changes
- Added download_status parameter to place_instance() in placement.py
- Added _get_node_download_fraction() to compute 0.0–1.0 download
progress per node/model
- Added _cycle_download_score() to sum download fractions across a
cycle's nodes
- Cycle selection now uses a (download_score, available_ram) tuple key —
download progress is the primary sort, RAM is the tiebreaker
- Passed self.state.downloads into place_instance() from master/main.py
## Why It Works
Python's tuple comparison gives download progress strict priority over
RAM, so a node with the model already downloaded will always be
preferred over one with more free RAM but no download.
## Test Plan
### Automated Testing
3 new tests cover: completed download preferred, higher partial progress
preferred, failed download not preferred over no-download node
## Summary
- **Add ISO date to verbose stderr format**: when running with `-v`, the
stderr timestamp changes from `HH:mm:ss.SSS` to `YYYY-MM-DD
HH:mm:ss.SSS`, matching the file log format. This makes it possible to
correlate entries across days when stderr is captured by launchd,
systemd, Docker, or file redirection.
- **File log respects verbosity**: `exo.log` now uses DEBUG level when
`-v` is passed, so the persistent log has the same detail as stderr.
Previously it was always INFO regardless of verbosity.
- **Startup banner with PID**: adds a visual separator and process ID to
the startup message, making it easy to identify session boundaries in
long-running logs (e.g. `grep "Starting EXO"`).
The non-verbose (default) stderr format is unchanged — end-user terminal
experience is not affected.
## Motivation
When exo runs as a service (launchd, systemd, Docker), stderr is
typically captured to a file. Without calendar dates in the timestamp,
it is impossible to tell which day a log entry belongs to. This caused
misidentification of log entries during a multi-day RDMA debugging
session on a 4-node cluster.
The file log (`exo.log`) already had dates and rotation, but only
captured INFO level — missing the DEBUG output needed for postmortem
analysis.
## Test plan
- [x] `basedpyright` — 0 errors, 0 warnings, 0 notes
- [x] `ruff check` — all checks passed
- [x] `pytest` — 249 passed, 1 skipped
Co-authored-by: user <user@m1.note>
## Motivation
Download progress in the dashboard was broken: mainly treating all
download statuses as ongoing
## Changes
- Backend (apply.py): Deduplicate download progress events by model_id
instead of full shard_metadata, preventing duplicate entries per node
- Dashboard (+page.svelte): Extract shared collectDownloadStatus()
helper that both getModelDownloadStatus and getInstanceDownloadStatus
use, eliminating ~100 lines of duplicated logic. Adds proper handling
for
DownloadCompleted/DownloadFailed events, uses a Map to deduplicate
per-node entries, and introduces a typed NodeDownloadStatus with
explicit status states (downloading/completed/partial/pending)
- ModelCard: Replace single aggregate progress bar with per-node
download bars, each color-coded by status. Instance preview now scopes
download status to participating nodes only
## Why It Works
- Deduplicating by model_id in apply.py ensures each node has exactly
one download entry per model
- The perNodeMap in the frontend keeps only the latest event per node,
preventing duplicate bars
- Handling DownloadCompleted allows the UI to show finished downloads
instead of dropping them
- Scoping instance previews to assigned nodes avoids showing irrelevant
download progress
## Motivation
Image generation/editing from the dashboard was broken. ChatForm
bypassed the parent's model-launch logic, so image requests fell through
to text chat.
## Changes
- Moved image routing logic from ChatForm.svelte to +page.svelte
(routeMessage())
- ChatForm now always delegates to parent via onAutoSend (made required)
- Fixed missing updateActiveConversation() call on message retry
## Why It Works
All sends now go through the parent's launch-then-route path, so image
models get launched before the request is dispatched to the correct
endpoint.
## Motivation
<!-- Why is this change needed? What problem does it solve? -->
<!-- If it fixes an open issue, please link to the issue here -->
This was to fix a small issue which was that the StepFun logo was not
included in the sidebar, and I also noticed it from an open issue:
- #1662
## Changes
<!-- Describe what you changed in detail -->
I added a condition to the FamilyLogos where if the family is "step"
then the logo will be included in the sidebar.
## Test Plan
Open exo, go choose a model, and scroll down the sidebar until you see
the step logo, which should be there.
### Manual Testing
<!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB,
connected via Thunderbolt 4) -->
<!-- What you did: -->
<!-- - -->
Check for the step logo in the sidebar.
## Summary
Fixes a bug where macmon metrics (GPU usage, temperature, power, RAM)
freeze permanently in the dashboard while non-macmon metrics (disk
usage) continue updating normally.
**Root cause: asyncio subprocess pipe transport flow control stall**
Observed on a 2-node Mac Studio M3 Ultra cluster (Python 3.13.12, anyio
4.11.0, macOS with kqueue) running EXO.app for ~36 hours. Diagnostics
confirmed:
1. **Only macmon monitoring is stuck** — disk metrics from the same
InfoGatherer continue updating, proving the Worker, EventRouter, and API
pipelines are healthy
2. **macmon IS producing data** — its stdout pipe is full at exactly
65536 bytes (64KB OS buffer), and macmon is blocked on write at 0% CPU
3. **The pipe read-end FD is still open** — the exo process holds it,
but asyncio isn't reading from it
4. **The stale GPU value (0.21) is wrong** — should be ~1.0 (matching
the other node under identical load)
This is consistent with asyncio's subprocess pipe transport getting
stuck in flow control: `pause_reading()` is called when the internal
buffer exceeds the high-water mark, but `resume_reading()` is never
called, permanently deregistering the FD from kqueue. The `receive()`
coroutine waits forever for data asyncio will never deliver.
```
$ lsof -p <macmon_pid>
macmon 74691 FD=1 PIPE SIZE=65536 ->0x5ae78ecf376ccd0e # full pipe
$ lsof -p <exo_pid> | grep <pipe_id>
exo 74689 FD=47 PIPE SIZE=65536 ->0xd129da474c1340f7 # read end held open, not consumed
```
Note: anyio 4.11's `Process.aclose()` already uses
`CancelScope(shield=True)` for cleanup during cancellation — this is NOT
an election cleanup issue (confirmed by @ciaranbor's testing).
**Fix:** Replace the `async for` iteration with an explicit `receive()`
inside `fail_after()`. If macmon produces no output for 10× its
configured interval (minimum 30s), `TimeoutError` fires, `open_process`
cleanup kills macmon and tears down the stale transport, and the loop
restarts with a fresh subprocess and fresh asyncio transport.
## Test plan
- [x] All 246 existing tests pass
- [ ] Verify macmon restarts after simulated pipe stall (e.g., `SIGSTOP`
macmon, wait for timeout, confirm restart and metrics resume)
- [ ] Long-running soak test on multi-node cluster to confirm the fix
prevents recurrence
## Motivation
The MLX version and git revision in nix/mlx.nix were hardcoded and had
to be manually kept in sync with uv.lock
## Changes
- flake.nix: Extract MLX git rev from uv.lock's source.git URL and pass
as uvLockMlxRev
- nix/mlx.nix: Use uvLockMlxVersion and uvLockMlxRev instead of
hardcoded values; remove version mismatch assertion
## Why It Works
uv.lock is already the source of truth — now Nix reads both version and
rev from it directly. The pinned fetchFromGitHub hash still guards
against unexpected changes.
## Problem
Running short completions (like `max_tokens=1` health check probes) can
finish so fast that `mlx_lm`'s internal `generation_time` rounds to
zero. When that happens, `BatchGenerator.stats()` in
`mlx_lm/generate.py` divides `generation_tokens / generation_time` and
throws a `ZeroDivisionError`, which kills the runner process.
EXO already handles this on its side — lines 289-293 in
`batch_generate.py` guard the TPS calculation with `if
generation_elapsed > 0`. But the call to `self._exo_gen.stats()` on line
294 goes into mlx_lm's *separate* timing code, which doesn't have the
same guard. Two different timers, only one is protected.
In my case, this was triggered by health check probes (content: `"a"`,
`max_tokens=1`). The generation completed in sub-microsecond time,
`generation_time` was exactly `0`, and the runner crashed. Since the
health check command stays in the queue and retries after recovery, it
created an infinite crash loop — every ~15 seconds the runner would load
the model, get the same health check, and die again.
## Fix
Wrap `self._exo_gen.stats()` in a `try/except ZeroDivisionError`. If it
throws, set `mlx_stats` to `None` and fall back to `0.0` for
`prompt_tps`. The only fields EXO reads from `mlx_stats` are
`prompt_tps` and `prompt_time` — losing them on a sub-microsecond
generation has no practical impact.
## Traceback
```
File "batch_generate.py", line 294, in step
mlx_stats = self._exo_gen.stats()
File "mlx_lm/generate.py", line 1224, in stats
self._stats.generation_tokens / self._stats.generation_time
ZeroDivisionError: division by zero
```
## Summary
- Dashboard now shows partial download progress for models that were
partially downloaded in a previous session, instead of showing 0%
- Both `getModelDownloadStatus()` and `getInstanceDownloadStatus()` now
handle `DownloadPending` entries that carry non-zero
`downloaded`/`total` bytes
Fixes#1042
## Root cause
When exo restarts with partially downloaded models, the
`DownloadCoordinator` emits `DownloadPending` events (because
`downloaded_this_session` is 0, even though real bytes exist on disk).
The main page dashboard only checked for `DownloadOngoing` entries, so
these partially downloaded models showed as 0%.
The dedicated `/downloads` page already handled this correctly — it
renders `DownloadPending` entries with a progress bar when `downloaded >
0`. This fix brings the same behavior to the main page.
## Changes
For both `getModelDownloadStatus()` and `getInstanceDownloadStatus()` in
`+page.svelte`:
- Accept `DownloadPending` in addition to `DownloadOngoing`
- For `DownloadPending` entries with `downloaded > 0` or `total > 0`,
synthesize a `DownloadProgress` object from the top-level fields (with
`speed: 0` and `etaMs: 0` since no active download is in progress)
- Skip `DownloadPending` entries where both `downloaded` and `total` are
0 (truly pending, not yet started)
## Test plan
- [ ] Partially download a model, quit exo, relaunch — dashboard should
show partial progress instead of 0%
- [ ] Fully downloaded models still show as complete
- [ ] Active downloads still show real-time progress with speed/ETA
- [ ] Models never downloaded show as not started (not falsely showing
progress)
- [ ] Dashboard builds without errors (`cd dashboard && npm run build`)
---------
Co-authored-by: Evan <evanev7@gmail.com>
## Motivation
The responses API often does not provide tool args nested under a
"function" field. Since we follow the chat completions format of tools
in the backend (for MLX chat templates), we need to normalise to this
format.
## Test Plan
### Manual Testing
Works on n8n!
<img width="3442" height="2076" alt="image"
src="https://github.com/user-attachments/assets/9e11d679-0102-4d83-9a8e-b0a7a5898708"
/>
## Problem
Models with fewer KV heads than nodes crash during tensor parallelism.
For example, Qwen3.5 MoE models have only 2 KV heads — trying to shard
across 4 nodes produces empty tensors and a reshape error at runtime.
The placement system already validates `hidden_size % num_nodes == 0`
but doesn't check KV heads, so it creates configurations that look valid
but blow up when the worker tries to split the attention heads.
Affected models include Qwen3.5-35B-A3B, Qwen3.5-122B-A10B,
Qwen3.5-397B-A17B, Qwen3-Next-80B-A3B, and Qwen3-Coder-Next (all have 2
KV heads).
## Changes
**Placement validation** (`src/exo/master/placement.py`):
- Combined KV heads divisibility check with the existing hidden_size
filter in a single pass
- Cycles where `num_key_value_heads % len(cycle) != 0` are now excluded
for tensor sharding
- Error message includes both constraints when no valid cycle is found
**Model card schema** (`src/exo/shared/models/model_cards.py`):
- Added optional `num_key_value_heads` field to `ModelCard` and
`ConfigData`
- Extracted from HuggingFace `config.json` (handles both top-level and
`text_config` nesting)
- Passed through in `fetch_from_hf()` for dynamically fetched cards
**All 68 inference model cards**
(`resources/inference_model_cards/*.toml`):
- Populated `num_key_value_heads` from each model's HuggingFace config
**Utility script** (`scripts/fetch_kv_heads.py`):
- Fetches `num_key_value_heads` from HuggingFace and updates TOML cards
- `--missing`: only fills in cards that don't have the field yet
- `--all`: re-fetches and overwrites everything
- Uses tomlkit for safe TOML editing and ThreadPoolExecutor for parallel
fetches
## Behavior
- Instance previews no longer show tensor options for models that can't
split their KV heads across the cluster size
- `place_instance()` rejects with a clear error instead of crash-looping
- Pipeline parallelism is unaffected
- 2-node tensor still works for 2-KV-head models (2 ÷ 2 = 1)
- Field is optional — existing custom cards without it continue to work
(validation is skipped when `None`)
## Problem
Emojis and other multi-byte UTF-8 characters are rendered as `\ufffd`
(Unicode Replacement Character) in batch streaming responses.
Byte-level BPE tokenizers (like Qwen's) can split multi-byte UTF-8
characters (e.g. a 4-byte emoji) across multiple tokens. The batch
generator decodes each token independently with
`tokenizer.decode([token_id])`, which produces U+FFFD for partial byte
sequences.
The sequential generator doesn't have this problem — it uses
`stream_generate()` from mlx_lm, which internally uses
`StreamingDetokenizer` to buffer incomplete bytes.
## Fix
Use `StreamingDetokenizer` in the batch generator, matching the
sequential path:
- Each `_EngineTask` gets its own `StreamingDetokenizer` instance
- `add_token()` buffers tokens, holding back incomplete UTF-8 byte
sequences until they form valid characters
- `last_segment` returns only complete, valid text
- `finalize()` flushes any remaining buffered bytes when generation
completes
Empty text segments during buffering are harmless — they're already
handled correctly by the downstream streaming pipeline.
---------
Co-authored-by: Ryuichi Leo Takashige <leo@exolabs.net>
ciaranbor