Qwen3.5/3.6-MoE checkpoints ship vision (model.visual.*) and multi-token-
prediction (mtp.*) weights; the partial shard loader assigned them into the
text-only Qwen3_5MoeForCausalLM and crashed with AttributeError 'mtp'.
Filter selected tensors against the built model's state_dict keys, matching
transformers' _keys_to_ignore_on_load_unexpected behavior.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Start background re-registration when the tracker is unreachable at startup so nodes do not stay permanently unregistered.
Co-authored-by: Cursor <cursoragent@cursor.com>
Unify --model and --model-id so catalog names use the tracker path, and allow --shard-start/--shard-end with --model instead of requiring --model-id.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resolve conflicts in dashboard.html (Call wall + live TPS/queue from remote) and server.py (proxy progress logging, request id forwarding, current_requests on node entries).
Co-authored-by: Cursor <cursoragent@cursor.com>
- cli.py: change default --quantization from bfloat16 to int8; saves
~50% VRAM/RAM for new nodes that don't specify a quantization
- startup.py: display memory budget and GPU info in GB (e.g. 124.9 GB RAM)
instead of MB; show remaining headroom after full model load
- test_tracker_routing.py: add test_shard_heal_cycle_surviving_node_covers_dead_peers_gap
— end-to-end proof that:
1. tracker purges expired node A and queues LOAD_SHARD for node B
2. node B receives directive on next heartbeat
3. TorchNodeServer.apply_tracker_directives hot-swaps the backend
4. node B re-registers covering the full model; coverage gap closed
Test runs in <1s with monkeypatched _load_backend (no GPU needed)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Nodes behind NAT (5G, WSL2, home routers) can now participate in
distributed pipeline inference over the internet via the relay server.
- torch_server: add module-level _relay_hop() that opens a WebSocket
to relay.../rpc/{peer_id}, sends the binary activation with
body_base64 encoding, and returns (status, headers, body)
- torch_server: _get_remaining_route returns list[dict] (was list[tuple])
preserving relay_addr from injected X-Meshnet-Route header and
from /v1/route slow-path node info
- torch_server: _run_downstream_pipeline dispatches via _relay_hop
when hop has relay_addr; falls back to direct HTTP on relay error
- tracker server: downstream_hops dicts include relay_addr when node
has one registered, so head node knows how to reach each peer
- relay_bridge: binary bodies (bfloat16 activations) use body_base64;
response preserves all X-Meshnet-* headers
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Nodes running inside WSL2 or Docker have a virtual network interface whose IP
(172.16-31.x.x) is not reachable from physical machines on the LAN. The UDP
socket probe for outbound IP returns this virtual IP, which gets registered with
the tracker — causing downstream pipeline hops to time out with "urlopen error
timed out".
_warn_virtual_network_ip() now prints a clear multi-line warning at startup
when the auto-detected advertise IP falls in 172.16.0.0/12, including the
fix: --advertise-host <LAN-ip>.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_select_route now prefers nodes with higher effective throughput
(benchmark_tokens_per_sec / (queue_depth + 1)) when multiple nodes
cover the same interval, breaking the tie after max-reach selection.
TorchNodeServer.route_timeout property added for external inspection.
Tests added:
- _select_route with overlapping shards (A:0-22, B:20-24) → correct X-Meshnet-Start-Layer protocol
- _select_route with gap → clear error message
- Fast node beats slow node when shards are equal
- Busy fast node vs idle slow node (queue_depth factor)
- Two-StubNodeServer integration via TrackerServer
- route_timeout property exposed on TorchNodeServer
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
