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>
When _select_route picks two nodes with overlapping registrations (e.g.
A:0-22 and B:20-24), the tracker now injects start_layer per hop so B
executes only layers 23-24, not 20-24.
- model_backend: forward_bytes + _run_layers accept start_layer offset;
clamped to shard_start to prevent out-of-bounds indexing
- torch_server: _handle_binary_forward reads X-Meshnet-Start-Layer header;
_run_downstream_pipeline sends it per hop; route is now list[tuple[str,int]]
- server: proxy injects {endpoint, start_layer} objects in X-Meshnet-Route;
/v1/route response includes start_layer per node in the nodes list
- test: fake backends accept start_layer=None kwarg
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Single-node mode now uses HF model.generate() instead of one-shot
decode_tail(), giving correct multi-token output with KV cache.
model_backend.py:
- generate_text(messages, max_new_tokens, temperature, top_p) — full
autoregressive generation via model.generate() with chat template
- generate_text_streaming() — yields token strings via TextIteratorStreamer
- _encode_messages() — applies chat template (tokenize=False then tokenize),
falls back to joining user messages; avoids BatchEncoding issues
torch_server.py:
- _handle_chat_completions: fast path when backend is head+tail — calls
generate_text() or generate_text_streaming() directly instead of the
single-token encode_prompt+decode_tail pipeline
- _stream_openai_response: new SSE streaming handler for token iterators
- Parses max_tokens, temperature, top_p from request body
- Distributed path (partial shards) unchanged
Verified: streaming and non-streaming both work with Qwen2.5-0.5B-Instruct.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
model_backend.py was using Python-style exclusive end (layers[start:end])
while all callers (CLI, tests, QUICKSTART) use inclusive 0-based indexing.
Result: 24-layer model with shard_end=23 ran only 23 layers and never
set is_tail=True, so decode_tail() was never called and responses were empty.
- is_tail: == total_layers → >= total_layers - 1
- _run_layers: layers[start:end] → layers[start:end+1]
- Validation: > total_layers → >= total_layers (was also wrong)
Inference confirmed: Qwen2.5-0.5B-Instruct now returns real LLM output.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
